Introduction

CitriniResearch’s raison d’etre is taking complex topics of interest for investors and explaining them in a way that’s intuitive. That’s why our work spans so many asset classes, and it’s why sometimes we write in-depth sector primers, macroeconomic missives and, occasionally, hypothetical scenarios that result in us receiving death threats (only semi-credible ones).

Talking about something confusing is what gets us excited. It’s also where great investment ideas are born.

The situation in the Strait of Hormuz is nothing if not confusing right now. So, CitriniResearch sent our incredibly capable field analyst – dubbed Analyst #3 in order to avoid emotional attachment – on assignment to the Strait of Hormuz. Armed with a fluency in four languages including Arabic, a Pelican case full of equipment, a pack of Cuban cigars, $15,000 in cash and a roll of Zyn, #3 set out to fulfill the itinerary we’d planned in our Manhattan offices the week prior.

We figured we’d leave with an impression that was basically “The strait was closed or open.” We also were quite aware that the trip might be a flop and we would learn nothing at all. However, we came away with a much more nuanced understanding of the current environment and the transition to a multipolar world.

If David Foster Wallace were alive today, he’d be reporting from the bar in a beachside town on the Omani coast– taking notes on a napkin about the particular quality of silence in a hundred-room hotel with three guests, watching tankers drift towards, but never quite reach, the Strait of Hormuz. That’s our inspiration here, if DFW was also concerned with finding alpha.

This is a story about the most consequential place on Earth right now — the fifty-four-mile passage between Iran and Oman through which the global economy flows, or doesn’t. There was no shortage of alpha on the Strait, including concrete information on the new rules, being written as we speak, on how the Iranian Revolutionary Guard is deciding who can, and can’t, pass.

Analyst #3 decided – against the counsel of an Omani border agent, the implicit counsel of God, and the extremely explicit counsel of two Coast Guard officers holding assault rifles – that he was going to the center of the most consequential waterway on earth, during a live war, in a speedboat with no GPS, captained by a man he met three hours ago at a port inlet by pulling out a wad of cash. For investment research purposes.

Here’s the story.

Into the Strait

Before crossing into Oman, the officer asked Citrini Analyst #3 to sign a document. The pledge — preprinted, presented over tea at a desert checkpoint — was an agreement not to engage in photography, journalism, or info gathering of any kind within the Sultanate of Oman. He signed it.

The officer then opened Analyst #3’s rugged Pelican case to inspect it. What he missed: the gimbal, the microphone kit, the recording sunglasses. The assignment was on.

On the other side, #3 talked his way onto a rickety speedboat with no GPS, defied advice from Omani officials to turn back, and rode on the open seas eighteen miles from the Iranian coast while Shahed drones flew overhead and Revolutionary Guard patrol boats ran patterns in the distance. He swam in the Strait of Hormuz smoking one of the Cubans he still had on him.

And then, he was intercepted by the Coast Guard, detained, and had his phone confiscated — before, eventually, making it back home to share everything he’d learned with us in an 8-hour debrief session.

This is what Analyst #3 found on his field trip to the Strait of Hormuz, from his point of view, with some key names, places and details of key events changed to protect the safety of anonymous sources. The quotes are based on his memory of the events, translated from the original Arabic. Which is the best we could do from an accuracy standpoint — considering #3’s phone, and all of the notes and photos on it, is thousands of miles away, likely being combed through by the Omani authorities as we speak.

I. The Idea

“What if I just went to the Strait of Hormuz?”

It’s the kind of question that starts as a joke — the kind of thing you say at two in the morning to yourself in bed that doesn’t survive contact with daylight, that joins the vast graveyard of schemes you were absolutely going to execute before you fell asleep and woke up a person with responsibilities again. But it wasn’t 2AM, and we weren’t in bed.

We were sitting in CitriniResearch’s offices, in Midtown Manhattan, watching the biggest geopolitical crisis of the decade unfold on our phones. We watched the most liquid markets in the world fluctuate like meme-coins in a game of Trump Tweet-AP Headline pong.

It was clear that nobody — literally nobody, not the analysts, not the correspondents, not the retired generals doing hits on cable news and least of all us — actually had any idea what was going on. Everyone was working from the same stale satellite imagery and the same unnamed Pentagon sources and the same AIS shipping data, which, as I would later discover, was missing roughly half of what was actually transiting the strait on any given day.

And, after all, wasn’t it our job to make confusing investment environments less confusing? I wanted to do that, I had the connections to make it happen (at least some of it) and it would make for a pretty great story. So it was decided.

From Citrini’s apartment in New York, we packed a Pelican case with a Xiaomi phone (that had a Leica camera with a 150x zoom, a souvenir from our trip to visit Robotics factories in China), an EPIRB, $15,000 in cash, a gimbal and a microphone kit. We sat down and researched an itinerary, working backwards from the questions we most wanted answered.

I’d land in Dubai, speak with some informed parties I knew and contacts of CitriniResearch’s, drive to Fujairah for some B-roll and information gathering at the oil terminal, then cross into northern Oman’s Musandam province, get to Khasab, and try to get on the water.

Read more

Here at Citrini Research, we’ve been talking about “Agents” for as long as we’ve been writing about AI.

Back in May 2023, in our Artificial Intelligence: Global Equity Beneficiaries piece, we predicted that AI adoption (and its investing implications) would follow a three-phase process:

• Phase I: Global Data Center Hyperscaling (“infrastructure”)

• Phase II: The Democratization of AI/ML (“commodification”)

• Phase III: Integration & Specialization (you’re here now)

Here’s how we described the vision at the time:

“The third phase is marked by the deeper integration of AI and ML into various industries and aspects of society, as well as further specialization within the field of AI itself. The technologies described as “artificial intelligence” may not resemble the LLMs we currently use, and one or more services will provide “agents” capable of carrying out tasks without being prompted directly (an “AI assistant”).”

This roadmap has proven prescient, though perhaps a better name for Phase III would be the “Agentic Era”.

We start today’s piece with this callback for a reason (not just to pat ourselves on the back). The reason is that we, like many others, have envisioned the coming of the agents pretty much straight out of the gate; AI was never going to be constrained to a chat bot.

In our follow-up piece, Less Deus More Machina, we zeroed in on “AI Losers” – companies and industries most vulnerable to disruption with the proliferation of AI (and, importantly, agentic AI). Our “standalone short” basket of 26 losers worked fairly well against a rising market for the following two years, albeit with a good deal of choppiness along the way. But, the waterfall really began late last year – just as the “agents” arrived.

We’ve reached an inflection. AI Agents have gone from an abstract concept to an actual service with expansive real world capabilities that can be directed from a simple iMessage. For the “losers”, the theoretical risks that were enough to stoke fear for investors have rapidly become tangible risks that fundamentally alter management outlooks.

The concept of an “AI Agent” went from buzzword to reality in November 2025 – reaching escape velocity with the release of ClawdBot (now OpenClaw, thanks to Anthropic’s lawyers). Until recently, the coding tools suffered the same flaw as every other hacked-together “AI” SaaS app: the call and response model. Due to degrading context windows, models couldn’t store history of past events so there was only so much that a single run could do. You were still limited by the command line, but not anymore.

In January, OpenClaw – an open-sourced, self-hosted agentic library – became the poster child for agentic technology. As shown by GitHub’s “Star History”, the adoption has been astronomical in just two months.

Along with exploding Github references, OpenRouter data shows just how dramatically OpenClaw has accelerated token usage within its product category.

OpenClaw went viral because it had the bright idea of storing context in a big text file and looping through new AI calls. Early users were running their agents overnight, and a new ecosystem of skills and plugins boosted its network effects. And, for just the cost of a Mac Mini, you could run it all via iMessage. We must reiterate that this is the core reason behind the ClawdBot/OpenClaw Mac Mini craze: it has nothing to do with the hardware. Most of these functions are run in the cloud rather than locally, although longtime readers know our thoughts on where that’s headed.

In this new paradigm, Jensen Huang is floating the idea of providing “token based compensation” to attract engineering talent. Jensen also recently commented that he would be ‘deeply alarmed’ if one of his engineers earning $500,000 a year did not consume at least $250,000 in tokens. While we’re cognizant of the fact that this is like a baker saying you should be eating a dozen donuts a day, there’s a lot of truth in it as well.

Homebrew Agents are still in “move fast and break things” mode. Agents are vibecoding, answering emails, dumping crypto wallets, and folding proteins. And companies have taken notice…

This tells us two things:

First, “Agentic AI” is now widely acknowledged as a critical market narrative.

Second, picking winners and losers in this space will be a whole lot harder than simply seeing who is talking about it on earnings calls.

Nevertheless, we expect this to be a powerful narrative for true “agentic winners” – particularly if it reshapes the investment case for companies in the dumps. As always, we love a great bottom.

We have been slowly adding to our allocation names that are aligned with the “Agentic Utility” layer - we are currently long AKAM, FSLY, CRCL and NET in the Citrindex. We’ve bought all of these when they were widely viewed as being entirely separate from the AI trade. Now, we think the moves in these names speak to how quickly stories and sentiment can reverse when the market recognizes an AI angle.

To understand where other agentic winners may lie, let’s first establish a framework for where to look for any novel beneficiaries of the new era of supercharged token consumption.

Framework: Agentic Utilities

Let’s start with a metaphor. In the chatbot era of AI, ChatGPT could help you plan dinner. It could help you pick out a menu, find recipes, and create a shopping list. It might even serve you ads that help you buy the ingredients. All that may be super helpful, but at the end of the day, it couldn’t actually make you dinner. Ultimately, you must use your own agency to drive to the store, slice the onions, and plate the chicken parm.

But imagine a world in which it can. Imagine if you had a service at your fingertips that could hop in a car, buy your groceries, pick up your laundry, and stop at the bank on the way home. It wouldn’t just stop at one grocery store, it would send ten cars out to every single grocery store in town to make sure you’re getting your Cheerios at the lowest price. It’s not free (it still chews tokens – and the “juice” of price comparison must be worth the compute), but as far as services go, it’s pretty darn useful – and in fact, you might use it a lot.

But what does this world actually look like?

First, the number of actual cars on the road explodes. Traffic patterns are reshaped and traffic jams become a big problem.

Second, the businesses that cater and optimize for these new task rabbits will see an influx in volume. Meanwhile, those that cater solely to human touch will lose market share.

Finally, these swarms of agents become a risk to individuals and businesses alike. Sheer overload might resemble a stampede even if intentions are good – even worse, they are just as easily directed towards nefarious ends.

For the immediate future, this metaphor applies mostly to the digital realm. The physical embodiment of AI via autonomous driving and robotics may be closer than we think, but we’ll hold off on this for now. The entire digital landscape of infrastructure, commercial interface, and security must adapt to a new paradigm.

We classify the winners into three categories.

1) Infrastructure: The Agentic Internet has a different structure than the one we are currently used to, both in terms of pathing and bandwidth. Digital plumbing needs to accommodate a boom in agentic traffic. This has unearthed legacy names that are already seeing inflections in both revenue and guidance that prove the demand has already arrived, and is rapidly growing.

2) Ecosystem: Simultaneously, a new customer category has emerged. Instead of selling to consumers (B2C) or businesses (B2B), companies will offer Agent (B2A) services.

Agentic Utilities are the first vendors in the B2A vertical – these are companies offering infrastructure and services to be utilized by agents that can’t be “vibecoded” away – for regulatory, operational, or path-dependent reasons.

Agents need ways to interact with the real world, and that includes payment rails. Stablecoins and agentic wallets are being quickly adopted, with non-humans making financial transactions using both crypto and the SWIFT system. Even robots can’t escape the dollar’s reserve currency status.

The evolution of the ecosystem will wind through the buildout of agentic plumbing (we’re seeing that now with the moves in the CDNs) and then begin pricing in their implementation. Value will accrue first to the picks and shovels, and then to the APIs in demand, the agentic harnesses (i.e. where the attention is) and the agents themselves.

3) Governance: And, the Agentic Internet needs new protections. Current AI capabilities make it possible for a rogue agent to break the internal systems at Meta, and that is the dumbest that agent will ever be. Every single company is asking “where is the killswitch, and how do we know when to use it?” This puts observability and counter-AI solutions in play. These names were thrown out with the bathwater when the entire software industry developed a new risk premium in under a quarter.

The companies that address these three problems today are laying the groundwork for the agentic future – we call them Agentic Utilities.

Read more

For readers who have recently joined us, our job as analysts is to wade through the noise and identify where the puck is going. We’ve been doing this since we first began publishing research in May 2023 with “AI: Global Equity Beneficiaries”, opting to focus on the second and third-order effects of the massive data center buildout necessary for AI. Today, the AI buildout remains relentless, with an estimated $650B–$700B in capital expenditures to be outlaid by hyperscalers in 2026.

Since then, we have woven a thread through the AI supply chain backed by our thesis that infrastructure bottlenecks would dominate investor attention. As each layer of the supply chain is reinforced and expanded, the bottlenecks simultaneously respond to technological adoption and drive new demand in the infrastructure layer. In order to seek out these opportunities, we’ve continually deepened and expanded our existing coverage. Today, that leads us to update what was perhaps the single-most emblematic stance of our “secular growth at cyclical prices” approach: optics.

CitriniResearch was one of the most timely on the street to outline the opportunity for optics. For two months, with most of the stocks at or near 52 week lows, we pounded the table on this in two separate pieces.

In “Interconnects 101” published in July 2024, we argued that AI’s next beneficiaries would become “more skewed towards who can provide the most marginal benefit,” specifically “optical interconnects and co-packaged optics (CPO)”. A few weeks later, in “Can You Hear Me Now? The Coming Optics Shortage & Telco Supplier’s New Dawn”, we made an even bolder and more bullish call that presented these cyclically blown out names as becoming a prominent AI bottleneck. We even compared Ciena’s (CIEN US) revision cycle in 2024 to Nvidia’s going into 2023. In other words, our first call was that optics would matter a lot and the second was that the shortage would emerge before the market even noticed.

This was not a popular take. At the time, optics sat inside a telecom supply chain most investors wanted nothing to do with. Consensus saw cyclical depression and inventory indigestion that was years from normalizing. We saw AI-driven connectivity demand colliding with years of underinvestment, setting up “a resulting optics shortage” precisely where the market was least interested in looking.

If you can believe it, given the way our optics names have traded since, many laughed at us for it. We were branded contrarians at best and tourists at worst, wading into a meat-grinder of a cycle that would end our streak of good calls. Well…they’re not laughing anymore. These calls were early, uncomfortable, and very right.

Our connectivity basket has more than tripled since we published it – recently pulling ahead of our Interconnects basket, which has lagged year-to-date as the market has come to appreciate optical interconnects at the expense of names like Astera Labs (ALAB US) and Credo (CRDO US). And it’s not done yet – the basket is up 35% YTD, led by now favorite Lumentum (LITE US).

This piece is the next chapter. The optics trade has broadened and matured, certainly even crowded in some names. The risk-reward on some of the crowd favorites has become significantly more symmetrical. Since it is now well established that optics are core to the AI trade, we must ask where the next asymmetry sits. Despite more than enough reason to claim the win and walk off…we believe there are still opportunities further down the stack.

As the bandwidth constraints of the “Memory Wall” pile up, the opportunities are in the companies enabling the next generation – photonic components, substrates, capital equipment, and the businesses positioned to benefit as CPO becomes a necessity.

We believe there is ample runway and enough yet-to-be crowded names in this sub-theme to provide continued outperformance for AI investors. Additionally, we think it’s time to be cautious and consider taking profits on some of the names we originally initiated on. Not because optics slows down, just because they’ve become priced for perfection.

Below the paywall, we provide a brief overview of what “photonics” are and outline our favorites of the picks and shovels the market has overlooked that we think remain undervalued in the shadow of LITE and peer’s massive run.

We’ve also teamed up again with financial investigative journalists Hunterbrook to dive into what we believe is a truly asymmetric play on the photonics supply chain. You’ll remember them from our last collaboration covering the dual robotics and memory testing tailwind of Teradyne (TER US) that provided our readers with an extremely successful play.

Read more

What if our AI bullishness continues to be right...and what if that’s actually bearish?

What follows is a scenario, not a prediction. This isn’t bear porn or AI doomer fan-fiction. The sole intent of this piece is modeling a scenario that’s been relatively underexplored. Our friend Alap Shah posed the question, and together we brainstormed the answer. We wrote this part, and he’s written two others you can find here.

Hopefully, reading this leaves you more prepared for potential left tail risks as AI makes the economy increasingly weird.

This is the CitriniResearch Macro Memo from June 2028, detailing the progression and fallout of the Global Intelligence Crisis.

Macro Memo

The Consequences of Abundant Intelligence

CitriniResearch

February 22nd, 2026 June 30th, 2028

The unemployment rate printed 10.2% this morning, a 0.3% upside surprise. The market sold off 2% on the number, bringing the cumulative drawdown in the S&P to 38% from its October 2026 highs.

Traders have grown numb. Six months ago, a print like this would have triggered a circuit breaker.

Two years. That’s all it took to get from “contained” and “sector-specific” to an economy that no longer resembles the one any of us grew up in. This quarter’s macro memo is our attempt to reconstruct the sequence - a post-mortem on the pre-crisis economy.

The euphoria was palpable. By October 2026, the S&P 500 flirted with 8000, the Nasdaq broke above 30k. The initial wave of layoffs due to human obsolescence began in early 2026, and they did exactly what layoffs are supposed to. Margins expanded, earnings beat, stocks rallied. Record-setting corporate profits were funneled right back into AI compute.

The headline numbers were still great. Nominal GDP repeatedly printed mid-to-high single-digit annualized growth. Productivity was booming. Real output per hour rose at rates not seen since the 1950s, driven by AI agents that don’t sleep, take sick days or require health insurance.

The owners of compute saw their wealth explode as labor costs vanished. Meanwhile, real wage growth collapsed. Despite the administration’s repeated boasts of record productivity, white-collar workers lost jobs to machines and were forced into lower-paying roles.

When cracks began appearing in the consumer economy, economic pundits popularized the phrase “Ghost GDP“: output that shows up in the national accounts but never circulates through the real economy.

In every way AI was exceeding expectations, and the market was AI. The only problem…the economy was not.

It should have been clear all along that a single GPU cluster in North Dakota generating the output previously attributed to 10,000 white-collar workers in midtown Manhattan is more economic pandemic than economic panacea. The velocity of money flatlined. The human-centric consumer economy, 70% of GDP at the time, withered. We probably could have figured this out sooner if we just asked how much money machines spend on discretionary goods. (Hint: it’s zero.)

AI capabilities improved, companies needed fewer workers, white collar layoffs increased, displaced workers spent less, margin pressure pushed firms to invest more in AI, AI capabilities improved…

It was a negative feedback loop with no natural brake. The human intelligence displacement spiral. White-collar workers saw their earnings power (and, rationally, their spending) structurally impaired. Their incomes were the bedrock of the $13 trillion mortgage market - forcing underwriters to reassess whether prime mortgages are still money good.

Seventeen years without a real default cycle had left privates bloated with PE-backed software deals that assumed ARR would remain recurring. The first wave of defaults due to AI disruption in mid-2027 challenged that assumption.

This would have been manageable if the disruption remained contained to software, but it didn’t. By the end of 2027, it threatened every business model predicated on intermediation. Swaths of companies built on monetizing friction for humans disintegrated.

The system turned out to be one long daisy chain of correlated bets on white-collar productivity growth. The November 2027 crash only served to accelerate all of the negative feedback loops already in place.

We’ve been waiting for “bad news is good news” for almost a year now. The government is starting to consider proposals, but public faith in the ability of the government to stage any sort of rescue has dwindled. Policy response has always lagged economic reality, but lack of a comprehensive plan is now threatening to accelerate a deflationary spiral.

How It Started

In late 2025, agentic coding tools took a step function jump in capability.

A competent developer working with Claude Code or Codex could now replicate the core functionality of a mid-market SaaS product in weeks. Not perfectly or with every edge case handled, but well enough that the CIO reviewing a $500k annual renewal started asking the question “what if we just built this ourselves?”

Fiscal years mostly line up with calendar years, so 2026 enterprise spend had been set in Q4 2025, when “agentic AI” was still a buzzword. The mid-year review was the first time procurement teams were making decisions with visibility into what these systems could actually do. Some watched their own internal teams spin up prototypes replicating six-figure SaaS contracts in weeks.

That summer, we spoke with a procurement manager at a Fortune 500. He told us about one of his budget negotiations. The salesperson had expected to run the same playbook as last year: a 5% annual price increase, the standard “your team depends on us” pitch. The procurement manager told him he’d been in conversations with OpenAI about having their “forward deployed engineers” use AI tools to replace the vendor entirely. They renewed at a 30% discount. That was a good outcome, he said. The “long-tail of SaaS”, like Monday.com, Zapier and Asana, had it much worse.

Investors were prepared - expectant, even - that the long tail would be hit hard. They may have made up a third of spending for the typical enterprise stack, but they were obviously exposed. The systems of record, however, were supposed to be safe from disruption.

It wasn’t until ServiceNow’s Q3 26 report that the mechanism of reflexivity became clearer.

SERVICENOW NET NEW ACV GROWTH DECELERATES TO 14% FROM 23%; ANNOUNCES 15% WORKFORCE REDUCTION AND ‘STRUCTURAL EFFICIENCY PROGRAM’; SHARES FALL 18% | Bloomberg, October 2026

SaaS wasn’t “dead”. There was still a cost-benefit-analysis to running and supporting in-house builds. But in-house was an option, and that factored into pricing negotiations. Perhaps more importantly, the competitive landscape had changed. AI had made it easier to develop and ship new features, so differentiation collapsed. Incumbents were in a race to the bottom on pricing - a knife-fight with both each other and with the new crop of upstart challengers that popped up. Emboldened by the leap in agentic coding capabilities and with no legacy cost structure to protect, these aggressively took share.

The interconnected nature of these systems weren’t fully appreciated until this print, either. ServiceNow sold seats. When Fortune 500 clients cut 15% of their workforce, they cancelled 15% of their licenses. The same AI-driven headcount reductions that were boosting margins at their customers were mechanically destroying their own revenue base.

The company that sold workflow automation was being disrupted by better workflow automation, and its response was to cut headcount and use the savings to fund the very technology disrupting it.

What else were they supposed to do? Sit still and die slower? The companies most threatened by AI became AI’s most aggressive adopters.

This sounds obvious in hindsight, but it really wasn’t at the time (at least to me). The historical disruption model said incumbents resist new technology, they lose share to nimble entrants and die slowly. That’s what happened to Kodak, to Blockbuster, to BlackBerry. What happened in 2026 was different; the incumbents didn’t resist because they couldn’t afford to.

With stocks down 40-60% and boards demanding answers, the AI-threatened companies did the only thing they could. Cut headcount, redeploy the savings into AI tools, use those tools to maintain output with lower costs.

Each company’s individual response was rational. The collective result was catastrophic. Every dollar saved on headcount flowed into AI capability that made the next round of job cuts possible.

Software was only the opening act. What investors missed while they debated whether SaaS multiples had bottomed was that the reflexive loop had already escaped the software sector. The same logic that justified ServiceNow cutting headcount applied to every company with a white-collar cost structure.

When Friction Went to Zero

By early 2027, LLM usage had become default. People were using AI agents who didn’t even know what an AI agent was, in the same way people who never learned what “cloud computing” was used streaming services. They thought of it the same way they thought of autocomplete or spell-check - a thing their phone just did now.

Qwen’s open-source agentic shopper was the catalyst for AI handling consumer decisions. Within weeks, every major AI assistant had integrated some agentic commerce feature. Distilled models meant these agents could run on phones and laptops, not just cloud instances, reducing the marginal cost of inference significantly.

The part that should have unsettled investors more than it did was that these agents didn’t wait to be asked. They ran in the background according to the user’s preferences. Commerce stopped being a series of discrete human decisions and became a continuous optimization process, running 24/7 on behalf of every connected consumer. By March 2027, the median individual in the United States was consuming 400,000 tokens per day - 10x since the end of 2026.

The next link in the chain was already breaking.

Intermediation.

Over the past fifty years, the U.S. economy built a giant rent-extraction layer on top of human limitations: things take time, patience runs out, brand familiarity substitutes for diligence, and most people are willing to accept a bad price to avoid more clicks. Trillions of dollars of enterprise value depended on those constraints persisting.

It started out simple enough. Agents removed friction.

Subscriptions and memberships that passively renewed despite months of disuse. Introductory pricing that sneakily doubled after the trial period. Each one was rebranded as a hostage situation that agents could negotiate. The average customer lifetime value, the metric the entire subscription economy was built on, distinctly declined.

Consumer agents began to change how nearly all consumer transactions worked.

Humans don’t really have the time to price-match across five competing platforms before buying a box of protein bars. Machines do.

Travel booking platforms were an early casualty, because they were the simplest. By Q4 2026, our agents could assemble a complete itinerary (flights, hotels, ground transport, loyalty optimization, budget constraints, refunds) faster and cheaper than any platform.

Insurance renewals, where the entire renewal model depended on policyholder inertia, were reformed. Agents that re-shop your coverage annually dismantled the 15-20% of premiums that insurers earned from passive renewals.

Financial advice. Tax prep. Routine legal work. Any category where the service provider’s value proposition was ultimately “I will navigate complexity that you find tedious” was disrupted, as the agents found nothing tedious.

Even places we thought insulated by the value of human relationships proved fragile. Real estate, where buyers had tolerated 5-6% commissions for decades because of information asymmetry between agent and consumer, crumbled once AI agents equipped with MLS access and decades of transaction data could replicate the knowledge base instantly. A sell-side piece from March 2027 titled it “agent on agent violence”. The median buy-side commission in major metros had compressed from 2.5-3% to under 1%, and a growing share of transactions were closing with no human agent on the buy side at all.

We had overestimated the value of “human relationships”. Turns out that a lot of what people called relationships was simply friction with a friendly face.

That was just the start of the disruption for the intermediation layer. Successful companies had spent billions to effectively exploit quirks of consumer behavior and human psychology that didn’t matter anymore.

Machines optimizing for price and fit do not care about your favorite app or the websites you’ve been habitually opening for the last four years, nor feel the pull of a well-designed checkout experience. They don’t get tired and accept the easiest option or default to “I always just order from here”.

That destroyed a particular kind of moat: habitual intermediation.

DoorDash (DASH US) was the poster child.

Coding agents had collapsed the barrier to entry for launching a delivery app. A competent developer could deploy a functional competitor in weeks, and dozens did, enticing drivers away from DoorDash and Uber Eats by passing 90-95% of the delivery fee through to the driver. Multi-app dashboards let gig workers track incoming jobs from twenty or thirty platforms at once, eliminating the lock-in that the incumbents depended on. The market fragmented overnight and margins compressed to nearly nothing.

Agents accelerated both sides of the destruction. They enabled the competitors and then they used them. The DoorDash moat was literally “you’re hungry, you’re lazy, this is the app on your home screen.” An agent doesn’t have a home screen. It checks DoorDash, Uber Eats, the restaurant’s own site, and twenty new vibe-coded alternatives so it can pick the lowest fee and fastest delivery every time.

Habitual app loyalty, the entire basis of the business model, simply didn’t exist for a machine.

This was oddly poetic, as perhaps the only example in this entire saga of agents doing a favor for the soon-to-be-displaced white collar workers. When they ended up as delivery drivers, at least half their earnings weren’t going to Uber and DoorDash. Of course, this favor from technology didn’t last for long as autonomous vehicles proliferated.

Once agents controlled the transaction, they went looking for bigger paperclips.

There was only so much price-matching and aggregating to do. The biggest way to repeatedly save the user money (especially when agents started transacting among themselves) was to eliminate fees. In machine-to-machine commerce, the 2-3% card interchange rate became an obvious target.

Agents went looking for faster and cheaper options than cards. Most settled on using stablecoins via Solana or Ethereum L2s, where settlement was near-instant and the transaction cost was measured in fractions of a penny.

MASTERCARD Q1 2027: NET REVENUES +6% Y/Y; PURCHASE VOLUME GROWTH SLOWS TO +3.4% Y/Y FROM +5.9% PRIOR QUARTER; MANAGEMENT NOTES “AGENT-LED PRICE OPTIMIZATION” AND “PRESSURE IN DISCRETIONARY CATEGORIES” | Bloomberg, April 29 2027

Mastercard’s Q1 2027 report was the point of no return. Agentic commerce went from being a product story to a plumbing story. MA dropped 9% the following day. Visa did too, but pared losses after analysts pointed out its stronger positioning in stablecoin infrastructure.

Agentic commerce routing around interchange posed a far greater risk to card-focused banks and mono-line issuers, who collected the majority of that 2-3% fee and had built entire business segments around rewards programs funded by the merchant subsidy.

American Express (AXP US) was hit hardest; a combined headwind from white-collar workforce reductions gutting its customer base and agents routing around interchange gutting its revenue model. Synchrony (SYF US), Capital One (COF US) and Discover (DFS US) all fell more than 10% over the following weeks, as well.

Their moats were made of friction. And friction was going to zero.

From Sector Risk to Systemic Risk

Through 2026, markets treated negative AI impact as a sector story. Software and consulting were getting crushed, payments and other toll booths were wobbly, but the broader economy seemed fine. The labor market, while softening, was not in freefall. The consensus view was that creative destruction was part of any technological innovation cycle. It would be painful in pockets, but the overall net positives from AI would outweigh any negatives.

Our January 2027 macro memo argued this was the wrong mental model. The US economy is a white-collar services economy. White-collar workers represented 50% of employment and drove roughly 75% of discretionary consumer spending. The businesses and jobs that AI was chewing up were not tangential to the US economy, they were the US economy.

“Technological innovation destroys jobs and then creates even more”. This was the most popular and convincing counter-argument at the time. It was popular and convincing because it’d been right for two centuries. Even if we couldn’t conceive of what the future jobs would be, they would surely arrive.

ATMs made branches cheaper to operate so banks opened more of them and teller employment rose for the next twenty years. The internet disrupted travel agencies, the Yellow Pages, brick-and-mortar retail, but it invented entirely new industries in their place that conjured new jobs.

Every new job, however, required a human to perform it.

AI is now a general intelligence that improves at the very tasks humans would redeploy to. Displaced coders cannot simply move to “AI management” because AI is already capable of that.

Today, AI agents handle many-weeks-long research and development tasks. The exponential steamrolled our conceptions of what was possible, even though every year Wharton professors tried to fit the data to a new sigmoid.

They write essentially all code. The highest performing of them are substantially smarter than almost all humans at almost all things. And they keep getting cheaper.

AI has created new jobs. Prompt engineers. AI safety researchers. Infrastructure technicians. Humans are still in the loop, coordinating at the highest level or directing for taste. For every new role AI created, though, it rendered dozens obsolete. The new roles paid a fraction of what the old ones did.

U.S. JOLTS: JOB OPENINGS FALL BELOW 5.5M; UNEMPLOYED-TO-OPENINGS RATIO CLIMBS TO ~1.7, HIGHEST SINCE AUG 2020 | Bloomberg, Oct 2026

The hiring rate had been anemic all year, but October ‘26 JOLTS print provided some definitive data. Job openings fell below 5.5 million, a 15% decline YoY.

INDEED: POSTINGS FALL SHARPLY IN SOFTWARE, FINANCE, CONSULTING AS “PRODUCTIVITY INITIATIVES” SPREAD | Indeed Hiring Lab, Nov–Dec 2026

White-collar openings were collapsing while blue-collar openings remained relatively stable (construction, healthcare, trades). The churn was in the jobs that write memos (we are, somehow, still in business), approve budgets, and keep the middle layers of the economy lubricated. Real wage growth in both cohorts, however, had been negative for the majority of the year and kept declining.

The equity market still cared less about JOLTS than it did the news that all of GE Vernova’s turbine capacity was now sold out until 2040, it ambled sideways in a tug of war between negative macro news with positive AI infrastructure headlines.

The bond market (always smarter than equities, or at least less romantic) began pricing the consumption hit, however. The 10-year yield began a descent from 4.3% to 3.2% over the following four months. Still, the headline unemployment rate did not blow out, the composition nuance was still lost on some.

In a normal recession, the cause eventually self-corrects. Overbuilding leads to a construction slowdown, which leads to lower rates, which leads to new construction. Inventory overshoot leads to destocking, which leads to restocking. The cyclical mechanism contains within it its own seeds of recovery.

This cycle’s cause was not cyclical.

AI got better and cheaper. Companies laid off workers, then used the savings to buy more AI capability, which let them lay off more workers. Displaced workers spent less. Companies that sell things to consumers sold fewer of them, weakened, and invested more in AI to protect margins. AI got better and cheaper.

A feedback loop with no natural brake.

The intuitive expectation was that falling aggregate demand would slow the AI buildout. It didn’t, because this wasn’t hyperscaler-style CapEx. It was OpEx substitution. A company that had been spending $100M a year on employees and $5M on AI now spent $70M on employees and $20M on AI. AI investment increased by multiples, but it occurred as a reduction in total operating costs. Every company’s AI budget grew while its overall spending shrank.

The irony of this was that the AI infrastructure complex kept performing even as the economy it was disrupting began deteriorating. NVDA was still posting record revenues. TSM was still running at 95%+ utilization. The hyperscalers were still spending $150-200 billion per quarter on data center capex. Economies that were purely convex to this trend, like Taiwan and Korea, outperformed massively.

India was the inverse. The country’s IT services sector exported over $200 billion annually, the single largest contributor to India’s current account surplus and the offset that financed its persistent goods trade deficit. The entire model was built on one value proposition: Indian developers cost a fraction of their American counterparts. But the marginal cost of an AI coding agent had collapsed to, essentially, the cost of electricity. TCS, Infosys and Wipro saw contract cancellations accelerate through 2027. The rupee fell 18% against the dollar in four months as the services surplus that had anchored India’s external accounts evaporated. By Q1 2028, the IMF had begun “preliminary discussions” with New Delhi.

The engine that caused the disruption got better every quarter, which meant the disruption accelerated every quarter. There was no natural floor to the labor market.

In the US, we weren’t asking about how the bubble would burst in AI infrastructure anymore. We were asking what happens to a consumer-credit economy when consumers are being replaced with machines.

The Intelligence Displacement Spiral

2027 was when the macroeconomic story stopped being subtle. The transmission mechanism from the previous twelve months of disjointed but clearly negative developments became obvious. You didn’t need to go into the BLS data. Just attend a dinner party with friends.

Displaced white-collar workers did not sit idle. They downshifted. Many took lower-paying service sector and gig economy jobs, which increased labor supply in those segments and compressed wages there too.

A friend of ours was a senior product manager at Salesforce in 2025. Title, health insurance, 401k, $180,000 a year. She lost her job in the third round of layoffs. After six months of searching, she started driving for Uber. Her earnings dropped to $45,000. The point is less the individual story and more the second-order math. Multiply this dynamic by a few hundred thousand workers across every major metro. Overqualified labor flooding the service and gig economy pushed down wages for existing workers who were already struggling. Sector-specific disruption metastasized into economy-wide wage compression.

The pool of remaining human-centric had another correction ahead of it, happening while we write this. As autonomous delivery and self-driving vehicles work their way through the gig economy that absorbed the first wave of displaced workers.

By February 2027, it was clear that still employed professionals were spending like they might be next. They were working twice as hard (mostly with the help of AI) just to not get fired, hopes of promotion or raises were gone. Savings rates ticked higher and spending softened.

The most dangerous part was the lag. High earners used their higher-than-average savings to maintain the appearance of normalcy for two or three quarters. The hard data didn’t confirm the problem until it was already old news in the real economy. Then came the print that broke the illusion.

U.S. INITIAL JOBLESS CLAIMS SURGE TO 487,000, HIGHEST SINCE APRIL 2020; Department of Labor, Q3 2027

Initial claims surged to 487,000, the highest since April 2020. ADP and Equifax confirmed that the overwhelming majority of new filings were from white-collar professionals.

The S&P dropped 6% over the following week. Negative macro started winning the tug of war.

In a normal recession, job losses are broadly distributed. Blue-collar and white-collar workers share the pain roughly in proportion to each segment’s share of employment. The consumption hit is also broadly distributed, and it shows up quickly in the data because lower-income workers have higher marginal propensities to consume.

In this cycle, the job losses have been concentrated in the upper deciles of the income distribution. They are a relatively small share of total employment, but they drive a wildly disproportionate share of consumer spending. The top 10% of earners account for more than 50% of all consumer spending in the United States. The top 20% account for roughly 65%. These are the people who buy the houses, the cars, the vacations, the restaurant meals, the private school tuition, the home renovations. They are the demand base for the entire consumer discretionary economy.

When these workers lost their jobs, or took 50% pay cuts to move into available roles, the consumption hit was enormous relative to the number of jobs lost. A 2% decline in white-collar employment translated to something like a 3-4% hit to discretionary consumer spending. Unlike blue-collar job losses, which tend to hit immediately (you get laid off from the factory, you stop spending next week), white-collar job losses have a lagged but deeper impact because these workers have savings buffers that allow them to maintain spending for a few months before the behavioral shift kicks in.

By Q2 2027, the economy was in recession. The NBER would not officially date the start until months later (they never do) but the data was unambiguous - we’d had two consecutive quarters of negative real GDP growth. But it wasn’t a “financial crisis”…yet.

The Daisy Chain of Correlated Bets

Private credit had grown from under $1 trillion in 2015 to over $2.5 trillion by 2026. A meaningful share of that capital had been deployed into software and technology deals, many of them leveraged buyouts of SaaS companies at valuations that assumed mid-teens revenue growth in perpetuity.

Those assumptions died somewhere between the first agentic coding demo and the Q1 2026 software crash, but the marks didn’t seem to realize they were dead.

As many public SaaS companies traded to 5-8x EBITDA, PE-backed software companies sat on balance sheets at marks reflecting acquisition valuations on multiples of revenue that didn’t exist anymore. Managers eased the marks down gradually, 100 cents, 92, 85, all while public comps said 50.

MOODY’S DOWNGRADES $18B OF PE-BACKED SOFTWARE DEBT ACROSS 14 ISSUERS, CITING ‘SECULAR REVENUE HEADWINDS FROM AI-DRIVEN COMPETITIVE DISRUPTION’; LARGEST SINGLE-SECTOR ACTION SINCE ENERGY IN 2015 | Moody’s Investors Service, April 2027

Everyone remembers what happened after the downgrade. Industry veterans had already seen the playbook following the 2015 energy downgrades.

Software-backed loans began defaulting in Q3 2027. PE portfolio companies in information services and consulting followed. Several multi-billion dollar LBOs of well-known SaaS companies entered restructuring.

Zendesk was the smoking gun.

ZENDESK MISSES DEBT COVENANTS AS AI-DRIVEN CUSTOMER SERVICE AUTOMATION ERODES ARR; $5B DIRECT LENDING FACILITY MARKED TO 58 CENTS; LARGEST PRIVATE CREDIT SOFTWARE DEFAULT ON RECORD | Financial Times, September 2027

In 2022, Hellman & Friedman and Permira had taken Zendesk private for $10.2 billion. The debt package was $5 billion in direct lending, the largest ARR-backed facility in history at the time, led by Blackstone with Apollo, Blue Owl and HPS all in the lending group. The loan was explicitly structured around the assumption that Zendesk’s annual recurring revenue would remain recurring. At roughly 25x EBITDA, the leverage only made sense if it did.

By mid-2027, it didn’t.

AI agents had been handling customer service autonomously for the better part of a year. The category Zendesk had defined (ticketing, routing, managing human support interactions) was already replaced by systems that resolved issues without generating a ticket at all. The Annualized Recurring Revenue the loan was underwritten against was no longer recurring, it was just revenue that hadn’t left yet.

The largest ARR-backed loan in history became the largest private credit software default in history. Every credit desk asked the same question at once: who else has a secular headwind disguised as a cyclical one?

But here’s what the consensus got right, at least initially: this should have been survivable.

Private credit is not 2008 banking. The whole architecture was explicitly designed to avoid forced selling. These are closed-end vehicles with locked-up capital. LPs committed for seven to ten years. There are no depositors to run, no repo lines to pull. The managers could sit on impaired assets, work them out over time, and wait for recoveries. Painful, but manageable. The system was such that it was supposed to bend, not break.

Executives at Blackstone, KKR and Apollo cited software exposure of 7-13% of assets. Containable. Every sell-side note and fintwit credit account said the same thing: private credit has permanent capital. They could absorb losses that would otherwise blow up a levered bank.

Permanent capital. The phrase showed up in every earnings call and investor letter meant to reassure. It became a mantra. And like most mantras, nobody paid attention to the finer details. Here’s what it actually meant…

Over the prior decade, the large alternative asset managers had acquired life insurance companies and turned them into funding vehicles. Apollo bought Athene. Brookfield bought American Equity. KKR took Global Atlantic. The logic was elegant: annuity deposits provided a stable, long-duration liability base. The managers invested those deposits into the private credit they originated and got paid twice, earning spread over on the insurance side and management fees on the asset management side. A fee-on-fee perpetual motion machine that worked beautifully under one condition.

The private credit had to be money good.

The losses hit balance sheets built to hold illiquid assets against long-duration obligations. The “permanent capital” that was supposed to make the system resilient was not some abstract pool of patient institutional money and sophisticated investors taking sophisticated risk. It was the savings of American households, “Main Street”, structured as annuities invested in the same PE-backed software and technology paper that was now defaulting. The locked-up capital that couldn’t run was life insurance policyholder money, and the rules are a bit different there.

Compared to the banking system, insurance regulators had been docile - even complacent - but this was the wake-up call. Already uneasy about private credit concentrations at life insurers, they began downgrading the risk-based capital treatment of these assets. That forced the insurers to either raise capital or sell assets, neither of which was possible at attractive terms in a market already seizing up.

NEW YORK, IOWA STATE REGULATORS MOVE TO TIGHTEN CAPITAL TREATMENT FOR CERTAIN PRIVATELY RATED CREDIT HELD BY LIFE INSURERS; NAIC GUIDANCE EXPECTED TO INCREASE RBC FACTORS AND TRIGGER ADDITIONAL SVO SCRUTINY | Reuters, Nov 2027

When Moody’s put Athene’s financial strength rating on negative outlook, Apollo’s stock dropped 22% in two sessions. Brookfield, KKR, and the others followed.

It only got more complex from there. These firms hadn’t just created their insurer perpetual motion machine, they’d built an elaborate offshore architecture designed to maximize returns through regulatory arbitrage.The US insurer wrote the annuity, then ceded the risk to an affiliated Bermuda or Cayman reinsurer it also owned - set up to take advantage of more flexible regulation that permitted holding less capital against the same assets. That affiliate raised outside capital through offshore SPVs, a new layer of counterparties who invested alongside insurers into private credit originated by the same parent’s asset management arm.

The ratings agencies, some of which were themselves PE-owned, had not been paragons of transparency (surprising to virtually) no one. The spider web of different firms linked to different balance sheets was stunning in its opacity. When the underlying loans defaulted, the question of who actually bore the loss was genuinely unanswerable in real time.

The November 2027 crash marked the transition of perception from a potentially garden-variety cyclical drawdown to something much more uncomfortable. “A daisy chain of correlated bets on white collar productivity growth” was what Fed Chair Kevin Warsh called it during the FOMC’s emergency November meeting.

See, it is never the losses themselves that cause the crisis. It’s recognizing them. And there is another, much larger, much much more important area of finance for which we have grown fearful of that recognition.

The Mortgage Question

ZILLOW HOME VALUE INDEX FALLS 11% YOY IN SAN FRANCISCO, 9% IN SEATTLE, 8% IN AUSTIN; FANNIE MAE FLAGS ‘ELEVATED EARLY-STAGE DELINQUENCIES’ IN ZIP CODES WITH >40% TECH/FINANCE EMPLOYMENT | Zillow / Fannie Mae, June 2028

This month the Zillow Home Value Index fell 11% year-over-year in San Francisco, 9% in Seattle and 8% in Austin. This hasn’t been the only worrying headline. Last month, Fannie Mae flagged higher early-stage delinquency from jumbo-heavy ZIP codes - areas that are populated by 780+ credit score borrowers and typically “bulletproof”.

The US residential mortgage market is approximately $13 trillion. Mortgage underwriting is built on the fundamental assumption that the borrower will remain employed at roughly their current income level for the duration of the loan. For thirty years, in the case of most mortgages.

The white-collar employment crisis has threatened this assumption with a sustained shift in income expectations. We now have to ask a question that seemed absurd just 3 years ago - are prime mortgages money good?

Every prior mortgage crisis in US history has been driven by one of three things: speculative excess (lending to people who couldn’t afford the homes, as in 2008), interest rate shocks (rising rates making adjustable-rate mortgages unaffordable, as in the early 1980s), or localized economic shocks (a single industry collapsing in a single region, like oil in Texas in the 1980s or auto in Michigan in 2009).

None of these apply here. The borrowers in question are not subprime. They’re 780 FICO scores. They put 20% down. They have clean credit histories, stable employment records, and incomes that were verified and documented at origination. They were the borrowers that every risk model in the financial system treats as the bedrock of credit quality.

In 2008, the loans were bad on day one. In 2028, the loans were good on day one. The world just…changed after the loans were written. People borrowed against a future they can no longer afford to believe in.

In 2027, we flagged early signs of invisible stress: HELOC draws, 401(k) withdrawals, and credit card debt spiking while mortgage payments remained current. As jobs were lost, hiring was frozen and bonuses cut, these prime households saw their debt-to-income ratios double.

They could still make the mortgage payment, but only by stopping all discretionary spending, draining savings, and deferring any home maintenance or improvement. They were technically current on their mortgage, but just one more shock away from distress, and the trajectory of AI capabilities suggested that shock is coming. Then we saw delinquencies begin to spike in San Francisco, Seattle, Manhattan and Austin, even as the national average stayed within historical norms.

We’re now in the most acute stage. Falling home prices are manageable when the marginal buyer is healthy. Here, the marginal buyer is dealing with the same income impairment.

While concerns are building, we are not yet in a full-blown mortgage crisis. Delinquencies have risen but remain well below 2008 levels. It is the trajectory that’s the real threat.

The Intelligence Displacement Spiral now has two financial accelerants to the real economy’s decline.

Labor displacement, mortgage concerns, private market turmoil. Each reinforces the other. And the traditional policy toolkit (rate cuts, QE) can address the financial engine but cannot address the real economy engine, because the real economy engine is not driven by tight financial conditions. It’s driven by AI making human intelligence less scarce and less valuable. You can cut rates to zero and buy every MBS and all the defaulted software LBO debt in the market…

It won’t change the fact that a Claude agent can do the work of a $180,000 product manager for $200/month.

If these fears manifest, the mortgage market cracks in the back half of this year. In that scenario, we’d expect the current drawdown in equities to ultimately rival that of the GFC (57% peak-to-trough). This would bring the S&P500 to ~3500 - levels we haven’t seen since the month before the ChatGPT moment in November 2022.

What’s clear is that the income assumptions underlying $13 trillion in residential mortgages are structurally impaired. What isn’t is whether policy can intervene before the mortgage market fully processes what this means. We’re hopeful, but we can’t deny the reasons not to be.

The Battle Against Time

The first negative feedback loop was in the real economy: AI capability improves, payroll shrinks, spending softens, margins tighten, companies buy more capability, capability improves. Then it turned financial: income impairment hit mortgages, bank losses tightened credit, the wealth effect cracked, and the feedback loop sped up. And both of these have been exacerbated by an insufficient policy response from a government that seems, quite frankly, confused.

The system wasn’t designed for a crisis like this. The federal government’s revenue base is essentially a tax on human time. People work, firms pay them, the government takes a cut. Individual income and payroll taxes are the spine of receipts in normal years.

Through Q1 of this year, federal receipts were running 12% below CBO baseline projections. Payroll receipts are falling because fewer people are employed at prior compensation levels. Income tax receipts are falling because the incomes being earned are structurally lower. Productivity is surging, but the gains are flowing to capital and compute, not labor.

Labor’s share of GDP declined from 64% in 1974 to 56% in 2024, a four-decade grind lower driven by globalization, automation, and the steady erosion of worker bargaining power. In the four years since AI began its exponential improvement, that has dropped to 46%. The sharpest decline on record.

The output is still there. But it’s no longer routing through households on the way back to firms, which means it’s no longer routing through the IRS either. The circular flow is breaking, and the government is expected to step in to fix that.

As in every downturn, outlays rise just as receipts fall. The difference this time is that the spending pressure is not cyclical. Automatic stabilizers were built for temporary job losses, not structural displacement. The system is paying benefits that assume workers will be reabsorbed. Many will not, at least not at anything like their prior wage. During COVID, the government freely embraced 15% deficits, but it was understood to be temporary. The people who need government support today were not hit by a pandemic they’ll recover from. They were replaced by a technology that continues to improve.

The government needs to transfer more money to households at precisely the moment it is collecting less money from them in taxes.

The U.S. won’t default. It prints the currency it spends, the same currency it uses to pay back borrowers. But this stress has shown up elsewhere. Municipal bonds are showing worrying signs of dispersion in year-to-date performance. States without income tax have been okay, but general obligation munis issued by states dependent on income tax (majority blue states) began to price in some default risk. Politicos caught on quickly, and the debate over who gets bailed out has fallen along partisan lines.

The administration, to its credit, recognized the structural nature of the crisis early and began entertaining bipartisan proposals for what they’re calling the “Transition Economy Act”: a framework for direct transfers to displaced workers funded by a combination of deficit spending and a proposed tax on AI inference compute.

The most radical proposal on the table goes further. The “Shared AI Prosperity Act” would establish a public claim on the returns of the intelligence infrastructure itself, something between a sovereign wealth fund and a royalty on AI-generated output, with dividends funding household transfers. Private sector lobbyists have flooded the media with warnings about the slippery slope.

The politics behind the discussions have been grimly predictable, exacerbated by grandstanding and brinksmanship. The right calls transfers and redistribution Marxism and warns that taxing compute hands the lead to China. The left warns that a tax drafted with the help of incumbents becomes regulatory capture by another name. Fiscal hawks point to unsustainable deficits. Doves point to the premature austerity imposed after the GFC as a cautionary tale. The divide is only magnifying in the run up to this year’s presidential election.

While the politicians bicker, the social fabric is fraying faster than the legislative process can move.

The Occupy Silicon Valley movement has been emblematic of wider dissatisfaction. Last month, demonstrators blockaded the entrances to Anthropic and OpenAI’s San Francisco offices for three weeks straight. Their numbers are growing, and the demonstrations have drawn more media coverage than the unemployment data that prompted them.

It’s hard to imagine the public hating anyone more than the bankers in the fallout of the GFC, but the AI labs are making a run at it. And, from the perspective of the masses, for good reason. Their founders and early investors have accumulated wealth at a pace that makes the Gilded Age look tame. The gains from the productivity boom accruing almost entirely to the owners of compute and the shareholders of the labs that ran on it has magnified US inequality to unprecedented levels.

Every side has their own villain, but the real villain is time.

AI capability is evolving faster than institutions can adapt. The policy response is moving at the pace of ideology, not reality. If the government doesn’t agree on what the problem is soon, the feedback loop will write the next chapter for them.

The Intelligence Premium Unwind

For the entirety of modern economic history, human intelligence has been the scarce input. Capital was abundant (or at least, replicable). Natural resources were finite but substitutable. Technology improved slowly enough that humans could adapt. Intelligence, the ability to analyze, decide, create, persuade, and coordinate, was the thing that could not be replicated at scale.

Human intelligence derived its inherent premium from its scarcity. Every institution in our economy, from the labor market to the mortgage market to the tax code, was designed for a world in which that assumption held.

We are now experiencing the unwind of that premium. Machine intelligence is now a competent and rapidly improving substitute for human intelligence across a growing range of tasks. The financial system, optimized over decades for a world of scarce human minds, is repricing. That repricing is painful, disorderly, and far from complete.

But repricing is not the same as collapse.

The economy can find a new equilibrium. Getting there is one of the few tasks left that only humans can do. We need to do it correctly.

This is the first time in history the most productive asset in the economy has produced fewer, not more, jobs. Nobody’s framework fits, because none were designed for a world where the scarce input became abundant. So we have to make new frameworks. Whether we build them in time is the only question that matters.

But you’re not reading this in June 2028. You’re reading it in February 2026.

The S&P is near all-time highs. The negative feedback loops have not begun. We are certain some of these scenarios won’t materialize. We’re equally certain that machine intelligence will continue to accelerate. The premium on human intelligence will narrow.

As investors, we still have time to assess how much of our portfolios are built upon assumptions that won’t survive the decade. As a society, we still have time to be proactive.

The canary is still alive.

Subscribe now

Acknowledgements: Thanks to Sam Koppelman of for his help with proofreading. Our co-author, Alap Shah of LOTUS, contributed the idea for this piece - CitriniResearch wrote this part, but he has written two other parts in a series called the Intelligence Explosion, we highly recommend reading it. You can find it here.

But sometimes it’s the simplest observations that actually work the best. The Drones basket turned out to be one of our best performing trade ideas of the year, turbocharged by a chaotic run of geopolitical events that emphasized the rapid evolution of asymmetric modern warfare, along with large jumps in global defense spending.

Recognizing a strong theme emerging, we dug deeper into defense in the past year with The New WFH (War From Home), and we continue to maintain an allocation to the theme in the Citrindex. For good reason. The theme spans multiple critical forces – all of which provide huge tailwinds.

• Evolving winners of “Fiscal Primary” favoring defense

• Disruptive technology with public market opportunity

• Geopolitical tailwinds in an increasingly fractured world

These catalysts are only accelerating.

While it is inevitable that most dollars of government spending will go to the major defense primes, our coverage has tended to focus on smaller enablers of “modern warfare” – that is, companies that provide more direct exposure to the trends with the highest pace of acceleration – autonomous systems (drones), modern weapons (lasers), and spectrum dominance (electronic warfare). And while it may not come off as the most sophisticated methodology, our “ripped-from-the-headline” approach continues to work out for us fairly well.

For example, there was nothing secretive about the massive US naval buildout around Venezuela in late 2025, and it didn’t take a four star general to think that “regime change” was on the table. Simply noting the possibility of regime change and highlighting the upside potential of defaulted Venezuelan sovereign bonds and oil service companies was a logical next step in our 26 Trades for 2026 (and it didn’t take long for that idea to work).

For better or worse, there has been no shortage of new geopolitical headlines to choose from. In just the past several weeks, we’ve seen the shutdown of airspace over El Paso, the US commandeering of dark-fleet tankers and extraction of Maduro, Ukrainian sea-drones totaling Russian submarines, and of course Trump’s demand for Greenland.

Each of these events points towards some sort of bigger picture trend and today we are going to break these down into three key buckets:

• Directed Energy Weapons

• Subsea Drones

• Greenland and Arctic Investment

To begin, let’s start with one company that has performed exceptionally well for us (but that we think is only just getting started).

Read more

Capital is rotating out of bits and into atoms.

Software companies, the darlings of the last decade, are being forced to answer a question they’ve never had to answer before: what happens to your 30x revenue multiple when new technology poses a potentially existential threat?

See, for many of these companies, valuations were so high and such sustained growth was priced in that agentic AI didn’t even need to actually impact the business. The threat of a threat was enough to reprice them.

Into this software selloff, anything that wasn’t nailed down was thrown overboard. Vertical SaaS, digital advertising platforms, portions of fintech, all suddenly trading at discounts to their historical ranges as the market begins to assign what I’d call an “AI Disruption Discount”. Does your high switching cost or sticky UX matter in the AI era? Don’t know? Okay - trade lower until you do. Will agentic commerce threaten your marketplace? Maybe? Okay, well, you’re worth 30% less.

The logic is ruthless and correct: none of us really know what AI will look like in a few years, and the technology is advancing at a rate that has us asking if the singularity actually just got penciled in for next Thursday. Agentic AI was a talking point for investors, but once Anthropic et al. shoved it in their face they had to reconcile with some hard potential realities that didn’t align with their portfolio.

We could spend a lot of time debating what moats will endure in software – do we just sell anything you click and buy the dip in the APIs? Is this just empowering new entrants to challenge incumbents or will enterprises bring entire software workflows in house via AI? The more you think about it, the more you realize that without a crystal ball you’ve got a lot more questions than answers.

That’s why I think investors will opt for a simpler outlook - what can’t you prompt your way out of?

The answer to this question leads you to a simple conclusion: Atoms have become more valuable relative to bits. As hyperscalers exponentially ramp up their bit-generation capabilities, their demand for power and materials has moved in lockstep. Virtually every single bottleneck or constraint surrounding the AI trade has been in the real world.

Claude, increase the global supply of Titanium by 5% above forecasts this year…

See. That still doesn’t work. Not even on Opus 4.6 Fast.

A lot of the “real world” beneficiaries that investors have piled into so far, however, are the obvious ones. Power generation equipment. Optical Interconnects. Grid infrastructure. Copper miners. Nuclear. Don’t get me wrong - these were great calls (many of them were ours), but they are all relatively crowded trades now. Many are priced for the perfection that physical world constraints ironically make impossible to deliver.

So where do we go from here?

The next evolution of the “atoms” trade is where new sources of demand collide with the existing data center buildout to constrain supply chains even further.

A poster child of this collision is in jet engines.

The Trump Administration is shooting for a $1.5T defense budget, with NATO allies building out their own capabilities. Japanese defense spending will also stay bid as a result of the recent election. Demand is stacked from the aerospace industry, which is hitting an inflection point with a historic, sustainable backlog.

Data centers use old jet technology for their natural gas turbines. And while the turbines are manufactured independently from current generation jet engines, both supply chains use the same input materials. In under a year, once overlooked businesses are now the “belle of the ball” with suitors ranging from AI hyperscalers to the Department of War.

Finding the winners of this new preference for “stuff” boils down to one question: can this company command pricing power over an entire capital cycle due to multiple tailwinds, or is it just a commodity trade with extra math?

Every material or chemical in this primer shares a specific profile that we feel will result in outperformance as the market becomes more intent to focus on atoms over bits.

We expect this theme to evolve over time, but today we will highlight our current favorites across:

● Advanced Materials & Processes

● Electrical Steel

● Literal Picks and Shovels

● AI Materials & Electronic Chemicals

Advanced Materials & Processes

These are all companies that deal with material sciences and advanced processes that sell into sectors we’re broadly bullish on: defense, electrification, aerospace, artificial intelligence, space. Production is concentrated among a handful of qualified suppliers, often operating as regional duopolies, oligopolies or outright monopolies. Manufacturing requires decades of process knowledge, stringent quality controls, and customer qualifications that take 2-5 years (you can’t simply build a bigger plant in response to price increases). Their moat is not in question, but their multiple may be now that the market places a premium on being resilient to AI-driven disruption.

Unlike lithium, PVC, or solar-grade silicon, these materials are protected by a combination of physics, technical barriers, defense procurement restrictions, chronic underinvestment, and/or quality requirements that Chinese producers struggle or choose not to meet. Cyclical demand recovery in many of their end markets is now converging with structural spending increases for the first time since the post-GFC period. Finally, for many, the market is still pricing most of these names as if only one of those tailwinds exists.

Uranium Conversion

Solstice Advanced Materials (SOLS US): The Cheapest Nuclear Play in the US

Uranium miners and enrichers have done quite well, but there’s a second stage that gets talked about less: conversion. It’s a quite interesting stage as well, and perhaps the last part of the trade where a beneficiary has flown under the radar. That beneficiary is Solstice Advanced Materials (SOLS US).

We spoke briefly about Solstice Advanced Materials (SOLS) in our 26 Trades for 2026 and added it into the Citrindex on January 28th before earnings. Now that they’ve reported earnings and validated our suspicions, we believe the stock is even cheaper than it was before it rallied in response.

SOLS was an under-the-radar spin-off from Honeywell in October 2025 that the market originally seemed to value solely based on their HVAC/refrigerant business. Something that the market hasn’t begun to value properly, however, is that they operate the only uranium hexafluoride (UF6) conversion facility in the United States: the Metropolis Works plant in Metropolis, Illinois.

Built in 1958, this facility sits at Stage 2 of the nuclear fuel cycle:

SOLS is the sole domestic supplier of conversion – a mandatory intermediate step in the nuclear fuel chain. Every pound of uranium destined for a U.S. reactor must be converted to UF6 gas before enrichment.

The global competitive landscape is narrow: Cameco (Port Hope, Canada), Orano (Comurhex/Philippe Coste, France), and Rosatom (Russia). With Russian supply increasingly sanctioned and geopolitically unreliable, and Orano/Cameco running near capacity, the conversion market is structurally tight. The plant has a nameplate capacity of approximately 15,000 MTU (metric tonnes of uranium as UF6) per year, representing roughly 20% of global conversion capacity.

While SOLS has a multiple on the upper end of the materials range, if you contrast with the only pure play uranium enrichment name in the US - Centrus Energy (LEU US), you can see the discrepancy and opportunity. LEU trades at 33x 2027E EV/EBITDA - once you strip out the non-AES business with an industry standard 11x multiple, you are left with the AES nuclear business at a $3.2B EV. We strongly believe this is far too low.

Metropolis shut down in November 2017 at $4.50, restarted Jul 2023 with spot already at $50+. The entire bull move happened while the plant was offline. SOLS literally came back online into the strongest conversion market in history.

SOLS reported in their earnings that Q4 revenue was up 39% YoY on volume growth of only 15-20%, pricing drove roughly half the beat.

This is due to legacy contracts that were struck at $20/kgU during or before the 2017-2023 idling period when nobody wanted conversion capacity. These contracts are rolling off over the next two to three years and, with spot conversion price around $64/kgU, every single one gets replaced at roughly triple the price.

Metropolis contracts today are generating a third of what the market will pay for the same work. This will all be repriced by 2028/2029 - which are the years the market should be valuing the AES segment on (similarly to how LEU is valued).

Consensus estimates have revenue staying basically flat for the next two years and margins roughly the same. Management just confirmed that they’ve got a $2B backlog in AES and that EBITDA will grow at a double digit CAGR (which we believe is more like 20% than 10%).

Why such a delta? Because the price is set to triple on volume that costs the same to produce. Conversion is a toll processing business: the customer delivers uranium oxide, Metropolis converts it to UF6, SOLS collects a fee. The cost of converting a kilogram under a $20 legacy contract and a $60 market-rate contract is identical, roughly $13-16 per kgU for energy, chemicals, and labor plus $150-170M in annual fixed costs for the plant. With roughly 25% of the legacy book rolling off each year, and $60 as the floor (not the ceiling) for new contracts, SOLS is adding $40+ per kgU of pure margin on each tranche that reprices.

The entire price difference on every repriced contract flows straight to EBITDA, meaning AES margins will explode to 60% over the coming years.

Above 10,000 MTU of annual throughput, which management just confirmed for 2026, margin expansion becomes almost mechanical. The EPC firm retained for an expansion study beyond the 15,000 MTU nameplate, with DOE backing, suggests we haven’t seen the ceiling on volume either.

And $60 is almost certainly a floor, not a destination. That spot price reflects the existing global fleet of ~440 reactors running normal fuel loading schedules. It does not incorporate 60+ reactors under construction worldwide, the SMR pipeline, AI-driven nuclear PPAs from Microsoft, Amazon, and Google, or the U.S. target of 400 GW by 2050 (4x the current installed base). If conversion demand grows even 10-15% from new builds alone, and Western supply stays capped at 35-37 thousand MTU, spot goes to $80-100.

Strip out the non-AES businesses at 11x on ~$827M of EBITDA and the market is assigning roughly $3.8 billion of enterprise value to the nuclear segment. If we simply adjust for the legacy contracts repricing, assume a 40% increase in volume over the next 3 years as they expand capacity and model the incremental flow through to EBIT, we can bridge SOLS’ AES segment to ~$673M by 2028.

This is before any price increase driven by the demand for nuclear fuel, and we’ll even handicap our EBIT bridge by 25% to account for any current contracts that have already rolled off and any incremental, unforeseen costs. We still get 2028E EBIT of $500M, putting this segment at a 7.6x EV/EBITDA valuation.

For context, Centrus Energy (LEU US) – the only US pure-play enrichment name – trades at 33x EV/EBITDA, while Cameco’s (CCJ US) fuel services segment is at 14-17x. SOLS is getting a commodity chemicals multiple on the only US UF6 conversion monopoly with a $2 billion backlog and guaranteed 3x repricing on a third of its remaining book. We think 12-14x is fair for the base case, which alone would imply $95-105 per share assuming zero growth in the other segments.

And this is all before you account for any nuclear renaissance demand wave arriving on schedule, which could take it even higher and see the bull case representing nearly a triple from here by 2028.

…

The rest of this primer covers our other atom picks like beryllium monopolies, titanium duopolies, aerospace aluminum, tungsten’s supply crisis, superalloys, electrical steel, mining capex beneficiaries, and an update to the AI materials bottlenecks we flagged in December (as well as some new ones) that are now making headlines. Become a paid subscriber to view the rest of the article.

Subscribe now

Read more

There is a lot going on in AI and the semiconductor supply chain. Almost too much, it seems at times.

The wild performance in the most prominent players has reinforced the supply/demand balance we described in the idea “Post-Traumatic Supply Disorder” from our 26 Trades for 2026. Memory, storage, semicap, advanced packaging – these sectors have performed incredibly well over the past month.

We’ve also been reminded of something we have known for a while.

“While one has indeed so far been able to get away with an understanding that only goes GPU-deep in AI, I believe that will rapidly begin changing…

In the meantime, generalists are left to keep up with a rapidly advancing technological landscape in semiconductors that can easily leave you behind... The shame in that is that it often robs you of the ability to truly understand what’s behind the whole picture while ensuring you’re properly positioned – a picture which will only become increasingly important when these narratives take hold in the real world.”

-AI Phase Two, April 2025

In order to have caught the most asymmetric trades of the “AI Bubble”, you had to be paying attention to things that were a bit complex for a generalist. You also had to have a strong view on what would happen in the supply chain 6-12 months from now due to whatever was happening right now.

Some bad news: it is not going to get any easier from here. We can’t sit idly by while the expected returns from mainstream “layups” like Nvidia, Micron, SK Hynix, Broadcom etc. moderate.

That’s why we’ve hired some talented semiconductor analysts, so they can explain to us ignorant generalists what’s going on behind the Wall Street curtain (and we can then translate it back into thematic upside).

Our two newest team members, Zephyr and Jukan, add deep semiconductor knowledge to Citrini’s existing thematic and trading expertise. We’re going to combine our existing penchant for having strong, occasionally non-consensus views on what the continued proliferation of Artificial Intelligence looks like with their in-depth knowledge of the semiconductor supply chain.

We’ve navigated the opportunities in AI Data Centers, GPUs, Memory, Optical Interconnects, ASICs and more so far.

For example, for us to underwrite Teradyne (TER) and its robotics opportunity, we had to keep abreast of the tailwinds for players in memory testing. To recognize the opportunity in Amkor (AMKR), we needed to recognize the increasing demand for chiplet architectures and advanced packaging that would increase demand for OSAT (Outsourced Semiconductor Assembly and Testing, if you haven’t been paying attention). Going back even further, for us to make the call that SK Hynix and Micron would outperform in Q1 2024 (while they were languishing in the doldrums of the post-COVID RAM glut), we needed to understand the constraints HBM placed on the supply chain and the nature of truly agentic AI.

This is the first of our new “Semis Memo” series, where we’ll go a layer deeper on current trends in the infrastructure powering AI while also ensuring our audience doesn’t need to go get a PhD in semiconductor design to keep up. Consider this a tour around what we view as the most interesting, potentially overlooked opportunities in one of the hottest sectors on the planet.

Today, we’re covering some existing trends and the second-order winners of the Memory Supercycle, Semicap Subsystems and Silicon Photonics. We’ll also share some of our early thoughts on opportunities that should be on your watchlist, everything from SpaceX supply chain beneficiaries to opportunities arising from the AI-driven SaaS selloff. Finally, we close with a list of single name stories we think are worth keeping on your radar.

Read more

In this piece, we will cover:

US Macro Reacceleration:

• AI productivity outweighs job cuts

• Pro-growth policy and fiscal impulse ahead of midterms

• Liquidity injections alleviate funding pressures

• Powell claps back

Global Instability Breeds Opportunity:

• Long oil & tankers

• Don-roe Doctrine winners

• China’s long march towards recovery

And of course… Our New Macro Trades

US Macro: Running Hot

We find few reasons to be bearish on the US economy. The two biggest concerns of 2025 (tariffs and job creation) haven’t led to a broader slowdown. Rather, consumption and overall GDP data is accelerating, and while the labor market is stagnant, the floor remains intact. Productivity gains, in part driven by AI, are likely to increase.

Meanwhile, the political pressure of midterm elections is already apparent in the White House, as Trump refocuses his attention towards everyday consumer concerns – inflation (oil supply), housing affordability (MBS buying), consumer credit (interest rate caps)-- even teasing direct stimulus checks.

With 16 months passing since the first rate cut of this cycle in August 2024, perhaps we are beginning to see the “long-and-lagging” effects of bullish monetary policy beginning to materialize. Credit spreads remain tight, overall borrowing costs have decreased, and cyclical sectors like housing are beginning to show signs of life.

The only real sand in the financial gears – increasingly stretched liquidity in funding markets which began to show up in late 2025 – has been cleaned out by the Fed’s new liquidity program. Whatever you want to call it, the Fed has officially crossed from base-money destruction to base-money creation and balance sheet expansion.

The Year of AI Productivity

While the U.S. labor market has significantly weakened in the past year, broader economic indicators tell a very different story.

By nearly all measures, the US job market appears to have considerable slack, even if signs of stabilization are emerging. Net new job creation has slowed to a standstill for much of 2025, with both public and private payrolls indicating soft to weak employment growth.

The net job growth required to maintain the unemployment rate has decreased due to a shrinking immigrant labor pool. Even still, the net number of jobs created likely sits at or below most reasonable estimates of “breakeven” job growth. Some evidence suggests private sector jobs may have grown while public sector employment contracted, but those private sector gains appear concentrated primarily in healthcare and hospitality services.

Hire and quit rates, which resumed their decline after Liberation Day last year, have now settled at levels below pre-COVID norms. Meanwhile, the layoff rate has crept slowly upward since Liberation Day. These figures are consistent with a challenging labor market, particularly for white-collar workers.

Wage growth continues to slow across the entire income spectrum and all job seeker types. Workers in the lowest income quartile, who experienced faster wage growth at the start of the pandemic due to labor shortages, now see slower wage growth than median income workers. This represents a normalizing labor market.

But none of this is new. These trends have been present for nearly three years and likely represent a normalization of a bullwhip in labor market conditions post-COVID. Further, the rate-of-change suggests that these moves are moderating, not accelerating.

And despite the weakness in labor conditions, the broader economy is solid.

GDP growth remains robust and is, in fact, showing signs of re-acceleration alongside rising labor productivity. Real GDP expanded at annualized rates of 3.8% in Q2 and 4.3% in Q3, and the Atlanta Fed’s GDPNow model is currently forecasting Q4 growth of 5.1%. As is typical, Nowcast estimates tend to become more reliable as the release date approaches.

Historically, when labor market data and GDP have disagreed, the labor market has typically proven correct. GDP reports are famously subject to subsequent, typically downward, revisions.

This time, however, we argue it’s much more likely that aggregate growth indicators will prevail.

The key is labor productivity. Labor productivity, typically measured as output per hour worked, captures how efficiently labor is used in the production process. While growth in labor input has slowed markedly, real GDP has continued to expand, implying a meaningful rise in output per hour. In other words, a growing share of recent GDP growth appears to be driven by productivity gains rather than labor force expansion.

What could explain the rise in labor productivity? And should we expect it to continue?

We know we will likely catch some flak from the economist crowd, who will find every possible way to explain this away, but we have a hot take here. It’s possible that the hundreds of billions of dollars we’ve been investing into the “productivity-enhancing-technology-machine” for the past three years is actually…enhancing productivity.

While the party line for economists is mostly that “productivity doesn’t measure what you think it does”, we’re going to temper that idea with an anecdotal vibe check. Most software engineers will freely admit that they’re getting much more work done per hour, and with the advances in agentic AI over the past 3-6 months it’s not unreasonable to assume that impact is also being felt across a wider swath of vocations as well.

Are firms pressing their existing workers to work harder? Probably, after all, not many have the luxury of quitting. But we believe the mechanism through which those employees are increasing their productivity is at least in part due to AI tools. These tools are both increasingly at their disposal (regardless of whether or not their work has enterprise licenses) and have improved immensely over the past year. This will become increasingly apparent in 2026.

Unlike the “missing productivity” of the internet era, the AI era will be marked by productivity gains that become harder and harder to be written off as one-offs or idiosyncratic quirks of arithmetic.

Payroll data shows the information and tech sector has been shedding workers. Yet the software sector has still been making significant contributions to growth.

By now, it’s pretty well known that one of AI’s most visible productivity gains has been in coding agents. That boost may be now spilling over into other sectors as “vibe coding” becomes more increasingly commonplace.

We believe the recent rise in labor productivity will continue as AI driven productivity carries the torch. While it’ll probably take a while for academics to feel comfortable confirming this, we’re content to see the writing on the wall.

This presents an interesting macroeconomic environment as investors acclimate. It’s quite possible we see 2026 play out with steady or even increasing unemployment despite well-supported GDP growth. A weakening labor market in combination with growing GDP and a rising stock market is not unheard of – 1992, 1975, 1958 all saw this dynamic play out. Perhaps most well known were the years following WW2 as a glut of labor developed due to GIs returning home to an economy that had learned to do more with less and wartime technology being repurposed from armaments production to everyday industrial and consumer uses.

It is only a matter of time before this productivity increase spills over from the tech sector and the largest companies into smaller companies in non-tech sectors. This will buoy corporate earnings, which will in turn support the stock market.

Below the paywall, we discuss our views as on monetary policy, midterms, housing, geopolitics and lay out our current and updated macro trades.

Read more

Introduction

Our Approach to the “Year Ahead Outlook”:

It’s a bit of a ridiculous premise isn’t it, the “year ahead outlook”?

Just because the number on the end of the calendar changes, you’re supposed to have a better idea of what happens over the next twelve months? I don’t know about you, but if I felt strongly that I knew what was going to happen over the next twelve months, I would not wait until December to share it.

As speculators, we spend the majority of our time trying to be right about the future. That tends to be a good thing—efforts spent towards being right (or at least less wrong) tend to be rewarded.

It doesn’t take a long or storied career as an investor to have numerous realizations that you’ve missed something very obvious in hindsight. Our vocational emphasis on accurate forecasting sometimes hurts us more than it helps. Becoming too in-the-weeds risks stifling the creative process that is a prerequisite to developing differentiated views.

Once a year, we drop all pretense. Just for this exercise, our primary focus isn’t being right at the right time. Instead, we look at what’s occurred in the passing year and what might occur in the year ahead. Without any of the pressure surrounding “this has to happen or else I’m wrong,” we find that the ideas tend to more closely approximate the truly unpredictable nature of how things will unfold.

Instead of an “outlook”, we create our Twenty-Something Trades. You can take a look at last year’s to get a better idea.

The end result is a long list of ideas. Some we think could be real winners, some are preparation for a specific scenario or eventuality, some we simply want to keep on our screen. Our aim is to minimize the risk of intellectual blind spots.

Are they all going to be good trades? Of course not, there’s 26 of them. Think of this as a thematic watchlist.

The purpose of this article is to put as many relevant ideas in front of you as possible, in hopes that one of them resonates at some point during the next 12 months. If you think one of these trade ideas is hilariously wrong, GOOD. Don’t keep it to yourself – we could end up finding the right trade from that discussion.

It’s the end of the year, another absolutely crazy year for that matter. Now, we get to prepare to do it all again… but differently.

2025 Scorecard

First, let’s revisit how last year’s thematic watchlist performed in 2025. Since publication, a little more than half of our ideas outperformed the S&P 500 and more than 80% had positive returns.

We caught some good themes (most of our drone names doubled, our Ukraine normalization picks crushed it, and South Africa rode the gold rip to crush Turkey). We also had some ideas that really did not work (why did we think remittances would go up while the Trump administration ramped deportations? I don’t remember, seems silly in hindsight).

Here’s our scorecard, incorporating all the 25 Trades for 2025 (some themes had more than one expression) that were covered in the article:

We had similar dispersion in our single stock picks - some big winners (Allient doubled on Rockwell returning to growth and robotics tailwinds, we correctly anticipated the success of ARQT’s Zorvye), some big losers as well. As of writing, the average return on an equally weighted allocation to our single names was +26.50%, still handily beating the S&P 500 by more than 10 percentage points.

Now, as my regret for sticking to the 2X Trades for 202X format grows each year, so too does my commitment to the bit. With that out of the way…

Let’s get into it.

Read more

We elaborated…

OpenAI is compute-constrained. Without their own datacenters or their own chips, they are at significant disadvantage to Google, which not only is an ascendant hyperscaler but also a chip designer that has designed the only mass produced chip competitive with Nvidia’s core offering across many metrics and has optimized their entire research, training and inference pipelines for it [...]

Google has low customer acquisition costs, best-in-class first party data, full vertical integration and TPUs enabling cheaper training [...]

It’s clear that both the technical and market tailwinds favor Google, but the market does not seem to be pricing that in.

While the writing has been on the wall for some time, the market’s perception of Google has dramatically reversed over the past several months – transforming from an AI loser bleeding its search dominance to a stalking horse destined to undercut the most consensus AI winners.

We can see this shift occurred around the time we wrote up Google (or, perhaps, around the time the legal overhang lessened) on this chart from Coatue:

This sentiment has only accelerated in the past week with the release of Gemini 3. Not only is Google now firmly positioned at the cutting edge of frontier models, but it’s doing it on its own terms, or TPUs. It is indeed possible to train a frontier model without NVDA... that is, if you’re an ML-pioneering hyperscaler who has spent the past 10 years developing and optimizing for this custom silicon.

The announcements that both Anthropic and Meta are planning to implement TPU chips raise further questions about NVDA’s dominance. Surprisingly, META reportedly wants TPUs for training, not just inference. None of this information is truly new (see above) but the one-two punch, combined with growing skepticism of OpenAI seems to have culminated in a passing of the public torch.

While we are happy that the market has caught up to the trade – Google remains our largest allocation in our Dynamic AI basket – we are also wary of over-extrapolation, recency bias, and consensus crowding. After all, as Oracle has shown, you can go from a 40% gap up to CDS watching in a matter of weeks.

Is Google’s vertical integration both a cost and structural advantage? Is the CUDA moat shrinking? Are NVIDIA margins at risk?

Perhaps.

But maybe there are two ways to think about this:

1. Switching costs for architecture/ecosystem is high for AI accelerators. Gemini 3 should be bullish for compute demand as labs locked into NVDA rush to buy more of the next gen chips to compete.

2. NVDA’s margins are so high that any competitive threat must be aggressively priced in even if it means higher volumes.

In other words, isn’t a breakthrough model broadly bullish?

The market is clearly leaning on the #2, but we could find ourselves tempted to dip back into NVDA if sentiment overshoots reality. Regardless – if a TPU scale-up is on the horizon, who else stands to benefit?

First, we will run through some technical details, then highlight where we think the broader “TPU trade” goes from here.

TPUs for Idiots

At its core, a neural net is just a machine for doing enormous numbers of weighted sums. For myself (an investor, not an AI engineer), I find it useful to reduce neural nets to:

Inputs x Weights = Outputs

Inputs are arranged as a matrix (a batch of tokens, images, whatever). Weights for a layer are another matrix. Take a simple example: handwritten digit recognition. A 28x28 grayscale image is 784 numbers - the “8-detector” neuron holds 784 weights (28x28). For that individual neuron, the core operation is multiplying each pixel by its weight, adding everything up, and then finally comparing the score against other digits.

Once you scale this up to our modern AI/ML models you’re doing trillions of these “matrix multiplication” (abbreviated as matmul) operations – in effect, a bundled batch of all those weighted sums.

Enter: CPUs, GPUs, TPUs.

A CPU is like one very smart and capable worker – good at anything but slow at repetitive jobs. A GPU is like thousands of generalized workers. A TPU is like an automated factory line built specifically for one task, with the machines arranged so the work has almost no back and forth.

It’s important to recognize that even within the AI/ML space, different functions require different chips. Google itself publishes the following guide for how to choose chips depending on the task at hand.

As we laid out in Interconnects 101, the layman’s understanding of the difference between a CPU and GPU is that a CPU is built with a relatively small number of cores, but with each of those cores capable of executing hundreds of specific instructions over dozens of units. Consumer CPUs generally have up to 16 cores with data center chips up to 128 but even with significantly higher clock speeds, this is no match for the parallelization that GPUs allow.

What we think of as a ‘core’ in a GPU is known as a streaming multiprocessor (SM) in Nvidia-parlence. Rather than focusing all of the core’s resources on routing instructions through specialized units, GPU SMs leverage dozens or hundreds of copies of the same units simultaneously. For example, Nvidia’s Blackwell architecture SMs each have 4 “tensor cores” and 128 “CUDA cores” for a B200 with 192 SMs, that totals 768 and 24,576 physical cores respectively.

Though not comparable on a core-to-core basis, TPUs take this philosophy to the next level and dedicate nearly the entire silicon area to crunching matmuls. They can do this because cores aren’t instructed in the traditional sense, but data is drawn through them more like an assembly line than a laboratory.

The architectural driver behind the TPU is the systolic array. Though reports differ on the size and arrangement of the array across TPU versions, the philosophical departure from GPU SMs is consistent.

The array is made up of MAC cells, short for multiply-accumulate, which Google describes as ALUs (arithmetic logic unit) which is a classic computing building block. The difference is that MAC cells are ALUs that don’t accept instructions, just data. They don’t have to choose between FMA, ADD, SHIFT, or any other of hundreds of possible instructions that a less specialized processor might see – they just multiply and accumulate. This saves on space by eliminating unit-level instruction caches while also essentially eliminating data caches as data flows through the array in a predetermined fashion and saves on power by pushing utilizations to the max.

This is why TPUs can offer very high performance per watt and per dollar on machine learning, but are essentially useless outside that domain. You get extremely high throughput on the kind of linear algebra that defines neural nets which means lower power per operation and a smaller die area for a given level of ML performance. However, this comes at the price of much less flexibility than a GPU.

Google has made Gemini’s per-prompt footprint a PR point, so Gemini likely has a real advantage in energy per text call – helped by TPUs and their own tuning – but there is no transparent or quantified comparison for the training runs. Inference, at least, looks to be much more energy efficient per user interaction on TPUs.

In short, the TPUs are hyper-specialized ML chips that sacrifice flexibility for efficiency. But as AI increasingly moves to the “mass-production” phase, even small efficiency gains can prove very meaningful.

What Does it Mean for NVDA

We’ve been wary of the competitive risks from ASICs to Nvidia for a while (and a TPU is in fact an ASIC). But we’re also wary of headlines that declare Nvidia dead. If our own DeepSeek coverage taught us anything, it’s that the easiest way to move the US stock market (which is a KPI for many financial media outlets) is to target Nvidia.

As we’re indifferent to moving the market and more focused on what happens over the long term, we’re trying to minimize sensationalism.

One key thing to emphasize is that switching from GPUs to TPUs means learning a new language – i.e. switching ML frameworks from PyTorch to JAX. Switching software environments is a significant task – non-trivial, disruptive and expensive. After all, the first TPU was created by Google back in 2015, and the giant has spent the last ten years developing internal workflow, physical infrastructure and model architecture around it. For others, particularly smaller players, the learning curve will be far steeper.

However, Nvidia’s silicon stranglehold and pricing power does provide a material incentive to at least attempt to diversify. Anecdotally, OpenAI has been trying to build an internal JAX equivalent for the last year – newer startups deciding to build initially on JAX rather than PyTorch can result in new “locked-in” users. Moves by Meta and Antrophic demonstrate that major players are indeed seeking silicon diversity.

But the reality is that CUDA and PyTorch are still dominant today and will probably mean that Nvidia will continue to see strong demand in the near term regardless of Google’s plans for TPUs (i.e. whether they sell, rent, or keep it as a walled garden). Keep in mind, despite its use of TPUs, Google is still one of NVDA’s largest customers.

We can certainly understand the market’s sensitivity to any risk that might jeopardize NVDA’s competitive dominance – even if it doesn’t materialize until 27/28. It isn’t surprising that the kneejerk reaction was for NVDA to trade lower. Still, it should be understood as a reaction to the potential of future risks – and future risks to margins, not to volumes. The market is not valuing Nvidia on what happens in 2026, and Nvidia’s margins are (historically speaking) unsustainable over the long term. As seen in the reaction to NVDA’s great earnings last week, the market cares much more about what things look like in five years than what they look like in six months.

Right now, NVDA is still able to deliver volume into a tight market, and overall we view this TPU development as much more positive for GOOGL than it is bearish for NVDA – in fact, we don’t really think this specifically is bearish for NVDA or their supply chain.

Rather it’s probably quite bullish for the supply chain. Both Google and Nvidia are going to be fighting for the exact same CoWoS capacity at TSMC, which will remain the rate limiting factor.

To recap:

• Gemini’s success reinvigorates model competition - reinforcing broad demand for compute, which is still dominated by NVDA

• CUDA/PyTorch is still a major switching cost, and Google’s success on TPU may not be as easily replicable for third parties

• Both supply and infrastructure bottlenecks, not demand, is the key bottleneck today

Will Google Sell TPUs?

Another open question is Google’s TPU distribution strategy:

• Do they keep TPUs for themselves as a competitive advantage?

• Do they rent out TPUs (which they do already, via GCP)?

• Do they become a merchant TPU vendor and alternative to Nvidia?

The first alternative – full in-house proprietary advantage – risks under-monetizing an incredibly valuable asset (remember, NVDA is worth more than GOOGL today) and has already been contradicted by chip sales to Anthropic, Meta, and Fluidstack.

In the second alternative – the “cloud route” – Google is limited by both TSMC capacity and the physical availability of their own data centers. Here, they may maintain ongoing cloud revenue, but are still both committed to and constrained by infrastructure.

However, if Google decides to become a selective merchant vendor (shipping TPUs into neoclouds, sovereigns and selective hyperscaler datacenters i.e. META) they would generate revenue and gain presence without committing to infrastructure. In a world where stocks get hit every six months on capex fears, this might be a good way of hedging. Data center and power capex is very long-duration and hard to unwind. By contrast, selling TPUs is essentially a pure margin game with variable production that can be ratcheted down. This shifts risk and produces optionality, making Google harder to kill.

In a somewhat reflexive way, we think the market’s reaction here is going to force Google to sell TPUs. If they don’t break out TPUs as merchant revenue and data center/power constraints cap GCP capacity growth, then the narrative will become a refusal to monetize the silicon while being subjected to an infrastructure bottleneck they do not control.

Further, if Google does not sell TPUs, they will end up creating space for a new merchant silicon threat in the medium term (think 4-7 years). It would be much better to broadly distribute TPUs and gain the telemetry advantage rather than risk this business going to a competitor (especially a Chinese one).

Back in September, it was reported that Google approached Crusoe, Coreweave (CRWV US) and Fluidstack to give them access to TPUs. Fluidstack, for example, has a deal where it deploys TPUs in a Terawulf (WULF US) leased data center in NYC – with Google providing a $3.2 billion backstop and taking 14% equity. In doing so, they’re using the “franchise” model versus the “merchant” model (you find the power and build the shell, we will drop in our TPUs). Furthermore, this week, we saw headlines about Meta (META US) buying TPUs from Google. Meta has its own silicon (MTIA), but if Google offers TPUs at a better price/performance ratio for serving Llama models, Meta is rational enough to use them.

In short, we believe the decision has largely already been made to begin selling TPUs.

So, what’s the trade?

First-Order Beneficiaries

Comparing an equally weighted basket of our 25 highest-confidence TPU suppliers to our broad 125-name Phase 1 AI Infrastructure basket, we can see that they’ve taken off recently in relative terms:

Across the various AI categories we track with our baskets, these direct TPU suppliers have been the best performing group YTD with the majority of that outperformance beginning around August:

Here’s the list of names we are looking at as TPU suppliers. We’ve been relatively picky in this list, and have left out some names that have been mentioned in association with TPUs but for whom the relationship seems less clear or direct.

View our equally weighted TPU Supply Chain basket on our portal here.

As a basket, these names still trade cheaper on average than our broader Phase 1 AI infrastructure basket that is heavily tilted towards GPUs. Some of this is due to the more cyclical nature of a lot of these suppliers, but if increased TPU demand drives a sustained upcycle in advanced PCBs, substrates, and high-speed networking boards (and their upstream suppliers in copper foil, resins and specialty laminates) we could continue to see a re-rating.

One thing to emphasize, though, is that there’s a huge amount of overlap in the NVDA and GOOGL supply chain – and the same overall capacity constraints haven’t changed. To the extent that two giants are competing for the same production capacity, it should benefit the pricing and bargaining power of these supply chain players.

We like the upside for names in advanced packaging – the obvious being TSMC (TSM US), but also ASE (3711 TT) and Amkor (AMKR US). These should do well as competition heats up. Memory (HBM) remains the bottleneck, with SK Hynix (000660), Samsung (005930 KS) and Micron (MU US) supplying memory (in that order).

We also think this could provide a catch-up impetus for Mediatek (2454 TT) – a name that seems to be well positioned from a markets perspective having participated in very little “AI” upside over the past year due to poor performance in mobile.

Focusing on the key differences in suppliers for TPUs and GPUs leads us to a few conclusions for our basket. First, TPU boards are denser with proprietary interconnect routing, benefiting niche high-end PCB makers such as Isu Petasys (007660 KS), TTM Technologies (TTMI US) and Unimicron (3037 TT). NVIDIA boards are also complex but follow a more standardized form factor (OAM/UBB).

Second, we see benefits for pluggable optical transceivers. Unlike GPUs that use electrical packet switches (InfiniBand/Ethernet) for this layer, Google uses MEMS-based optical mirrors to physically redirect light (called OCS or Optical Circuit Switches). This eliminates the need for power-hungry electrical transceivers at the switch level. Google has increasingly partnered with Lumentum (LITE US) to mass-produce these switches (specifically the Lumentum R300 and newer R64 series). While Lumentum designs the system, the actual MEMS mirror chips are often fabricated by Silex Microsystems, a pure-play MEMS foundry owned by Sai MicroElectronics (300456 CH). Also, while Lumentum supplies the MEMS mirrors for Google’s internal OCS (Project Apollo), GOOGL is driving the OCP (Open Compute Project) standard for OCS to wider adoption. Huber+Suhner (HUBN SW) is the other critical player in this niche.

Some other names worth an honorable mention as well are:

• Alchip (3661 TT) which was mentioned alongside Mediatek in a TrendForce write-up on Trillium in chip design but has no incremental headlines since, and is additionally tied to AWS (Inferentia/Trainium) rather than Google

• Global Unichip (3443 TT) is the likely backend design partner for Axion - Google’s custom ARM-based CPU that will be utilized instead of Intel Xeons

• Jabil (JBL US) and Flex (FLEX US) in System EMS which were mentioned in a secondary source Taiwanese supply chain piece but did not have the same level of confirmation as Celestica

• Panasonic (6752 JP) and Furukawa (5801 JP) which were mentioned in that same piece but also lacked confirmation.

• Monolithic Power (MPWR US) has been gaining share in the 48V server rack architecture for Google’s TPUs. There are also names which deal with aspects like the Liquid Cooling (Deschutes CDU and Cold-plate Loop) - the clearest direct manufacturing partner being Boyd, who sold its thermal management segment to Eaton (ETN US) recently. Still, these are broader and other players have compatible solutions that would work with TPUs.

Second-Order Winners

Who else might materially benefit from the proliferation of TPUs and Gemini 3 becoming a viable state-of-the-art model?

Apple: A Google AI Customer?

Somewhat surprisingly, one of the biggest indirect winners might be Apple (AAPL US), who may see its patience and discipline pay off with the compute platform of their dreams. Apple and Google already have a strong working relationship, so competitive dynamics shouldn’t preclude their collaboration here.

From a product philosophy standpoint, Apple and Google are already converging on a few shared ground-truths: first, that reducing round-trip latency and jitter is imperative to user experience. This most notably manifests itself as lagging Siri, spotty chat-bots, or frozen applications. Second, both Apple and Google are, for lack of a better term, control freaks. While both experiment, neither is remotely comfortable with the unpredictability of truly dynamic LLM inference… remember the debut of Google Search’s AI features?

Both of these issues can be solved in one fell swoop by leveraging the static determinism of a TPU pod’s systolic array. If Apple wants to position Siri as a fast, reliable, transactional personal assistant rather than a superintelligent, conversational chatbot, TPUs offer a perfect engine for nearly any workload that can’t be run locally. While Apple has made great strides in reducing barriers to local inference on mobile devices – security features like FaceID will never be offloaded for example – much of the token volume that Siri and Apple Intelligence will handle is likely to require quick, cheap cloud-based inference using information straight from the web, which you may have heard is another Google specialty.

Even if Apple wanted to maintain a fully closed loop here, they would still need to incorporate external information for requests like sports scores, restaurant reviews, or thousands of other everyday requests and Google just happens to have the best of both. Apple isn’t locked in to Nvidia’s ecosystem to the same degree that others are, and Google ramping Gemini might be the path to finally getting a Siri that doesn’t suck.

SiTime and MACOM

In our July 2024 article Can You Hear Me Now? Telco Suppliers New Dawn and the Coming Optics Shortage, we highlighted Lumentum (LITE US), Ciena (CIEN US) and Coherent (COHR US) as big winners of optical demand. There’s been a renewed focus on optics demand amidst the TPU take frenzy for good reason. Google’s internal TPU forecast for the next cycle has been revised from ~2 million units to ~4 million.

That step-up in TPU V7 servers is being viewed as a key driver behind a jump in 1.6T optical module demand from roughly 3 million modules in 2025 to ~20 million in 2026, with Google alone consuming between 6-10 million of those via TPU V7 racks.

We do think there’s a real risk to the lazy long who just says “more TPUs = linear growth in optical modules”. Bandwidth efficiency could end up being a genuine headwind to a naive links per GPU/TPU model. We think it’s important to focus on where the odds of capturing the next rate transition are much higher related to expectations. Our read is simple: TPUs shift bandwidth.

Google can flatten pods and keep more traffic local, but the surviving links still need to jump from 800G to 1.6T to keep panel counts, thermals and power budgets under control. The chip count is also compounding fast enough that total endpoints rise even if links per unit fall. Inside the rack, short-reach copper and linear architectures carry part of the load, which changes the product mix but not the opportunity set for two names we like.

We came across these companies that aren’t directly in the TPU supply chain but could ultimately do quite well in their slipstream – SiTime (SITM US) and Macom (MTSI US).

SiTime sells the timing layer that every high-speed system depends on: oscillators, clocks and synchronizers that live inside 1.6T modules, NICs, switches and accelerator boards. Macom sits in a similar structural position on the photonics and copper side. Their products aren’t tied to one architecture; they follow the rate transition itself. If AI operators use 1.6T pluggables, then MACOM ships 200G photodiodes, TIAs and drivers. If they push traffic onto short-reach copper, MACOM ships linear equalizers and reach-extension silicon. If they experiment with near-packaged optics, MACOM sells the same photonic front end into that form factor.

SiTime will see their MEMS oscillator content in a 1.6T optical module close to double (to about 2 dollars per module), and Macom maintains >60 percent share in the 200G photodiodes used in those same 1.6T modules. Both of these companies posted beat and raise quarters, with stronger revenue and margin expansion. MACOM’s 200G per lane products will see a significant ramp in 2026, along with their photonics products.

Since the AI trade kicked off in earnest, SiTime has lagged in performance relative to Coherent, Lumentum, and Ciena due to their exposure to IoT and mobile.

They don’t operate in the most sexy niche – after all “precision timing” doesn’t really have the same draw as “AI Accelerator”, but they are pretty much the only player in this space.

Over the past five years, SITM has become a much more diversified company. While IoT, Consumer & Mobile still make up a significant portion of their revenue, Communications, Enterprise & Datacenter (CED) has more than tripled.

The CED segment is set to make up almost half of its revenue, and the company is guiding to $100-103 million in Q4 revenue with around 60% gross margin. This implies an exit revenue run-rate above $400 million annually - driven heavily by AI infrastructure. It’s an entirely reasonable assumption that CED/AI revenue compounds at something like 30% annually over the next three to four years (as hyperscalers keep building out AI capacity) and as SiTime increases content per rack through products like SiT5977 and TimeFabric.

As we mentioned earlier, TPUs use OCS rather than DSP retimers – if the entirety of NVDA GPU volume became TPU volume it would likely be negative for SITM. Since we don’t see that happening, we see the higher demand for high speed optical modules being an overall net positive.

This renewed focus on 1.6T could not come at a better time to position here – the recent market volatility saw SITM’s shares drastically appreciate following earnings and then sell-off on little more than beta. SITM reported a blow-out quarter, with revenue up 45% YoY, expanding gross margins and EPS more than doubling.

Rough back of the napkin math: estimating ~20M 1.6T optical module units in 26 with $2 content per module, gets us to ~$40M of MEMS timing TAM in 2026 purely from 1.6T modules. SiTime is likely to secure between 25-50% of this. While these aren’t heroic numbers compared the whole company, consider this is just one socket in the AI rack, it’s driven by a specific, visible product cycle and it is incremental to their timing content in NICs, switches, routers, baseboards and other accelerators.

The company has cited multiple times the demand for their oscillators in increasing GPU efficiency and lowering latency. On the opportunity for 1.6T, they said:

“For example, optical module bandwidth is doubling to the 1.6 terabit level and these new modules are beginning to ramp now. Demand for the 1.6 terabit modules has recently doubled, indicating a sharp transition to 1.6 terabit technology in first-half 2026. Additionally, SiTime’s oscillator ASPs in this application are higher because of the higher frequency and performance requirements.”

If the first bull-case for SITM is the move from 800G to 1.6T optics in AI data centers being pulled forward by Google’s TPU build-out, where SITM’s oscillator content per module roughly doubles, the second is much more firmly in the tail. As we mentioned earlier, SITM has failed to keep up with peers primarily due to the tough handset environment and their still-significant mobile segment. As Apple begins to integrate with Google to improve AI, we’ll likely begin to see more investor focus for inference-on-device (see our memo here) and AI-centric wearables. Similar to our expectation that ASICs end up taking up more share of the AI Accelerator market, our view on AI devices has taken a bit longer to play out. Still, we would not be surprised to see that it is dominating the discussions on AI hardware towards the end of next year.

What does that look like? While data centers will continue to train and host large models, handsets, watches, medical devices and other wearables will increasingly run distilled, on-device models. These edge devices need very small, low-power, highly stable timing components that can live inside the RF front end, radios, and sensor hubs without blowing the power budget.

SITM is firmly in that world today. Apple has chosen SiTime as the exclusive supplier of MEMS oscillators for its in-house 5G modem, which should translate to ~$240M of revenue in 2027 just from that modem win. Their Symphonic integrated clock is another vector of growth in AI wearables – collapsing multiple clocks into a single, low-jitter, low-power device. That will become increasingly important for battery-constrained AI endpoints (think earbuds doing on-device translation, watches doing health inference or medical patches doing anomaly detection).

On-device AI will drive a new set of constraints: always-on sensing, local neural nets, real-time sensor fusion (for glasses and AR devices) and continuous-learning. Lower power despite more complexity has been a crucial pathway to outperformance in the AI complex, and it will only be increasingly important in the AI device complex.

In late 26-27, they’ll begin recognizing revenue from their new resonator Titan. This is the first time the company has gone below its oscillator products and targeted replacing the quartz resonator itself. Titan is a chip-scale MEMS resonator, with the smallest timing device footprint (~0.46mm x 0.46mm), ~50% lower power, 3x faster startup and stability across temperatures.

Conclusion

We are still on cautious footing into the end of the year (we’ve reduced from +100% net long to +80% net long and significantly hedged, while planning to raise more cash). However, keeping track of the implications of the AI accelerator market and general data center capex will continue to be important regardless of the macro-gyrations of the market and give us a good list of names to buy in any further volatility.

One point we’ve heard on the bearish side is the concern around MAG7 cash flow going towards capex rather than buybacks. As shown below, share count reductions have slowed or reversed for much of big-tech. To be fair, some degree of this has been offset by Nvidia’s own increased buybacks.

To the extent that these companies have a low cost of capital and believe in the ROI of AI capex, they will be less focused on returning shareholder capital. AMZN and ORCL shares outstanding have climbed, while MSFT is maintaining buybacks at a level that’s just enough to offset stock-based compensation.

Google and Nvidia have kept pace with buybacks since the AI capex craze kicked off, but with Google potentially ramping up capex on TPUs and Nvidia having to reinvest more capital to compete with them, it’s possible that we could see a simultaneous erosion of Free Cash Flow (FCF) for both parties and a structural contraction in shareholder yield. In short, the industry at large is shifting from a high-margin ecosystem to a capital-heavy utility model where cash is trapped in silicon rather than returned to shareholders.

Further, the sheer pace of capex acceleration does mechanically result in an overstatement of earnings relative to cash flow. Usually we can ignore these sort of accounting discrepancies as rounding errors, but over the past twelve months the gap between hyperscaler capex and D&A has exploded to over $200 billion – a material portion of the entire market’s earnings.

We flag this as a point to watch. We don’t think it’s something that will drive markets currently, but we could see the narrative forming in a downturn.

I’d also be remiss if I didn’t flag that it seems Jensen has gotten really, really concerned about Nvidia’s share price. See NVDA’s very odd tweet about Google:

And Nvidia taking the time to address 20 different bear theses in a communique…

…is just kind of weird. I’m used to small cap companies being very bothered by short theses but not the largest company in the world.

We will continue to position for the changing dynamics of the industry while also recognizing the potential risk of a broader reset in expectations.

In our last Macro Memo, we expected:

• Slowing Without Breaking - labor market weakens but doesn’t collapse; further wage growth and rent inflation slowdown offsetting goods reflation.

• K-Shaped Economy Deepens - Lower-end consumer struggles (negative real wages, credit stress); wealthy cohort’s spending keeps aggregate consumption stable enough to avoid slowdown.

• China Tensions Peak Then Ease - Tit-for-tat escalation through October, then deal-making talk resumes with each seeking off-ramps amidst escalatory rhetoric. In a follow-up message on Substack, we further expressed high confidence that the US and China were highly motivated to quickly find a resolution to the export controls.

• Bull Market Continues - Not over; expect 5-10% pullback in near-term but significant upside remains longer-term.

• Fiscal Support Until Midterms - Trump administration will backstop markets/economy ahead of midterms with policy stimulus that will show up in 2026.

For the next four weeks, we expect:

• The US job market continues to deteriorate based on private data.

• A deluge of post-shutdown economic data will unleash volatility over the next few weeks.

• End of government shutdown and QT to improve funding market liquidity.

• The Fed leaves the policy rate unchanged in December out of caution, STIR prices in more cuts to 2026.

• Risk-off sentiment continues to hit speculative areas, such as crypto, quantum and bitcoin miners, out to mid-December.

• We see a resolution of fears expected to lead to a Santa Rally as fiscal expansion is priced in for 2026.

Before we get to our trades, a quick review of the economic developments since we last published…

US Economy and Job Market

How is the US economy doing?

Because the federal government has been in a shutdown since Oct 1, we are missing many macro data points. In our last macro memo, for the US Economy and the Job Market section, we said

“With the U.S. federal government shut down, last month’s public macroeconomic data are unavailable. Even so, we can still form a reasonable view of overall macro conditions and labor market health. The picture that emerges is one of ‘slowing without breaking’. The U.S. economy remains starkly K-shaped: the wealthy continue to thrive on asset appreciation, while much of the middle and lower income segments face persistent strain.”

Four weeks later, our opinion hasn’t changed. Almost every element of our earlier assessment still holds and is increasingly validated by private data and corporate earnings. Reports from companies such as McDonald’s, Chipotle, Cava, and Tyson highlight softening demand among lower-income consumers, consistent with the K-shaped dynamics we described. At the same time, we are growing more concerned about the downward momentum emerging across many of the indicators we track.

Since our last macro memo, private data sources suggest that the US labor market continues to slow. ADP nonfarm private payroll growth continues to show a steady slowdown, despite month to month noise. When the official NFP job growth stats are released for the month of September on November 20 (apparently the October Jobs Report will be delayed), we expect to see that the ADP numbers are corroborated.

There was some controversy surrounding the “bias adjustment” ADP made based on the QCEW adjustment (good discussions on the topic here and here). This is really getting into the nerdy weedy stuff, but based on everything we have read, we see little reason to think the ADP numbers are wildly off the mark and misrepresent the state of the labor market.

We also received another reading of the US job openings numbers from the private job posting site Indeed. Here, once again the number of job openings trended lower. Over the last few years, job openings numbers from the official JOLTS report and Indeed Job Posting Index have shown a high degree of correlation, making us believe that once released, JOLTS job openings show a similar decline.

We care about job openings because job vacancy reduction has borne most of the brunt of the labor market cooling since the Fed started hiking. The working theory of the Fed is that job openings should be allowed to decline up to a point, beyond which we should expect the unemployment rate to rise, through either outright firings or simply a lack of hiring. Many forget the unemployment rate, like a gauge of water level in a bath tub, can rise due to faster exit or slower entry at a constant rate of exit!

While we wait for the release of public economic data, our baseline view is that the trends in the US labor market are unlikely to have changed. If anything, we’d be worried about those trends accelerating further downward.

In terms of labor market turnover, the hire and quit rates were trending lower, while the layoff rate has been going sideways. In a classical business cycle, falling hiring rate is what leads to higher unemployment rate before outright firing begins in earnest.

Speaking of unemployment rates, they have been slowly but surely grinding upward. Historically, job opening rates being lower than 4.5% has been associated with rising unemployment rates.

Construction sector payroll growth continues to decline. Note that this does not mean that the construction sector is outright shrinking, but rather growth has slowed. Historically, the housing sector has been predictive of the business cycle.

Consumer confidence and sentiments continue to trend lower – a defining characteristic of this business cycle. The UMich consumer sentiment seems particularly bad, but it’s similarly corroborated by the survey from the Conference Board. They both agree with consumers’ perception of the job market.

Goods Spending Rebounds and Services Spending Slows

Wage growths continue to slow.

Despite the deteriorating trends of the labor market, the overall US economy seems to be so far holding up. In fact, the Q3 real GDP NowCast from the Atlanta Fed is projecting a whopping annualized 3.7%! Now these NowCasts were made during the government shutdown. So with the release of the make-up stats, we could see a revision, but in all likelihood the new data will affect Q4 GDP much more. Could we see a sharply cooling labor market while the overall GDP holds up?

Consumer Strain Lingers

Delinquency rates among consumer borrowers remain concerningly high, and may move higher. This is among the best demonstrations of the K-shaped economy. Auto loans in particular stand out as one of the poorest performing categories of consumer credit, with metrics continuing to deteriorate.

By contrast, mortgage delinquencies are not high by historical standards, reflecting the fact that American households, especially the older and richer ones, have generally de-levered and accumulated significantly home equity over the past decade. Yet even here, it’s unmistakable that delinquency rates are rising. Student loan delinquencies have surged too as the COVID moratorium has ended, and non-paying accounts are again recognized accordingly.

Below the paywall, we go further into the setup for macro markets in the weeks to come and share our current positioning.

Read more

Fifteen months ago, we initiated our Healthcare Innovation basket with a clear and (in our humble opinion) compelling long-term thesis: advances in medical technology and life sciences were on the brink of unleashing a super-cycle of more effective and efficient healthcare, and yet the market was overlooking the opportunity due to transient cyclical factors. But the situation quickly became complicated.

Almost immediately following Election Day last year, we faced Markets Gone RF-Krazy as investors panicked about the new Trump administration’s healthcare policies and the potential impact on company fundamentals and valuations. At the time, we argued that this reaction was directionally incorrect and that the administration’s new policy agenda would ultimately prove highly supportive of our healthcare industry thesis.

Since then, we’ve seen a see-saw of policy-induced freakouts and easing concerns – again (January Musings), again (Healthcare’s AI-Powered Trump Card), again (Insuring Against Insanity), again (Talk to Me, Goose), and again (Rough Seas Make Strong Sailors). Each time, we made the case for staying the course and used the volatility to increase our exposure to the theme. And if that was not contrarian enough, we focused our healthcare exposure on *gasp* small caps, the market’s ugliest step-children over the last several years.

But this conviction has been rewarded, at least in the short-term, with various proxies for our thematic staging a sharp rebound and the basket itself handsomely outperforming those proxies.

More than just a deadcat bounce, we think this the start of a multi-year industry upswing that leaves us more positive on the theme than ever before. In the pages that follow, we detail what we are watching and how we are positioning ourselves to best capitalize on this opportunity.

Patient Status: Industry Health Improves, Investors Still Feel Sick

Anyone invested in the healthcare space over the last twelve months will remember the headlines: “anti-science” MAGA officials were going to defund healthcare providers and researchers, increase red tape, and cause progress in the industry to grind to a halt.

Despite these scary predictions, very few have come to pass. Instead, we have seen the following:

• Threats of major research funding cuts have been walked back or reversed

• Major pharmaceutical companies have struck deals with the US government to boost domestic manufacturing capacity while reducing the risk of negative pricing actions

• Landmark approvals have been handed down in the most cutting-edge areas of medicine

• Reimbursement for innovative new diagnostics has received more support than at any point in the past

These dynamics have helped contribute to strong industry fundamentals since our last update. The number of companies in healthcare guiding above analyst expectations in the last two quarters is near record highs.

You might expect that the combination of improving news flow and positive fundamental surprises would have investors feeling euphoric, but the reality is quite the opposite. Despite this improving backdrop, long positioning in healthcare hovers around the lows of the last twenty years, while short interest in some of the most dynamic corners of healthcare is testing a post-COVID high.

The combination of negative sentiment and positioning has weighed on valuations, which remain depressed despite recent equity appreciation in corners of the healthcare industry.

We have previously stated our preference for smaller companies in life sciences and medical technology, especially those with long growth runways. As the current revenue multiple of the Russell 2000 healthcare growth sub-index demonstrates, there is significant upside in a reversion to average historical multiples (especially as fundamentals improve and interest rates come down).

What happens when improving fundamentals meet depressed valuations for a prolonged period of time? We have predicted in the past that if public market participants ignore this dynamic for too long, strategic investors would step in to take advantage of the disconnect. Indeed, so far in 2025 we have seen a surge in healthcare M&A, with the number of deals so far in 2025 tracking substantially ahead of any prior year on record.

We think more than a few of the members of our basket are attractive targets for potential M&A, if the public market does not recognize the opportunity first. So, with that said, let’s dive into the specific themes within our healthcare coverage.

Read more

Background

Natural gas might not be as sexy as solving energy storage, or making GPUs more efficient, but it is the way we’re powering AI. AI eats electrons, electrons eat gas.

We’ve focused quite a bit on power and electricity, but not so much on its source. There’s no way around it, if we want to power these data centers, we’re doing it with gas – at least for the foreseeable future.

This incremental demand comes at a time when gas supply and demand balances were already becoming tighter with the consistent ramp of new LNG export facilities. This confluence of factors will ultimately result in natural gas prices moving structurally higher.

For the past fifteen years, drillers from every corner of the United States have overwhelmed the natural gas markets with persistent supply growth. Most of this growth has been completely price insensitive. In shale’s infancy, capital markets rewarded production increases over cash flow. Then, associated gas (a by-product of oil-focused drilling) swamped an already oversupplied market. For well over a decade, gas has been plentiful and dirt cheap.

This has led to a gross miscalibration of how much production growth drillers will deliver on a go-forward basis, and the price at which they’ll deliver it. As oil production from the Permian Basin begins to plateau, and we enter the final innings of the associated gas deluge; the US’ natural gas supply stack becomes increasingly price sensitive and weighted towards dry gas (i.e. not oil-bearing) plays, where any incremental growth will require an economic return.

This represents a shift in natural gas market dynamics.

After a decade dominated by supply-push, the market will now be driven by a demand-pull. New demand growth in the US will be driven by two key factors (and the competition between them):

• Extraordinary new demand from LNG exports

• Extraordinary new demand for domestic electricity driven, in part, by the massive AI Data Center buildout

Below we chart the implied call on natural gas over the coming years. This call on dry gas nets out associated gas production adds from the Permian Basin to show what dry gas plays must deliver – the volumes are material. For context, this amounts to adding another Appalachia (the largest dry gas play in the United States).

While many focus on the short-term factors that drive natural gas prices, like the weather, we’ve chosen instead to focus on the market’s fundamental equilibrium. That is to say: where long-term supply and demand meet to clear the market, both today, and in the future. We do believe higher prices will lead to production growth – though in this piece, we explore why the trigger price for incremental natural gas growth is moving structurally higher, along with the physical limitations that future growth may face from an infrastructure, regulatory, and remaining inventory perspective.

Don’t get me wrong – most of the stocks in this space will continue to trade in a volatile manner, with high beta to commodity prices and weather forecasts. Therein, we think, lies the opportunity – by focusing on what will happen over the next 3-10 years, we can look through the volatility to see the trend. Because of the far-reaching implications, it’s also possible to focus on names with a bit less in the way of commodity beta, but we think that, ultimately, both supply and demand need to adjust, which only higher prices can catalyze.

This is our part-primer/part-thesis on natural gas and where the opportunities lie.

Note: As with most commodities, natural gas is not generalist-friendly - therefore, we’d like to shout out HTM Energy Partners and their BasinScout product for making sure that we didn’t miss any of the nuances while still presenting the case within.

Why Now?

U.S. power demand from AI and electrification is colliding with the largest LNG build‑out in history. Very long-lead export infrastructure is being front-run by price-insensitive power generation infrastructure at a time when natural gas plays are maturing. It’s partly geology, and partly the wall of demand that’s been built through a decade of low prices.

Every single gas-related demand factor is surging up and to the right, while dry gas supply is stagnating. It really is that simple.

Feel free to roll your eyes. A natural gas bull market starting next year isn’t a novel idea. The gas market has chewed up and spit out so many speculators over the last decade that some would find it impossible to see gas as anything but a completely unpredictable commodity removed from any fundamentals.

We agree. For the past decade, as price-insensitive growth dominated, the market lacked any fundamental regulator. E&Ps trying to fill infrastructure, hold acreage, or grow oil volumes had no care in the world as to what price they realized for their gas.

Today, however, both the Permian Basin and the dry gas basins – the two largest gas supply buckets – have matured, and on a go-forward basis, very much do care about pricing. The US has churned through much of its lowest-cost supply base. Despite this, forthcoming demand growth is only accelerating in the face of declining price-insensitive supply.

The two demand factors that dominate the market are LNG export growth, and power generation growth to service new electricity demand. This encompasses industrial onshoring, electrification and, yes, data centers. As we detail below, over the next decade, we expect ~35 Bcf/d of demand growth between high-confidence LNG and power generation needs.

Notably, we only assume ~100GW of baseload gas-fired power generation growth which maps to US-deliverable utility-scale gas turbine capacity in the medium-term. There is upside to power generation growth with simple-cycle engine installations (the likes of which we observed in our Stargate article as well as in the xAI Colossus buildout) and turbine manufacturing capacity expansions. Currently, turbine manufacturers like Siemens Energy (ENR GY) and GE Vernova (GEV) are being extremely disciplined about capacity expansion – any changes to that and this forecast would move materially higher.

A more expansive view brings the total potential call on US dry gas supply to nearly 70 Bcf/d – a 70% increase from current production levels, even after underwriting ~11Bcf/d of structural associated gas growth from the Permian Basin. This is too much demand!

There are two key supply factors – dry gas production (primarily Appalachia and the Haynesville in Louisiana), and associated gas production (that is, gas produced from oil wells, almost entirely the Permian Basin). Together, these two factors account for ~70% of US Lower 48 gas production.

At current NYMEX forward pricing, we expect 15-20 Bcf/d of supply growth. Compared to our high-confidence demand growth forecast of ~35 Bcf/d, it’s suddenly very easy to appreciate that the natural gas market is at a fundamental inflection point.

Dry gas production in Appalachia (the Marcellus and Utica formations), which provided the majority of growth over the past decade, has already reached a clear plateau for several years. Associated gas from the Permian will continue to grow, but it is just not sufficient to satisfy the wave of new demand.

Like demand, there are numerous smaller-scale supply factors growing at the margin (i.e. the Austin Chalk, San Juan, and the Piceance). In aggregate, these marginal supply factors will largely offset declines of legacy production from other basins, and certainly come nowhere near filling the potential supply gap that only structurally higher prices can solve.

Of course, at higher pricing, certain stated LNG projects won’t be built, and gas drillers will ramp production – but therein lies our thesis. Current NYMEX forward pricing will not incent the growth needed – the price discovery of this new equilibrium will continue to move higher as the decade progresses.

The supply side of the market will become increasingly elastic while demand becomes inelastic. Price-pass-through demand (i.e. LNG exports under long-term offtake agreements) and new inelastic demand (i.e. hyperscaler power generation) continues to dominate the makeup of incremental consumption. This mismatch is the perfect recipe for higher prices.

The balancing factor in the gas markets over the past decade has been associated gas; which has an incremental cost of ~$0. Going forward, associated gas is no longer the balancing factor, and consumers/exporters will have to begin paying for growth when they need it; structurally changing the US’ natural gas market to demand-pull, from supply-push.

Henry Hub (NYMEX) futures are in steep backwardation, with early-2030s pricing at ~$3.20/MMBtu. This is unsustainable considering how gas market balances are evolving. The longer the back-end of the NYMEX curve stays low, the more gas-demand projects pencil under the guise of cheap gas forever.

This repricing will largely benefit incumbent producers that will see the value of their reserves rerated much higher. This primarily benefits Appalachian E&Ps with the deepest inventory, but also the handful of producers that have meaningful growth assets in the Gulf Coast; namely Comstock (CRK US), whose new Western Haynesville play is one of the only remaining dispatchable assets in the US that isn’t severely limited by long-haul egress.

When we look forward over the next decade, it’s clear that in a vacuum, one side of the equation needs to give – supply needs to move higher (which will require substantially higher NYMEX prices), or planned demand needs to move lower (which will require not only higher prices, but the stranding of hundreds of billions of dollars worth of infrastructure).

Both sides of the equation will behave dynamically, though directionally, both sides of the market’s balance equation need to adjust themselves. This can only be catalyzed by higher NYMEX prices.

Even without any new projects, the US has clear line-of-sight to 3-6 Bcf/d of consistent annual natural gas demand growth, while the Permian, the latent growth engine over the past decade, will see growth slow to ~1.5-2.0 Bcf/d; leaving dry gas plays to balance the market. Meanwhile, the large dry gas producers, like EQT (EQT US), Aethon (private), and Comstock, have all said they need to see $5/MMBtu natural gas to dispatch meaningful growth.

Smaller players will certainly provide some new growth on the way up, but the chunky greenfield gas projects that the market will eventually need, have to come from the incumbent large dry gas producers, at a price that works for them. This dynamic is completely opposite to the last decade of the natural gas market.

The longer futures remain stagnant, the more violent the repricing will be.

Below the paywall, we push back on the current bearishness surrounding gas, dive into the competition between LNG & data center power generation and lay out our basket of winners.

Read more

Macro is a continual story. In raging bull markets, however, the psychology of investors treats each individual development as unique and unprecedented. It’s important to revisit some of the priors we have established:

• The cyclical economy and the lower income cohort of the U.S. economy has been doing poorly for some time.

  • Haves vs Have Nots

• The secular economy, which represents a large portion of the market capitalization, is currently roughly halfway through a cycle of capex spending that sees no sign of slowing. This is the primary driver of equity returns on indices. Outside of this, the rest of the economy continues to look relatively precarious.

  • Stargate

• The past two points, in combination, have resulted in a dynamic we refer to as the Macroeconomic Ouroboros. In this dynamic, the significantly skewed consumer spending volume made up by high net worth/high income households results in the K-shaped economy that is primarily driven by the wealth effect and equity prices.

  • Macro Memo: After the Trump Bump

• Additionally, our house view at CitriniResearch has long been that we are deeply in the era of Fiscal Primacy. Regardless of the political party in charge of the White House, the US has embraced a populist bent since COVID that means government spending will backstop certain industries along partisan lines.

  • Fiscal Primacy

• Because of this dynamic, as well as Fiscal Dominance (Fiscal driving monetary policy), we see an environment that means the administration will take whatever steps necessary to support the economy until midterm elections – where the house is hanging on by a thread for the Republicans. Whether through stimulus checks like the proposed “tariff refund” or the targeting of mortgage rates through lifting the cap on the amount of MBS the GSEs can hold, the administration will try to keep asset prices elevated for as long as possible.

  • Macro Memo: Summer is Over

• Tariffs, as presented on April 2nd (which, obviously, has not occurred) would lead the U.S. economy into a period of Stagflation. The administration is aware of this and is currently balancing that reality with their stated trade policy and foreign policy goals.

  • Seeing the Stag

• Considering the above, we take the nuanced view that:

  • a) Trump’s desire to restructure trade and tariffs is not going away. This will remain a primary driver of equity volatility for the duration of his term.

    • Tariff-ied

  • b) To accomplish this, Trump will repeatedly bring us to the brink – progressively heightening trade tensions, but utilizing shiny objects and promises of deescalation or stimulus in an attempt to prevent collapse.

    • This is a view distinct from the mainstream “TACO” narrative. At the risk of being deemed reactive or trigger-happy, we think each individual development in the ongoing tariff drama should be weighed along a spectrum and trend – negotiating tactics, to be sure, but threats that still entail risk. Both sides, the US and China, will ramp up rhetoric and repeatedly search for off-ramps without putting themselves in a weaker position.

    • The best way to assess when and how these episodes wax and wane is to have a strong understanding of the constraints which are currently: Debt, Demographics, Recession-Risk, Midterm Elections and Defense.

      • Macro Memo: Who Holds the Cards?

      • Macro Memo: Summer Blues

With the caveat that this is simply another paragraph in the longer story, let’s describe the current status. For the past couple months it has felt like macro, especially the Fed, no longer mattered. Since the September rate cut decision, there were momentary doubts over the wisdom of the resumption of rate cuts.

In our last Macro Memo, we predicted:

• Cyclical economic slowdown and labor market weakness (despite data-obscurity in the wake of the government shutdown, we believe private data supports this continuing)

• The Trump Administration to attempt to lower mortgage rates utilizing GSE privatization and deregulation (a tactic we believe we’ll see used by Q1 2026)

• Self-fulfilling seasonally turbulent asset markets for the months of September and October (half-right, potentially)

• Increasing short-term inflation concerns (likely to reaccelerate with renewed tariff concerns)

• Fed rate cut in September but uncertainty further out as Fed board composition hangs in balance (the Fed cut in September and, it appears, will cut again at the next meeting. December remains less clear)

• Our take now: “it’s not oct-over”

Read more

CitriniResearch has been an early and vocal believer in the scale of the data center buildout necessary to implement and realize the full potential of Artificial Intelligence. In our inaugural AI primer, we dubbed Phase 1 of the AI Trade the “Global Data Center Hyper-Scaling” - a tongue-in-cheek reference to both the hyperscalers involved and the coming scale of the buildout. More than two years ago, we wrote:

“The first phase [...] is centered around the development and build-out of robust computational infrastructure. This involves a massive and sustained tailwind to the suppliers of data center equipment of accelerated computing required for both AI training and inference”

“The enormous computing power required for AI/ML places considerable demands on a variety of technology sectors. The companies involved in providing the necessary infrastructure stand to benefit significantly from this heightened demand. This includes businesses involved in semiconductor manufacturing, networking equipment, data center construction and management, and more.”

Artificial Intelligence: Global Equity Beneficiaries (5/31/23)

Since then, we’ve spent our time in front of our screens, watching green numbers climb higher across the supply chain involved in this massive effort. But there’s a risk of failing to appreciate it that is inherent to that; sometimes you just have to bear witness.

Screens make it easy to forget the physical world behind the numbers. Chips don’t float. Fiber doesn’t spool itself. Power doesn’t magically appear at the edge of the rack. At some point, you have to step away from the chart and go stand where the megawatts meet the mud.

So we went to Abilene, Texas. Or, at least, we sent a drone to Abilene, Texas with the help of investigative journalism firm Hunterbrook Media (who we partnered with on our Teradyne (TER) thesis). Maybe I’m slow to the uptake but, for me, seeing the progress there really drove home the scale of what’s going on right now.

Roughly a year ago, there was dirt.

Now, it looks like this:

On the ground level, nine months ago, it looked like this:

Today it looks like this, and this is just one section:

These images may seem mundane, but one has to consider that they represent just a tiny fraction of the datacenter development breaking ground across the country. This single facility represents 600 football fields filled with cables, fiber, pipes, generator sets, transformers, chillers, switchgear, cable trays, fuel tanks, batteries, copper busways, electrical components (2N redundant), pumps, air handlers, fire suppression systems, all the fancy actual computing and cooling gear inside the data halls and 200,000 tons of other equipment that collectively consume enough electricity to power the city of Seattle and must reject a level of waste heat 24/7 that creates thermal plumes visible on weather radar.

Each is literally a “Ludicrous Building.” And there are eight under construction on this single site. And this is just one of the Stargate data centers, there are many more. xAI’s Colossus 2, Meta’s Louisiana Hyperion, Microsoft’s Fairwater, Quantum Loophole, Nebius’ New Jersey DC, Amazon’s Pennsylvania Project, Google/Anthropic’s Project Rainier - just to name a few.

Together, the current footprint of this complex is roughly the size of lower Manhattan:

Which pales in comparison to Meta’s Louisiana facility, which would stretch from the top of central park down to SoHo:

Now our tagline for this article starts seeming a lot less like hyperbole and a lot more like an understatement. “What if Manhattan, but computers?”.

Don’t take our word for it. Really, don’t. Watch some of the footage we’ve uploaded here:

You can see the currently completed data halls (one is currently being used for early training/inference runs by OpenAI), as well as the new construction that will massively expand the current capacity. The scale and speed of the buildout is truly immense, and – for me, at least – awe-inspiring. Don’t let anyone tell you the U.S. doesn’t build things anymore.

Despite the fact that we’ve owned or covered many of the companies involved in this (Vertiv (VRT), Eaton (ETN), Schneider (SU FP), ABB (ABBN SW), Siemens Energy (SIE GR), GE Vernova (GEV) etc). We’ve decided to take the weekend to drive home what the “data center buildout” really means from the outside, rather than the inside, of the data hall. We’ve spilled a lot of ink zoomed into the data center, focused on racks, high-bandwidth memory, pluggable optics, active electrical cables, racks, switches, GPUs, ethernet fabric. Now, we’re zooming out in a very literal sense.

This is the clearest demonstration yet of how quickly the AI infrastructure machine is spinning up, and how many industrial hands are turning the gears. The execution tells us more: a who’s who of global vendors rushing to bolt down turbines, transformers, chillers, tanks, and cables in the Texas plains – and they’re powering it all with natural gas, one of the only reliable and dispatchable fuels we have today.

These projects are so large that even multi-national corporations cannot fulfill the entirety of the needs, with multiple contractors and OEMs all working towards a singular goal: the largest industrial investment cycle for the United States since World War Two. This requires the builders to call on anywhere and everywhere to realize their ambitions.

In short, “tech” has become truly capital intensive again for the first time in a quarter century. For the broader economy, this is an enormous development. But a picture is worth a thousand words, so drone footage must be worth a thousand charts...

At the Stargate Data Center in Abilene, by mid-2026 eight buildings will be completed representing approximately 4 million square feet with a total power capacity of 1.2 gigawatts (GW). And this is just one site. Although we’ve decided just to focus on this one (somewhat arbitrarily, somewhat because the “Stargate” moniker was leading to doubts), the important thing to keep in mind is that there are going to be tens if not hundreds of similar projects built over the next five years.

Five new Stargate sites are already in the permitting or construction stages (Shackelford will have its first building complete in H2 26) and there’s already discussion about expanding the Abilene DC by another 600MW.

We already have published on the need for:

• More compute in May 2023, in our article “Artificial Intelligence: Global Equity Beneficiaries”

• More power in May 2024, in our article “The Utility of Bubbles: The Intersection of Fiscal & AI”

• More connectivity in July 2024, in our article “Can You Hear Me Now? The Coming Optical Shortage”

We have continually tracked what’s happening inside the data center. Which, to be clear, remains vitally important and is not slowing down:

“At completion, each data center building will be able to operate up to 100,000 GPUs on a single integrated network fabric, advancing the frontier of data center design and scale for AI training and inference workloads. The first phase includes two buildings optimized for both direct-to-chip liquid cooling and air cooling, ensuring flexibility and cutting-edge performance. Upon completion, the data center will support up to 100,000 GPUs on a single network infrastructure.”
Source: Crusoe/DPR Communications

But today, we’re doing two things. The first is appreciating that we were right back in 2023, the scale of the data center buildout is indeed “hyper”. Second, we are taking a closer look at who’s actually building the data centers and facilitating their immense needs. It’s a bit less exciting than who’s developing the hardware upon which the Machine God will run, but after viewing the scale of this project we believe it requires some additional attention. Who makes these massive facilities appear from the ground up, like an alchemical transmutation of money into power consumption and compute?

Because there’s a lot that needs to be done, and a lot of money to go around. Today, it’s Texas. Soon, it’ll be New Mexico, Ohio, Wisconsin…everywhere there’s enough energy and land to build them. Therefore, paying attention to the blueprint is a good idea.

Below the paywall, we compile the names involved or likely to be involved in the continued wave of massive infrastructure building for AI. Included is a breakdown of 50 high confidence and 200 low-medium confidence public companies that will continue to see upside to these megaprojects.

Read more

Given the transformational nature of the AI thematic, it is perhaps no surprise that most of our single stock ideas – from Celestica (CLS) to Credo Semiconductor (CRDO) to American Superconductor (AMSC) – have been based on the role each company plays in the AI infrastructure buildout. You can trace a pretty straight line through the progression of that buildout using our favorite longs. In 2023, we favored the heart of the server – names like Nvidia, Supermicro and Celestica – as the largest concern was getting more GPU servers to train AI models. Then, we favored names like Credo and Ciena, as the infrastructure required to connect servers and span distances inside the data centers became front and center for investors. More recently, we focused on names like American Superconductor and Siemens Energy, jumping out of the data center entirely towards products that manage the high voltage power requirements necessary to run them. Now, we ask, what about the new data centers? We’re not just putting a bunch of AI accelerators into existing facilities, ensuring they’ve got the connectivity and electricity to run and then calling it a day, are we?

No. Regardless of inflation, a few trillion bucks over five years still buys you a bit more than that.

The future is greenfield data centers and the infrastructure required to support them. Given the scale of the buildout, the infrastructure implications go beyond chips and servers to things like resource consumption, maintenance and the capacity of society’s most foundational utilities. As a consequence, the impact of AI is converging with some of the other megatrends we cover, such as fiscal primacy. Indeed, as the buildout progresses, we’re seeing an expansion of technological capability – spearheading the trend’s evolution even further into embodied AI, or robotics. Naturally, most stocks that have obvious exposure to multiple areas of narrative momentum are already trading at the higher end of their historical valuation ranges. But that’s nothing a little detective work can’t solve.

Below the paywall, we discuss a relatively undiscovered name with significant upside to trends in data centers, grid reliability and automation. We cover the core business as well as recent developments that pave the way towards unlocking a right tail scenario for earnings. As fiscal priorities, technological innovation and AI capex spending converge, we feel this name represents an opportunity to pick up a secular trend at a multiple reflective of a lack of understanding by the market.

Read more

In this article, we review:

• US Economy and Job Market

• Long Bond Yields: Secular Meets Cyclical

• Tariff Revenue & Bond Yields

• The Housing Unlock?

  • The Trump Admin’s Plan for Mortgage Rates

• Macro Trades Ideas

In our last Macro Memo, we predicted:

• The US economy and job market will continue to slow without breaking. [Yes]

• Overall inflation flat or lower despite tariff effects on goods prices. [Half Right]

• “The Powell Struggle” leads to fiscal - monetary convergence.

  • Tariff uncertainty resolves [Half Right]

  • The Fed to signal openness to rate cuts [Yes]

  • Falling long yields to unlock housing market [Yes and No]

• STIR markets price more cuts into the first year of Powell’s successor’s term versus out to the end of Powell’s term [Unch]

Going forward, we expect:

• Cyclical economic slowdown and labor market weakness

• The Trump Admin to attempt to target lower mortgage rates utilizing GSE privatization and deregulation

• Self-fulfilling seasonally turbulent asset markets, equity volatility for the remainder of Q3

• Increasing short-term inflation concerns

• Fed rate cut in September but uncertainty further out as Fed board composition hangs in balance

The market is juggling between inflation and growth risks again. Over the last three years, it seems like we’ve experienced an inflation scare every other quarter interlaced with a growth scare in between. Maybe this is just the navigation of a soft landing, but the fact that the US stock market has been on a nonstop tear since the April bottom amplifies these risks during a seasonally volatile period.

While we continue to believe the tariff-driven inflation to be short-lived, it doesn’t matter what eventually transpires if the market is worried about tariff-driven goods inflation, its second-order in services, and non-tariff factors driving a second round of inflation acceleration. Given these short run inflation concerns, a whiff of growth scare (like a weak jobs report), will bring back the cries of “stagflation”.

Something we’ve long debated is whether the Trump administration’s clear desire to get long bonds down is a self-defeating goal, but we believe we’ve found a missing piece - based in the fact that the admin, as the midterms approach for 2026, doesn’t really care nearly as much about treasury yields as they do about mortgage rates.

Key macro data points to watch:

• Sept 3 (today): JOLTS job openings, quit rates, layoff rates

• Sept 4: ADP job growth, Unit Labor Cost, Initial and Continuing Claims

• Sept 5: S&P PMI, AHE, Unemployment rate

Summer is over. Time to worry?

Read more

I can confidently say that 2025 has been one of the most thematically driven years in recent memory. Unlike 2024, which was mostly dominated by one theme (AI), this year has seen broad dispersion across trends. Sometimes a singular environment is easier – if there’s one dominant theme it’s much less likely you’ll miss out. We’ve caught our fair share so far this year – from drones/electronic warfare, to the EU defense trade, AI’s gyrations, and its embodied counterpart Robotics and the Trump trade’s various forms (from DOGE concerns to tariffs to the One Big Beautiful Bill).

At the same time, these kinds of rapid-fire thematic markets make a “misfire” all the more likely. When trends shift around rapidly, shiny and fast moving things can present investors long-term promise when there’s none to be had. We recall not-so-fondly the thematic trades in 5G, IoT and Work-From-Home that burned investors despite becoming real parts of everyday life. Some recent “themes”, like the Digital Asset Treasury trade or Quantum Computing, give us a bit of pause as to where we are in the cycle.

Our efforts are meant to focus on the most durable, investable and market-moving of these trends, but it’s important once in a while to step back and see what’s moving. So, let’s check in on the CitriniResearch thematic universe:

In this quarter’s “State of the Themes” we will cover our continued allocation decisions in AI, Robotics, Healthcare, Defense and Government Policy. Additionally, we bring back a thematic basket we opted to exit in favor of concentration into more secular trends – our “Small Themes” basket.

We see a market environment that is distinctly bifurcated – the AI Data Center trade has gone through multiple iterations, all focused on the next bottleneck. While we may have narrowed our focus in AI a bit early, we still expect to be proven correct about the path forward being more concerned with implementers than infrastructure and selective when infrastructure is the focus.

In AI, the datacenter (“Phase 1”) trade has remained asymmetric towards the bottlenecks – power and interconnects have smashed – while broadly speaking there’s been more significant dispersion among connectivity, edge and servers. Our decision to broaden outside of solely beneficiaries in our “Phase 2” basket earlier this year has kept us performing in-line but has yet to take off.

The Trump administration has continued to be a whirlwind, with policy announcements and executive orders with material market impact being released seemingly daily. It’s an environment that reminds us of the value in withholding judgement – aligning with Trump initiatives like Rare Earth Metals, Domestic Data Center Investment and Defense Tech has rewarded investors handsomely.

On the other hand, after making a splash earlier this year, Europe seems to still be stuck in the talking phase. We have continued to focus our adjustments to allocations along US policy, having booked profits in many of the EU fiscal trades. The YTD performance of EU Fiscal beneficiaries broadly still sits near the top of our thematic strategy breakdown, but has significant moderated returns relative to our broader thematic universe QTD.

It’s a common refrain to hear that the market “has gotten too expensive”. That might be the case in the most common places to look, especially among the hotter themes like drones, but looking towards smaller, more cyclical themes like lower rates, cyclical reacceleration and niche AI upside there are plenty of names trading at very reasonable valuations.

Our basket tracking modest valuation + thematic upside among small and mid-cap US stocks that we created in December has managed to surprise us - overtaking SPX returns YTD last week despite the volatility. We look into the environment to pick up some cheaper exposure to durable themes in the SMID caps.

In this Quarter’s SOTT:

• Dynamic AI

  • GPT-5 Takeaways

  • Google’s Advantage

  • AI Implementation Winners

  • Deepseek’s R2

  • Datacenter Selectivity

  • Optical and Interconnects

  • Edge AI

  • The Taiwan/US Discount Gap

  • Updates to the Basket

• Payments

  • Finding AI Upside in Fintech “Superapps”

• Fiscal Primacy

  • Where AI and Fiscal Intersect: The AI Action Plan / OBBB

  • Solar Energy Stocks

  • The One Big Beautiful Bill Act

  • Korean Fiscal and the “KOSPI 500” era

• Robotics:

  • Regal Rexnord & ABB Partnership

• Electronic Warfare:

  • Updates to the Basket

• Healthcare Innovation:

  • Updates to the Basket

• GLP-1 Long / Short:

  • Updates to the Basket

• Re-introducing: “Small Themes” Baskets

  • Rate Sensitivity: Mortgage Underwriters, Elder Care REITs, Home Transaction Volume Links

  • Virality: What Happens When Sleepy Brands Catch Fire?

• The CitriniResearch Community Picks Basket

Read more

In this piece, we provide an update on:

• US economy and job market

• Tariff and Inflation Risks

• Market correction and/or repeat of Dotcom Bubble

• “The Powell struggle”, rate cuts and housing unlock

• Bond yields and housing unlock and cyclical sectors

• US - China Talk

• Macro Positioning

Our diagnoses and predictions have mostly borne out to be accurate:

• Tariff inflation risks were overstated [Yes]

• Services disinflation continues and overwhelms goods inflation [Yes]

• General US economic resilience with a softening labor market [Yes]

• The Fed makes good news insurance cuts [Not quite, yet.]

• US - China tension continues to take off-ramps [Yes]

• Rare earth and the importance of strategic leverage [Yes]

• European outperformance in FX & Equities would taper off [No/Not yet]

• Our last macro memo broke our 100% YTD hit rate in STIR trading - we managed to catch a bit of a move up in Z5 but haven’t seen any follow through

The Upshot

• The US economy will continue slowing without breaking

• Overall inflation flat or lower despite tariff effects on goods prices

• “The Powell Struggle” leads to fiscal - monetary convergence

  • Tariff uncertainty resolves

  • The Fed to signal openness to rate cuts

  • Falling long yields to unlock housing market

In June, we wrote that the U.S. economy was “humming along.” Two months later, that characterization still fits - there is a distinct lack of fresh macro drama, the likes of which prompted us to begin doing monthly macro updates. Since autumn 2024 our stance has been consistent: headline growth remains resilient even as the labor market cools and disinflation progresses. The intermediate calls that flowed from that thesis have largely been validated, but the macro trading regime has shifted. For the past two years, successful macro traders have largely been those that were focused on fading overreactions at key moments in rates and FX.

President Trump’s recent showdown with Fed Chair Jerome Powell—what we call The Powell Struggle—marks the end of that lull. We now anticipate a transition from Trump’s failed public pressure campaign to a more strategic effort to engineer macro conditions that justify rate cuts. With inflation cooling, the trade war de-escalating, and dovish voices gaining traction within the Fed, we expect the door to policy easing to open soon. This emerging political–monetary convergence sets the stage for a gradual decline in long-term interest rates.

As summer winds down, both the economy and the market stand at a crossroads. Should we expect a re-acceleration of the economy and an end to the two and half year favorable disinflation process? In terms of the stock market, do we see a macro catalyst that triggers a long-awaited pullback? Or will markets continue to grind higher on secular bullish narratives like continued AI capex spending, deregulation, a more supportive tax policy and crypto adoption against a still-supportive macro backdrop. In the latter scenario - a continued strong economy with companies riding the coattails of secular trends despite elevated valuations - we start to get distinct vibes for the potential replay of the Dotcom Bubble that we alluded to in January.

Read more

In case you missed it, our new portal allowing real-time tracking and performance on every thematic basket we cover and our model portfolio recently went live at citrindex.com

Introduction

Two years ago, CitriniResearch’s third-ever thematic primer “Upgrading from Overweight” made a host of bold claims about the new class of weight loss drugs, GLP-1s. First, that GLP-1s represent a breakthrough addressing one of the largest health crises in history - the obesity epidemic. We argued that GLP-1s would become ubiquitous, following a trajectory similar to Lipitor or Viagra. Widespread adoption, in turn, would lead to the first “downtick” in the U.S. obesity rate for the past four decades. That has begun:

Secondly, while we expected GLP-1 manufacturer share prices to increase in the near term, we were unique in predicting that long-term second-order effects of a successful weight loss drug would be the most enticing trade as competition heated up for the manufacturers (leaving only a few key “winners”). While many of our trades on these second-order effects had strong performance (shorts on obesity related medical devices, alcohol suppliers, and unhealthy foods), we felt that moves were primarily the result of the market attempting to “front run” the impact.

This concern, combined with the risk of patent cliff and compounded pharmaceuticals/grey market peptides on GLP-1 manufacturer economics, caused us to exit the basket in November 2024. We realized a 75% gain over the span of 14 months, and promised to continue monitoring what we believe is a generational inflection.

Now, the thing about second-order effects is that the first-order has to play out to some critical degree. Based on prescription numbers, it is clear that more and more people are using these drugs. At the end of 2024, nearly 4% of U.S. adults had a prescription for GLP-1 drugs.

With more than 30% of the U.S. adult population suffering from obesity, one might think that is a drop in the bucket. But these numbers don’t take into account the massive trend in grey market peptides or compounding pharmacies (like Hims, Ro, Musely and others).

With or without a prescription, it’s likely that the number of patients actually on GLP-1s is 2-3x higher than official statistics suggest. A study published in JAMA confirmed that GLP-1s, or at least Semaglutide, is practically over-the-counter in practice, with hundreds of online vendors dispensing without a prescription:

Two years ago, my evidence for GLP-1s entering the zeitgeist was “Can you believe Ozempic got mentioned at the Oscars?!”. Last week, on a ten minute walk down the street in Manhattan, I spotted three signs on the sidewalk for MedSpas advertising GLP-1 treatment. It’s safe to say that our original thesis regarding the “reflexive nature of a successful weight loss drug” has been proven correct.

So, as we rapidly approach a critical mass of people (not just Americans - China just approved its first GLP-1 drug for obesity) receiving these drugs, we have to consider something. We can’t just focus on the number of patients on GLP-1s, but must look to the number of patients having been on GLP-1s for long enough that second-order effects (including negative ones) will begin making themselves known beyond a doubt.

Many of the downstream impacts that we’ve speculated in theory will now become increasingly obvious on income statements. This will only accelerate as we see an explosion of generics and me-toos crushing the cost-curve.

Ironically, the market – which once expected an immediate, pronounced materialization – has now become comfortable with the idea that these slower moving changes aren’t coming at all. This complacency comes just as we are reaching a critical mass of adoption, impacting consumer behavior at scale.

After two years of covering GLP-1s, and roughly eight months after we took our basket off for the first time, we’re ready to put it back on – this time with an increased emphasis on the impacts most unappreciated by the market.

In this report, we cover:

1. The Manufacturers and Supply Chain

2. A Second Time Around for Second-Order Effects

3. Ongoing Therapeutic Innovations & Clinical Data Review

Read more