Citrini, the earliest bull on AI hardware, has significantly cut its semiconductor positions, saying changes in financing structure have made sentiment a fundamental

The viral chip research firm Citrini Research was once one of the earliest and loudest bulls on AI hardware trades. Today, Citrini has substantially cut its semiconductor positions and even holds some shorts. Behind this shift is his judgment about a fundamental change in the financing structure of AI infrastructure: when incremental capital switches from the free cash flow of hyperscale cloud providers to equity and debt markets, market sentiment itself becomes part of the fundamentals.

On October 21, Citrini's founder James van Geelen recorded this conversation on the Monetary Matters podcast. At that time, semiconductor stocks had just suffered a nine-day losing streak—host Jack noted that this was the longest streak since 2006. James's judgment was that this was not yet enough to shake AI construction spending, but the nature of the risk had changed. He said: "If market sentiment takes a sharp turn for the worse, that incremental capital will dry up—because this is no longer a company spending its own money; it is decided by the market."

The conversation covered why he rotated out of semiconductors, the logic behind being long Meta and short Google, what DeepSeek's roughly 80% inference gross margin means, and what signals would make him go long chips again. At the same time, he disclosed that the bottleneck in new drug R&D is "a shortage of monkeys," not compute, and that the real investment opportunity lies in clinical outsourcing, not AI concept stocks.

When the giants' cash flow turns negative and the financing structure changes, sentiment can kill the fundamentals

To understand why Citrini cut AI hardware, one must first understand what he calls "reflexivity" risk—this is the core mechanism of the entire judgment.

In the first three to four years of the AI trade, hyperscale cloud providers paid for AI infrastructure with their own operating cash flow. This meant that no matter how market sentiment fluctuated—whether tariff shocks, unwinding of yen carry trades, or any other panic—as long as companies like Meta and Microsoft kept spending, semiconductor companies' fundamentals had a "floor" underneath them. Sentiment could push share prices down, but it could not crush revenue.

Now this mechanism has failed, and the cash flow of tech giants is rapidly drying up or even turning negative. Citrini James's judgment is that almost every incremental investment in AI infrastructure today comes from equity financing (whether venture capital or public markets) or debt financing.

This is what he calls "reflexivity": a downturn in sentiment can directly cut off the funding chain, thereby affecting actual construction spending, not merely depressing valuations.

He further quantified this pressure point. Supporting current infrastructure valuations requires AI labs' annual recurring revenue (ARR) to reach about $1 trillion by the end of 2028. Based on the current ARR base of about $13.6 billion, that requires about $50 billion in net new ARR every month.

If you are relying on (AI) labs' combined ARR reaching $500 billion to support your assumption of capital flowing back into infrastructure, then if it ends up at $300 billion, that is a disaster.

James listed several factors dragging on ARR.

First, "the consumer side is good enough." His own experience using frontier models is that everyday needs are "99% covered" by existing models, and stronger capabilities can hardly elicit new payments.

Second, "price cuts have not brought proportional usage." He cited an estimate from the Anthropic Economic Index: "A 10% reduction in task cost corresponds to only a 3% increase in usage."

In his understanding, this data point shows that the Jevons paradox (efficiency gains actually push up total consumption) may not necessarily apply to tokens in the short term. After enterprises use AI to set up a process that is good enough, call volume may even decline: perhaps when it is not easy to use, token demand is actually higher, because you have to keep revising it.

Third, the open-source share is rising. He mentioned that the share of open-source models in new ARR has, by his estimate, risen from about 1% a year ago to about 18%; buyers pay per task, and as long as the minimum capability is sufficient: they don't drive fighter jets to work; they drive cars.

He believes that the capability gap between Chinese and American models still exists, so a price shock is not what he worries about most. The criterion he values most is: this is not a question of whether AI is useful or whether AI will get better, but whether frontier spending is actually economical. He believes that in the short term, "white-collar work" must become that killer app and bear the vast majority of token demand in the coming years.

"Cheapest" is exactly what the top of the cycle looks like

The host pointed out that on an earnings basis, these tech companies now look cheaper than in 2023 because earnings have surged—so why cut positions instead? Citrini James applied the rule of cyclical stocks and replied:

This is the cycle; cyclical industries always peak when they look cheapest on next year's expectations.

In other words, James believes cheapness is an illusion created by temporarily inflated earnings, as is always the case with cyclical stocks. Cyclical stocks' earnings rise and fall with the business cycle. When the cycle is strongest, company profits hit a high, the denominator grows, and the price-to-earnings ratio calculated on "next year's expected earnings" becomes very low, making them look cheap. But this profit is the profit at the peak of the cycle; when the cycle turns down, earnings fall and share prices fall with them. So the moment when something "looks cheapest" is often exactly the moment when earnings are highest and the cycle is near its top.

The host added an analogy: no one buys an oil company when its profits are fattest and its P/E ratio is only 14 times.

This logic explains his position adjustment. He said that if he had still maintained a heavily weighted infrastructure portfolio in April and May, the overall portfolio "would probably have fallen 50%." And when semiconductors experienced a two-to-three-month drawdown in September and October 2023, he was not in a hurry to find substitutes, because valuations were low at the time and downside was limited. Now it is different:

The downside, when negative results or negative sentiment truly emerge, is very large.

He frames his move as a portfolio adjustment, not a bearish call on AI as a whole: he remains bullish on Nvidia and believes the more asymmetric opportunities lie in what he calls "agentic utilities" (that is, industries that truly use AI and benefit from AI adoption). He also said that demand still exceeds supply at present, and the problem lies in the future.

Long Meta, short Google, waiting for the signal to return to chips

In terms of specific holdings, Citrini James's logic is equally clear.

The core thesis for going long Meta is the potential monetization space of Muse (Meta's AI assistant product). His calculation framework is: Meta's current average revenue per user (ARPU) is a little over $20, and its data assets cover Instagram, WhatsApp, and Facebook. "If Muse achieves any meaningful adoption rate, you can draw a straight line from that data to a 4x increase in ARPU by 2030." He said he turned strongly bullish this summer, and Muse's launch confirmed that judgment.

The logic for shorting Google comes from his concern about the brain drain at DeepMind. He believes the market underestimates the negative impact of this wave of talent departures, and that this is happening precisely in a window when the search business still faces structural threats. "I myself use Google search much less now. Will search be completely killed? I don't know, maybe not. But the risk remains, and at the same time they are still experiencing a talent drain." He said he holds a stop-loss, but currently maintains the "long Meta, short Google" pair trade.

As for when to return to semiconductors, he gave clear trigger conditions: sustained growth in paid usage, and the second derivative of ARR growth re-accelerating, so that construction can truly be financed through customer cash flow. If current demand is in fact still constrained by supply, and the current period is merely a gap between a leap in model capabilities and the launch of new products, and once the products launch users will more easily pay for new capabilities, then his concerns do not hold, and this pullback in infrastructure stocks is instead a buying opportunity.

His judgment on CoreWeave is relatively mild—he believes the probability of default is about 60%, but does not expect an actual default. His view of the downward path of the AI cycle is that it will "end with a whimper rather than a bang": not a crash triggered by some company defaulting, but lab ARR expectations being revised down from $400 billion to $360 billion, with this incremental difference spreading outward because it has already been fully priced in, dragging down the entire sector.

On DeepSeek, he cited data from DeepInfer (DeepSeek's internal inference framework): it serves 155 trillion tokens per day, with revenue of about $38 per GPU-hour, while the market rental cost of a GPU-hour is about $5 to $8, implying an inference gross margin of about 80% and a GPU payback period of about 81 days. "This means token pricing may have further downside, especially in the non-frontier model space."

Designing scorpion venom antibodies with Claude: what AI compresses and what it cannot

Citrini James also disclosed a rare experiment: his team worked with Anthropic to use Claude to design 211 candidate scorpion venom antibody molecules in 64 hours, 17 of which successfully bound to the target in the lab. This result is both astonishing and led him to a conclusion that is not friendly to AI bulls: AI is "commoditizing" drug design, but the real bottleneck is in the lab, in monkeys, not in computing power.

To figure out exactly what AI can and cannot do in drug discovery, Citrini James's team did something unusual: they contacted Anthropic's life sciences team (Nathan Frey, Amir Chennisa, and Steven Molina) and, inside Anthropic's offices, using its internal research model, spent 64 hours designing scorpion venom antibodies.

Throughout the process, Claude invoked about 300 sub-agents, generated 180 designs, and ultimately submitted 211 candidate molecules for testing. Lab results showed that 17 molecules successfully bound to the target, 7 of which were against the specific Mexican scorpion venom CN2, and the best candidate, CN2B1, also formed a competitive binding with a known neutralizing antibody in the expected region.

"AI generated these molecules, and 17 of them bound in the lab—that is quite astonishing."

But James immediately offered what he considers the more important other side. What AI compresses is the time from idea to experimental protocol—a process that originally took months now takes only days. But synthesis, testing, and obtaining control data still took 39 days. "If data moves at the speed of life, the arrival of AI cannot make cells in a petri dish grow faster. That is where the bottleneck lies."

He estimates that across the entire process from drug design to clinical trials, AI can compress about 20% of the time, while the other 80% is incompressible. This judgment points directly to his investment conclusion: go long on experiments, go long on companies that can make experiments faster, rather than going long on AI drug discovery concept stocks.

He explicitly named Charles River Laboratories (CRL) as his highest-conviction long position. About 60% of CRL's revenue comes from early discovery and safety assessment; it is at the trough of a cyclical recovery and has ample capacity to take on new orders, with each new contract flowing directly into profit.

He also pointed to a bottleneck the market has not yet fully priced in: the monkey shortage. Non-human primates are the closest-to-human, irreplaceable resource in preclinical testing, with a breeding cycle of about four years, making supply expansion extremely difficult. "This may be the most important supply constraint in preclinical testing." Risk Disclosure and Disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.