Citrini Research, an influential chip research firm, was once among the earliest and loudest bulls on the AI hardware trade. Now, Citrini has significantly reduced its semiconductor exposure and even holds some short positions. This shift stems from its 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 founder James van Geelen recorded this conversation on the Monetary Matters podcast. At the time, semiconductor stocks had just experienced nine consecutive days of declines — the longest losing streak since 2006, as host Jack pointed out. James' assessment was that this alone was not enough to shake AI buildout spending, but the nature of the risk had changed. He said: "If market sentiment takes a sharp turn, that incremental capital will dry up — because it's no longer a company spending its own money, but something determined by the market."
The conversation covered why he rotated out of semiconductors, the logic behind his long Meta Platforms, Inc. (META) and short Alphabet (GOOGL) positioning, the implications of DeepSeek's roughly 80% inference gross margin, and what signals would make him go long chips again. He also revealed that the bottleneck in new drug development is "a shortage of monkeys" rather than compute power, and that the real investment opportunity lies in clinical outsourcing rather than AI concept stocks.
Tech Giants' Cash Flow Turns Negative, Financing Structure Has Changed, and Sentiment Can Kill Fundamentals
To understand why Citrini is reducing its AI hardware exposure, one must first understand what he calls "reflexivity" risk — the core mechanism behind the entire thesis.
In the first three to four years of the AI trade, hyperscale cloud providers financed AI infrastructure with their own operating cash flow. This meant that regardless of how market sentiment fluctuated — whether from tariff shocks, yen carry trade unwinds, or any other panic — as long as companies like Meta Platforms, Inc. (META) and Microsoft kept spending, semiconductor companies' fundamentals had a floor. Sentiment could push stock prices down, but it couldn't crush revenue.
Now that mechanism has broken down, and tech giants' cash flow is rapidly depleting or even turning negative. Citrini's James assesses that virtually every incremental dollar of AI infrastructure investment currently comes from equity financing (whether venture capital or public markets) or debt financing.
This means AI can become more useful and usage can continue to expand, but if market sentiment takes a sharp turn, that incremental capital will dry up — because it's no longer a company spending its own money, but something determined by the market.
This is what he calls "reflexivity": a sentiment downturn can directly sever the funding chain, thereby affecting actual buildout spending, not just compressing valuations.
He further quantified this pressure point. Supporting current infrastructure valuations requires AI labs' annual recurring revenue (ARR) to reach approximately $1 trillion by the end of 2028. Based on the current ARR base of roughly $13.6 billion, this requires approximately $50 billion in net new ARR per month.
"If you're relying on (AI) labs' combined ARR reaching $500 billion to support your assumptions about reinvested infrastructure capital, then if it ends up being $300 billion, that's a disaster."
James listed several factors weighing on ARR.
First, "the consumer side is satisfied." His own experience using frontier models is that daily needs are "99% covered" by existing models, and stronger capabilities struggle to generate new paid adoption.
Second, "price cuts haven't generated proportional usage increases." He cited estimates from the Anthropic Economic Index: "A 10% reduction in task cost corresponds to only a 3% increase in usage."
In his interpretation, this data point suggests that the Jevons paradox (where efficiency gains actually increase total consumption) may not apply to tokens in the short term. After enterprises build an adequate AI workflow, call volumes may actually decline: perhaps you need more tokens when it doesn't work well, because you have to keep revising.
Third, open-source share is rising. He mentioned that open-source models' share of new ARR has risen from about 1% a year ago to approximately 18% by his estimate, with buyers paying per task as long as the minimum capability is sufficient: "They're not driving fighter jets to work — they're driving cars."
He believes the capability gap between Chinese and American models still exists, so price disruption isn't his biggest concern. His most important criterion is: this isn't about whether AI is useful or whether AI will get better — it's about whether frontier spending is actually economical. He believes that in the short term, "white-collar work" must become that killer application, absorbing the vast majority of token demand for the next few years.
"Cheapest" Is Exactly What a Cyclical Top Looks Like
The host pointed out that based on earnings, these tech companies now appear cheaper than in 2023 because earnings have surged dramatically — so why reduce positions? Citrini's James applied the rules of cyclical stocks to answer:
This is how cycles work — cyclical industries always peak when they look cheapest on next year's expected earnings.
In other words, James believes cheapness is an illusion created by temporarily inflated earnings, as is always the case with cyclical stocks. Cyclical stock earnings rise and fall with the business cycle. At the peak of the cycle, company profits surge to highs, the denominator expands, and the price-to-earnings ratio calculated on "next year's expected earnings" becomes very low, making the stock look cheap. But that profit is peak-cycle profit — when the cycle turns, earnings decline and stock prices fall with them. So the moment something "looks cheapest" is often precisely when earnings are at their highest and the cycle is nearing its top.
The host added an analogy: nobody buys an oil company when its profits are fattest and its P/E ratio is only 14x.
This logic explains his portfolio adjustment. He said that if he had maintained a heavily infrastructure-weighted portfolio in April and May, the overall position would have "probably dropped about 50%." When semiconductors experienced a two-to-three-month pullback in September and October 2023, he wasn't in a hurry to find replacements because valuations were low and downside was limited. Now it's different:
The downside, when negative outcomes or negative sentiment actually materialize, is substantial.
He framed his actions as portfolio adjustment, not a bearish call on AI as a whole: he remains bullish on NVIDIA (NVDA) and sees more asymmetric opportunities in what he calls "agentic utilities" — industries that actually use AI and benefit from AI deployment. He also noted that demand still exceeds supply currently; the question is about the future.
Long Meta, Short Google, and Waiting for Signals to Return to Chips
On specific positions, Citrini's James is equally clear.
The core thesis for going long Meta Platforms, Inc. (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 in the low $20s, with data assets spanning Instagram, WhatsApp, and Facebook. "If Muse achieves any meaningful adoption rate, you can draw a straight line from that data to a 4x ARPU increase by 2030." He said he turned strongly bullish this summer, and Muse's launch confirmed that judgment.
The logic for shorting Alphabet (GOOGL) comes from his concerns about DeepMind talent attrition. He believes the market underestimates the negative impact of this talent exodus, which is occurring precisely during a window when the search business still faces structural threats. "I myself use Google search far less now. Will search be completely killed? I don't know, maybe not. But the risk remains, and they're simultaneously experiencing talent attrition." He said he maintains a stop-loss but currently holds the "long Meta, short Google" pair trade.
On when to return to semiconductors, he gave clear trigger conditions: seeing sustained growth in paid usage, and the second derivative of ARR growth re-accelerating, truly supporting financing buildout through customer cash flow. If current demand is actually still supply-constrained and this is merely a gap between a model capability leap and new product launches — where users find it easier to pay for new capabilities once products ship — then his concerns would be invalidated, and this pullback in infrastructure stocks would instead be a buying opportunity.
His assessment of CoreWeave is relatively measured — he believes the probability of default is about 60%, but he doesn't expect an actual default. His view on the path of an AI cycle downturn is that it will "end with a whimper, not a bang": not a company default triggering a crash, but lab ARR expectations being revised down from $400 billion to $360 billion, with that incremental difference spreading outward because it's already fully priced in, dragging down the entire sector.
On DeepSeek, he cited data from DeepInfer (DeepSeek's internal inference framework): serving 155 trillion tokens per day, generating approximately $38 per GPU-hour in revenue, while market rental costs for GPU-hours are about $5 to $8, implying an inference gross margin of roughly 80% and a GPU investment payback period of about 81 days. "This means token pricing may have further downside, especially in non-frontier model segments."
Designing Scorpion Venom Antibodies with Claude: What AI Compresses and What It Cannot
Citrini's James also disclosed a rare experiment: his team collaborated with Anthropic to use Claude to design 211 scorpion venom antibody candidate molecules within 64 hours, 17 of which successfully bound to targets in the lab. The result was both astonishing and led him to a conclusion that isn't friendly to AI bulls: AI is "commoditizing" drug design, but the real bottleneck is in the lab, on the monkeys, not in compute power.
To figure out what AI can and cannot do in drug discovery, Citrini's James team did something unusual: they contacted Anthropic's life sciences team (Nathan Frey, Amir Chennisa, and Steven Molina) and, inside Anthropic's office, used its internal research model to spend 64 hours designing scorpion venom antibodies.
Throughout the process, Claude deployed approximately 300 sub-agents, generated 180 design proposals, and ultimately sent 211 candidate molecules for testing. Lab results showed that 17 molecules successfully bound to targets, 7 of which were against the Mexican-specific scorpion venom CN2, and the best candidate CN2B1 formed competitive binding with a known neutralizing antibody in the expected region.
"AI generated these molecules, and 17 of them bound in the lab — that's quite astonishing."
But James immediately presented what he considers the more important other side. What AI compresses is the time from idea to experimental protocol — a process that once 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, AI's arrival doesn't make cells in a petri dish grow faster. That's where the bottleneck is."
He estimates that across the entire process from drug design to clinical trials, AI can compress approximately 20% of the time, with the other 80% being incompressible. This judgment leads directly to his investment conclusion: go long on experiments, go long on companies that 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. CRL derives about 60% of revenue from early discovery and safety assessment, is at a trough of cyclical recovery, and has ample capacity to take on new orders, with each new contract flowing directly to the bottom line.
He also pointed to a bottleneck the market hasn't fully priced in: a shortage of monkeys. Non-human primates are the most irreplaceable resource closest to humans in preclinical testing, with a reproduction cycle of about four years, making supply expansion extremely difficult. "This may be the most important supply constraint in preclinical testing."
Comments