Cyclical stocks always peak when they're “cheapest” — Citrini, the “chip influencer” who was the earliest to be bullish on AI hardware, is now quietly exiting

The influencer chip research institution Citrini Research used to be one of the earliest and most vocal bulls in AI hardware trading. Today, Citrini has significantly reduced its semiconductor holdings and even holds some short positions. Behind this shift is his judgment on a fundamental change in the financing structure of AI infrastructure: when incremental funding shifts from hyperscale cloud vendors' own cash flow to the equity and debt markets, market sentiment itself has become 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 nine consecutive declines—the host Jack pointed out that this was the longest losing streak since 2006. James's judgment: this was still not enough to shake AI spending, but the nature of the risk had changed. He said: "If market sentiment takes a sudden turn, those incremental funds will dry up—because it is no longer a company spending its own money, but rather decided by the market."

This conversation covered why he rotated out of semiconductors, his logic for going long Meta and short Google, the implications of DeepSeek's roughly 80% inference gross margin, and what signal would make him go long chips again. At the same time, he revealed that the bottleneck in new drug development is the "scarcity of monkeys" rather than computing power, and the real investment opportunity lies in clinical outsourcing, not in AI concept stocks.

Cash Flows Turn Negative for Giants, Financing Structure Changes, Sentiment Can Kill Fundamentals

To understand why Citrini reduced their AI hardware positions, one must first understand the "reflexivity" risk he mentions—this is the core mechanism of his judgment.

In the first three to four years of the AI trade, hyperscale cloud providers paid for AI infrastructure with their own operating cash flows. This meant that regardless of market sentiment fluctuations—be it tariff shocks, yen carry trades being unwound, or any other panic—as long as Meta, Microsoft, and similar companies kept spending, there was a "floor" supporting the fundamentals of semiconductor companies. Sentiment could drag down stock prices, but not revenues.

Now this mechanism has failed, and cash flows of tech giants are rapidly depleting and even turning negative. Citrini James judges that almost every incremental investment in AI infrastructure today comes from equity financing (be it venture capital or public market) or debt financing.

This means that AI can become more useful and usage can continue to expand, but if market sentiment sours rapidly, those incremental funds will dry up—because it's no longer a single company spending its own money but is determined by the market.

This is what he means by "reflexivity": negative sentiment can directly cut off the funding chain, thus affecting actual construction spending, not just compressing valuations.

He further quantified the stress point. To support current infrastructure valuations, AI labs' annual recurring revenue (ARR) must reach about $1 trillion by the end of 2028. With the current ARR base at about $13.6 billion, roughly $5 billion in net new ARR must be added every month.

If you're counting on AI labs reaching a combined ARR of $500 billion to support your infrastructure capital recycle assumptions, but in the end it's only $300 billion, then that's a disaster.

James listed several factors dragging on ARR.

First is "the consumer side has enough." His own experience using cutting-edge models is that 99% of daily needs are covered by existing models; stronger capabilities are unlikely to drive much new paid demand.

Second, "lower prices haven't generated proportional usage increases." He cited the Anthropic Economic Index estimates: "A 10% drop in task cost only leads to a 3% increase in usage."

In his view, this suggests the Jevons Paradox (efficiency boosts driving higher total consumption) may not immediately apply to tokens. Once companies build functional AI processes, their call volumes might even decline: perhaps when the models perform poorly, you actually use more tokens because you constantly have to make corrections.

Third, open-source share is rising. He noted that open-source models' share of incremental ARR, by his estimation, has risen from about 1% a year ago to around 18% now, with buyers paying per task so long as minimal capability is met: they're not going to commute in a fighter jet, just a regular car.

He believes the gap in capabilities between US and Chinese models still exists, so price shocks are not his main concern. His key criterion: it's not about whether AI is useful or will become better, but about whether frontier expenses are truly economical in the short term. He thinks in the immediate future, "white-collar work" must become the killer app, shouldering the vast majority of token demand in the next few years.

“Cheapest” Is Exactly What the Cycle Top Looks Like

The host pointed out that in terms of profits, these tech companies look even cheaper than in 2023 due to soaring earnings—so why underweight them now? Citrini James borrowed the logic of cyclical stocks to reply:

This is the cycle: cyclical industries always top out when they look cheapest on next year's forecasted earnings.

In other words, James thinks "cheapness" is an illusion caused by temporarily inflated profits—this is always the case with cyclicals. Profits in cyclical stocks rise and fall with the business cycle. When times are best, profits soar, the denominator grows, and the forward P/E looks low—it appears cheap. But those are peak-cycle profits. When the cycle rolls over, profits fall, and the stock price drops accordingly. So the time when things "look cheapest" is often the profit peak, just before the cycle top.

The host added an analogy: nobody would buy oil company stocks at their most profitable, with a P/E of only 14.

This logic explained his portfolio rotation. He said if he had kept the heavy infrastructure positions through April and May, the portfolio as a whole "would likely have fallen 50%." And when semiconductors experienced two or three months of decline in September and October 2023, he was not in a rush to find substitutes, since valuations were low and downside was limited. Now it's different:

The downside, when real negative outcomes or sentiment show up, is significant.

He characterizes his action as a portfolio adjustment, not being bearish on all of AI: he still likes Nvidia and believes that the more asymmetric opportunities are in what he calls "agentic utilities"—the industries that truly use and benefit from AI. He also remarks that, for now, demand still exceeds supply; the issue is the future.

Long Meta, Short Google, Waiting for the Signal to Return to Chips

Regarding specific positions, Citrini James’s logic is just as clear.

The core case for going long Meta is the potential monetization of Muse (Meta's AI assistant product). His framework: Meta’s current average revenue per user (ARPU) is just over $20, with data assets covering Instagram, WhatsApp, and Facebook. "If Muse gets any meaningful adoption, you can directly extrapolate from these data to a 4x increase in ARPU by 2030." He said he turned strongly bullish last summer, and Muse’s launch validated this view.

The logic for shorting Google comes from concern over DeepMind's talent exodus. He thinks the market has underestimated the negative impact of this wave of departures, which is happening just as the search business faces structural threats. "Personally, I use Google search a lot less now. Will search be completely disrupted? I don't know, maybe not. But the risk remains, and they're also losing top talent." He said he's keeping tight stops, but currently maintains this "long Meta, short Google" paired trade.

As for when to return to semiconductors, he offered clear triggers: he wants to see sustained paid usage growth, and the second derivative of ARR growth reaccelerating—meaning customers’ cash flows are truly funding infrastructure. If current demand is actually still supply-constrained, and if now is just a lull between model capability increases and the next product launches (after which users happily pay for new capabilities), then his concern doesn't hold—and this correction in infra stocks would actually be a buying opportunity.

He's relatively mild in his assessment of CoreWeave—thinks there’s about a 60% default probability but isn’t expecting an actual default. He views the downside path of the AI cycle as "ending with a whimper, not a bang:" not with one company’s default triggering a collapse, but with AI lab ARR expectations cut from $400 billion to $360 billion. This incremental change, having already been priced in, would spread and drag down the whole sector.

Regarding DeepSeek, he cited data from DeepInfer (DeepSeek’s internal inference framework): it serves 1.55 quadrillion tokens daily, with about $38 in revenue per GPU hour, while GPU rental costs in the market are around $5–8 per hour, implying an inference gross margin of about 80% and a GPU payback period of roughly 81 days. "This means token pricing could fall further, especially outside the frontier model segment."

Designing Scorpion Toxin Antibodies with Claude: What AI Compresses and What It Can't

Citrini James also disclosed a rare experiment: his team collaborated with Anthropic, using Claude to design 211 candidate scorpion toxin antibodies in 64 hours, with 17 successfully binding targets in the lab. This result was both shocking and led him to an arguably bearish conclusion for AI bulls: AI is commoditizing drug design, but the real bottleneck lies in the laboratory—with monkeys—not in computing power.

To clarify 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), used Anthropic's internal research models at their office, and spent 64 hours designing scorpion toxin antibodies.

Throughout the process, Claude deployed around 300 sub-agents and generated 180 design proposals, eventually testing 211 candidate molecules. Lab results showed that 17 molecules successfully bound their targets, including 7 targeting Mexican-specific scorpion toxin CN2; the best candidate, CN2B1, also formed a competitive bind in the expected region with a known neutralizing antibody.

"AI generated these molecules, a full 17 of which bound in the lab—that’s pretty shocking."

But James quickly added what he considered the more important reality. What AI compresses is the time from idea to experimental protocol—a process that once took months now takes days. But synthesis, testing, and control data generation still took 39 days. "If data moves at the speed of biology, AI's arrival can’t make the cells in a petri dish grow faster. That’s where the bottleneck is."

He estimates that in the entire process from drug design to clinical trials, AI can cut about 20% of the time; the other 80% is not compressible. This point leads to his investment takeaway: go long experiments, go long companies that can accelerate experiments, not AI drug discovery concept stocks.

He explicitly named Charles River Laboratories (CRL) as his highest-conviction long. About 60% of CRL’s revenue comes from early discovery and safety evaluation; it is at a cyclical bottom, has plenty of capacity to take on new orders, and each new contract can go straight to profit.

He also pointed to a bottleneck not yet fully priced in by the market: a shortage of monkeys. Non-human primates are an irreplaceable resource in preclinical testing, with breeding cycles of about four years, making supply expansion extremely difficult. "This may be the most important supply constraint in preclinical testing."