Citrini Research has cut back its semiconductor exposure and added selective short positions, marking a sharp reversal for a firm that had been an early and outspoken supporter of the AI hardware trade. Founder James van Geelen said the decision reflects a more fragile funding model for AI infrastructure: new data-center spending increasingly depends on debt and equity markets rather than the operating cash flow of the largest cloud companies.
Speaking on the Monetary Matters podcast on Oct. 21, van Geelen said that change leaves chipmakers more exposed to shifts in market sentiment. Semiconductor stocks had just recorded nine consecutive daily declines, their longest losing streak since 2006, providing a reminder that expectations for AI spending can change quickly even when long-term demand remains strong.
The concern is less about whether AI requires substantial computing capacity than about who funds the next round of capacity. During the first phase of the AI buildout, hyperscalers such as the largest cloud platforms could use internally generated cash to finance purchases of chips, servers and data centers. That arrangement gave suppliers a degree of protection from broader macroeconomic volatility.
Van Geelen argued that the buffer is weakening. As cash flow at major technology companies tightens and AI capital expenditure rises, more of the incremental funding for infrastructure is being connected to capital markets. Under that model, a fall in equity values or a more restrictive debt market could limit actual construction budgets, rather than merely reducing the valuation multiples applied to hardware companies.
Revenue targets behind the AI infrastructure buildout
Van Geelen framed the issue through the revenue growth AI companies would need to justify the scale of infrastructure now being deployed. He estimated that AI labs would need to generate roughly $1 trillion in annual recurring revenue by the end of 2028, compared with about $13.6 billion today.
That trajectory would require approximately $50 billion in net new recurring revenue each month, based on his calculation. Such figures illustrate the gap between the infrastructure already being financed and the paid software and AI-service demand needed to support it.
Citrini’s revised positioning rests partly on the view that demand may not expand fast enough to close that gap. Van Geelen said existing models already meet an estimated 99% of everyday user requirements, reducing the urgency for customers to pay for increasingly capable systems. He also cited Anthropic’s Economic Index, which found that a 10% decline in the cost of a task was associated with only a 3% increase in usage.
The relationship suggests that lower prices alone may not create demand at the pace required by aggressive infrastructure forecasts. AI providers may be able to serve more tasks cheaply, but lower unit economics could also complicate the effort to turn rising usage into the revenue growth implied by current spending plans.
Open-source models present another pressure point. Van Geelen estimated that their share of net new annual recurring revenue rose from about 1% a year earlier to roughly 18%. If businesses choose the least expensive model capable of completing a given task, premium model providers may have less room to raise prices or preserve margins.
Why cheap chip valuations can mislead
Van Geelen also warned against relying too heavily on forward price-to-earnings ratios when assessing semiconductor shares. Chip stocks can appear inexpensive late in a cycle because earnings forecasts are strongest precisely when demand and margins have already expanded, pushing the earnings denominator higher.
If the cycle cools and earnings estimates are cut, an apparently low valuation can quickly prove less attractive. This dynamic has repeatedly shaped semiconductor trading, where production lead times and inventory cycles can leave earnings trailing changes in end-market demand.
He contrasted the recent environment with the pullback in semiconductor shares during September and October 2023. In that earlier period, he said valuations had already fallen enough to reduce downside risk after two to three months of declines. By comparison, he said a portfolio that remained heavily allocated to infrastructure names through April and May would have declined by about 50%.
Citrini has not abandoned AI-linked equities altogether. Van Geelen described a long position in Meta, based on the possibility that its average revenue per user, currently a little above $20 in his model, could rise substantially by 2030 if AI products such as Muse gain meaningful adoption across Instagram, WhatsApp and Facebook.
He paired that trade with a short position in Google. His rationale included talent departures from DeepMind and the uncertain effects of changing search behavior as AI-generated answers become more common. The pairing reflects a preference for companies with consumer distribution and existing monetization channels over businesses whose value depends more directly on continued infrastructure expansion.
Conditions for a renewed hardware bet
Van Geelen said he would consider increasing semiconductor exposure if paid AI usage accelerates enough for infrastructure spending to be funded increasingly through customer cash flow. Stronger recurring revenue growth would give AI labs and cloud providers a more durable basis for financing servers and data centers without relying as heavily on receptive equity and debt markets.
He also left room for a more bullish outcome if present AI demand is constrained by limited computing supply. In that case, improvements in model capabilities could translate into new paid products once developers launch applications that users find worth purchasing. The decisive evidence, in his view, would be sustained growth in monetized usage rather than rising model benchmarks or infrastructure announcements alone.
His discussion of CoreWeave captured the uncertainty around heavily financed AI infrastructure providers. Van Geelen estimated roughly a 60% probability of default for the company, while adding that he did not expect default to occur. He described a potential downside scenario in which AI-lab revenue expectations are revised downward — for example, from $400 billion to $360 billion — creating repricing pressure across suppliers, data-center operators and financing structures tied to those forecasts.
Van Geelen also pointed to potentially attractive inference economics, citing figures he attributed to DeepInfer, an internal DeepSeek inference framework. He said the system processes 155 trillion tokens a day and could generate about $38 in revenue per GPU hour against estimated rental costs of $5 to $8, implying an inference gross margin near 80% and an 81-day hardware payback period. Those economics, if sustained, would also leave room for further price cuts, particularly for applications that do not require frontier-model performance.
AI’s gains meet physical bottlenecks
AI may compress the design stage of scientific work without eliminating slower laboratory processes. Van Geelen cited an Anthropic life-sciences test in which Claude was used over 64 hours to design 211 scorpion-toxin antibody candidates. Seventeen candidates bound to the intended target in laboratory testing, including seven directed at the Mexico-specific CN2 toxin, he said.
The synthesis, testing and control work took 39 days after the design phase. Van Geelen estimated that AI could reduce total drug-development timelines by about 20%, while laboratory processes account for the remaining 80% of the schedule. He identified the supply of non-human primates for preclinical research as one physical constraint, citing breeding cycles of roughly four years.
That view led him toward clinical outsourcing companies rather than treating AI as a complete substitute for laboratory capacity. He named Charles River Laboratories, saying about 60% of its revenue comes from early discovery and safety assessment and that the company has capacity to accept additional orders.
Citrini’s repositioning places the AI hardware trade on a tougher test: paid demand must increasingly validate the financial structure built around it. The next phase of the cycle may depend less on the scale of announced compute projects and more on whether AI services can generate enough recurring cash flow to pay for the infrastructure already planned.
To navigate AI-driven volatility beyond semiconductors, explore today’s AI market moves and refine your macro-informed trading approach.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
