Citrini: Monkeys Are The Next Bottleneck | Citrini on AI Drug Discovery Experiment & Reflexive Financing Risk to AI Buildout & Hardware Trade

Citrini has largely shifted since March from AI hardware and semiconductor stocks toward AI deployment and “agentic utilities,” while remaining bullish on Nvidia and infrastructure. He sees more asymmetric upside in companies using AI than in owning only the infrastructure providers. His central risk is that the next phase of AI investment is increasingly financed through equity issuance and debt rather than hyperscalers’ operating cash flow. If sentiment weakens, that funding could contract even as AI use grows, creating a financing-driven reversal in infrastructure expectations.

The key uncertainty is whether model capability is converting into paid usage quickly enough. One estimate suggested a 10% reduction in task costs might produce only a 3% increase in usage, and a mature workflow may need fewer inference calls than one still being debugged. The speakers discussed estimates of about $1 trillion in AI annual recurring revenue by the end of 2028, but the exchange did not clarify the units or period attached to a separate “13.6” figure. Open models’ share of net new ARR was said to have risen from about 1% to 18% in a year. Citrini’s main concern was inadequate capability-to-revenue conversion, not Chinese open-weight models eroding US pricing power.

Citrini’s drug-discovery experiment tested how much of early drug design investment analysts could accomplish with AI, and what still required specialist expertise and laboratory work. Using an Anthropic internal research model and existing protein-design tools, the team pursued a scorpion-toxin mini-binder project informed by published Baker lab work. It generated 180 designs over 64 hours using 308 sub-agents, then sent 211 designs to two contract research organizations. After 39 days of lab work, 17 designs bound the target; seven bound CN2. The leading candidate, CN2B1, bound in both labs and competed near the intended region with an established neutralizing antibody.

That result did not establish that CN2B1 neutralizes the toxin: the experiment tested binding, not neutralization. Further design work, binding-affinity tests and a patch-clamp study were proposed. The team also lacked the safeguards required by a 60-question biosecurity review and proposed doing the work at Anthropic under supervision. The broader conclusion was that AI compressed literature review, ideation and dry-lab work that could take months into days, but synthesis, testing and controls remained time-consuming. The speakers expect AI to expand early-stage candidates and preclinical experiments, not to deliver autonomous cures soon or drive exponential compute demand over the next 12–24 months. One estimate was that AI could reduce about 20% of the path from drug design to clinical trials, while biology and the need for experimental data constrain the rest.

Dave identified scarce research monkeys as a major bottleneck on the path to clinical trials, though not one directly caused by AI demand. Breeding takes about four years, China had stopped exporting monkeys and was described as sourcing them internationally, and researchers may need a model close to humans. Scarcity could create pricing power, but the speakers noted there is no monkey price index. Computer-based toxicology may eventually reduce animal testing, yet reliable simulations still need experimental data, including animal data. Dave said Charles River Laboratories (CRL) has about 50% of revenue in preclinical work and about 60% in early-stage discovery and safety assessment, leaving capacity for additional orders and potential higher-margin earnings flow-through if activity grows.

Cheaper candidate generation could weaken the value of preclinical or early-clinical assets, particularly “me-too” companies built around one asset in a well-understood mechanism. The speakers saw more durable advantages in distinctive causal-biology expertise, new druggable targets or other lasting capabilities. Eli Lilly was cited as better positioned because it has advanced acquired molecules to FDA approval and uses AI, including TuneLab, which trains on proprietary data and requires users to contribute designs. A cited AI drug-discovery experiment used 10,000 GPU hours and was estimated to have cost about $50,000, versus $2,000–$4,000 at the time of the discussion; a Baker Lab result showing an 86% reduction in tokens needed to design a cortisol binder was offered as evidence of cost deflation. Lower costs may spur experimentation, but the resulting demand for compute remains uncertain.

The financing risk could unfold through gradual estimate cuts: one scenario reduced combined AI-lab ARR expectations from $400 billion to $360 billion, while Citrini said $300 billion instead of an assumed $500 billion would be disastrous for infrastructure expectations. Semiconductor stocks had nine consecutive days of declines in October, depending on the index, but sentiment had not yet materially affected the buildout. The speakers described the decline as a possible correction to unrealistic early-Q1 expectations; in one scenario, an infrastructure-heavy portfolio held in April or May could be down about 50%. They expected more moderate semiconductor returns and a longer consolidation unless new products or capabilities make it easier for customers to spend. On CoreWeave, a speaker rejected a 60% default probability by 2032, but corrected an equity-cushion figure from $45 billion to $35 billion and warned that it could be used quickly.