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

NVIDIA

Across the shows: 37 speakers in other episodes, 32 bullish and 12 bearish calls, past 30 days

• CitriniBullish

  Still pretty bullish on Nvidia from here, though more asymmetric opportunities sit elsewhere.

  Citrini's firm has shifted away from being broadly long semiconductors since the momentum drawdown began in late May or early June. He says that does not mean he is done with semis. He remains pretty bullish on Nvidia from here, and sees opportunities in AI hardware, but the more asymmetric ones are in what he calls agentic utilities. He notes that the infrastructure trade requires underwriting significant expectations. The support for the Nvidia view is thin in this passage, with no specific numbers given.

  “I'm still pretty bullish on NVIDIA from here.”at 3:00

Eli Lilly and Company

Across the shows: 3 speakers in other episodes, 3 bullish and 0 bearish calls, past 30 days

• CitriniBullish

  Eli Lilly is a likely AI-drug-discovery beneficiary: strong acquisition record and a proprietary AI design model.

  As AI commoditizes phase one assets, large acquirers that bring molecules to FDA approval should benefit. Eli Lilly has a strong track record of acquiring molecules and taking them through approval. It is also an AI pioneer with TuneLab, an AI drug-design model trained on decades of its proprietary data. Anyone can use it, but users must contribute designs to improve it, so Lilly collects data that keeps the model improving. Citrini expects significant dispersion across pharma as AI improves, with less defensible companies that rely on a single asset in one mechanism losing out.

  “But companies like Eli Lilly has a great track record of executing on acquiring these kind of companies, acquiring molecules, and bringing them through to FDA approval.”at 55:33

Charles River Laboratories International

• CitriniBullish

  Charles River is Citrini's highest-conviction long: AI-driven experiment demand and a scarce monkey supply should lift it.

  AI lowers the barrier to creating drug candidates and compresses literature review and dry-lab work, so the market needs far more experimental data. Charles River is a preclinical CRO. About 50% of its revenue is preclinical, and about 60% comes from discovery and safety assessment early-stage experiments. It has spare capacity for new orders, unlike smaller CROs, so each new contract is higher margin and flows straight to earnings, and Citrini thinks earnings growth from this alone could inflect. Monkeys are irreplaceable for preclinical testing, and they take about four years to breed. China cut monkey exports and is now seeking international demand, so the monkey supply is the most supply-constrained part of the path to clinical trials. Citrini does not directly link AI to monkey demand, but sees biotech funding improving. The sell-side-style lab automation plays are a side point.

  “Yeah. We've laid out the whole thesis on CRO, but that... That's our highest conviction long from this.”at 1:03:39

Meta Platforms

Across the shows: 43 speakers in other episodes, 43 bullish and 19 bearish calls, past 30 days

• CitriniBullish

  Meta is a bullish call: Muse and its user data should drive ARPU up roughly fourfold by 2030.

  Meta holds vast data on its users from Instagram, WhatsApp and Facebook, with ARPU around $200 and some. If Muse, its AI model, sees meaningful adoption, a straight line runs from that data to ARPU rising about fourfold by 2030, with plenty of room to run on its existing AI infrastructure. Citrini turned bullish this summer on this thesis, before Muse was released, and says the release confirms it. Citrini holds a long Meta position.

  “What's the thesis? I'm bullish on Meta because I think that they're really killing it with Muse.”at 1:07:30

Alphabet

Across the shows: 12 speakers in other episodes, 10 bullish and 1 bearish calls, past 30 days

• CitriniBearish

  Alphabet is underpriced risk: DeepMind talent exodus and lingering search disruption threaten it. Short it against long Meta.

  Citrini thinks the market underestimates the negative impact of the DeepMind talent exodus to competitors, which is happening as the AI race intensifies. Google sold off in 2025 on fears that AI would hurt search, and Citrini took a contrarian long then, which worked, but search risk persists. Many people, including Citrini, now search Google far less. Citrini runs a long Meta, short Google trade and says it is the trade on now.

  “my take is that I think people are underestimating the negative impact from this deep mind exodus that's occurred.”at 1:08:42

CoreWeave

Across the shows: 45 speakers in other episodes, 11 bullish and 41 bearish calls, past 30 days

• CitriniBullish

  CoreWeave's downside is overblown; Citrini has turned mildly bullish on its bonds and does not expect default.

  Citrini thinks the bearish CoreWeave narrative is overblown. The market-implied default odds are high at 60% by 2032, but Citrini does not expect CoreWeave to default. Citrini has turned a little bullish on CoreWeave bonds. CoreWeave has around $45 billion of equity that could be tapped on the way down, which provides a cushion. If the negativity proves justified, the AI cycle would turn with a whimper, through lower lab revenue estimates rippling out, not a CoreWeave default.

  “I think the CoreWeave stuff, I think, was, like, kind of way overblown. Actually, we kind of got a little bullish on the CoreWeave bonds specific.”at 1:12:06

DeepSeek

• CitriniBullish

  DeepSeek's inference margins look near 80%, suggesting token pricing has more room to fall.

  DeepSeek's in-house inference framework claims about 155 billion tokens a day from one cluster, which Citrini converts to roughly $38 of revenue per GPU hour. Renting compute costs about $5 to $8 per GPU hour, so margins are around 80%, with an 81-day payback on GPUs. That is unusual for Chinese industry, which Citrini usually sees as a race to the bottom. The high margin suggests pricing could keep falling, especially for non-frontier models, given competitive cloud API pricing.

  “But DeepSeek, I mean, that would be like 80% ish margins, like an eighty one day payback on their GPUs. So that's pretty crazy.”at 1:15:09

Pharma & biotech

• Citrini

  AI compressed months of drug design work into days, but wet-lab synthesis and testing still take months.

  Citrini's team used Claude to design antivenom mini-binders for scorpion toxin. In about 64 hours, with roughly 308 sub-agents, it generated 180 designs, and the team sent 211 designs to the labs. Seventeen bound the target, and the best candidate, CN2B1, bound in both labs and competed with a known neutralizing antibody at the intended site. The neutralization test was not run. The AI compressed what would have been months of scientist work into days. But synthesis, testing, and getting controls still took months, and the wet-lab step took 39 days. Growing cells in a dish runs at the speed of life, so physical experiments remain the bottleneck.

  “the AI kind of compressed what would be months of work for scientists into days. ... But the synthesis, the testing, getting the controls, that still took months.”at 38:51

• Citrini

  Investment thesis: be long experiments, especially tools and CROs that speed up wet-lab work.

  The team's central investment conclusion from the experiment is to be long experiments, meaning anything that makes experiments faster. The idea is that AI shrinks the idea-to-experiment cycle dramatically, while physical experiments remain the slow step. Citrini frames this as a cyclical recovery trade: life sciences tools and CROs, which were bought at cyclical trough pricing after the COVID-era overbuild, were starting to tick up despite rising rates. Drug development is long-duration and tends to be correlated with rates, which makes the signal notable. He hints at a secular component too, comparing it to the early AI infrastructure trade in 2022-2023.

  “our central conclusion as far as investment thesis go, which is we really want to be long experiments, right? And anything that can make experiments...”at 39:52

• Citrini

  AI is commoditizing drug design, so the value of early-stage and preclinical biotech assets will fall.

  Citrini says drug design, at least in causal biology, is largely solved by AI, including AlphaFold and specialist tools, and will commoditize further. Drug candidates in preclinical testing have risen sharply, while the number of FDA-approved molecules has been flat for five years. Citrini estimates AI can reduce only about 20% of the drug-design-to-clinical-trial process, with 80% irreducible, so AI will not speed the full pipeline much. The value of a preclinical or early clinical asset will probably fall over time. Differentiated companies with real causal biology insight or a new druggable target should keep value. Me-too companies with a single asset in a well-understood mechanism will struggle.

  “Drug design is pretty much, at least in causal biology, it's pretty solved.”at 45:09

• Citrini

  The US has fallen behind China and Australia on phase one trial speed; US pilots aim to close the gap.

  Citrini cites the time to complete a phase one study, where the US has fallen significantly behind China and Australia. In response, the FDA and HHS have launched pilot programs to cut early-stage development time so US phase one trials are more competitive globally. Citrini notes an expedited IND pilot program, called Operation Trailblazer, whose application closes at the end of the month, so it is still early. The point is that these efforts are an action to speed trials, not a proven outcome.

  “The US has fallen significantly behind China and Australia, interestingly enough, in like the time it takes to complete a phase one study.”at 1:06:27

Frontier AI labs & model race

• Citrini

  Lab revenue lags capability gains; combined lab ARR at $300B instead of $500B would be a disaster.

  The pattern has been that model capability improves and lab ARR follows, but recently there is a big jump in capability without a commensurate increase in ARR, or a moderating ARR increase. The AI infrastructure build is being underwritten on assumptions that combined lab ARR reaches about $500B, which is what funds reinvestment into infrastructure. Consensus estimates reportedly imply roughly $1T in AI ARR by 2028, and Citrini says current run-rates need to grow sharply to get there. If combined lab ARR ends up around $300B, the shortfall would be a disaster for the financing assumptions behind the buildout.

  “if you're relying on the combined lab ARR to be $500,000,000,000 and that's where you're underwriting your kind of assumptions of what gets funneled back into infrastructure, If it ends up being 300, then that's a disaster.”at 7:13

• Citrini

  AI labs and AI startups will become major buyers of experimental data to improve their models.

  Before AI, demand for experimental work came mainly from pharma and biotech companies. Now AI labs and AI startups in biology become customers because they need experimental data to improve their models. Anthropic is a leading example, with biology a major focus. Citrini sees AI-generated experiments feeding back into models as a source of new demand for lab work, and the model-improvement loop cannot be sped up past the time experiments take.

  “the AI labs themselves and AI startups in this area become customers because they need to generate the data, the experimental data, to improve their models, right?”at 51:26

• Citrini

  AI labs have weak moats because switching costs are near zero; defensibility will come mainly from capability.

  Citrini does not accept the view that AI labs have no moat, but he argues the switching costs are close to zero. Users get used to how a model behaves, yet a competitor such as ChatGPT can release a model that is much better overall or for a specific task, and users can move. He sees specialisation as a possible source of moats, for example a biology-focused model that clearly beats a rival's. He expects any moats that do develop to be offshoots of capability, so the competition still comes down to which model is best.

  “But the switching costs are zero, right? If [ChatGPT] comes out with a model that's way better or is way better for a specific thing.”at 1:24:05

AI infrastructure & data centers

• Citrini

  AI buildout now relies on equity and debt, so falling sentiment could dry up funding for infrastructure.

  Hyperscaler operating cash flow used to fund AI capex, giving drawdowns a floor because fundamentals kept improving regardless of tariffs or other scares. Now more infrastructure is financed against extraordinary expectations, and nearly every incremental AI dollar comes from selling equity (VC or public markets) or from debt. That makes the buildout reflexive: AI usage can keep expanding, but if market sentiment turns, incremental funding dries up because spending is market-determined rather than company-determined. Citrini says the sentiment-driven downside is now significant; a portfolio of the infrastructure-heavy names held in April or May would be down about 50%. He would become much more constructive on infrastructure only if paid usage growth re-accelerated enough for customer cash flows to support the financing. He says sentiment has not yet turned enough to hit the buildout.

  “probably every incremental dollar that's being spent in AI is from selling equity, whether that's venture capital or in the public markets, and from financing, from debt.”at 5:55

• Citrini

  Hyperscaler and AI-infrastructure weakness may be a temporary lull, but returns could moderate and consolidation may last longer.

  Hyperscalers have delinked from the AI trade, and Citrini thinks the market partly believes AI-trade weakness could lift hyperscaler free cash flow. Citrini thinks hyperscalers would not come out unscathed and rejects that view. Demand is currently constrained by supply. If the pullback is a temporary lull between capability jumps and a new release that makes spending easier, the bullish case holds and this is a dip to buy. Returns could still moderate and consolidation could run longer, which would invalidate the bullish case. Bearish signals to watch are CoreWeave-style stress.

  “And I mean, it's possible that this is a good one too, but it's also possible that the returns kind of moderate from here and they're just not as insane.”at 1:11:34

AI pricing & unit economics

• Citrini

  Cheaper tokens won't drive proportional usage growth near term, so Jevons-style demand is doubtful.

  Citrini raises questions about Jevons paradox, which says cheaper resources raise total consumption, not really applying to tokens in the short term. He cites an Anthropic Economic Index finding, which he extrapolates, that a 10% reduction in the cost of a task correlates with only about a 3% increase in usage. Demand for compute depends on how quickly new demand arrives and how much computation each task requires. As enterprises build efficient AI systems that work well, they may need fewer inference calls, since earlier token demand was partly driven by fixing things that did not work. He calls these short-term concerns but says they need to be taken seriously given the buildout's financing.

  “questions about Jevan's [Jevons] paradox, at least in the short term, not really applying to tokens.”at 11:59

• Citrini

  Token and AI-experiment costs are falling fast, roughly 20x to 25x cheaper in months, showing deflation.

  Citrini's own drug-discovery experiment, with about 10,000 GPU hours, cost about $50,000 at the time. Redoing it now would cost roughly $2,000 to $4,000, a steep fall in months. A Vega lab release showed an 86% reduction in tokens needed to design a cortisol binder. Citrini calls this commoditization or deflation, which is already happening in AI pricing.

  “and that cost about $50,000 with the drug discovery thing. By the way, crazy statistic. From the time we did that to the time we were able to publish because we had the data, the cost, like if we were to have done that now, it would have cost somewhere between 2,000 to $4,000”at 1:14:41

Consumer AI products

• Citrini

  Current consumer assistants already cover nearly all ordinary needs, capping usage growth.

  Citrini uses the current consumer models, which he says are very good, as an example. As an ordinary user he finds that about 99% of what he needs is covered by these tools. Demand for more tokens from consumers would only grow if the models kept commoditizing and became cheaper or more useful. This means consumer usage alone is unlikely to drive the next leg of ARR growth, so the demand has to come from an economy that needs to support it.

  “Because we use Astra Fable as a regular person, 99% of what you'll ever need to do is pretty much covered by those.”at 9:42

Open-source AI models

• Citrini

  Not very worried about Chinese open-weight pricing pressure; the US capability gap has held.

  Competition is part of the picture, but the capability gap between US frontier models and open-source Chinese models has remained defended, so Citrini is not as concerned about open-weight models crushing pricing. He would change his mind if lab ARR growth re-accelerated sharply. He also notes that open models' share of net new ARR has risen from about 1% a year ago to about 18%, and that buyers pay per task and need only minimum capability, so cheaper models can win work where top capability isn't required. That share gain is a tension with the view that the gap is holding, and it is the part of the picture he flags as the competitive threat.

  “I do think that the gap in capability between The US models and open source Chinese, it's remained defended. So I'm not as concerned about that.”at 16:45

AI bubble & tech valuations

• Citrini

  AI infrastructure looks cheap on earnings, but cyclical earnings peak when multiples look lowest.

  Jack argues AI infrastructure stocks are cheaper now because earnings rose sharply, even if they were not cheap two years ago. Citrini agrees that cyclical businesses tend to peak when they look cheapest on next year's estimates. He compares this to not buying an oil company at 14 times earnings because it has been making a lot of money. He says it remains to be seen whether these infrastructure companies are over-earning, which would make the apparently low valuations a trap.

  “That's the cycle though. Like the cyclical businesses always top when they look the cheapest on... next year's estimates?”at 18:33

Enterprise AI adoption & AI software

• Citrini

  White-collar work will be AI's killer app and the bulk of token demand for several years.

  The next pool of paid demand is harder for investors to see than the agentic wave was in 2025. Biggest near-term threat to the trade is that the economy and enterprises are slow to adopt what models can already do. Enterprise workflows need to be reorganized, and complex use cases have bottlenecks laypeople don't understand. Advanced industries such as biology will take time to become a large share of token demand and compute usage, so white-collar work must be the main driver over the next few years. Citrini says it is partly priced in, and he is unsure whether people are over- or underestimating how much coding and enterprise work there will be.

  “white collar work is really going to have to be the killer app. It's really what's going to account for an overwhelming amount of the demand, at least for the next few years.”at 21:28

US rates & the Fed

• Citrini

  Fed hikes are largely priced in, but one could still be a tantrum hike, so rate risk remains.

  Citrini thinks the market has priced in enough hikes, with about three expected over the next twelve months. Still, Citrini sees a higher chance that one of the hikes, if they are delivered, becomes a tantrum hike that surprises markets. Citrini thinks it is better to get hikes out of the way now rather than later. Jack Farley adds that the market prices about a 20% chance of an October hike.

  “Like, we've priced in enough hikes at this point. And I think that there's there's a higher chance, like, if if we actually go through with all these hikes, like, one of them ends up being like a tantrum hike.”at 58:28

US housing

• Citrini

  Homebuilders may be a decent buy as rates are priced in, with some names trading below book.

  Citrini says home prices have stalled after a roughly 125% COVID-era appreciation, and at 8% mortgage rates homeowners cannot keep the gains made when rates were zero. Citrini's view on homebuilders depends on rates, and Citrini thinks the hikes are largely priced in. Broad exposure to homebuilders could be a decent time to buy, and some smaller names trade below book value. Jack Farley says homebuilders are too hated and could rebound.

  “At Home Builders, I think, like, if you're gonna buy them broadly, it's probably a decent time to buy a thing. But you can find some good ones”at 58:37

Software & SaaS

• Citrini

  Market underprices AI disruption of software; many software firms survive, but business models must change and investors need company-level adaptation views.

  Citrini argues investors have grown complacent about how disruption gets priced in. Disruption usually unfolds as a sequence of repricings, not a straight line down (citing BlackBerry after the iPhone and Macy's after Amazon), and buying beaten-down names can pay once the threat becomes consensus. He still expects many software companies to survive, but says some business models will have to change significantly and others will stop working. He cites a capability gap, where AI can do far more than people currently use it for, and expects labs to close it by releasing products that pull users into those capabilities, which threatens incumbents. If the labs become public companies with quarterly targets, they will have more incentive to ship such products. He says a lot of air has come out of software valuations that were inflated in 2021. Investors therefore need a strong view on how each software company they own will adapt, and should not be complacent about the threat.

  “I personally think that there's still going to be a lot of software companies... But I think you really have to have a strong view on how the company that you own in software is going to adapt.”at 1:19:18

AI & jobs

• Citrini

  AI-driven job losses are needed to fund frontier AI spending; if unemployment shows no AI impact by late 2027, his thesis is wrong.

  Citrini argues that for frontier AI spending to be economic, enterprises must spend more. Consumers cannot supply that growth because they top out in how much they will spend. Some of the extra enterprise spending would have to come from AI doing work that would otherwise require hiring people, and he cannot say how much of the return would come from productivity alone. He cites current capability as evidence: AI completed a fix across roughly 45 to 50 figures in his report without a single error, and he says the number of people whose main job is adjusting PowerPoint slides is no longer zero. He does not want this scenario to happen. He sets a test for late 2026 to 2027: if AI has not affected unemployment by the middle or end of 2027, he will consider the piece wrong.

  “...I think that's the period of time where we'll see for sure. If it's the middle or the end of twenty twenty seven, and there's been no impact on unemployment from AI, then I'll be pretty comfortable being like, No, the piece was wrong.”at 1:21:44