Oct 25, 2026 · 1h 27m
AI drug design hits a bottleneck money cannot quickly fix
Citrini: Monkeys Are The Next Bottleneck | Citrini on AI Drug Discovery Experiment & Reflexive Financing Risk to AI Buildout & Hardware Trade
The episode tests whether AI’s next investment opportunity lies in real-world experimentation rather than ever-larger hardware buildouts.
- 1AI may commoditize molecule design while laboratories, automation, CROs, and research monkeys remain scarce.
- 2Citrini sees adoption and customer cash flows, not Chinese competition alone, as the key test for AI infrastructure valuations.
- 3Meta looks better positioned than Alphabet, while CoreWeave illustrates how quickly financing risk can spread through the AI buildout.
Don't miss
Citrini explains how his Claude-assisted scorpion antivenom experiment generated toxin-binding proteins while exposing the slow, expensive laboratory bottleneck.
The brief
James Citrini argues that the AI trade is moving from hardware toward applications, enterprise workflows, and companies that capture actual usage rather than promised capacity.
The central tension is whether falling token costs unlock enough demand to justify infrastructure valuations, or instead expose spending that customer cash flows cannot support.
Citrini’s Claude-assisted scorpion antivenom experiment produced proteins that bound to toxins, but the long, costly validation process showed where AI-generated biology meets reality.
As molecule design becomes cheaper, testing capacity becomes more valuable: CROs, automation, laboratory consumables, and research monkeys could constrain the next wave of drug discovery.
The conversation then widens to Meta, Alphabet, Chinese inference economics, software disruption, FDA reform, homebuilders, and CoreWeave’s exposure to a financing shock.
What was said on this episode
43 statements · 21 positive · 20 negative · 2 mixed
Citrini remains bullish on NVIDIA.
“I'm still pretty bullish on NVIDIA from here.”
Listen at 3:00
Incremental AI spending increasingly relies on equity and debt financing.
“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.”
Listen at 5:56
AI infrastructure sentiment now has greater fundamental impact.
“sentiment can have a much more pronounced impact”
Listen at 6:18
AI biology models need vastly more data to discover new targets or breakthroughs.
“in very specialized areas like biology, for example, where you have a model that can do literature review or it can go off the existing data, but it's not like it's going to come up with... new targets or kind of like breakthroughs without and really, you know, orders of magnitude more data.”
Listen at 8:16
AI will provide good long-term returns on investment.
“I don't have any doubt, I've never had any doubt that AI would provide a good return on investment and that the returns to intelligence would be good over the long term.”
Listen at 8:51
AI demand currently exceeds supply.
“The demand still currently exceeds supply.”
Listen at 14:57
US AI models retain a capability gap over Chinese open-source models.
“the gap in capability between the US models and open source Chinese, it's remained defended.”
Listen at 16:47
Faster paid-usage growth would make Citrini more constructive on AI infrastructure.
“I would become much more constructive on infrastructure if we saw it's the same growth in paid usage. And that second derivative kind of speeding up again”
Listen at 18:53
Current estimates imply roughly $1 trillion of AI annual revenue by 2028.
“you're looking at about a trillion dollars in AI AR by 2028.”
Listen at 20:18
Average drug development takes 10–15 years from idea to FDA approval.
“Average drug takes between 10 to 15 years from, hey, we've got an idea to, okay, here's our FDA approved drug.”
Listen at 27:00
AI drug discovery still requires money, infrastructure, and specialized expertise.
“we wanted to experience what AI could accelerate. We wanted to experience what the bottlenecks still were, what still required money and infrastructure and especially what required real expertise.”
Listen at 32:12
Seventeen AI-designed proteins bound the scorpion-toxin target.
“We got 17 of the designs that bound”
Listen at 37:16
AI compressed some drug-discovery work from months into days.
“the AI kind of compressed what would be months of work for scientists into days.”
Listen at 38:45
Synthesis, testing, and controls still take months despite AI assistance.
“the synthesis, the testing, getting the controls, that still took months.”
Listen at 38:53
Citrini recommends exposure to experiments and experiment-enabling businesses.
“we really want to be long experiments”
Listen at 39:55
AI-driven cancer research will not drive much compute demand within two years.
“if you're concluding that the demand from AI over the next year or two is going to come from how much compute it's going to take to cure cancer, you're probably going to be disappointed.”
Listen at 41:38
Causal biology is a primary bottleneck for AI drug discovery.
“the bottleneck is in, first off, like causal biology.”
Listen at 42:03
Predictive AI toxicology requires vastly more causal-biology data.
“we would need like some there are like patrick sue is working on this with the virtual cell atlas there are efforts to kind of create these perturbational maps of like causal biology in order to like really understand it to like the degree where you're getting an ai that can that can predict say toxicology we're going to need orders of magnitude more data”
Listen at 42:30
AI has largely solved drug design within causal biology.
“drug design is pretty much at least in in causal biology it's like pretty solved”
Listen at 45:09
AI advances will further commoditize large-molecule drug design.
“it's going to commoditize even more”
Listen at 45:32
AI may reduce roughly 20% of drug-design-to-trial work; 80% remains irreducible.
“our estimate i think is overall in the entire like drug design like to clinical trial thing you can reduce about 20 of it and 80 of it is just irreducible”
Listen at 45:44
AI-designed drugs will enter clinical trials.
“There will be AI designed drugs that enter clinical trials.”
Listen at 48:33
AI will most strongly benefit preclinical experimentation through more experiments.
“the most pronounced impact will be on taking advantage or benefiting from this explosion of experiments especially in pre the pre-clinical space”
Listen at 49:38
AI lowers barriers and speeds creation of drug candidates ready for laboratory testing.
“the biggest impact that AI is having is in creating additional candidates, like reducing the barrier to entry, reducing the amount of time it takes to get from, I have an idea to, I have a molecule that's ready to send to the lab.”
Listen at 50:27
Preclinical and early-clinical drug assets will lose value over time.
“the value of having either a preclinical or early clinical stage asset is probably going to go down over time.”
Listen at 52:35
Eli Lilly has successfully acquired molecules and advanced them to FDA approval.
“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.”
Listen at 55:36
Eli Lilly will benefit from commoditization of phase-one assets.
“that that's a example of a company that is going to benefit from this commoditization of phase one assets.”
Listen at 56:24
At 8% mortgage rates, prior 125% housing appreciation is unlikely to persist.
“mortgage rates are 8%. You do not get to keep the 125% you made when they were zero.”
Listen at 57:57
Citrini is relatively bullish on interest rates.
“I'm relatively bullish.”
Listen at 58:26
Some homebuilders may be attractive because they trade below book value.
“home builders, I think like, if you're going to buy them broadly, it's probably a decent time to buy a thing, but you can, you can find some good ones that are like trading below book.”
Listen at 58:42
Research monkeys are in severe shortage for preclinical testing.
“there's a huge shortage of monkeys”
Listen at 1:01:42
Breeding research monkeys takes approximately four years.
“The time it takes, I think it's four years to breed these monkeys.”
Listen at 1:02:53
China stopped exporting research monkeys.
“China originally cut off the supply that they stopped exporting monkeys.”
Listen at 1:03:04
The United States trails China and Australia in phase-one trial completion speed.
“the U.S. has fallen significantly behind China and Australia, interestingly enough, in like the time it takes to complete a phase one study.”
Listen at 1:06:25
Citrini is bullish on Meta because of Muse's performance.
“I'm bullish on Meta because I think that they're really killing it with Muse.”
Listen at 1:07:31
Meaningful Muse adoption could quadruple Meta ARPU by 2030.
“if Muse sees any meaningful degree of adoption, you can draw a straight line from that and that data to ARPU increasing by like 4X by 2030.”
Listen at 1:08:00
AI-driven behavior changes pose continuing risk to Google Search.
“there's still some degree of risk to search. Like a lot of people have changed the way that they search, myself included. I do a lot less Google searching now.”
Listen at 1:09:14
Citrini holds a long-Meta, short-Google trade.
“long meta short Google is a trade that I have on”
Listen at 1:09:42
A spending-boosting AI product would invalidate current infrastructure concerns.
“if what we're seeing right now is better capabilities preceding a new product or some new kind of a mock that really jacks up spending, then... then all those concerns are moot and we should just full hog with the AI infrastructure trade.”
Listen at 1:11:02
Citrini does not expect CoreWeave to default.
“I'm not expecting CoreWeave to default.”
Listen at 1:12:53
DeepSeek's high inference margins could enable continued token-price deflation.
“it seems like they have an 80% margin there. So that's like, so if they're running 80% margins, you could kind of imagine like maybe there's going to be a continual cheap deflation in token pricing.”
Listen at 1:17:56
No AI employment impact by late 2027 would falsify Citrini's scenario.
“If it's like the middle or the end of 2027 and there's been no impact on employment from AI, then I'll be pretty comfortable being like, no, the piece was wrong.”
Listen at 1:21:51
AI model switching costs are effectively zero when a superior model appears.
“the switching costs are zero, right? If chat GPT comes out with a model that's way better or is way better for a specific thing.”
Listen at 1:24:08
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.