Case one: the lab demand that AI created
Recursion designs drugs on an NVIDIA supercomputer, yet its stock fell -31.4%, -39.5% and -18.3% across 2024-26, three years running. Over the same period the supply side selling that compute surged. That gap is the heart of the techbio case. On any frontier, the infrastructure demand re-rates before the application's payoff is ever proven (the full grammar was set in Every frontier re-rates infrastructure before application), and here it splits most sharply.
Techbio reverses the direction of the chain. Most companies this wiki tracks (its nodes) sell AI infrastructure; techbio buys that infrastructure to make drugs and diagnostics. Here AI is not something sold on a shelf. It is a factor of production. And so the case splits the grammar cleanly in two. The application's (drug) payoff is still a story, but the infrastructure demand that application generates is already measured.
Techbio is a demand node that buys AI
Recursion trains foundation models on 23PB of bio and chemical data, and that training runs on the BioHive-2 supercomputer built from NVIDIA GPUs. NVIDIA put $50M in during 2023, so the two are bound by capital as well. Schrodinger runs physics simulation in the cloud and licenses software to pharma (software ACV of $198M); Tempus buys sequencing and compute to build multimodal medical data and sells it to AstraZeneca and Pathos on a $200M deal. AbCellera designs antibodies with AI and hands them to Eli Lilly and Jazz. The forms differ, but all four buy compute as a line-item cost.
The application's payoff is still a story
Yet this demand side has entirely lagged the infrastructure it consumed. On top of Recursion's three straight down years, Schrodinger, after -46.1% in 2024, never recovered, at -7.3% and -9.7% in 2025 and 2026. Over the same period the supply side of the compute they bought climbed hard. Why did the gap open? Infrastructure revenue is booked the moment a GPU ships, but the value of the drug application hangs on the probability that a candidate molecule clears the clinic, years away from realization. That the promise is not yet earnings shows in the numbers: Recursion's Q4 2025 revenue was $35.5M, and AbCellera's full 2025 revenue $33.7M.
But the infrastructure demand is measured
Separate from whether the application's payoff is a story, the demand that application generates is already a number. Tempus's 2025 revenue rose 83% to $1.27B, of which diagnostics were $955M. Long before any drug clears the clinic, the data and diagnostics demand is already realized. Beneath them sits the picks-and-shovels layer, Illumina. The world No.1 in DNA sequencing, at 92% share, it makes the source data AI-bio firms train on. The bigger the application side grows its models, the more sequencing it needs, so Illumina holds the seat the GPU holds for AI-bio. Illumina fell -46.9% in 2022 and -31.1% in 2023, then turned to +35.4% in 2026 (Q4 revenue $1.16B, FCF $931M). The infrastructure demand is measured before the application's payoff, and here it shows plainly. While an application like Recursion still runs losses, the layer selling it data has already turned profitable.
AbCellera begins turning the promise into a number
The grammar has a next phase too: the moment the application's promise first becomes a number. AbCellera crawled the floor at -29.2%, -43.6% and -48.7% across 2022-24, then, against a clinical entry (ABCL635 Phase 2), 104 pipeline programs and $700M of liquidity, turned to +16.7% in 2025 and +94.7% in 2026. If NVIDIA holding equity in Recursion is the circular financing of AI invests in itself reaching into the bio layer, then AbCellera's rebound looks like the application at the far end of that circuit beginning, for the first time, to show up in earnings.
Verdict
Techbio was the first test of the grammar, and the result matched the prediction. The infrastructure demand (compute, sequencing, data) is already measured, and the application's (drug) payoff is still a story. Infrastructure first, applications not yet. This lag must not be mistaken for a cycle. As many clocks shows, infrastructure and applications run on different clocks even under the one word AI.
One case could be coincidence. So a second is needed, this time a chain outside AI. Was it really EVs that revived the battery, or does the same grammar hide here too? It was ESS, not the EV, that revived the battery takes up case two.