The Recursion Bet

There is a metaphor that Chris Gibson, Recursion Pharmaceuticals’ CEO, has used repeatedly to describe what his company is trying to do: treat drug discovery like Intel treats semiconductor manufacturing. Rigorous process. Massive scale. Predictable output. Make the factory reliable enough, and the molecules will follow.

It’s a compelling pitch. It raised over $900 million before the company went public in 2021. It attracted Nvidia as a strategic investor. It enabled the acquisition of Exscientia in early 2025 in a deal that made Recursion, at least on paper, the largest AI drug discovery company in the world by number of programs.

The metaphor is also, in a specific and important way, wrong — and understanding why it’s wrong tells you something useful about the broader ambition of industrialized biological discovery.

What Recursion Actually Built

The core of Recursion’s approach is phenomics at scale. Rather than targeting a specific disease mechanism and designing molecules to disrupt it, Recursion runs perturbation experiments — silencing genes with CRISPR, adding compounds, inducing disease states — and photographs cells in vast numbers, capturing morphological changes that might indicate biological activity. Their imaging platforms capture roughly 1,000 features per cell across millions of cells per experiment. The resulting dataset is enormous by any biological standard.

The theory is that these cellular phenotypes (the appearance of cells under perturbation) encode information about the underlying biology that is too complex for humans to parse directly but accessible to machine learning. If you teach a model on enough perturbation-phenotype pairs, it can find patterns that predict which compounds might have effects on which diseases, without requiring you to fully understand the causal mechanism.

Related reading

This is not a crazy idea. High-content screening has been part of pharmaceutical R&D for decades. What Recursion added was scale (their Cell Painting operation runs roughly 2.2 million experiments per week), machine learning on top of the imaging data, and a business strategy built around the premise that the data advantage would compound over time.

The Pipeline Reality

By mid-2026, Recursion has reported Phase II data on REC-994 for cerebral cavernous malformation — a rare disease where the blood vessels in the brain form malformations that can bleed. The Phase II results were negative. Primary and secondary endpoints were not met. The program, which had been one of Recursion’s highest-profile validations of their platform, is discontinued.

REC-2282, a candidate for NF2-related schwannomatosis (tumors of the peripheral nervous system), showed more promising Phase II signals — preliminary data in 2025 suggested tumor volume reduction in a subset of patients, and Phase II enrollment continues. This is a rare disease with limited treatment options, which means the clinical bar is relatively accessible if the biology holds.

The broader pipeline — which Recursion claimed to contain over 40 programs as of their 2025 annual report — is at various stages of preclinical and early clinical development, and nearly all of it is in rare diseases and oncology. There are no programs in common diseases. There is no cardiovascular program. There is no metabolic disease program. The factory, to the extent it is producing output, is producing output in the disease categories where orphan drug designations provide regulatory advantages and smaller patient populations mean less expensive trials — not in the areas where industrialized discovery would have the greatest public health impact.

This is not a criticism unique to Recursion. It’s an observation about the economic logic of drug development at scale: you go where the regulatory path is clearest, not necessarily where the unmet medical need is greatest.

Why Biology Is Not a Fab

Semiconductor fabrication works because silicon is a deterministic material. Under precisely controlled conditions, the same process reliably produces the same result. Intel’s factory is valuable because the variance is extremely low — a chip that passes quality control behaves predictably in the field.

Biology is stochastic and context-dependent in ways that semiconductor physics is not. The same gene silenced in HeLa cells (the immortal human cervical cancer cell line that has been in laboratories since 1951) produces a different phenotype than in primary lung epithelial cells from a human patient, which produces a different phenotype in a patient with a specific genetic background, which produces yet another phenotype in the context of a tumor microenvironment. The cellular phenotype that Recursion’s imaging captures is not a ground truth — it’s a measurement of one biological context. The translation from that measurement to a patient population is the leap that industrialization cannot guarantee.

This is the fundamental problem with phenomic drug discovery at scale: it generates enormous datasets, but the information density per experiment may be lower than it appears, because the biological context is only partially controlled and partially understood. A pattern that predicts biological activity in Recursion’s cell lines may predict nothing in a clinical trial. The history of pharmaceutical R&D is littered with compounds that worked beautifully in carefully controlled in vitro systems and failed entirely in animals, let alone humans.

The Exscientia Integration

Related reading

The 2025 merger with Exscientia brought a different computational approach — generative chemistry and AI-designed molecules — into Recursion’s phenomics-first paradigm. In theory, the combination is powerful: use phenomics to identify targets and validate biology, use generative chemistry to design molecules against those targets. In practice, integrating two large organizations with different scientific cultures, different computational platforms, and a history of competing for the same narrative is hard.

The merged entity shed roughly 25 percent of its combined workforce in the 18 months post-merger — not unusual in large pharma mergers but significant for companies that built their identities around being science-driven platforms rather than traditional drug companies managing portfolios. The platform narrative has become quieter in investor communications; the program-by-program clinical narrative has become louder. This is what happens when a platform company runs out of the luxury of judging itself on process rather than output.

The Honest Accounting

Recursion is not a failure. Its pipeline has advanced compounds to clinical stage that might not have been found through traditional screening, and it has built genuine infrastructure that differentiates it from conventional biotechs. The merger has created a company with more shots on goal than either predecessor.

What it is not, at least not yet, is a validation of the semiconductor factory metaphor. The deterministic precision that makes Intel’s fab valuable — the same inputs reliably producing the same outputs — doesn’t translate cleanly to biology, because biology is not a deterministic material. The factory produces many experiments efficiently. Whether those experiments reliably produce useful drugs at scale is still the empirical question.

The useful benchmark will come in the next two to three years, when enough Recursion-generated compounds will have completed Phase II to assess whether their success rate materially exceeds historical baselines. If it does, the factory metaphor is vindicated — the machine learned to navigate biological stochasticity reliably. If it doesn’t, then industrialized discovery will need to revise its theory of what, exactly, scale can and cannot fix in drug development.

Gibson’s bet is still live. The semiconductor analogy was always a pitch as much as it was a hypothesis. The trials will decide which.

The Scale Problem Nobody Talks About

There’s a subtler issue with Recursion’s approach that doesn’t surface in investor communications: the relationship between the scale of their data generation and the informativeness of that data. Running 2.2 million experiments per week sounds like an overwhelming competitive advantage. But if many of those experiments are systematically correlated — measuring similar perturbations in similar cell lines under similar conditions — the effective information content of the dataset may be much smaller than the raw experiment count implies.

This is the difference between running 2 million diverse experiments and running 100,000 experiments each repeated 22 times. Biological datasets can suffer from exactly this kind of hidden correlation, because the perturbation space is not uniformly sampled. Common gene families, well-known signaling pathways, and commercially available compound libraries are heavily represented; genuinely novel biology is underrepresented. A model trained on this distribution learns the well-sampled regions of biology reliably and remains uncertain about the regions that matter most for finding truly novel drugs.

Recursion has published analyses suggesting their dataset covers biological space broadly, and their academic collaborators have validated the diversity of their perturbation profiles. But the translation from dataset diversity to model performance on held-out drug targets — particularly targets that are genuinely novel — is the empirical question that phenomic screening cannot answer internally. It requires prospective programs with novel targets to proceed to human trials, which is exactly what’s happening now.

Related reading

The Partnership Model

One adaptation Recursion has made to the platform-company narrative is aggressive partnership with large pharma. Their Roche/Genentech partnership (announced 2022, expanded 2024) and Bayer collaboration involve Recursion applying its platform to targets those companies bring, in exchange for milestone payments and royalties. This is a different business model than running all programs internally — it outsources the later-stage development risk to partners with the capital and infrastructure to manage Phase II and III trials — while providing near-term revenue that sustains the platform operation.

Whether this model proves that the platform creates value, or merely that large pharma is willing to pay for access to a novel data-generation capability on an exploratory basis, will depend on whether the partnered programs produce clinical results. The first Roche-Recursion collaboration to reach Phase II will be closely watched. Large pharma partnerships have a long history of producing option value for biotechs without producing drugs — the partnership validates the technology in a reputational sense while the clinical data remains years away.

The semiconductor fab comparison may ultimately need to be replaced by a different metaphor: Recursion as a specialized analytical instrument company, providing data and models that pharmaceutical companies consume as inputs to their own programs. That would be a valuable and sustainable business. It’s just not the same story as industrialized discovery — and the distance between the two matters for how the company’s scientific ambitions are evaluated.

What the Recursion story illustrates about the broader project of industrializing biological discovery is that the factory metaphor has genuine appeal — it captures something real about the potential for scale and process discipline to improve productivity — while eliding the fundamental difference between manufactured goods and biological knowledge. Semiconductors improve at rates described by Moore’s Law because the laws of physics governing semiconductor behavior are stable, knowable, and compressible into engineering practice. Biology improves our understanding of itself at rates that are governed by the pace of experimental discovery, which is subject to serendipity, translational failure, and the resistance of living systems to being fully characterized by any method that doesn’t involve actually studying them. The factory can run at Intel speeds. The biology still moves at biology speeds.