Recursion

Recursion Pharmaceuticals: The Company That Photographs Sick Cells for a Living — and Why That Might Be a Trillion-Dollar Idea

July 24, 2026 | healthcare

↓ .md Take this into your AI — the full analysis + graph as markdown, ready to paste into ChatGPT, Claude, Gemini or any AI.

Based on 29 related nodes across 8 research explorations in the AI sector

What Does Recursion Actually Do?

Imagine you’re trying to figure out which key opens a lock, but instead of trying each key one at a time, you take a photograph of millions of locks simultaneously and teach a computer to recognize which ones look like they’ve been opened. That’s roughly what Recursion does — except the locks are diseased human cells, and the keys are potential drugs.

Every week, Recursion’s machines generate more than 10 terabytes of photographs of cells — what they look like when they’re healthy, when they’re sick, and when various chemical compounds are introduced to them. Over years, this has produced one of the most detailed biological image libraries ever assembled. The company then uses AI to find patterns in those images that suggest which compounds might work as medicines.

This approach is called “phenomics” — studying what things look like and how they behave (their “phenotype”) rather than just studying their genetic code.

The Mapmaker Analogy

Here’s a useful way to think about Recursion’s structural position: they are the mapmakers.

When European explorers were navigating the world, the companies that made accurate maps didn’t need to own the ships, the trade routes, or the ports. They just needed everyone who did own those things to buy their maps. Recursion’s phenomics data functions similarly. Other AI drug discovery efforts — including the most sophisticated protein-folding models in the world — need biological context that their tools alone cannot provide. Recursion generates that context at industrial scale.

The research graph underlying this analysis shows one of the strongest relationships in the entire dataset pointing in Recursion’s direction: the emerging concept of “virtual cells” — computer simulations that could replace animal testing entirely — depends on Recursion’s platform as a primary data source. That’s not a partnership or a contract. That’s a supply chain relationship. Recursion doesn’t need to win every drug discovery competition; it needs to be the raw material supplier to whoever does win.

Why Their Data Is Hard to Copy

The most non-obvious finding in this analysis is that Recursion’s real competitive advantage is not their AI models. AI models are becoming increasingly easy to build, fine-tune, and replicate. The advantage is the data those models are trained on.

You cannot reverse-engineer a decade of cellular imaging experiments from public scientific literature. You cannot scrape it from the internet. You cannot generate synthetic substitutes that carry the same biological validity. If you want to build a virtual cell model — which the pharmaceutical industry is increasingly desperate to do, especially as regulators move away from requiring animal testing — you need training data that looks like Recursion’s library.

This is why the data is described as a “moat” — something that protects a castle from attack. The castle is Recursion’s business. The moat is the irreplaceable, proprietary, continuously growing dataset that took years and hundreds of millions of dollars to accumulate.

The Closed Loop: What the Exscientia Merger Changed

In 2025, Recursion merged with a company called Exscientia. Before that merger, Recursion was excellent at identifying which biological targets might respond to a drug — but they needed partners to then design and synthesize the actual drug molecules.

After the merger, they can do both. The full cycle now works like this: photograph sick cells, find patterns suggesting a target, design a molecule computationally, synthesize it in-house, test it against the cells, observe the result, update the models, repeat. This loop runs continuously without depending on outside companies at critical steps. In manufacturing terms, this is called vertical integration. In drug discovery terms, it means the timeline from “interesting pattern in cells” to “compound ready for animal-free testing” compresses dramatically.

The Strengths

The data flywheel compounds over time. The more data Recursion collects, the better their models get. The better their models get, the more useful their platform is to pharmaceutical partners. The more pharmaceutical partners use the platform, the more data gets generated. This self-reinforcing cycle is hard to interrupt once it reaches a certain scale, and Recursion has been running it longer than almost anyone.

They are inside the pipeline, not competing with it. Recursion’s platform integrates with AlphaFold3 — the dominant protein structure prediction tool, owned commercially by Google’s Isomorphic Labs — rather than trying to replace it. Structure prediction tells you the shape of a protein. Phenomics tells you what happens to a living cell when you perturb it. These are complementary, not competing, questions. Recursion has positioned itself as a required component of the combined answer.

Regulation is moving in their direction. The FDA announced in 2025 that it is phasing out the requirement for animal testing in certain drug development stages. This is enormously significant for Recursion. If regulators accept AI-generated cellular data as evidence of safety — a transition that is now happening in policy terms — Recursion’s library moves from “useful research tool” to “required regulatory filing input.” That changes the monetization calculus entirely.

The Vulnerabilities

They have demonstrated the problem, not just solved it. This is the most uncomfortable finding in the analysis. The research graph describes Recursion’s own platform as something that “demonstrates” the clinical translation gap in AI drug discovery. That gap refers to a pattern that has emerged across the industry: AI systems that produce brilliant predictions in the lab consistently underperform when those predictions are tested on actual patients in clinical trials.

The concern is not that Recursion is uniquely bad at this. The concern is that the gap is real and difficult to close — and Recursion, despite having one of the most sophisticated platforms in existence, has not yet proven they can consistently bridge it.

Their computers depend on a supply chain with one fragile point. Recursion runs on NVIDIA graphics processing units. NVIDIA chips are manufactured almost exclusively on specialized equipment in Taiwan. The analysis traces a direct chain from geopolitical risk in the Taiwan Strait, through NVIDIA’s supply constraints, to Recursion’s ability to generate 10 terabytes of data per week. If that supply chain is disrupted — by trade war, conflict, or export controls — Recursion’s data flywheel stops. Structure-based competitors, who require less continuous raw compute, would be comparatively unaffected.

A better-funded competitor is building toward the same destination. Xaira Therapeutics launched with $1 billion in committed capital and a Nobel Prize-winning co-founder. Their explicit goal is to build the same kind of virtual cell foundation model that Recursion’s data currently underpins. Recursion has a first-mover advantage in data accumulation, but Xaira has more capital and scientific prestige. The question of how much of a lead Recursion’s head start actually represents is genuinely unresolved.

Bull Case: Why This Could Work Spectacularly

The optimistic scenario connects three things happening at once.

First, regulators formally accept AI-generated cellular evidence as a substitute for animal testing. This is already moving from policy statement toward implementation. Second, virtual cell models mature enough that pharmaceutical companies begin licensing or training on them at scale. Third, Recursion’s phenomics library is established as the primary training data source for those models.

If that chain completes, Recursion transforms from a drug discovery company — which must prove clinical results to generate value — into a biological data infrastructure company, where every pharmaceutical company building an AI platform is a potential customer for the underlying data. This is a fundamentally more durable and scalable business than winning individual drug races.

The analogy would be becoming the satellite imagery provider that every mapping application depends on, rather than building one mapping application and hoping it wins.

Bear Case: Why This Could Go Wrong

The pessimistic scenario is more straightforward. Recursion’s AI finds patterns in cells. Those patterns suggest drugs that should work. The drugs then fail in human trials, repeatedly, because human biology at the whole-organism scale is vastly more complex than cell images capture.

If several high-profile Recursion-derived programs fail clinically over the next few years, the narrative that phenomics data is the key to AI drug discovery becomes damaged. Pharmaceutical partners become skeptical. The data moat argument loses force if the data doesn’t produce clinically translatable insights.

This failure mode is not hypothetical — it has already happened to other AI drug discovery companies. BenevolentAI’s highly publicized atopic dermatitis program failed in clinical trials despite strong computational predictions. The pattern is established enough that the research graph treats Recursion’s own platform as an illustration of the risk rather than a solution to it.

The worst-case version combines a clinical setback with a compute disruption — forcing Recursion to raise capital at a depressed valuation while their data generation is temporarily impaired.

The Leverage Points Worth Watching

The single highest-leverage question is whether the FDA formally establishes a pathway for phenomics-derived safety data to substitute for animal testing results. If that happens on a clear timeline, Recursion’s position crystallizes quickly. If it remains ambiguous, the value of their data moat remains theoretical.

The second question is whether Xaira builds competing phenomics infrastructure or focuses on different approaches. A $1 billion commitment going toward protein design and generative chemistry is a threat Recursion can absorb. That same capital going toward building a competing cellular imaging library would be an existential challenge to the moat thesis.

Bottom Line

Recursion has built something genuinely unusual: a proprietary library of biological data that other people’s AI models increasingly need. They occupy the mapmaker position in an industry where the maps are becoming essential infrastructure. Their data moat is real, their regulatory tailwinds are real, and their integrated pipeline is more complete than almost any competitor’s.

The central uncertainty is whether the data translates to medicine. A library of beautiful, high-resolution photographs of sick cells is only worth what it predicts — and the prediction that matters most, whether a drug will work in a human being, remains unsolved by everyone in this space, Recursion included. The bull case is that solving this problem is a matter of data volume and model sophistication, and Recursion is positioned to win that race. The bear case is that the problem is harder than that, and the confidence that large proprietary datasets solve clinical translation is the same confidence that has repeatedly disappointed the industry.

This is a high-conviction structural position with a genuine unresolved question at its center. That question — whether phenomics data bridges the gap from lab to clinic — is what the next five years of Recursion’s pipeline will answer.

Based on a research map of 29 related concepts and 208 connections across eight separate research runs in the AI sector.

Structural Position

Recursion sits at the center of the self-driving-lab, closed-loop research paradigm — not just as a participant, but as the company the research treats as definitionally representative of it. The link between the self-driving-lab concept and Recursion’s OS Phenomics Platform runs in both directions at the highest strength recorded in their immediate research area: the lab concept is implemented by Recursion’s platform, and the platform in turn exemplifies the concept.

That position has two distinct layers.

Upstream, Recursion supplies critical data. Virtual Cell Foundation Models — the most ambitious near-term vision in AI drug discovery, aiming to replace wet-lab experiments with computational cell simulation — depend directly on Recursion’s phenomics platform, one of the strongest dependency relationships found anywhere in the research. That dependency was reinforced by the 2025 Exscientia merger, which extended Recursion’s platform from imaging into generative chemistry and in-house synthesis, closing the design-make-test loop internally.

Downstream, Recursion is a required module, not a standalone alternative. The research describes an emerging industry-standard “integrated pipeline” for AI drug discovery by 2030, which incorporates Recursion’s platform alongside tools like AlphaFold3 and generative chemistry systems. Recursion isn’t positioned as competing with that pipeline — it’s positioned as a piece of it.

Recursion also shows heavy, roughly equal exposure (ten connections apiece) to the two opposing poles that define this sector’s central tension: the promise of dramatically compressed drug discovery timelines and costs, and the risk that AI-driven predictions fail to survive clinical trials. That balance confirms Recursion is maximally exposed to both the sector’s upside and its central risk.

Key Strengths

1. Industrial-scale phenomics data (durable). Recursion generates more than 10TB of cell-imaging data per week across millions of cellular perturbation states. The research identifies a broader pattern in AI drug discovery: language models and molecular AI are rapidly becoming commoditized, but proprietary, high-quality, well-annotated biological experiments are not. Recursion’s phenomics data is the clearest real-world example of that non-commoditizable moat — it can’t be reverse-engineered from public sources or easily replicated by new entrants.

2. Upstream position in the virtual-cell supply chain (durable). The dependency linking Virtual Cell Foundation Models to Recursion’s platform is among the strongest of its kind found anywhere in the research. As virtual cell models mature from a research concept into a tool regulators are willing to accept — a shift the FDA’s 2025 move to phase out animal testing is actively accelerating — Recursion’s training data becomes a required input to an increasingly valuable output. It’s a supply-chain position that gains value passively as the rest of the field develops around it.

3. Closed-loop integration since the Exscientia merger (moderately durable). The merger gave Recursion in-house generative chemistry, closing the design-make-test-analyze cycle. The research shows Recursion’s self-driving-lab feedback loop closing the loop for generative molecular design, with the phenomics platform itself feeding directly into that design process. Together, these suggest Recursion can now carry a phenotypic discovery from hit to synthesizable compound without relying on outside partners — reducing how much of the discovery cycle needs external collaboration.

4. A hedge across drug-discovery paradigms (moderately durable). Recursion’s phenomics-based foundation model competes with AlphaFold3’s structure-prediction approach, but only moderately — the two approaches are differentiated enough to coexist rather than being mutually exclusive. More tellingly, Recursion’s platform is shown integrating with the AlphaFold3 pipeline rather than fighting it. Recursion’s phenotypic approach supplies biological context that structure-based models can’t generate on their own, making it complementary to the AlphaFold3/Isomorphic paradigm at the pipeline level even while the two compete at the foundation-model level.

Structural Vulnerabilities

1. Recursion’s platform is used as evidence of the field’s central failure mode (immediate, only partly within Recursion’s control). This is the single most significant liability the research surfaces. Recursion’s phenomics foundation model is linked to the AI Drug Discovery Clinical Translation Gap — not as something that mitigates or targets that problem, but as something that demonstrates it. That gap is defined in the research as AI’s consistent failure to beat placebo in clinical trials after clearing computational benchmarks, and the strength of this link is well above peripheral — this is described as a well-established characterization, not a minor aside. Read plainly: the research treats Recursion as having generated high-quality in-silico predictions that did not translate into clinical outcomes.

2. A double dependency on NVIDIA (long-term, outside Recursion’s control). Two separate NVIDIA infrastructure offerings — its biological-AI infrastructure and its drug-discovery stack — both enable Recursion’s platform, at identical, substantial strength. Layered on top of this is a broader “chokepoint” risk pattern and NVIDIA’s GPU market dominance, which show up elsewhere in Recursion’s research connections. That dominance itself depends on a single vulnerable point in the chip supply chain: TSMC’s manufacturing, which carries direct exposure to a Taiwan Strait crisis. In short, the geopolitical risks that threaten NVIDIA’s supply chain flow straight through to Recursion’s core operations.

3. Recursion is a significant player, not the sector’s most heavily weighted one (long-term, partly within control). In the research’s overall strength ratings, Recursion’s platform ranks below both Xaira Therapeutics and the 2030 integrated pipeline concept. Xaira, a rival AI-first drug company, is explicitly building toward the same virtual-cell endpoint that represents Recursion’s most important upstream relationship — and Xaira brings superior funding (over $1 billion committed) and a Nobel laureate co-founder to that race.

4. Adjacent to GLP-1, but without ownership (medium-term). Recursion’s platform shows three connections each to the GLP-1 AI-drug-discovery feedback loop and to the broader GLP-1 multi-indication market opportunity — the most capital-intensive drug class in pharma today. But those connections reflect general platform relevance, not ownership of named clinical programs. This is positioning without capture: Recursion is in the room, but hasn’t claimed a seat at the table.

Competitive Dynamics

Isomorphic Labs (AlphaFold3 / IsoDDE). Recursion’s primary paradigmatic rival. The competitive link between Recursion’s phenomics foundation model and AlphaFold3’s structure-prediction approach is moderate rather than existential — the two are differentiated enough to coexist. Isomorphic holds exclusive commercial rights to AlphaFold3 and has Alphabet’s balance sheet behind it. But Recursion’s phenotypic signal (what drugs do to cells) is orthogonal to Isomorphic’s structural signal (how molecules bind proteins), and the research shows integration rather than displacement — Recursion’s platform is folded into the AlphaFold3-based pipeline. In pharma partnerships that need both structural prediction and phenotypic validation, Recursion looks like a complement to the AlphaFold3 ecosystem, not a replacement for it.

Xaira Therapeutics. Rated higher overall than Recursion’s platform, better capitalized, but earlier-stage. Xaira is explicitly building toward the same virtual-cell endpoint that anchors Recursion’s most valuable upstream relationship. If Xaira succeeds in building competing data-generation infrastructure, it would directly erode that dependency — Recursion’s single most structurally important asset. The research doesn’t make clear how far along Xaira actually is, so Recursion’s head start in accumulating phenomics data may or may not be a durable lead.

NVIDIA (infrastructure dependency, not a direct competitor). NVIDIA’s drug-discovery computing stack is shown competing with quantum-computing approaches to drug discovery economics, signaling NVIDIA’s own ambitions in this space. But its two enabling relationships with Recursion reflect dependency, not exclusivity — NVIDIA serves Recursion, but doesn’t serve only Recursion. Any pharma company running on the same NVIDIA infrastructure gets comparable compute access, which means Recursion’s real differentiation lies in its proprietary data, not in its compute stack.

Regulatory Exposure

FDA’s 2025 animal-testing phase-out (a positive tailwind). The FDA’s move away from animal testing feeds directly into demand for virtual cell models, which in turn depend heavily on Recursion’s platform — a regulatory shift that increases demand for Recursion’s core output through an indirect but strong chain. As rodent toxicology studies are phased out, drug sponsors need alternative mechanistic evidence of safety, and Recursion’s cellular perturbation data is a strong candidate to fill that gap.

FDA accelerated-approval pathways. A secondary regulatory benefit: virtual-cell models — trained on Recursion’s data — could serve as evidentiary support for sponsors pursuing accelerated approval on plausible-mechanism grounds. That would turn Recursion’s data asset into a direct input for regulatory filings, not merely a research tool.

No direct negative regulatory exposure. The research surfaces no enforcement or restriction actions aimed at Recursion’s core operations — cellular imaging and AI-driven phenotypics. What compliance risk exists is indirect: through the NVIDIA compute dependency, and through the sector-wide clinical translation gap, which is an industry risk rather than one specific to Recursion.

Strategic Leverage Points

1. Deepen the dependency that Virtual Cell Foundation Models have on Recursion’s platform. This is the single highest-leverage relationship the research identifies. Any investment that increases data volume, broadens coverage (more cell types, more perturbation types), and improves annotation quality would simultaneously deepen this dependency, strengthen Recursion’s proprietary-data moat, and help close the clinical translation gap — one action addressing several problems at once.

2. Convert GLP-1 adjacency into actual pipeline ownership. The three-connections-each relationship with GLP-1 discovery and market-opportunity concepts represents unmonetized potential. The research points to drug repurposing — using AI-driven biomedical knowledge to find new indications for existing GLP-1 compounds — as amplifying the GLP-1 market opportunity, suggesting repurposing may be the fastest route for Recursion to convert its phenomics data into real ownership, bypassing early-stage de-risking work entirely.

3. Move fast on the FDA’s animal-testing phase-out. This creates roughly a two- to five-year window in which regulators are actively looking for alternative safety-testing methodologies, and Recursion’s phenomics data fits that need precisely. Proactively investing in regulatory science — working directly with the FDA to get phenomics-derived safety data accepted as a standard — would give Recursion a first-mover position that would be structurally hard for competitors to copy.

4. Reduce single-vendor reliance on NVIDIA. Given the multiple links between NVIDIA’s GPU dominance and Recursion’s operations, and the broader chokepoint pattern connecting that dominance to geopolitical risk, investing in compute diversification — such as custom chip alternatives, which elsewhere in the research are shown undermining similar compute dependencies in other sectors — would reduce the biggest structural risk that sits outside Recursion’s own control.

Bull Case

Thesis: Recursion’s phenomics data moat compounds into the defining infrastructure position of the AI drug discovery era, as virtual cell models become the primary regulatory-accepted alternative to animal testing.

The chain runs from the FDA’s animal-testing phase-out, which enables virtual cell models, which depend on Recursion’s phenomics platform — all three links rated at solidly high strength. If virtual cell models win regulatory acceptance (a path the FDA’s policy shift is already accelerating), the training data needed to build them becomes a supply-constrained input. Recursion generates over 10TB of that data weekly, and the research is explicit that datasets like this don’t commoditize. Every pharma company building or licensing a virtual cell model becomes, in effect, a Recursion data customer — turning a research platform into an infrastructure company.

Recursion’s position inside the emerging standard drug-discovery pipeline reinforces this: its platform is a required component of the 2030 integrated pipeline, and it integrates directly with the AlphaFold3-based pipeline too. Recursion doesn’t need to beat Isomorphic Labs at foundation models — it just needs to remain the phenomics layer that every structure-based approach needs for biological validation.

The Exscientia integration closes the loop further: the platform now spans target identification, phenotypic screening, and generative chemistry, with the self-driving-lab feedback loop closing the design-make-test cycle for generative molecular design. That positions Recursion to run continuous, autonomous discovery campaigns without external dependencies.

What has to hold true: the FDA accepting virtual cell data as an evidentiary standard; Xaira staying focused on protein design rather than building competing phenomics infrastructure; and NVIDIA’s supply chain staying intact through 2030.

Bear Case

Thesis: Recursion’s platform generates impressive data but has demonstrated, rather than solved, the clinical translation gap — and its compute costs scale faster than its therapeutic output.

The critical relationship: Recursion’s phenomics foundation model is treated in the research not as a solution to the clinical translation gap but as an instantiation of it — a well-established characterization, not a passing detail. The translation-gap concept itself holds that AI excels at predicting molecular properties, but biology is far more complex than any current model captures, citing BenevolentAI’s failed atopic dermatitis program as a precedent for this pattern. If Recursion’s phenomics-derived hits fail similarly at the clinical stage, its data moat becomes a liability — an expensive dataset producing confident predictions that never translate.

The compute dependency compounds the risk. NVIDIA depends on TSMC, which is itself the central node in a broader chokepoint pattern, and a Taiwan Strait crisis is shown triggering that chokepoint at the highest strength recorded anywhere in this research. A 12-18 month disruption to that supply chain would halt Recursion’s data-generation flywheel, while structure-based competitors — who need much less continuous compute for inference than for training — would be comparatively unaffected.

Meanwhile, Xaira’s explicit push toward the same virtual-cell endpoint, backed by over $1 billion in committed capital and stronger founder credentials, represents a credible path to eroding the upstream dependency that is Recursion’s most valuable structural asset.

Most likely negative scenario: a string of clinical failures undermines the narrative that phenomics data produces clinically translatable hits, cooling pharma partnership interest right when Recursion most needs to monetize its data. Most severe scenario: a Taiwan supply-chain disruption halts data generation at the same time as a clinical setback, forcing a capital raise at a depressed valuation.

Regulatory Stress Test

FDA animal-testing phase-out (positive, low risk of harm). Full enforcement on the stated timeline benefits Recursion — if regulators require non-animal alternatives for Phase 1 safety data, phenomics-derived evidence becomes a required filing element. The risk is that regulators instead favor organ-on-chip or patient-derived organoid models over AI-predicted cellular responses, in which case Recursion’s data may not map cleanly onto whatever format becomes required. Overall: manageable, and more likely to help than hurt.

FDA accelerated-approval mechanisms. Virtual cell models depend on this pathway staying open. If the FDA tightens accelerated approval or demands confirmatory trial data before launch, that closes off a key downstream use case for Recursion’s data. This wouldn’t eliminate Recursion’s value, but it would slow how quickly its data reaches clinical application. Overall: manageable, medium likelihood.

Data privacy / patient-data regulation. Not something the research addresses directly, but worth noting: Recursion’s imaging data comes from laboratory cell lines, not patient records, which likely reduces its exposure to health-data privacy rules relative to competitors working from electronic health records or genomic databases. This is a plausible compliance advantage that the underlying research doesn’t explicitly capture.

Compute export controls (indirect exposure). Broader connections around Chinese rare-earth and chip-related geopolitical tension suggest that US-China semiconductor friction reaches Recursion’s operating environment through GPU availability and cost. Export controls that constrain Chinese competitors also tighten upstream chip supply generally, likely raising Recursion’s own compute costs. Overall: a medium-term cost pressure, not an existential threat.

Open Questions

1. What does it actually mean that Recursion’s platform “demonstrates” the clinical translation gap? This could reflect published research that characterized the problem, or it could reflect clinical programs that actually failed. The distinction matters a great deal: one positions Recursion as a scientific contributor diagnosing the problem, the other as a cautionary tale. The research doesn’t resolve which.

2. How is the Exscientia integration actually going? The 2025 merger is described as expanding the platform, but nothing in the research addresses integration risk or realized synergies. Failed M&A integration — harmonizing model architectures and data pipelines — is a common failure mode in this space that simply isn’t captured here.

3. Is the GLP-1 connection real ownership or just platform exposure? Three connections each to the GLP-1 feedback loop and market-opportunity concepts show relevance, not ownership. Whether Recursion has actual clinical-stage GLP-1 programs, or is simply a platform pharma partners use to find GLP-1 indications, isn’t something the research settles.

4. How does Recursion actually make money from its data? The proprietary-data-moat thesis implies licensing value independent of Recursion’s own drug pipeline, but nothing in the research indicates whether Recursion licenses its phenomics data, earns milestone payments from partners, or uses it purely for internal discovery. This is the biggest open financial question the research leaves unanswered.

5. How far along is Xaira, really? Xaira is described as “building toward” the virtual-cell endpoint — language that implies an in-progress trajectory, not an achieved capability. Without a clearer timeline, it’s impossible to say how much of a head start Recursion’s earlier phenomics data actually buys it.