Recursion was founded to discover drugs, not to sell foundation models.
The company built automated laboratories, ran biological experiments at industrial scale, and photographed how human cells responded to diseases, genes, and potential treatments. Machine learning helped its scientists turn those images into predictions about which compounds might work.
Alongside drug candidates, that system produced proprietary data, software, and models that captured how Recursion worked.
Now one of those models has a customer. Under an agreement signed this month, Tempus will pay Recursion $12 million for a two-year license to TxFM, a foundation model for RNA-sequencing data. Tempus can use it in oncology research, diagnostics, and clinical applications.
The license points to a larger path for AI-native companies. An enterprise begins by building models to improve its own work, while its operations keep producing proprietary data. As the models become more specialized, useful, and difficult for an outside vendor to reproduce, the system built to run the company can become something it sells to its own industry.
Recursion is an early example of that progression. It remains a drug-discovery company, but its model has become a product.
The model inside the company
Recursion has used machine learning since its earliest experiments. Its founding idea was to replace subjective judgments about whether diseased cells looked healthier with repeatable computational measurements. It built automated laboratories and a continuous learning loop: experiments generated standardized data, models learned from the results, and those models helped choose the next experiments. For years, the model was infrastructure. The commercial output was expected to be a medicine or pharmaceutical partnership.
Foundation models widened that possibility. TxFM grew from Recursion’s work across cellular imaging, chemistry, and gene expression. The public paper describes a version trained on 1.4 million curated public RNA-sequencing samples; Recursion has separately said the model used public and proprietary data.
In 2023, Recursion agreed to pay as much as $160 million for limited access to more than 20 petabytes of Tempus oncology data. This month, it replaced two scheduled $42 million annual fees with three $14 million payments, accepted a lower cap on unique records, and surrendered its right to terminate for convenience. Tempus traded the possibility of larger payments for $42 million of longer-term commitments.
In a separate agreement signed the same day, Tempus licensed TxFM for $12 million and agreed to provide additional de-identified pathology records linked to clinical data. The deal captures why unique datasets are becoming so valuable. Models will become easier to build and distribute, but the proprietary records created inside a company’s daily work, together with the outcomes needed to evaluate them, are far harder to reproduce. That data can give an enterprise a durable advantage in building models tuned to the problems it understands best.
At first, those models will serve the company’s own use cases by lowering costs, improving decisions, or strengthening its core product. Once they perform well inside their workflows, the company can sell access to other organizations facing the same problems but lacking the same data, expertise, or operating history. Recursion’s agreement with Tempus is an early example of that progression. TxFM began as part of Recursion’s drug-discovery machinery; it’s now a product another healthcare company will pay to use. The long-term opportunity is to turn what proprietary data teaches into models sold beyond the walls of the company that trained them.
Partnerships & Deals
Compute, Energy & Infrastructure
Anthropic and Akamai signed a seven-year, $11.6 billion cloud agreement for CPU workloads. The commitment could expand by another $9 billion, while an Akamai warrant issued to Anthropic can vest into as much as roughly 5 percent of the cloud company’s common stock.
NetApp agreed to acquire PEAK:AIO to add independently scalable metadata services and parallel file architecture for AI clouds operating trillions of files and multi-exabyte storage environments.
Sunrun and SPAN expanded their partnership to combine residential solar and battery systems with SPAN’s distributed data-center nodes, targeting behind-the-meter AI compute deployments at gigawatt scale.
Data, Knowledge & Retrieval
TinyFish and 15 data providers launched the Data Partners Alliance, connecting agents to licensed market, company, identity, legal, research, and specialized data through TinyFish’s web infrastructure. Founding members include Alpha Vantage, Databento, Crunchbase, Similarweb, Tracxn, Enigma, OpenAlex, and Trellis Law.
Progress Software completed its $400 million acquisition of substantially all of Domo’s AI and data platform business, adding more than 2,400 customers and technology for connecting, governing, and activating enterprise data.
Models, Training & Developer Tools
Snorkel AI raised a $350 million Series E at a $3.5 billion valuation to expand its training- and evaluation-data business for frontier labs, enterprises, and government agencies.
Ando raised a $20 million seed round from Accel, Index Ventures, and Emergence Capital to launch a messaging platform where people and AI agents share channels, context, identities, and permissions.
Enterprise Deployment & Distribution
BNP Paribas and Google Cloud signed a five-year partnership covering infrastructure, Gemini models, and Gemini Enterprise. Initial plans include integrating Gemini into the bank’s internal LLM@CIB assistant and deploying agents for corporate credit memos and other investment-banking workflows.
Accenture invested in Within and formed a delivery partnership around Within’s system for capturing undocumented processes, exceptions, and workarounds as context for enterprise agents.
Physical AI & Robotics
Cognex agreed to acquire RealSense for approximately $500 million in cash, adding 3D depth cameras and robotic-perception technology used in autonomous mobile robots, industrial automation, quadrupeds, and humanoids. The transaction is expected to close in the fourth quarter.
Science & Healthcare
Enveda raised a $311 million Series E led by Catalio Capital Management to advance three clinical-stage medicines, bring additional programs into trials, and expand its AI platform for finding drug candidates in natural chemistry.
Harell Data and CoreWeave signed a multi-year agreement to run training, fine-tuning, and inference against proprietary biotech datasets without releasing the raw data. Harell says data owners will receive a share of each training run while model builders retain their models and can charge for their use.