Miles Wang Disputes Reported $2 Billion Terms for Planned AI Drug Startup

• Miles Wang is reportedly leaving OpenAI to form an AI drug-discovery company with several researchers expected to join him.
• Wang disputes both the reported $200 million raise at a $2 billion valuation and the description of the venture; negotiations remain unfinished.
• His OpenAI work offers relevant wet-lab evidence, but not yet a disclosed drug platform, therapeutic pipeline or clinical operation.
• Billion-dollar launch funding has precedent, while established rivals already claim pharma adoption, data, therapeutic programs or clinical capabilities.

The consequential fact is not a completed $2 billion valuation. It is that investors are reportedly discussing one before Wang has publicly defined the company—and that Wang says the figures and business description are wrong.

The headline terms remain contested

Wang is leaving OpenAI to start a company developing AI models for drug discovery, according to a report based on four people with knowledge of his plans. Several other OpenAI researchers are expected to join, those people said.

Two of the people said Wang was in talks to raise about $200 million at a $2 billion valuation, with Lightspeed discussing a lead role. The negotiations were ongoing and could change. Wang disputed both the financing figures and the description of the company but did not provide alternatives; Lightspeed did not comment.

That attribution chain matters. One follow-on account and another summary trace the financing and strategy back to the original report; they do not independently settle Wang's objections. The retained reporting identifies no disclosed company name, product, completed round or therapeutic pipeline.

The proposed strategy is also tentative. A couple of the original report's sources said the startup may build models to find new uses for approved drugs and possibly candidates that failed in earlier trials. Those are not equivalent pools: an approved medicine comes with an established safety profile, while a failed candidate's evidence and reason for failure would have to be assessed case by case.

Wang has laboratory evidence, not yet a drug operation

Wang left Harvard's computer-science program and joined OpenAI's reinforcement-learning team in March 2024, according to the retained accounts. His work has included alignment and evaluations as well as AI for science. That makes him more than a generic software founder entering biotech, but it does not by itself demonstrate the capabilities needed to develop medicines.

The most concrete evidence is a molecular-cloning study Wang co-authored. OpenAI said GPT-5 proposed protocol changes over repeated experimental rounds while scientists performed the physical work and returned the results. The final combination yielded 79 times as many sequence-verified clones as the baseline for a fixed amount of input DNA.

The scope is narrower than the headline number suggests. The improved assembly reaction delivered a 2.6-fold gain and a separate transformation procedure delivered a 36-fold gain; together they produced the 79-fold result. OpenAI said the improvement was specific to its model system, still required scientists to set up and run protocols, and was not expected to displace the simpler, robust cloning method for most uses.

The study also shows a possible route around the physical-execution constraint. A robotic system reproduced the ranking of a baseline protocol and an AI-modified protocol, achieving 89% of the human-performed relative improvement. But its absolute colony counts were about tenfold lower than manual execution. That is promising automation evidence and a reminder that laboratory reliability, not just model reasoning, will determine whether the approach scales.

A later industry analysis framed the same mismatch more broadly: Wang's background is in reinforcement learning and evaluation, while much drug-discovery work relies on specialized methods such as diffusion models, graph neural networks, protein-structure prediction and molecular dynamics. It also identified experience with regulators, pharmaceutical companies and contract research organizations as capabilities a new team would need to add.

Large checks already fund broader systems

The reported terms are aggressive, but large launch rounds and multibillion-dollar financings already exist in AI-led biotechnology. The more revealing comparison is what was attached to the capital.

These are imperfect comparisons. Wang's figure is a reported valuation under negotiation, while the others are announced funding events. Chai and Isomorphic describe their own adoption, technical progress and pipelines, rather than providing independent proof of clinical success. Xaira shows that investors will fund an AI-biotech platform at enormous scale from inception, but it also shows the operating apparatus that can accompany such a launch.

The comparison therefore limits two easy conclusions. The reported $2 billion price is not unprecedented for the category, and Wang is not being valued against a blank competitive field. Rivals say they are using capital to expand compute, data, drug-industry relationships, therapeutic programs and clinical talent. Wang's venture has not yet disclosed which of those layers it intends to own.

Repurposing changes the risk, not the need for proof

Starting with approved drugs could avoid some costs and time associated with a molecule designed from scratch. The FDA says existing knowledge, including a drug's safety profile, can help accelerate the availability of treatments.

But a new use must still be supported by safety and effectiveness data. The agency is separately examining cases where scientific evidence may support a new use but limited commercial incentives discourage anyone from pursuing approval. That makes indication selection, evidence ownership, trial sponsorship and the path to updated labeling part of the economics of repurposing.

Nor does faster candidate selection remove the clinical bottleneck. An analysis of AI in pharmaceutical development noted that annual FDA drug approvals had remained at about 50 through the AI boom. Formation Bio CEO Ben Liu argued that trials—not discovery—are the limiting factor because they can take years and cost hundreds of millions of dollars.

Formation offers a useful alternative deployment model. The company says AI can cut as much as half the time from administrative and analytical work such as patient recruitment, regulatory filings and matching drugs to diseases. It does not claim to shorten the treatment period when a drug is actually tested on patients. That distinction is the central economic constraint for any company promising that better models will bring medicines to market faster.

The next evidence must define the company

The first unresolved event is the financing itself: whether a round closes, at what size and valuation, and with which lead investor. Even a completed round would establish investor demand, not a functioning drug platform.

The disclosures that would test the business are more concrete:

• an official company description that resolves whether drug repurposing is actually the initial strategy;
• named leaders with drug-development, regulatory, clinical and commercial experience alongside the AI researchers;
• access to biological and clinical data, plus reproducible wet-lab capacity for testing model output;
• a therapeutic program or partner that shows who will fund trials and control development, evidence and commercialization.

Until then, the $2 billion figure remains a disputed negotiating term attached to an undisclosed plan. The meaningful next question is not whether Wang can produce an impressive model result; the cloning study already shows that possibility. It is whether the new company can assemble and control the expensive system that turns such results into clinical evidence.