OpenAI researcher Miles Wang is negotiating the launch of a startup that will leverage artificial intelligence to accelerate drug discovery, with an initial valuation hovering around $2 billion. Lightspeed Venture Partners is in talks to lead the investment round, according to sources close to the matter.

The news comes as generative AI is reshaping the limits of scientific research. Drug discovery, protein folding, molecular design — fields that until a few years ago were the exclusive domain of wet labs and traditional HPC simulations — are now increasingly blending with transformer architectures and language models trained on biological sequences. It is no coincidence that Wang comes from OpenAI: the ability to work with Large Language Models at scale and to adapt them to vertical domains is a rare asset, and the market is pricing it accordingly.

A billion-dollar valuation before the startup even has a public product is not just a bet on Wang as an individual. It is a structural signal: venture capital believes that the prototyping phase of pharmaceutical AI is ending, and that the time has come for massive investments to bring the first “AI-discovered” drugs to clinical trials. Who stands to gain? First and foremost, holders of patents and proprietary data — big pharma companies that collaborate with these startups will access much broader pipelines of candidate molecules, reducing upstream risk. Those who risk falling behind are research centers that lack the computational infrastructure or the talent to compete on this playing field.

There is an aspect that directly touches our observatory on on-premise infrastructure. In pharmaceuticals, the molecular screening data and clinical records on which these models are trained are often subject to sovereignty and confidentiality constraints. Not every company is willing to upload trade secrets to public clouds. A self-hosted deployment — on local clusters with high-VRAM GPUs, possibly in air-gapped configurations — becomes no longer an academic laboratory exception but a business requirement. Demand for specialized hardware for training and inference of LLMs applied to biology could therefore grow asymmetrically: not only from large hyperscalers, but also from private data centers of pharmaceutical and biotech companies.

The choice of Lightspeed as lead investor adds a piece: the fund has already bet on startups operating at the intersection of AI and hard sciences. This round, if confirmed, reinforces the trend of building complete stacks — from base models to therapeutic distribution — while internally controlling the most critical resource: data and compute.

Ultimately, the Wang story is a bellwether for anyone tracking the evolution of AI beyond chatbots. The next generation of deep tech startups is being born with unicorn valuations at the very first round, but it needs solid, often on-premise computing architectures to deliver on its promise. The open question is whether hardware will keep pace, or whether the real bottleneck will become GPU availability for those who do not have the scale of a Google or a Microsoft.