Microsoft Discovery goes GA, with BHP screening half a million molecules for copper

Microsoft Discovery is now generally available on Azure, and Microsoft is positioning it less as a model and more as an R&D operating layer: teams of specialized AI agents that handle literature review, hypothesis generation, simulation, and iterative learning, with hooks into high-performance compute and, per Microsoft, physical labs, robotics, and instrumentation.

The flagship receipt is BHP. Working with Microsoft and computational chemists at Prescience Insilico, BHP says it screened more than 500,000 candidate molecules for copper leaching, running "tens of thousands of quantum chemistry calculations and simulations" before narrowing to a shortlist for wet-lab testing in Australia. That is the shape of the pitch to builders: use agents to triage an intractable search space with quantum chemistry in the loop, then hand a small set of survivors to humans at the bench.

What's actually new for builders

• A platform, not a model. Microsoft Discovery is described as an Azure-hosted system of role-specialized agents, designed to "mimic the scientific method" and break the silos between literature, simulation, and experiment.
• HPC + AI as one surface. The emphasis on quantum chemistry calculations at scale suggests the value proposition is orchestration across model inference and traditional scientific compute, not just LLM calls.
• Lab integration claims. Microsoft says the platform can integrate with lab instrumentation and robotics. The BHP example stops short of describing closed-loop autonomous experimentation; the shortlist still goes to human scientists.

What we don't know

Microsoft hasn't shared, at least in this post, which foundation models sit underneath, pricing, rate limits, what the agent SDK looks like, how customers bring their own simulators, or how data isolation works for proprietary chemistry. There are also no published benchmarks for hit rate versus traditional virtual screening, and no disclosure yet of whether BHP's shortlist has produced a validated reagent. The 500,000-molecule figure is a throughput number, not an outcome.

Still, the strategic read is clear. Microsoft is staking out vertical agentic AI for science as a distinct product category from Copilot, and it wants the early reference customers to be heavy industry — mining, materials, pharma — where the unit economics of shaving years off R&D are obvious. Expect competitive responses from the other hyperscalers and from the specialist players who have been quietly building agentic chemistry stacks for a while. The interesting question for developers is whether Discovery becomes an opinionated workflow you adopt, or a set of primitives you can wire into your own.