AI for Science across Big Tech: A 2025–26 Landscape
"AI for Science" covers so much ground now that the label barely tells you how differently each lab actually works. This post is a map: it sorts the major players by what they're really selling — a model, a drug, or an infrastructure layer — and links the flagship work behind each. Treat it as a hub; papers I read closely get their own posts, linked at the bottom.
00Three archetypes
Build the models
DeepMind, Microsoft, Meta FAIR, ByteDance — structure prediction, generation, materials, and climate, shipped open or semi-open.
Make the drugs
D. E. Shaw, Xaira, Isomorphic — physics simulation or large wet-lab loops that carry candidates all the way to the clinic.
Lay the floor
NVIDIA BioNeMo — serving everyone else's models on GPUs, competing on the stack rather than the benchmark.
01Google DeepMind & Isomorphic Labs
Google DeepMind & Isomorphic Labs
protein · genome · materials · climate- – AlphaFold 3 — complexes of proteins with DNA, RNA, ligands, and ions (a diffusion module replaces AF2's structure module)
- – AlphaProteo — de novo binder design, 3–300× better affinity across seven targets
- – AlphaMissense for coding variants and AlphaGenome for the non-coding genome (1 Mb in, regulatory tracks out)
- – GNoME for crystals; GraphCast / GenCast for weather
DeepMind treats AlphaFold as a template for bringing all of science "to digital speed." The commercial side is walled off: AlphaFold Server is free for academics, but the rights to use AlphaFold 3 commercially belong to Isomorphic Labs, its drug-discovery spinout, now working with Eli Lilly and Novartis.
02Microsoft Research
Microsoft Research — AI for Science
materials · DFT · protein · climate- – MatterGen — generates crystals conditioned on a target property; asked for 200 GPa bulk modulus, it produced TaCr₂O₆, later measured at 169 GPa
- – MatterSim and Skala — a universal interatomic potential, and a learned exchange–correlation functional for DFT itself
- – BioEmu — protein equilibrium ensembles, thousands per hour on a single GPU
- – Aurora — an atmospheric foundation model spanning weather, air quality, and ocean waves
It spans more domains than any other lab, and ships most of it openly. The stated direction for 2026 is agentic "AI lab assistants" that propose hypotheses and help run the experiments.
03Meta FAIR Chemistry
Meta FAIR — Chemistry
molecules · materials · catalysisThe clearest thesis in the field: scale the data, then train one universal atomic model on top of it. Because the releases are open — commercial use included — academic groups are shifting from training their own models to fine-tuning or distilling these.
04NVIDIA — BioNeMo & Clara
NVIDIA — BioNeMo & Clara
platform · drug discovery- – NIM microservices that serve models like AlphaFold2, ESM-2, DiffDock, and MolMIM as APIs
- – Blueprints — reference workflows for multi-step generative pipelines
- – Evo 2, a genome foundation model built with the Arc Institute
A picks-and-shovels play. Rather than competing on benchmarks, NVIDIA sells the lab-in-the-loop flywheel — predict, synthesize, measure, retrain — and keeps the whole ecosystem running on its GPUs.
05D. E. Shaw Research
D. E. Shaw Research (DESRES)
physics-first · molecular dynamics- – Anton 3 — a special-purpose MD supercomputer, roughly 100× faster than general-purpose hardware
- – Physics as a data source: MD trajectories now train ML models such as diffusion cofolders
- – Several candidates have reached clinical trials
The clearest case of a physics-first approach reaching the clinic. Simulation is used to get around how scarce and expensive experimental data is — though here physics generates the data rather than acting as a constraint inside the learning objective.
06Xaira Therapeutics
Xaira Therapeutics
virtual cell · protein design- – X-Atlas/Orion — one of the largest public Perturb-seq datasets, spanning every protein-coding gene
- – X-Cell — a multi-billion-parameter diffusion model for perturbation response, with zero-shot transfer to unseen cell types
- – Antibody and binder design in the RFdiffusion / ProteinMPNN lineage
Spun out in 2024 from the David Baker / IPD lineage. It folds protein design, perturbation biology, foundation models, and wet-lab into one loop, and bets that biology has scaling laws too.
07ByteDance Seed
ByteDance Seed — AI for Science
protein · open · reproducibleOpen, reproducible, benchmark-driven. Together with Tencent and DP Technology (Uni-Mol), it's part of a broad Chinese effort to compete with Western labs through open foundation models.
08Side by side
09Cross-cutting trends
Task-specific models are giving way to universal ones — UMA, Evo 2, virtual cells.
Proprietary data at scale (OMol25's 100M calculations, genome-wide Perturb-seq) is becoming the real barrier to entry.
Predict, synthesize, measure, retrain — where NVIDIA, Xaira, and MSR all converge.
AI "scientists" that propose hypotheses and drive experiments — the theme of 2026.
10Open questions
11Deep dives
As I write closer reads of individual papers, they'll show up here:
- Evo 2 — a DNA foundation model that reads and writes genomes across all of life read →
- – AlphaGenome — architecture and regulatory-track prediction (coming soon)
- – MatterGen — property-conditioned diffusion for crystals (coming soon)
- – BioEmu — emulating equilibrium ensembles (coming soon)
A living landscape post; details and links are current as of mid-2026 and will be updated.
