Coming up: Sat 31 Oct · Frontier Paper Club: DeepSeek-R1: reasoning learned through reinforcement learning; Wed 4 Nov · Build Night: Agents for Science: Build a literature agent that cites its sources; Sat 7 Nov · Frontier Paper Club: AI Co-Scientist: multi-agent hypothesis generation; Sat 14 Nov · Frontier Paper Club: Language agents that synthesise scientific knowledge; Sun 15 Nov · GPU 101: From Notebook to Accelerator: Profile a PyTorch training loop on a free GPU; Wed 18 Nov · Build Night: Agents for Science: Score genetic variants with Evo 2
What changed in compute, models, and AI for science over the last two years, and why each change matters for a research lab. Summaries of public announcements, written by the team.
Latest on the board
Jan 2026
Vera Rubin enters production
NVIDIA claims up to 10x lower inference cost per token than Blackwell. Cheaper tokens are what make always-on research agents affordable.
14 milestones shown
At CES 2026, NVIDIA said its Vera Rubin platform, the successor to Blackwell pairing Rubin GPUs with Vera CPUs, was in full production, with partner systems arriving in the second half of 2026.
Why it mattersNVIDIA claims up to 10x lower inference cost per token than Blackwell. Cheaper tokens are what make always-on research agents affordable.
NVIDIA NewsroomGoogle released Gemini 3, pushing reasoning and multimodal understanding across text, images, video, and code.
Why it mattersScience is multimodal (figures, gels, scans, sequences), and models now read them together.
GoogleOpenAI released GPT-5, which routes between fast and deep-reasoning modes, and in the same month published gpt-oss, its first open-weight models since GPT-2.
Why it mattersBoth closed and open options now offer reasoning, so labs can choose based on privacy, cost, and control.
OpenAIAnthropic released Claude Opus 4 and Sonnet 4, built for long-running coding and agent tasks with tool use, and later the 4.5 generation in the autumn.
Why it mattersAgents that can keep working on a task for hours are the building block for automated literature review and analysis pipelines.
AnthropicUnder the IndiaAI Mission, Sarvam AI was selected to build an Indian foundation model, the first of several teams chosen to develop models for Indian languages.
Why it mattersModels that understand Telugu and other Indian languages make AI usable for far more people in our region.
IndiaAIAt GTC 2025, NVIDIA announced Blackwell Ultra, the Dynamo inference software, and a public roadmap: Vera Rubin in 2026, Rubin Ultra in 2027, then Feynman.
Why it mattersInference, not just training, became the headline workload as reasoning models and agents consume far more tokens per answer.
GTC keynoteArc Institute and NVIDIA released Evo 2, an open model trained on trillions of DNA bases from across the tree of life, able to score the likely effect of genetic variants.
Why it mattersVariant effect prediction from sequence alone feeds directly into our genomic variant impact research track.
Evo 2 on GitHubAt CES 2025, NVIDIA announced Project DIGITS (later DGX Spark), a desktop machine built on the GB10 Grace Blackwell chip, alongside Cosmos world foundation models for robotics.
Why it mattersSerious model development started to fit on a desk, which matters for small labs that cannot rent clusters for every experiment.
NVIDIA NewsroomDeepSeek released R1, an open-weights reasoning model under the MIT licence that performed close to leading closed models, along with a paper describing how it was trained.
Why it mattersStrong reasoning became something any lab could download, study, and fine-tune, and it reset assumptions about how much compute frontier results require.
arXiv paperThe Chemistry prize went to David Baker, Demis Hassabis, and John Jumper for protein design and structure prediction; the Physics prize to John Hopfield and Geoffrey Hinton for foundations of neural networks.
Why it mattersAI for science moved from promising to prize-winning, and funding and talent followed.
nobelprize.orgOpenAI released o1, a model trained to spend more computation “thinking” through a problem before answering, with large gains on maths and science benchmarks.
Why it mattersTest-time reasoning became a second scaling axis next to model size, and it is what makes multi-step research agents practical.
OpenAIGoogle DeepMind and Isomorphic Labs published AlphaFold 3, which predicts the joint structure of proteins with DNA, RNA, ligands, and ions.
Why it mattersMoving from single proteins to complexes brings structure prediction much closer to drug design, the heart of our computational biology pillar.
Nature paperThe Government of India approved the IndiaAI Mission with an outlay of over ₹10,000 crore, including shared GPU compute for researchers and startups.
Why it mattersSubsidised national compute lowers the barrier for labs outside the metros, including ours, to train and test models.
IndiaAIAt GTC 2024, NVIDIA announced Blackwell (B200) GPUs and the GB200 NVL72 rack, which links 72 GPUs into one large accelerator for training and serving very large models.
Why it mattersRack-scale systems changed the unit of AI compute from a single chip to a whole rack, and set the pace for frontier model training over the next two years.
NVIDIA NewsroomGet the next milestone explained in your inbox, every two weeks.
Subscribe free