Allen Institute brings $152M cluster online
- Ai2 turned on the first live compute from its $152 million NSF-Nvidia science AI project, making the OMAI system operational on May 7. - The cluster runs Nvidia Blackwell Ultra hardware near Austin, managed by Cirrascale, and is meant to build fully open models others can inspect. - That matters because academic labs usually cannot afford frontier AI compute — and closed systems rarely share the costly exploratory work.
Compute is the story here. Not a flashy chatbot launch, not a new consumer app — a big shared machine for training scientific AI models. Ai2, the Allen Institute for AI, said on May 7 that the first compute from its Open Multimodal AI Infrastructure for Science project is now live. That matters because a lot of cutting-edge AI research has drifted behind corporate walls, where universities can’t afford the hardware and can’t inspect the full training process anyway. (allenai.org) ### What actually came online? Ai2 brought online a new Nvidia-powered cluster as part of OMAI — short for Open Multimodal AI Infrastructure for Science. The project was announced in August 2025 with $75 million from the National Science Foundation and $77 million from Nvidia, for a combined $152 million. Ai2 leads it, with academic partners including the University of Washingt(allenai.org)mpshire, and the University of New Mexico. (allenai.org) ### Where is this system, exactly? The hardware sits outside Austin, Texas, and it runs on Nvidia Blackwell Ultra systems. Cirrascale Cloud Services manages the infrastructure, which tells you this is not just a lab toy in a back room — it is meant to operate like serious shared research infrastructure. Ai2’s earlier announcement also named Supermicro as a platform supplier for the broader buildout. (geekwire.com) ### Why does “open” matter so much? Because in science, “you can use the model” is not the same as “you can verify the work.” Ai2’s pitch is that it will release not just model weights, but the data, code, evaluations, and training methods needed to analyze, modify, fine-tune, or even retrain systems from scratch. Ba(geekwire.com)e. (allenai.org) ### What problem is this trying to fix? The bottleneck is compute. Training serious multimodal models now costs far more than most university labs or federally funded researchers can handle. That pushes frontier work toward a small number of companies with giant GPU budgets. NSF framed the partnership as a way to give U.S. researchers access to advanced AI models and infrastructure for ma(allenai.org)ithout depending entirely on closed commercial systems. (nsf.gov) ### Why not just use commercial AI tools? You can — but the catch is that closed systems waste value from a science point of view. Ai2 argues that a lot of training effort goes into experiments, false starts, and intermediate checkpoints that never leave the company. It says some training runs spend 82% of effort on exploratory work rather than th(nsf.gov) something, which turns each GPU hour into more than one product. (allenai.org) ### Has the project produced anything yet? Yes. Ai2 said work supported by the project has already fed into upgrades to its OLMo and Molmo model families, including Molmo 2 for video understanding and OLMo Hybrid, a more compute-efficient language model architecture. So the cluster is not arriving before the research plan exists — it is landing into a pipeline that is already producing artifacts. (geekwire.com) ### Why is Nvidia part of this? Because open models still need very expensive hardware. Nvidia is contributing infrastructure as well as money, and the system uses its Blackwell Ultra generation. The bigger point is that this is a public-private bet: federal money sets the scientific mission, and Nvidia supplies the kind of compute stack that universities usually cannot assemble on their own. (nsf.gov) ### Bottom line? Ai2 did not just announce a plan — it switched the machine on. If OMAI works, more of the most expensive part of AI research could happen in public, where other scientists can actually build on it instead of just renting access to the result. (allenai.org)
