When compute is locked behind corporate walls, the work stops.
In a moment when a handful of corporations have quietly come to govern which scientific questions get asked and which go unasked, the Allen Institute for AI has brought online a federally backed computing cluster designed to return that power to the broader research community. The $152 million NSF OMAI initiative, drawing on NVIDIA's latest hardware and partnerships with universities across the country, is less a technical announcement than a philosophical one: that openness, reproducibility, and shared infrastructure are not handicaps in the race for AI progress, but its most durable foundations. The models already emerging from this cluster — leaner, more efficient, and outperforming systems many times their size — suggest the argument may be more than idealism.
- A small number of corporations now control the compute that shapes which AI research gets done, creating a structural imbalance that distorts science itself.
- The NSF OMAI cluster went live last week, injecting $152 million in federally backed infrastructure into the open research ecosystem and directly challenging that concentration.
- Early results are striking: an 8-billion-parameter model is outperforming its 72-billion-parameter predecessor, and a language model is matching prior benchmarks at half the training cost.
- Universities in Hawaii, New Hampshire, New Mexico, and Washington are already woven into the effort, broadening the geography of who gets to participate in frontier AI development.
- The deeper wager is whether open, transparent, reproducible infrastructure can hold its own against closed systems — and whether the U.S. can lead not by locking knowledge away, but by sharing it.
The Allen Institute for AI brought a new piece of infrastructure online last week — a $152 million federally backed computing cluster called the Open Multimodal AI Infrastructure for Science, or NSF OMAI. Powered by NVIDIA's Blackwell Ultra GPUs and supported by the National Science Foundation, it was built around a premise that feels almost countercultural in 2026: that serious AI development shouldn't require membership in an exclusive club of three corporations.
The concentration of compute among a few companies has real consequences. It shapes which models get built, which research gets accelerated, and which questions simply don't get asked. Ai2 won the NSF funding last August to lead a different kind of effort — one that brings in the University of Hawaii at Hilo, the University of New Hampshire, the University of New Mexico, and the University of Washington, and keeps its models open for inspection, adaptation, and reuse. Principal investigator Noah A. Smith framed the distinction simply: closed compute produces one product and stops; open infrastructure keeps generating value as others build on it.
The cluster is already producing evidence for that argument. Molmo 2 added video understanding and object tracking, with an 8-billion-parameter version outperforming the original 72-billion-parameter model on key benchmarks. A follow-up model replaced text-coordinate outputs with a token-based grounding system and reached state-of-the-art accuracy on spatial reasoning. On the language side, Olmo Hybrid matched previous models while using roughly half the training data — not incremental progress, but a demonstration of what open experimentation makes possible.
NSF's Wendy Nilsen described the investment as being fundamentally about replicability and transparency — the conditions that make science trustworthy. When models are open, other teams can verify them and understand their limits. That strengthens the whole field, and it positions the U.S. not as the country hoarding the most powerful proprietary systems, but as the one building the infrastructure others can innovate on. Cirrascale Cloud Services is managing deployment, the machinery is in place, and the question now is what the broader research community will do with it.
The Allen Institute for AI flipped a switch last week, and suddenly there was a new way to do things. The cluster came online Thursday—a $152 million piece of federally backed infrastructure called the Open Multimodal AI Infrastructure for Science, or NSF OMAI. It's powered by NVIDIA's Blackwell Ultra GPUs and backed by the National Science Foundation. And it exists for a reason that feels almost quaint in 2026: to prove that you don't need to be one of three companies to build serious AI systems.
The concentration problem is real. A handful of corporations now control most of the compute that matters, which means they control which questions get asked, which models get built, which research gets accelerated. The White House noticed. So did the NSF. Last August, Ai2 won the funding to lead this project, and now the infrastructure is live, ready to serve researchers across materials science, biology, energy—fields where open models and shared resources could actually change what's possible. The University of Hawaii at Hilo, the University of New Hampshire, the University of New Mexico, and the University of Washington are all part of the effort.
Noah A. Smith, the principal investigator at NSF OMAI and senior research director at Ai2, put it plainly: when compute is locked behind corporate walls, you get one product and then the work stops. When it's open, the value keeps flowing. Scientists can inspect it, adapt it, build on it. That's the philosophy driving this cluster. It's not about having the biggest machine in the room. It's about how efficiently the capacity gets used and shared.
The infrastructure is already producing results. Molmo 2, a multimodal model from Ai2, added video understanding and object tracking to the family—and an 8-billion-parameter version outperformed the original 72-billion-parameter model on key benchmarks. That's the kind of efficiency gain that matters. A follow-up called MolmoPoint replaced text-coordinate outputs with a token-based grounding mechanism and hit state-of-the-art accuracy on spatial reasoning tasks. On the language side, Olmo Hybrid combined transformer attention with linear RNN layers, matching prior models while using roughly half the training data. These aren't just incremental tweaks. They're demonstrations of what becomes possible when you have the infrastructure to experiment openly.
Wendy Nilsen, the deputy directorate head for NSF's computer and information science and engineering directorate, framed it as an investment in replicability and transparency—the things that make science actually work. When models are open, other teams can verify them, reproduce them, understand why they work or don't. That strengthens the entire field. It also positions the U.S. differently in the global race for AI leadership, not as the country with the most proprietary models, but as the one building the infrastructure for everyone else to innovate on.
Cirrascale Cloud Services is handling the deployment and management. Ai2 already works with them to distribute Olmo, Molmo, and Tulu. So the machinery is in place. The cluster is live. And now the question becomes what researchers will build with it—and whether this model of open, shared infrastructure can actually compete with the closed systems that have dominated the last few years.
Citas Notables
At a time when access to advanced AI systems is increasingly concentrated among a small number of companies, bringing this hardware infrastructure online represents a critical step for us.— Noah A. Smith, principal investigator at NSF OMAI and senior research director at Ai2
By investing in open, shared resources, we are enabling scientists and researchers across disciplines to build, test, reproduce, validate, and advance AI systems.— Wendy Nilsen, deputy directorate head for NSF's computer and information science and engineering directorate
La Conversación del Hearth Otra perspectiva de la historia
Why does it matter that this cluster is federally backed rather than built by a private company?
Because a private company builds for profit. They train a model, release it, and move on. The NSF is building for science—for the long-term value of having infrastructure that researchers can use, modify, and build on. That's a fundamentally different incentive structure.
But doesn't NVIDIA benefit from this? They're providing the GPUs.
They do, yes. But the difference is that the models and the research outputs are open. NVIDIA gets hardware sales; researchers get access to tools they can actually study and improve. It's not zero-sum.
The narrative mentions that an 8-billion-parameter model outperformed a 72-billion-parameter one. How is that possible?
Better architecture, better training data, better techniques. It's not just about scale anymore. If you're smarter about how you build the model, you can do more with less. That's what open research enables—you can see what works and iterate on it.
What happens if this cluster becomes the standard? Does that change the power dynamics in AI?
It redistributes them. Right now, three companies control the compute, so they control the narrative. If universities and smaller labs have real infrastructure, they can ask different questions, build different kinds of models. You get more voices in the conversation.
Is this a threat to the companies that currently dominate?
Not directly. They have more resources than this cluster will ever have. But it's a threat to their monopoly on what's possible. And that matters more than you'd think.