AI2 Fires Up $152M Blackwell Cluster for Open Science

The Allen Institute for AI just brought its federally backed compute cluster online, and the first wave of models trained on it's already on Hugging Face with weights, training code, and data all public.

What's In the Rack

The system sits outside Austin and runs on Nvidia's HGX B300 platform - the Blackwell Ultra generation that NVIDIA started shipping in January 2026. Each B300 SXM GPU carries 288 GB of HBM3e memory, 8 TB/s of memory bandwidth, and 15 petaFLOPS of dense FP4 compute. Nodes are air-cooled to liquid, eight GPUs per baseboard, with 800 Gb/s of external connectivity via ConnectX-8 SuperNICs.

The Hardware Stack

Cirrascale Cloud Services provisions and manages the infrastructure. Cirrascale already ran compute for AI2's earlier Olmo, Molmo, and Tulu model families, so this isn't a greenfield buildout - it's a significant scale-up of an existing relationship. The cluster location outside Austin puts it in a region with competitive energy pricing and proximity to Cirrascale's existing AI Innovation Cloud footprint.

Nvidia is contributing $77 million of the $152 million total, with the other $75 million coming from the National Science Foundation through the OMAI cooperative agreement awarded in August 2025 as part of the White House AI Action Plan. Four universities join AI2 as research partners: University of Washington, University of Hawai'i at Hilo, University of New Hampshire, and University of New Mexico.

What Makes This Different From Renting

The usual path for academic AI research is to apply for credits on AWS, Azure, or Google Cloud, wait weeks for approval, and then work within token budgets that don't survive the training runs of serious foundation models. OMAI is meant to fix that structural problem by giving researchers dedicated cluster access for the life of the project rather than metered cloud time.

AI2's own analysis found that 82% of a typical training run goes toward exploratory work rather than the final model. When that exploratory work is done privately and the only artifact released is the final checkpoint, every downstream lab pays that cost again from scratch. Open infrastructure changes the arithmetic.

High-density GPU clusters like the OMAI system pack Blackwell Ultra nodes with direct liquid cooling. Source: unsplash.com

What the Cluster Is Already Running

The compute didn't sit idle after it came online. AI2 says research supported by the project has already produced upgrades across three model families.

OLMo 3 - The Language Foundation

The OLMo 3 family delivers what AI2 is calling the strongest fully open base and thinking models to date. OLMo 3-Base 32B leads among fully open base models, and OLMo 3-Think 32B is the strongest fully open thinking model on AI2's OLMES evaluation suite, which covers 20 benchmarks across knowledge recall, commonsense reasoning, and mathematical reasoning. OLMo 2 32B was the first fully open model to beat GPT-3.5 Turbo and GPT-4o mini on a suite of academic benchmarks without any proprietary data or closed training runs.

Molmo 2 - Video and Spatial Grounding

The efficiency numbers here are what matter most. Molmo 2's 8-billion-parameter version beats the original Molmo 72B on pointing and grounding benchmarks. That's a 9x parameter reduction with better task performance - the kind of number that gets infrastructure engineers to actually pay attention. In head-to-head tests on certain video understanding tasks, Molmo 2 performs competitively with Google's Gemini 3. The embodied reasoning variant, Molmo 2-ER, scores 63.8 out of 100 across 13 benchmarks, ahead of GPT-5, Gemini 2.5 Pro, and Qwen3-VL-8B.

MolmoAct 2, released May 5, extends the Molmo family into robotics - a foundation model for real-world robot manipulation tasks, again fully open. AI2's MolmoWeb showed this direction earlier; MolmoAct 2 is the physical-world version of that bet.

You can pull any of these models directly from Hugging Face:

# OLMo 3 - 32B thinking model
pip install transformers accelerate
python -c "
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained('allenai/OLMo-3-Think-32B')
model = AutoModelForCausalLM.from_pretrained('allenai/OLMo-3-Think-32B', device_map='auto')
"

Running the full 32B model requires either multiple A100s or a single H100/B200. The 7B variant runs on a single 24 GB GPU.

AI2 releases the full research stack - not just model weights but training data, code, and evaluation tools. Source: unsplash.com

The Infrastructure Access Table

Why the "Fully Open" Claim Holds Up

The phrase "open source" has been so abused by AI labs that it barely means anything anymore. Meta releases weights but not training data. Mistral releases weights but withholds the instruction-tuning details. Most "open" models are open in the same way a compiled binary is open if you can read the instruction set.

OMAI's commitment is different: weights, training data, code, and documentation all released. The distinction matters for reproducibility. A researcher who wants to freeze a model version for a long-running study, modify the architecture for a specific biology dataset, or audit training decisions for safety purposes can do all three. With a weights-only release, they can only do the first.

Jensen Huang framed the partnership in explicit terms: "AI is the engine of modern science - and large, open models for America's researchers will ignite the next industrial revolution." That's a marketing statement, but the release practice backs it up. AI2 has shipped every training data artifact for every model in this family. That isn't the norm.

Where It Falls Short

The GPU count isn't disclosed. AI2 confirmed Blackwell Ultra hardware and Cirrascale as manager, but the total node count, aggregate compute budget in FLOPs, and cluster topology haven't been published. For a project explicitly about transparent AI infrastructure, that's a gap.

The four partner universities get dedicated allocations, but the broader research community doesn't. Access for external researchers runs through the Tulu API or model downloads - both legitimate but not equivalent to cluster access. A grad student at a non-partner institution trying to train a domain-specific variant of OLMo 3 on specialized scientific data still faces the same compute wall as before this project launched.

The timeline for the "initial releases later in 2026" language in NVIDIA and NSF announcements is also vague. The cluster is running, and OMAI-supported model updates are shipping, but the explicit multimodal scientific models - the ones for materials science, biology, and energy research that the project was funded to build - don't have a firm ship date yet.

None of these gaps make the project less significant. They make it less complete than the announcement language suggests.

The practical picture: AI2 has a Blackwell Ultra cluster running in Austin, and it's already producing models that beat larger predecessors on specific benchmarks with full training transparency. That's a higher standard than almost any corporate open-source release in the last two years. Whether the broader scientific community gets meaningful access beyond model downloads depends on what partner expansion and external allocation programs look like in the next six months.

Sources:

• Open by design: Ai2 brings fully open AI infrastructure online with NSF OMAI
• AI2 Launches $152M Compute Cluster for Scientific AI
• NVIDIA, National Science Foundation Support Ai2 Development of Open AI Models
• NSF and NVIDIA Spending $152 Million to Build Open AI for Science
• Cirrascale Powers Ai2's $152M Push for Open, Scientific AI
• Molmo 2: State-of-the-art video understanding, pointing, and tracking - Ai2
• MolmoAct 2: An open foundation for robots that work in the real world
• NVIDIA HGX B300 Explained: Specs, Memory and NVLink