By ByteBot
NVIDIA just committed $5 billion to a startup with no products, no published papers, and a $32 billion valuation built almost entirely on one researcher’s reputation. That startup is Safe Superintelligence. In return, NVIDIA gets access to SSI’s closely guarded research — and SSI gets a 10x compute upgrade via NVIDIA’s next-generation Vera Rubin platform. It is a deal that tells you more about where AI infrastructure is heading than almost anything else announced this year.
What SSI Actually Is
Ilya Sutskever cofounded Safe Superintelligence Inc. in June 2024, less than a month after leaving OpenAI. He brought along Daniel Levy (AI researcher) and Daniel Gross (former Apple AI head). The mission is blunt: build safe superintelligence, nothing else. No interim APIs, no consumer products, no enterprise deals until the primary goal is reached.
Two years later, SSI has raised over $6 billion, reached a $32 billion valuation, and published exactly zero models and zero papers. The valuation rests on Sutskever’s track record — AlexNet, sequence-to-sequence learning, the GPT lineage, OpenAI’s o1 reasoning systems — plus undisclosed internal research milestones that nobody outside the lab has seen. It is the most expensive bet on a single researcher’s judgment in tech history.
The Deal Structure
The official announcement describes a “long-term strategic partnership.” Bloomberg reported the figure at $5 billion. What each side gets matters more than the number.
SSI gets compute. Specifically, access to NVIDIA’s Vera Rubin platform — the next generation after Blackwell — which expands their capacity by an order of magnitude. Sutskever put it plainly: “We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so.” SSI had previously been running on Google Cloud TPUs. That relationship is now over.
NVIDIA gets something harder to price: a window into Sutskever’s research before it becomes public knowledge. The partnership includes collaboration on NVIDIA’s future compute platforms, using SSI’s “unique insights into the future of AI.” Jensen Huang called it excitement about discovery. Analysts call it purchasing pre-publication intelligence on where AI bottlenecks are heading next.
What Vera Rubin Is
Vera Rubin is NVIDIA’s post-Blackwell rack-scale compute system. The full platform spans seven chip types across five rack configurations. The flagship NVL72 packs 72 R100 GPUs and 36 Vera CPUs into a single liquid-cooled rack, delivering 3.6 EFLOPS of inference compute — five times faster than Blackwell at one-tenth the per-token cost.
The R100 GPU runs on TSMC 3nm, carries 336 billion transistors (1.6x Blackwell), and ships with 288GB HBM4 memory per chip. Vera Rubin NVL72 enters partner delivery in H2 2026. SSI gets it now.
For SSI, 10x more compute is not incremental. It is the difference between validating a hypothesis and running the experiment at scale. Whatever SSI has been building quietly for two years, they can now build far more of it.
The Infrastructure Angle
NVIDIA controls roughly 80% of the AI GPU market. Its CUDA platform creates developer lock-in at the software layer that is independent of any hardware comparison. Google is pushing TPUs. Anthropic committed to 1 million Google Ironwood chips. Microsoft and Amazon are diversifying to Trainium and custom silicon. Every major AI lab is quietly trying to reduce its NVIDIA dependency.
SSI is doing the opposite. They are coupling their entire infrastructure to one vendor at the exact moment the rest of the industry is hedging. That is either a strong signal about Vera Rubin’s technical edge, or it is the cost of accessing $5 billion in funding you could not get elsewhere.
The more consequential part of this deal is what NVIDIA gains structurally. Competing silicon vendors design hardware against published benchmarks and released models. NVIDIA is now designing against whatever Sutskever is working on before it is published. That is a compounding advantage that no benchmark comparison can capture.
What to Watch
SSI has made no commitment on whether its outputs will be open or closed. If they eventually publish models, will developers have access? Will they be licensed? Restricted to NVIDIA infrastructure? No answers yet.
What is already clear: AI infrastructure is consolidating faster than the model layer. The compute tier has established winners, and those winners are now locking in the research talent capable of shaping the next hardware generation. For developers building on AI, the access terms on that compute layer matter — and so does knowing who controls it.
This deal is not NVIDIA giving SSI faster GPUs. It is NVIDIA buying a structural position at the frontier of AI research and making sure that position deepens with every Vera Rubin rack SSI deploys.
