NVIDIA Backs Ilya Sutskever's SSI With $5B as Two-Year Stealth Lab Scales 10x
Safe Superintelligence has broken its two-year silence. Ilya Sutskever’s lab, founded in mid-2024 after he left OpenAI, announced a long-term strategic partnership with NVIDIA on July 27, with an approximately $5B commitment attached to the deal. The partnership adds a GPU compute path alongside SSI’s existing Google Cloud TPU arrangement.
SSI’s statement is spare by design: “We reached the point where our research is worth scaling and with this partnership we will be able to.” The company says compute will increase 10x within 12 months.
What Makes This Unusual
SSI has done nothing publicly since it raised $1B in June 2024 at a $5B valuation, then added more capital to reach a $32B post-money valuation in April 2025. No papers. No models. No product. The lab’s entire public signal has been that it exists and that Sutskever, Daniel Gross, and Daniel Levy are running it.
NVIDIA’s willingness to put ~$5B behind a lab with no output is a statement about who Sutskever is as much as what SSI is building. He was OpenAI’s co-founder, chief scientist, and the architect of the training runs behind GPT-2 through GPT-4. He was also part of the board that briefly fired Sam Altman in November 2023.
The Compute Economics
The investment reflects NVIDIA’s standard play in 2026: deploy capital into AI labs that will spend it on NVIDIA hardware. AMD’s $5B deal into Anthropic, CoreWeave’s billion-dollar loan facilities, and NVIDIA’s own $40B equity investment portfolio all follow the same structure: vendor financing that guarantees GPU demand.
For SSI, 10x compute growth on a base that already included Google TPU capacity positions the lab to run the kind of long-horizon training runs that Sutskever has publicly described as necessary for “safe superintelligence”, a research agenda he has characterized as fundamentally different from incrementally scaling RLHF-tuned chat models.
What Comes Next
SSI has still published nothing. The conventional interpretation is that the lab is training a very large model, wants to do it quietly, and is now ready to spend. The less charitable read is that two years of stealth with no verifiable output makes any capability claim impossible to assess.
A valuation of $32B for a lab with no product is only credible if investors believe the team can build something that doesn’t exist yet. NVIDIA’s $5B bet is a vote that the team can.