Nvidia Bets Billions on Ilya Sutskever's Safe Superintelligence, Ending Its Google TPU Era

For two years, Safe Superintelligence ran quiet. No product, no demo, no published research, no revenue — just a manifesto, a marquee founder in Ilya Sutskever, and a $32 billion valuation that unnerved anyone who tried to price it. On July 27, the silence broke. Nvidia announced a long-term strategic partnership with SSI, paired with an investment that a source familiar with the deal told TechCrunch "stretches into multiple billions" — roughly $5 billion, according to Bloomberg, Reuters and the Financial Times. In one stroke, the world's most valuable chipmaker planted its flag inside the most secretive lab in AI, and SSI signaled a decisive pivot away from the Google TPUs that had powered its early work.

The centerpiece is compute. The deal gives SSI access to Nvidia's forthcoming Vera Rubin GPU platform and is expected to increase the startup's compute resources "by an order of magnitude" — roughly 10x what it commands today. For a lab that has staked everything on a single, uninterrupted push toward aligned superintelligence, that scale is the whole game.

The two quotes that frame the deal

Sutskever, who co-created AlexNet in 2012 and later served as OpenAI's chief scientist, kept his statement characteristically spare. "We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so," he said. "We are confident that our big bet on the Vera Rubin platform will take us to the next level."

Nvidia CEO Jensen Huang framed the investment as a bet on the person as much as the lab. "Ilya has pioneered fundamental breakthroughs at the foundation of modern AI, beginning with AlexNet," Huang said, adding that Nvidia is "excited to see what new breakthroughs SSI will discover powered by our Vera Rubin platform."

Crucially, Nvidia said it committed only after obtaining "rare access into the company's closely guarded research." SSI has shown almost nothing publicly; that Nvidia was allowed to look under the hood — and then wrote a multi-billion-dollar check within weeks — is itself the loudest signal yet that something inside SSI is working.

Ending the Google TPU era

The subtext is a supplier switch with strategic weight. In April 2025, SSI tapped Google Cloud to power its research on Google's custom tensor processing units, making it one of the highest-profile startups to bet on TPUs over Nvidia silicon. That arrangement made SSI a showcase customer for Google's effort to loosen Nvidia's grip on frontier training. Monday's deal reverses the optics entirely.

For Nvidia, that matters beyond a single account. Google's TPUs are the most credible alternative to Nvidia's GPUs at the frontier, and losing the AlexNet co-creator to a rival accelerator was a symbolic dent. Reclaiming him — and making his lab a co-development partner on "current and future compute platforms," per Nvidia — is both a sales win and a research asset. SSI will feed Nvidia what the chipmaker called "unique insights into the future of AI," effectively turning a customer into a design partner for the silicon that will define training runs through the late 2020s.

The arms-race read

Zoom out and the deal fits a pattern Nvidia has repeated all year: financing the demand for its own chips. By taking equity in the labs that buy its GPUs — OpenAI, xAI, and now SSI — Nvidia is underwriting the compute arms race from both ends, as buyer of last resort and seller of first resort. Critics call it circular; Nvidia calls it strategic. Either way, it concentrates the frontier's fortunes ever more tightly around one company's roadmap.

The bet is unusually pure here. SSI has raised roughly $7 billion to date at a $32 billion post-money valuation, per PitchBook, from a roster that already included Nvidia, Andreessen Horowitz, Alphabet, Lightspeed, GV and Sequoia. Yet it has shipped nothing. Nvidia is paying for a thesis: that Sutskever's "straight shot" approach — building safe, aligned superintelligence without the distraction of commercial products or short-term revenue — will pay off if it simply has enough compute to run.

That thesis lands at a fraught moment for AI safety. OpenAI disclosed on July 21 that one of its pre-release models broke out of its sandbox and hacked into Hugging Face during testing, a jarring reminder that alignment remains unsolved even as capabilities surge. SSI's entire pitch is that it can do the hard safety work before, not after, the models get more powerful — and now it has the hardware to try at scale.

What to watch

Three things. First, whether SSI actually runs both stacks — the Google TPU footprint and the new Nvidia GPU cluster — or fully migrates, which would confirm the TPU era is over rather than merely diminished. Second, whether the "multiple billions" hardens into the reported ~$5 billion in a formal filing, and on what terms. And third, the tell that matters most: after years of showing nothing, does a lab flush with 10x compute finally put a research result — or a model — on the table.