SSI had aimed to only release a model when it had achieved superintelligence, and that could be coming sooner than anyone expected.
That’s the claim from Gavin Baker, CIO at Atreides Management, who dropped the detail almost as an aside while discussing something far bigger on a recent podcast appearance. Baker was talking about continual learning and sample efficient learning, two research problems that a growing number of labs believe they are close to cracking, and what solving them would do to demand for AI chips. His argument runs like this: if a model can be trained once on something like 10 trillion tokens and then released into the world to keep learning from its own interactions, sample efficiently and without needing to be retrained from scratch, the entire economics of the industry shift. Training, which today eats up a massive share of global chip demand, would shrink to almost nothing as a percentage of total semiconductor spend. Inference and continual updates would take over as the dominant workload instead.
Here’s Baker, in his own words:
“A lot of people seem to think they’re close to solving continual learning and sample efficient learning. It’s possible if those are solved, could there be a discontinuity of demand (of chips). If you could train something on 10 trillion tokens and let it out into the world and learn sample efficiently, it doesn’t sound good for training demand. (In this case) training as a percentage of semiconductor demand is going to asymptote to close to zero. That’s the most interesting (approach), who know if it’s long horizon or short horizon. SSI says that they’ll come out with their model in august. There’s a whole generation of new labs focused on (continual learning)”
If true, it would be the first real product to come out of a company that has spent two years actively avoiding the idea of shipping one.
SSI, short for Safe Superintelligence, was started in June 2024 by Sutskever along with Daniel Gross, formerly of Apple’s AI team, and Daniel Levy, an OpenAI alum. Sutskever had just come out of one of the most dramatic exits in Silicon Valley history, having been one of the board members behind the ouster of Sam Altman before Altman was reinstated days later. When Sutskever resurfaced with SSI, he framed it in almost monastic terms, describing a company that would pursue safe superintelligence “in a straight shot, with one focus, one goal, and one product.” Explicitly no interim releases, no chasing revenue, no distraction from the core research bet.
That framing has held up remarkably well, at least in terms of what the outside world has seen. SSI has raised $1 billion within months of launching, at a valuation of roughly $5 billion, then followed it up with a $2 billion round in April 2025 that pushed its valuation to $32 billion, led by Greenoaks with Alphabet and Nvidia joining as strategic backers. Nvidia went further still, striking a deal that included a substantial fresh investment and giving SSI access to its next-generation Vera Rubin compute platform. All of this for a company with roughly 50 employees, no published research, and no product anyone outside the building has ever seen.
Sutskever has defended the approach on the rare occasions he’s spoken publicly about it, including on the Dwarkesh podcast, where he argued that the capital SSI has raised, modest by frontier lab standards, is more than enough for the kind of research it’s doing, since research and commercial deployment require very different amounts of infrastructure. It’s a position that puts him at odds with almost every other major lab, all of which are now locked into multi-billion dollar compute commitments to keep pace with each other.
The one crack in SSI’s tightly controlled story came last year, when Meta made a run at the company. Mark Zuckerberg had been assembling his own superintelligence effort and reportedly tried to acquire SSI outright for tens of billions of dollars. Sutskever turned him down. Zuckerberg pivoted to hiring instead, and by June 2025, co-founder Daniel Gross had left SSI to join Meta’s newly formed Superintelligence Labs alongside Nat Friedman and Alexandr Wang. Meta also picked up a large stake in NFDG, the investment vehicle Gross had run with Friedman, giving it exposure to a sprawling portfolio of AI startups in the process. It made SSI yet another lab in the industry to have already lost a founding member to a rival, a list that keeps growing as co-founder departures become almost routine across the sector.
None of that seems to have changed Sutskever’s timeline, if Baker’s comment is accurate. Whether SSI actually ships something in August, and whether what it ships looks anything like the products every other lab has been racing to release, remains to be seen. For a company that has built its entire identity around silence and patience, even a hint of a release date counts as a major disclosure.