Safe Superintelligence is a mysterious company, and one I struggle to forecast. Ilya Sutskever founded it in June 2024 after leaving OpenAI, where he was co-founder and chief scientist, together with Daniel Gross and Daniel Levy. It has raised about $8 billion, most recently $5 billion from Nvidia in July, holds priority access to Nvidia's next-generation Vera Rubin platform, employs roughly 50 people, and in more than two years has shipped nothing. No model, no product, no paper.
To some degree it is the purest of the "neolabs," the wave of billion-dollar AI startups founded by researchers who left the big labs, and on my overall neolab forecast it is the most credible challenger to the frontier labs, the least on observable outputs over the coming years.
Last week, Gavin Baker rumored/leaked on Invest Like the Best that SSI would launch a model in August 2026. That seemed like a significant update.
So here I go into more detail on what I was able to find on SSI, and why I forecast better outcomes for them than other AGI startups.
| Question | Forecast |
|---|---|
| SSI publicly releases or announces anything in 2026 | 38% |
| What form its first public artifact takes, by end of 2028 | Research publication or demo, 35% (nothing at all, 22%) |
| SSI's reported valuation at the end of 2027 | $58B ($30B to $207B) |
I anchored my forecasts to this timeline of events:
- June 2024. Sutskever, Gross, and Levy found SSI with offices in Palo Alto and Tel Aviv, promising "one product: a safe superintelligence" and no interim releases.
- September 2024. First round, $1 billion at a $5 billion valuation.
- April 2025. $2 billion more at $32 billion, led by Greenoaks, still the last valuation anyone has put a number to.
- June 2025. Meta reportedly tries to acquire SSI outright and is rebuffed. Co-founder Daniel Gross leaves for Meta's superintelligence lab. Sutskever takes the CEO role and Daniel Levy becomes president.
- July 2026. Nvidia invests $5 billion and grants priority access to Vera Rubin hardware. Both companies say the deal raises SSI's compute by an order of magnitude. When I forecast SSI's compute for the bench, that deal was most of why SSI's 2028 median came out at 0.65 GW, second of the six neolabs.
- July 2026. Sutskever gives the Dwarkesh interview, softening the no-releases stance.
- August 2026. The Baker launch rumor, so far uncorroborated.
- Total shipped in that time: nothing.
One caution for anyone researching SSI. The vendor profiles that rank for "SSI funding" contradict each other, $3B, $6B, $7B, even a fabricated $14.1B, and one auto-generated page invents "$1.2M revenue" for a company with no product. The arithmetic is public and it is the three rounds above.
The forecast puts 38% on SSI publicly releasing or announcing anything at all before 2027. Not 38% on the August rumor, 38% on the whole year. The reasoning is blunt. Twenty-five months of total silence is a powerful base rate, the Baker claim is secondhand and SSI itself has said nothing, and this question resolves only if SSI actually publishes, not if it gets reported on. What pulled the number up from very low is Sutskever himself. In a recent Dwarkesh interview he softened the straight-shot stance, saying even that scenario involves gradual release, and he justified the Nvidia deal by saying SSI has "research that is worthy of scaling up." A lab that just bought 10x compute for a new training run is usually at the start of something, which is also why the forecast leans 2027 rather than never.
Conditional on breaking the silence by the end of 2028, what comes first? A research publication or public demo at 35%, a hosted model or developer API at 29%, and the product paths far behind, 8% consumer and 6% enterprise. The forecast still gives 22% to SSI staying completely dark through 2028. I find the ordering believable. A technical report is the cheapest possible artifact for a lab that promised no product cycles, and it is how a safety-first lab shows progress without handing anyone a model. If the Baker rumor turns out directionally right, note that a model launch and a technical report are not mutually exclusive. The report is usually how the launch happens.
The valuation forecast is the strangest number on this page. The median for SSI's most recently reported mark at the end of 2027 is $58B, nearly double today's $32B, and the p90 is over $200B. But the distribution has a lump in it. Roughly a third of the probability sits exactly at $32B, because SSI has already shown it can take $5 billion from Nvidia without disclosing a valuation. So the honest reading is not that SSI will be worth $58B. It is that if we ever see a new number it will be much bigger, and we may simply never see one.
The 2026 question resolves in under five months, which by SSI standards is practically a day trade. If Baker is right and something ships in August, I will grade myself quickly and happily. The deeper forecast, from the bench, is that SSI reaches a frontier-class model around January 2029 without ever showing the public anything along the way, resolving the way Anthropic's unreleased Mythos did, through leaks and expert assessment. Watch the tone of the reporting, not the launch page.
Forecast these yourself in the FutureSearch app , the moment the rumor resolves, or the moment SSI raises again.