Ilya Sutskever's SSI: The No-Product Safety Bet Against OpenAI

Safe Superintelligence Inc. has raised $6 billion at a $32 billion valuation with roughly 20 researchers, no commercial product, and no published papers. For Ilya Sutskever, the absence is the argument.

When OpenAI's Superalignment Team Dissolved

OpenAI formed its Superalignment team in mid-2023 with Sutskever and Jan Leike as co-leads, and a stated four-year mission to solve the problem of aligning superintelligent AI. The team lasted less than a year. Bloomberg reported on May 17, 2024 that the team was disbanded shortly after Sutskever announced his resignation.

Leike's exit was accompanied by a direct critique. He wrote that OpenAI had put "shiny products" ahead of "safety culture and processes," per the same Bloomberg report. William Saunders, a former Superalignment researcher, told a US Senate committee in September 2024: "Today, that team no longer exists; its leaders and many key researchers resigned after struggling to get the resources they needed to be successful," as covered by Bloomberg.

OpenAI's response was to fold the superalignment work into its wider research organization, describing the move as a way to embed safety concerns across all technical teams rather than siloing them. The implicit logic is that safety knowledge accumulates through deployment: ship products, observe failures, iterate. Sutskever's SSI is the explicit counter-position. A company with no products to ship has no trade-off to make between safety margin and release velocity. The constraint is structural rather than cultural.

The Scaling Thesis Behind the Research Bet

SSI is not only a reaction to OpenAI's organizational choices. It rests on a specific technical thesis about what comes after the scaling era.

At NeurIPS 2024, Sutskever told a packed research audience that "pre-training as we know it will unquestionably end," arguing that the internet's text corpus is finite and has largely been consumed by frontier models. In a November 2025 interview with Dwarkesh Patel, highlighted by the EA Forum, he stated the position more bluntly: "The data is finite. There is only one internet. Pre-training as we have known it is over."

Sutskever has described AI development in three distinct phases: a research era from 2012 to 2020, a scaling era from 2020 to 2025 in which adding compute and data produced reliable capability gains, and a new research era beginning in 2026 in which algorithmic innovation, not additional GPU hours, drives progress. If that thesis holds, then alignment approaches built on top of scaled pre-training may not transfer cleanly to what comes next. SSI is betting that the lab which treats alignment as a technical design constraint from the start, rather than adding it post-training, has a structural advantage in the new era.

How SSI's Structure Differs from Anthropic and OpenAI

Both Anthropic and OpenAI describe safety as central to their missions. The structural difference lies in the relationship between safety research and commercial products.

Anthropic launched the Claude API roughly 23 months after its April 2021 founding. OpenAI's first commercial API access to GPT-3 came approximately 54 months after the lab's December 2015 founding. Both labs used that commercial surface to generate revenue that funds continued research, and both describe the feedback from deployed models as a necessary input to safety work. Anthropic drew in several former Superalignment researchers, including Leike himself, suggesting those researchers found its model, which maintains dedicated safety research alongside a commercial product line, a workable middle ground.

SSI has made no analogous commitment. Its founding statement, reported by Bloomberg in June 2024, read: "This company is special in that its first product will be the safe superintelligence, and it will not do anything else up until then." The business consequence is that SSI runs entirely on investor capital. The valuation arc from $5 billion at its first close in September 2024, to $30 billion after a Greenoaks Capital-led round in March 2025, to $32 billion after a $2 billion raise in late 2025 reflects investor conviction, not revenue growth, per reporting by Globes and Calcalist.

The competitive pressure showed in 2025. Meta attempted to acquire SSI in the first half of the year, per The Information. Sutskever declined. In July 2025, co-founder Daniel Gross departed to join Meta's new Superintelligence Labs division; Sutskever became CEO, tightening his operational control. Former OpenAI colleague Mira Murati took the inverse path, raising capital for Thinking Machines Lab and shipping a product within roughly a year of leaving OpenAI.

What It Means

The three largest safety-oriented AI labs now embody three distinct theories of how safety gets built. Anthropic integrates commercial deployment with a dedicated alignment research track. OpenAI embeds safety across product teams, using real-world deployment as a feedback mechanism. SSI holds the line that no product ships until the alignment problem is solved by design. Sutskever's bet is unusual in Silicon Valley not because it is cautious but because it is structurally uncompromising: roughly 20 researchers, $6 billion in capital, and no commercial milestone to hit before the mission is complete. Which theory survives long enough to reach its goal, and which proves its safety claims in practice, remains the central unresolved question in frontier AI development.

Sources

• Bloomberg: OpenAI Dissolves Key Safety Team After Chief Scientist Ilya Sutskever's Exit
• Bloomberg: OpenAI Forms Board Safety Committee After Criticism, Ilya Sutskever Departure
• Bloomberg: Ilya Sutskever Has a New Plan for Safe Superintelligence
• The Information: SSI Co-founder Ilya Sutskever Becomes CEO, Addresses Acquisition Talks
• Calcalist: Ilya Sutskever's Safe Superintelligence Raises $2B at $32B Valuation
• Globes: Ilya Sutskever's SSI Raising Money at $30B Valuation
• EA Forum: Highlights from Ilya Sutskever's November 2025 Interview with Dwarkesh Patel

Editorial standards: every claim is sourced. Tips: [email protected]