29 July 2026

^{NVIDIA Vera Rubin Platform (photo: NVIDIA)}

 

Safe Superintelligence Inc., the artificial intelligence research company founded by former OpenAI chief scientist Ilya Sutskever, has spent its first two years telling the world remarkably little. It has released no chatbot, published no model specifications and offered no public demonstration of the technology it is developing.

Now Nvidia has provided the clearest indication yet that something potentially significant is happening behind its closed doors.

The companies have entered a long-term strategic partnership that gives Safe Superintelligence, known as SSI, access to Nvidia’s next-generation Vera Rubin computing systems. They say the infrastructure will allow SSI to increase its available computing power by an order of magnitude.

Nvidia, which participated in an earlier SSI funding round, has also made an additional equity investment. The companies did not disclose its size, but Reuters, citing a person briefed on the transaction, reported that Nvidia is investing $5 billion.

“We have research that is worthy of scaling up, and having access to a big Nvidia computer will let us do so,” Sutskever said in the announcement.

 

What is Safe Superintelligence?

 

Sutskever founded Safe Superintelligence in June 2024 with former Apple AI executive and investor Daniel Gross and former OpenAI researcher Daniel Levy. The American company has offices in Palo Alto and Tel Aviv.

From the outset, SSI presented itself as a different kind of AI company. Its declared objective was not to build a family of commercial models or develop products while gradually working towards more capable systems. It described itself as the world’s first “straight-shot SSI lab”, with one goal and one product: safe superintelligence.

Superintelligence generally refers to an artificial intelligence system whose capabilities substantially exceed those of humans. AI alignment concerns the separate but closely related problem of ensuring that such systems continue to behave in accordance with human intentions and constraints.

SSI argues that capability and alignment should be developed together as parts of the same technical challenge. Its founding statement promised to advance AI capabilities as quickly as possible while keeping safety ahead of them.

The company’s structure was designed around that mission. Without commercial products, customer commitments or frequent releases, its researchers would theoretically be insulated from the product cycles and short-term competitive pressures affecting other AI laboratories.

This also makes SSI an extraordinary investment proposition. Its backers are not financing a product whose adoption or revenue can be measured. They are backing a team, a largely undisclosed research thesis and, above all, Sutskever’s record.

 

Ilya Sutskever’s role in modern AI

 

Sutskever has been central to several of the developments that produced the current generative AI boom.

As a student of Geoffrey Hinton at the University of Toronto, he worked with Hinton and Alex Krizhevsky on AlexNet, the neural network that achieved a decisive victory in the 2012 ImageNet computer-vision competition. Its performance helped demonstrate that deep neural networks trained with large datasets and graphics processing units could outperform established approaches.

Sutskever subsequently contributed to important work on sequence-to-sequence learning, which helped machines process and generate ordered information such as language. He also contributed to research behind AlphaGo and OpenAI’s generative pre-trained transformer, or GPT, models.

In 2015, he became one of OpenAI’s co-founders and later served as its chief scientist. He was an early and influential advocate of scaling: the principle that model performance could continue improving as developers increased data, computing power and model size.

That conviction helped establish the technical direction followed by OpenAI and much of the wider AI industry.

Towards the end of his time at OpenAI, Sutskever became increasingly focused on controlling systems that might eventually become more intelligent than their creators. He co-led the company’s Superalignment programme, which investigated how future superhuman models might be kept aligned with human intentions.

His departure followed OpenAI’s turbulent attempt to remove chief executive Sam Altman in November 2023. Sutskever initially supported the board’s decision before saying he deeply regretted his participation in its actions. Altman returned within days, while Sutskever formally left OpenAI in May 2024.

SSI emerged one month later.

 

How did SSI reach a $32 billion valuation?

 

Investors moved quickly. In September 2024, Safe Superintelligence raised $1 billion from backers including Andreessen Horowitz, Sequoia Capital, DST Global and SV Angel. Reuters reported that the three-month-old company was valued at $5 billion.

A subsequent $2 billion funding round led by Greenoaks valued SSI at $32 billion. Alphabet and Nvidia also invested in the company during this period, according to Reuters.

The increase reflected neither product sales nor user growth. It represented a wager that Sutskever and his team could discover a path towards more capable AI that other laboratories had missed.

The company then experienced an early leadership disruption. In 2025, Meta reportedly attempted to acquire SSI as part of Mark Zuckerberg’s campaign to strengthen its superintelligence research. SSI rejected the approach, but Gross subsequently left to join Meta.

Sutskever became chief executive, with Levy serving as president. Addressing concerns about SSI’s future, Sutskever said the company remained independent and was focused on completing its work.

“We have the compute, we have the team, and we know what to do,” he wrote at the time.

The Nvidia partnership makes the first part of that claim considerably more substantial.

 

Is SSI moving beyond conventional AI scaling?

 

Safe Superintelligence has not revealed its model architecture, training process or alignment methodology. Nevertheless, Sutskever’s public comments provide some clues about the problem it may be trying to solve.

The current generation of large language models is largely built through pre-training: exposing a model to enormous quantities of text, images and other data so that it learns underlying patterns. Additional training and feedback are then used to refine its behaviour.

This approach has produced striking results, but Sutskever has argued that conventional pre-training is approaching its limits. High-quality human-generated data is finite, while larger training runs do not automatically produce systems that reason reliably or learn as efficiently as people.

He has increasingly emphasised generalisation: the ability to understand unfamiliar situations and acquire capabilities from comparatively limited information. Humans can learn new tasks from relatively few examples. Current AI systems often require vastly more data while still performing unevenly outside the conditions represented in their training.

Sutskever has described the industry as moving from an “age of scaling” towards an “age of research”. That does not mean computing power has become irrelevant. It means that more compute may need to be applied to new technical ideas rather than simply used to produce a larger version of an established model.

His assertion that SSI now has research “worthy of scaling up” therefore carries particular weight. It implies that the company believes it has identified such an idea. Until SSI publishes evidence or releases a system, however, outsiders cannot assess its novelty, effectiveness or safety.

 

What Nvidia’s Vera Rubin systems give SSI

 

SSI previously partnered with Google Cloud to use its tensor processing units, or TPUs, specialised chips designed for AI workloads. Reuters reported in 2025 that SSI was primarily using TPUs rather than graphics processing units for its research.

The Nvidia agreement does not say whether that relationship will change or whether Vera Rubin will replace Google’s infrastructure. It does, however, give SSI access to a new generation of Nvidia systems while allowing the company to increase its overall compute tenfold.

Vera Rubin succeeds Nvidia’s Blackwell generation and is designed as an integrated, rack-scale AI computing platform. It combines Rubin graphics processors, Vera central processing units, high-speed memory, networking, security and supporting software. The architecture is intended to train, refine and operate increasingly large AI models without bottlenecks elsewhere in the system.

For SSI, the arrangement extends beyond buying more chips. Nvidia says the two companies will collaborate on the technical development of its current and future computing platforms, using SSI’s insights into where advanced AI is heading.

Nvidia said it entered the partnership after obtaining “rare access” to SSI’s closely guarded research. The chipmaker therefore knows more about the company’s progress than the investors and outsiders being asked to judge it from public statements alone.

 

Can an independent AI safety laboratory remain independent?

 

The Nvidia–SSI partnership illustrates a wider reality about frontier AI. Advanced research requires enormous computing clusters, energy supplies, high-speed networking and billions of dollars. Even a company specifically designed to avoid conventional commercial pressures cannot operate outside the industrial system supporting AI development.

SSI rejected reported acquisition interest and continues to describe itself as independent. Nvidia’s investment is not an acquisition. But a reported $5 billion equity commitment, privileged access to infrastructure and collaboration on future computing systems create a relationship that goes considerably beyond an ordinary supplier contract.

For Nvidia, the partnership also provides another strategic connection to a frontier AI developer whose future infrastructure requirements could be enormous. Investing in model companies can support demand for Nvidia systems while giving the chipmaker insight into the workloads its future hardware will need to accommodate.

SSI remains a multibillion-dollar promise built around one scientist’s record and a technical programme the public cannot inspect. Nvidia’s backing does not prove that Sutskever has found a route to safe superintelligence. It does show that, after gaining unusual access to SSI’s research, the world’s most important AI chipmaker has seen enough to place a very large bet on scaling it.

 

Further reading on MoveTheNeedle.news:

Can OXMIQ’s OxPython turn CUDA against Nvidia?

Verda bets on NVIDIA Rubin as demand for AI computing power accelerates