NVIDIA backs Mira Murati's new lab with a gigawatt of Rubin chips

Just a few months after losing half of its founding team back to OpenAI, Mira Murati's new lab just locked down an astronomical mountain of compute.

On Tuesday, Nvidia and Thinking Machines Lab announced a multiyear strategic partnership to deploy at least one gigawatt of Nvidia's next-generation Vera Rubin AI systems. The rollout is targeted to kick off early next year, alongside a fresh financial investment from Nvidia into the young AI lab.

One gigawatt to silence the doubters

Deploying a full gigawatt of data center capacity is a threshold usually reserved for tech giants like Microsoft, Meta, or hyperscale clouds. For a company founded in February 2025, it is an eye-popping scale.

Thinking Machines Lab is already valued at north of $12 billion after raising more than $2 billion from heavyweight backers like Andreessen Horowitz and Accel. Even AMD's venture arm threw money into the pot earlier, though it is Nvidia that is now walking away with the real architectural keys.

The massive compute infusion comes at a crucial moment. The startup recently took a string of heavy personnel hits: co-founder Andrew Tulloch bolted for Meta last October, and earlier this year, three other co-founders—Barret Zoph, Luke Metz, and Sam Schoenholz—abruptly decamped to return to OpenAI.

Despite the brain drain, Murati is pushing ahead. The startup rolled out its first API product, dubbed Tinker, last fall to tackle reproducible AI outputs. Now, Nvidia and Thinking Machines say they will jointly optimize training and serving software specifically for Nvidia hardware, while opening up frontier models to enterprise and research clients.

Why it matters

We think this partnership proves that Jensen Huang is not just selling chips; he is actively picking winners and hedging his bets across the frontier AI landscape.

For Thinking Machines, the deal is an essential lifeline of confidence. Losing your technical co-founders to your fiercest rival is usually a death knell for an early-stage lab. By locking in one gigawatt of Rubin hardware, Murati sends an unmistakable signal to enterprise clients and talent: the lab has the infrastructure muscle to train frontier-class models, no matter who walked out the door. If you are an enterprise looking to build custom models, Tinker and its upcoming platform just earned a serious look.

Yet the paradox is clear. Compute is only half the equation. Having access to one gigawatt of bleeding-edge silicon does not automatically solve the problem of missing research leadership. Extracting real frontier capabilities from Rubin requires world-class systems researchers, and replacing the talent that went back to OpenAI will not be easy.

For the wider industry, this cements the terrifying new table stakes of frontier AI. Jensen Huang predicted companies would pour up to $4 trillion into AI infrastructure by the end of the decade. By trading early hardware access and capital for deep architectural integration, Nvidia continues to dictate the pace of the race while ensuring emerging challengers stay completely dependent on its stack.