Thinking Machines Lab Signs Multibillion-Dollar Google Cloud Deal for GB300 Compute, Adding a Second Hyperscaler to Its Stack
Mira Murati's startup expands its Google Cloud relationship with a non-exclusive multibillion-dollar deal centered on GB300 systems and reinforcement-learning workloads behind its Tinker fine-tuning API.
Overview
Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, has signed a multibillion-dollar agreement with Google Cloud for compute infrastructure built on Nvidia’s GB300 chips, TechCrunch reported on April 22, 2026. The deal, valued in the single-digit billions, is non-exclusive and adds Google as a second hyperscaler partner alongside the company’s existing Nvidia relationship.
The agreement reflects an industry-wide pattern: frontier labs are increasingly committing to multi-vendor compute stacks rather than relying on a single provider, while cloud companies race to lock in the next wave of AI customers before they grow too large to switch.
What We Know
Under the deal, Thinking Machines will gain access to Google Cloud infrastructure built atop Nvidia’s GB300 chips, according to TechCrunch. The contract also bundles Google Cloud services that support model training and deployment.
TechCrunch describes the arrangement as non-exclusive, meaning Thinking Machines can continue to use other providers — a notable structural choice at a moment when frontier-AI cloud deals have increasingly featured exclusivity or strong preference clauses. TechCrunch characterizes the agreement as the first time the lab has struck a deal with a cloud services provider, though TechCrunch’s earlier reporting noted that Thinking Machines had previously struck a smaller, undisclosed Google Cloud arrangement to power its AI models. The new agreement is the company’s first formal multibillion-dollar cloud-services contract.
Google disclosed the partnership in a blog post and emphasized that its cloud platform can support the reinforcement-learning workloads that underpin Thinking Machines’ product roadmap, TechCrunch reported. The only on-the-record customer quote in the announcement came from Myle Ott, a founding researcher at Thinking Machines, who said “Google Cloud got us running at record speed with the reliability we demand.”
The compute is intended to feed Tinker, Thinking Machines’ first commercial product. VentureBeat reported at Tinker’s October 2025 unveiling that the platform is an API for distributed fine-tuning of open-weight language models. Users submit jobs through the API to fine-tune supported open-weight models via supervised learning or reinforcement learning, and Thinking Machines runs those jobs on its internal training clusters. The product uses Low-Rank Adaptation, or LoRA, to keep costs down and was launched in private beta with free usage during the beta period, with usage-based pricing slated to follow.
Early Tinker users include academic groups whose work was highlighted by VentureBeat: Princeton’s Goedel Team used Tinker and LoRA with 20 percent of the data to match the performance of full-parameter supervised fine-tuned models, with the resulting model reaching 88.1 percent pass@32 on the MiniF2F theorem-proving benchmark; Stanford’s Rotskoff Lab applied reinforcement learning on top of Llama 70B for chemistry reasoning, lifting IUPAC-to-formula conversion accuracy from 15 percent to 50 percent; and Berkeley’s SkyRL group ran async off-policy multi-agent reinforcement learning loops with multi-turn tool use.
Thinking Machines itself is barely more than a year old. According to TechCrunch’s earlier reporting, the San Francisco-based company closed a $2 billion seed round in July 2025 led by Andreessen Horowitz at a $12 billion valuation, with participation from Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street. TechCrunch noted at the time that the round was one of the largest seed funding rounds in Silicon Valley history, raised when the lab was “less than a year old” and had yet to ship a product. Murati founded the lab after departing OpenAI as its chief technology officer, and the team includes former OpenAI researchers John Schulman, Barret Zoph and Luke Metz.
What We Don’t Know
The specific dollar value of the Google Cloud contract has not been disclosed beyond the “single-digit billions” framing in TechCrunch. Neither Google nor Thinking Machines has publicly stated the duration of the agreement, the volume of GB300 capacity reserved, or the milestones tied to payment. The size and terms of the earlier Google Cloud arrangement referenced in TechCrunch’s July 2025 reporting are also not public.
It is also unclear how the deal compares in size to Google’s previously announced AI cloud commitments — such as its TPU and infrastructure deals with Anthropic and Meta — because the relevant contract terms remain confidential in each case. The non-exclusive nature of the Thinking Machines arrangement leaves open whether Murati’s company will sign additional hyperscaler agreements with AWS or Microsoft Azure in the months ahead.
Finally, while Tinker has accumulated visible academic adoption since its October 2025 launch, neither Thinking Machines nor Google has disclosed the size of Tinker’s paying customer base, its revenue, or the share of Thinking Machines’ overall compute footprint that the new Google Cloud allotment is expected to absorb. Those metrics will determine whether the multibillion-dollar commitment is a forward bet on growth or a sizing of demand that is already in hand.
Analysis
The agreement reinforces a pattern that has defined frontier-AI infrastructure procurement over the past year: labs increasingly distribute their compute across multiple cloud and silicon providers rather than concentrating it in a single relationship. By layering Google Cloud’s GB300 capacity on top of its prior Nvidia partnership, Thinking Machines is following the playbook used by larger competitors that have stitched together multi-cloud stacks to reduce supply risk and to preserve negotiating leverage as next-generation accelerators come online.
For Google, the deal is part of a broader push to position Google Cloud as the default home for AI workloads built on Nvidia’s newest chips, even as the company simultaneously markets its own TPU silicon. Anchoring a high-profile customer like Thinking Machines — whose Tinker product runs the kind of reinforcement-learning workloads that are becoming central to post-training pipelines — gives Google Cloud a reference deployment that it can cite as it pursues other frontier labs.