Mira Murati's Thinking Machines Lab Signs Multibillion-Dollar Google Cloud Deal for Nvidia GB300 GPUs (April 2026)

Mira Murati's Thinking Machines Lab on April 22, 2026 announced a new multibillion-dollar agreement with Google Cloud to expand its use of Google's AI Hypercomputer, including A4X Max virtual machines powered by Nvidia's latest GB300 NVL72 GPUs. The deal — unveiled at Cloud Next '26 in Las Vegas — is valued in the single-digit billions of dollars, according to a source familiar with the matter cited by TechCrunch.

What Happened

Google Cloud and Thinking Machines Lab issued a joint announcement on April 22, 2026 stating the lab will become "one of the first Google Cloud customers to utilize NVIDIA GB300 NVL72" via the new A4X Max VM family. In early testing, Thinking Machines reported a 2× increase in training and serving speed on A4X Max compared with prior-generation Nvidia GPUs, with Google's Jupiter network handling the "near-instantaneous weight transfers required for TML's reinforcement learning workloads."

The contract bundles raw GPU compute with Google Cloud's wider stack — GKE for orchestration, Spanner for transactional metadata, Cloud Storage, Anywhere Cache, and Cluster Director for automated remediation. Thinking Machines says the integrated package will power both its frontier model research and its first commercial product, Tinker — an API that automates the creation of custom fine-tuned models, launched in October 2025.

Key Details

• Deal value: Single-digit billions of dollars, per TechCrunch's source; neither party disclosed an exact figure.
• Hardware: A4X Max VMs running Nvidia's GB300 NVL72 systems on Google's Jupiter network fabric.
• Performance: 2× training and serving speed-up versus prior-generation GPUs, per Thinking Machines' own early-testing benchmarks.
• Non-exclusive: Thinking Machines retains the right to use other cloud providers; the lab also has an existing relationship with Microsoft Azure.
• Founder context: Murati left OpenAI as CTO in September 2024 and founded Thinking Machines in February 2025, raising a record $2 billion seed round at a $12 billion valuation from Andreessen Horowitz, Accel, Sequoia and others.
• Headcount: The lab includes ex-OpenAI, Meta and Mistral talent and recently hired a Meta veteran to lead infrastructure, per reporting in Implicator.ai.

What Developers and Industry Are Saying

Reaction in the AI infrastructure community has framed the deal as a compute-war signal rather than a routine cloud purchase. Implicator.ai noted that "Google Cloud locks in" one of the most-watched independent labs at a moment when Anthropic just signed a 5-gigawatt power deal with Amazon and Google itself recently agreed to supply Anthropic with custom TPU chips. eWeek called the agreement evidence that "the bottleneck for frontier labs has shifted from talent to compute."

Andrew Tulloch, a co-founding researcher at Thinking Machines, said in the press release: "This seamless integration of high-performance compute, fast storage, GKE orchestration, and automated remediation via Cluster Director has allowed us to focus on the unique aspects of the stack like Tinker and reinforcement learning." Mark Lohmeyer, VP & GM of AI and Computing Infrastructure at Google Cloud, called Thinking Machines' research and product offerings "very exciting."

What This Means for Developers

For developers building on Thinking Machines' Tinker fine-tuning API, the practical impact is throughput. The 2× speed-up on A4X Max should translate into faster iteration on custom RL fine-tunes, lower wall-clock training cost, and tighter feedback loops for production deployment alongside training — Thinking Machines specifically calls out continuous-training-while-serving as a use case the new infrastructure unlocks.

For everyone else, the deal is another data point that frontier model training is now infrastructure-bound, not algorithm-bound. Multibillion-dollar single-customer cloud commitments are becoming routine for top-tier labs (OpenAI–Microsoft, Anthropic–AWS, Anthropic–Google, xAI–Oracle, and now TML–Google), and the GB300 generation appears to be the new floor for serious training workloads.

What's Next

Thinking Machines has not announced a public timeline for its next model release, but the additional capacity is widely read as preparation for one. The lab's research blog Connectionism remains the canonical place to track output, and Tinker's API documentation is at thinkingmachines.ai/tinker. Google Cloud said additional GB300-powered customer announcements would follow Cloud Next '26 over the coming weeks.

Sources

• Google Cloud Press Corner — Thinking Machines Expands Use of Google Cloud AI Hypercomputer — primary press release with technical specifics on A4X Max, GKE, Spanner and Anywhere Cache.
• TechCrunch — Exclusive: Google deepens Thinking Machines Lab ties — broke the deal value ("single-digit billions") and the non-exclusive terms.
• eWeek — Mira Murati's Thinking Machines Lab Lands Billion-Dollar Google Cloud Deal — analyst context on the compute-bottleneck thesis.
• Implicator.ai — Thinking Machines Signs Multi-Billion Google GB300 Deal — industry analysis framing the deal as a compute-war move.
• Datacenter News — Google Cloud expands AI deal with Thinking Machines Lab — datacenter-industry coverage with infrastructure detail.
• Thinking Machines Lab — official site — primary source on Tinker, the lab's mission and team.