Google Cloud inks multi-billion-dollar AI infrastructure deal with Mira Murati's Thinking Machines Lab.

What Happened

Former OpenAI CTO Mira Murati’s new venture, Thinking Machines Lab, has secured a multi-billion-dollar infrastructure agreement with Google Cloud. The deal provides the startup with massive compute resources, specifically targeting Google's upcoming AI systems powered by Nvidia’s next-generation GB300 GPUs, alongside core infrastructure services like storage, Google Kubernetes Engine (GKE), and Cloud Spanner.

Technical Details

Unlike Anthropic's recent multi-gigawatt TPU deal with Google and Broadcom, Thinking Machines Lab is explicitly targeting Nvidia's GB300 architecture. The GB300 represents Nvidia's bleeding-edge silicon, offering significant memory bandwidth and interconnect improvements over the H100 and B200 generations. By bundling this raw compute with GKE and Cloud Spanner, Google is providing a highly scalable, globally distributed control plane for model training and deployment. This indicates Thinking Machines is building a massive distributed training cluster that will likely require Spanner's strict serializability for robust checkpointing and metadata management across regions.

Why It Matters

From an engineering perspective, this deal highlights a dual-track strategy in the AI cloud wars. While Google heavily pushes its custom TPU v5p and v6 infrastructure to edge out competitors and improve margins, it remains entirely willing to act as a massive Nvidia reseller to secure high-profile workloads. For Thinking Machines Lab, opting for Nvidia GB300s over TPUs ensures their training stack remains CUDA-compatible and hardware-agnostic, avoiding vendor lock-in to Google's XLA/TPU ecosystem. It also underscores the immense capital required to train frontier models today; a single-digit billion-dollar commitment is now the baseline for new entrants attempting to compete with OpenAI and Anthropic.

What to Watch Next

Monitor the deployment timeline of the GB300 clusters, as Nvidia's supply chain will dictate the actual realization of this compute power. Additionally, watch for how Thinking Machines Lab leverages GKE for orchestration—managing tens of thousands of GB300s requires novel approaches to fault tolerance and network topology, likely relying heavily on Google's Jupiter optical circuit switched network. Finally, keep an eye out for potential direct equity investments from Google into the startup, which frequently accompany these massive cloud commitments.