The AI developer community spent the last few days in a state of high-velocity speculation. On X and GitHub, the conversation shifted abruptly from model weights to infrastructure as news leaked regarding the mysterious movements of Thinking Machines Lab. Founded by Mira Murati, the former Chief Technology Officer of OpenAI, the startup had largely operated in stealth, leaving the industry to guess whether Murati was building a direct competitor to her former employer or something entirely different. That ambiguity vanished this week with the announcement of a massive, multi-billion dollar partnership with Google Cloud, signaling that the company is no longer just planning for the future but is actively scaling for a massive compute-heavy launch.
The Infrastructure Play and the GB300 Advantage
The scale of the agreement is staggering, with reports indicating the contract value sits in the single-digit billions of dollars. This is not a standard cloud subscription but a strategic resource grab. Thinking Machines Lab has secured priority access to Nvidia GB300-based systems, the latest iteration of AI accelerators designed to handle the most demanding workloads in the industry. According to Google, these GB300 systems deliver a two-fold increase in both training and serving speeds compared to the previous generation of GPUs. This performance leap is critical for a company attempting to compress the development cycle of a frontier model.
Beyond the raw silicon, the deal integrates Thinking Machines Lab into the broader Google Cloud ecosystem. The infrastructure package includes Google Kubernetes Engine for containerized application management and Spanner, Google's globally distributed database service, ensuring that the startup can scale its data pipeline across regions without latency bottlenecks. This mirrors a broader trend where cloud providers are bundling high-end AI hardware with enterprise-grade orchestration tools to lock in the next generation of AI labs. Earlier this month, Anthropic followed a similar trajectory, partnering with Google and Broadcom to secure several gigawatts of TPU capacity to sustain its own scaling needs.
The financial foundation of Thinking Machines Lab is equally aggressive. Established in February 2025, the company entered the market with a seed round that valued the firm at 12 billion dollars, raising 2 billion dollars in initial capital. This valuation is nearly unprecedented for a seed-stage entity, reflecting the market's confidence in Murati's leadership and the potential of the company's first product, Tinker. Launched in October, Tinker is designed to automate the creation of custom frontier AI models, effectively attempting to build a factory for intelligence rather than a single static model.
The Reinforcement Learning Compute Tax
While the hardware specifications are impressive, the true significance of the deal lies in the specific type of workload Thinking Machines Lab is preparing for. Google explicitly stated that the partnership is designed to support reinforcement learning workloads. This is the pivotal detail. Reinforcement learning, the process of training a model through trial and error to optimize for specific outcomes, has become the gold standard for achieving breakthroughs in reasoning and complex problem-solving, as seen in the latest trajectories of DeepMind and OpenAI. However, reinforcement learning is notoriously inefficient in terms of compute, requiring vast amounts of iterative processing that can bankrupt a startup without a massive capital cushion.
By securing the GB300s, Thinking Machines Lab is addressing the compute tax inherent in the Tinker architecture. If Tinker's goal is to automate the generation of frontier models, it must run thousands of reinforcement learning loops simultaneously to find the optimal configurations for different use cases. The multi-billion dollar price tag of the Google deal is not merely for storage or hosting; it is a direct payment for the raw energy and silicon required to fuel these RL loops. The two-fold increase in speed provided by the GB300 is not just a marginal gain—it is the difference between a model taking six months to train or three.
This move also places Thinking Machines Lab in the center of a high-stakes proxy war between cloud giants. We are seeing a pattern where frontier labs play providers against one another to secure the best possible terms. Anthropic, for instance, maintains a delicate balance, utilizing Google while simultaneously securing up to 5 gigawatts of capacity from Amazon for the training and deployment of Claude. Thinking Machines Lab had previously established a partnership with Nvidia that included direct investment, but the move to Google Cloud marks its first major commitment to a specific cloud ecosystem. Because the contract is non-exclusive, Murati retains the flexibility to diversify her infrastructure, but Google's early capture of the lab gives them a strategic foothold in the automation of frontier models.
In an era where the ceiling of artificial intelligence is determined by the volume of available compute, the ability to secure the fastest chips at a multi-billion dollar scale transforms a startup into a sovereign power. The intersection of massive capital, elite leadership, and next-generation silicon suggests that the speed of innovation at Thinking Machines Lab will likely outpace the traditional release cycles of the established big tech incumbents.
