Google’s multibillion-dollar bet on Thinking Machines shows the new AI moat: compute and talent

Google’s new agreement with Thinking Machines Lab is less a routine cloud renewal than a stake in a frontier AI pipeline. TechCrunch reported that the deal is worth a single-digit billions amount and gives the lab access to Google Cloud systems powered by Nvidia’s GB300 chips, while Google said the startup is one of the first customers on its GB300-powered hardware ^[1]. SiliconANGLE and Implicator both frame the contract as part of Google’s larger push to bundle AI compute with storage, databases, and orchestration services rather than selling bare metal alone ^[4][6].

That matters because the deal is explicitly non-exclusive, which means Thinking Machines can spread workloads across providers over time ^[1]. But even a non-exclusive contract can function like a supply reservation in a market where frontier labs are racing to secure scarce capacity before the next chip generation arrives. Google is not simply hosting a startup; it is trying to become the default operating environment for companies that need extreme scale, uptime, and fast access to the latest hardware ^[4][6].

The technical core of the deal is reinforcement learning, which Google said its infrastructure can support for Thinking Machines’ Tinker product ^[1]. Tinker, launched in October, automates the creation of custom frontier models, and that kind of workflow tends to multiply compute demand because weights and updates move across large GPU fleets repeatedly during training ^[1][6]. In other words, this is not a generic inference contract; it is a bet on a workload that gets more expensive as the model and experiment count grow.

Google’s pitch is speed and reliability. The company says its GB300-based systems deliver roughly a 2x improvement in training and serving compared with prior-generation GPUs, and Myle Ott, a founding researcher at Thinking Machines, said Google Cloud got the team running at record speed ^[1][6]. That combination of faster hardware and lower operational friction is exactly what reinforcement learning teams need, because the bottleneck is often not model design but the ability to keep large-scale experiments moving without network or orchestration failures ^[6].

Thinking Machines is not signing into a quiet market. Earlier this month, Anthropic struck a multi-gigawatt agreement with Google and Broadcom for TPU capacity, even as it also expanded with Amazon for up to 5 gigawatts to train and deploy Claude ^[1]. That means Google, Amazon, and other hyperscalers are now competing not just on price or tooling, but on who can pre-sell the most future capacity to the most important AI labs. Implicator notes that Thinking Machines is now the third frontier customer lined up for Google’s Blackwell and TPU silicon this month ^[6].

The economics help explain the urgency. Implicator cites Alphabet’s 2026 capital expenditure guidance near $200 billion, a Google Cloud backlog that more than doubled to $240 billion last year, and an annualized cloud revenue run rate around $72 billion ^[6]. Those numbers suggest the cloud market has evolved from a generalized enterprise utility into a capital-intensive supply chain for model builders. The labs that can reserve chips early get a performance edge; the hyperscalers that can keep those chips full lock in recurring revenue and deeper platform stickiness ^[4][6].

Thinking Machines has moved unusually fast for a startup that was founded only in February 2025, after Murati left her role as OpenAI’s chief technologist ^[1]. The company raised a $2 billion seed round at a $12 billion valuation, then shipped Tinker in October, while keeping much of its roadmap secret ^[1]. That secrecy may be strategic, but it also makes the Google deal unusually revealing: the infrastructure choice hints at the kind of model-training and reinforcement-learning work the lab is pursuing ^[1].

The timing is awkward. Meta has reportedly hired five founding members of the lab, including, by one report, a $1.5 billion engineer, while another account says seven founders have now been poached, including Tinker’s lead engineer ^[3][6]. The exact headcount is less important than the pattern: model labs are now defending themselves on two fronts at once, against compute scarcity and against talent raids. For a young frontier company, losing either one can weaken the other, because researchers and infrastructure plans are tightly coupled in this stage of the AI race ^[3][6].

The practical lesson for AI teams is that frontier infrastructure is becoming portfolio management. Thinking Machines’ non-exclusive contract suggests sophisticated buyers will keep multiple clouds in play, but the deal also shows that each provider is trying to win a larger share of the workload by bundling GPUs, networking, databases, and orchestration into one operating stack ^[1][4]. That makes vendor choice a technical decision about throughput, failover, and data movement, not just a finance decision about unit cost.

The second lesson is that capability is no longer just a function of model architecture. The companies that can combine reliable chip supply, fast interconnects, and research talent will move faster than rivals that treat any one of those as interchangeable ^[6]. For practitioners, that means planning for capacity reservations, portability across clouds, and stronger retention strategies for scarce researchers. The AI winners of this cycle may be the teams that can negotiate both compute and people with equal discipline ^[1][3][6].