Google sells so much TPU capacity its own researchers face queues
▼ Summary
– Google’s AI infrastructure, including custom TPUs and cloud business, is so successful that internal researchers now compete for computing resources sold to external customers like Anthropic and Meta.
– Alphabet committed up to $40 billion to Anthropic, including five gigawatts of TPU capacity over five years, while Meta signed its own TPU deal, locking up capacity unavailable to Google’s internal teams without queueing.
– DeepMind CEO Demis Hassabis cited hardware bottlenecks, like high-bandwidth memory from limited suppliers, and research throughput needs as constraints, with internal allocation as a separate issue.
– Researchers, including long-tenured DeepMind contributor Ioannis Antonoglou, have departed for startups as compute access became harder to secure, with Oren Etzioni noting rationing by managerial seniority.
– Google faces a delicate balance: it needs external TPU customers to validate the technology against Nvidia, while maintaining enough internal capacity for Gemini and DeepMind research, a tension unresolved in upcoming quarters.
For more than a decade, Google quietly built what many consider the most formidable AI infrastructure in the industry: a thriving cloud business, proprietary custom chips, and supply agreements that position its TPUs as the primary alternative to Nvidia for major external clients. Yet the very success of that strategy has created an internal friction the company did not foresee.
According to a report from Bloomberg’s Julia Love, Google’s own AI researchers,including teams within Google DeepMind,are now competing for access to the same computing resources the company is selling to external partners like Anthropic and Meta. The structural root of this tension is straightforward.
Google has committed up to $40 billion to Anthropic under a deal that includes five gigawatts of TPU capacity over five years and access to up to one million seventh-generation Ironwood chips. A separate supply line, mediated by Broadcom, will provide an additional 3.5GW of TPU capacity for Anthropic starting in 2027, building on the 1GW already being delivered in 2026. Anthropic itself has publicly described the Google TPU stack as central to its training and serving roadmap.
Meta, the other major TPU customer named by Bloomberg, signed its own agreement earlier this year. The capacity locked up by these external commitments is capacity that Google’s internal model teams cannot access without waiting in a queue.
DeepMind CEO Demis Hassabis acknowledged earlier this year that the constraint cuts both ways. Part of the bottleneck is hardware: “a few suppliers of a few key components,” he said, with high-bandwidth memory from Samsung, Micron, and SK Hynix cited as the most common choke point. Another part is research throughput, because, in Hassabis’s words, researchers “need a lot of chips to be able to experiment on new ideas at a big enough scale.” While the hardware constraint is partly outside Google’s control, the internal-allocation constraint is not.
The numbers behind this are enormous. Alphabet has guided a capex range of $175 billion to $185 billion for 2026, part of a combined Big Tech AI infrastructure spend that has surpassed $650 billion this year. Google, by its own commentary, has been bringing well over a gigawatt of new AI compute capacity online in 2026.
The decade-long bet on TPUs is finally delivering the kind of unit-economics advantage that allows the company to sell its chips, host its competitors’ models, and run its own frontier research on the same fabric. That fabric, however, is no longer large enough to serve all three uses simultaneously.
Bloomberg’s reporting highlights two specific signals of this tension. Researchers including Ioannis Antonoglou, a long-time DeepMind contributor, have left for startup roles in the past 18 months,a pattern that has accelerated as compute access inside Google has become more difficult to secure. Oren Etzioni, former CEO of the Allen Institute for AI, has framed the dynamic as the predictable result of an internal market where compute is rationed by managerial seniority rather than by the unit-cost economics that govern external customer contracts.
Over the past 18 months, Google has walked a delicate line: it needs its TPU programme to show volume traction with named external customers to validate the technology against Nvidia, while reserving enough internal capacity for Gemini training runs and DeepMind research. A four-partner inference-chip supply chain involving Broadcom, MediaTek, and Marvell is a hedge designed to relieve the constraint by adding capacity downstream of TPU training. It has not yet shipped at the scale demand requires.
Google did not dispute Bloomberg’s internal-allocation framing on the record, instead pointing to its broader infrastructure investment posture and the fact that compute constraints are a category-wide condition rather than a Google-specific one. That is true on the evidence: every major model provider is, based on the cleanest reading of Q1 2026 earnings, compute-constrained relative to its own research aspirations.
What makes Google’s version newsworthy is the juxtaposition: the company has, at the same time, become its main competitors’ largest infrastructure supplier. Whether it can keep selling the asset and still use it is the question the next several quarters will settle.
(Source: The Next Web)