Google has sold so much TPU capacity that its own researchers are queueing for the rest

Alphabet runs the most enviable AI infrastructure stack in the industry. The success of the third-party deals with Anthropic and Meta is the reason internal access has become a competitive resource.

Google has spent a decade quietly building the most enviable position in AI infrastructure: a healthy cloud business, its own custom chips, and supply deals that make its TPUs the default alternative to Nvidia for major external customers.

The success of that strategy has produced an internal problem that the company did not anticipate.

Bloomberg’s Julia Love reported on Monday that Google’s own AI researchers, including teams inside Google DeepMind, are now jockeying for access to the computing resources their employer is also selling to Anthropic and Meta.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

The structural cause is straightforward. Google has agreed to invest up to $40bn in Anthropic on 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, Broadcom-mediated supply line covers a further 3.5GW of TPU capacity for Anthropic from 2027, building on the 1GW the company is already receiving in 2026. Anthropic itself has publicly described the Google TPU stack as central to its training and serving roadmap.

Meta, the other commercial-scale TPU customer Bloomberg names, signed its own deal earlier this year. The capacity those commitments lock up is capacity not available to Google’s internal model teams without queueing.

DeepMind’s chief executive Demis Hassabis said earlier this year that the constraint cuts in two directions. Some of the bottleneck is hardware: ‘a few suppliers of a few key components’, as he put it, with high-bandwidth memory from Samsung, Micron and SK Hynix the most-cited choke point.

Some of it 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’. The hardware constraint is partly outside Google’s control. The internal-allocation constraint is not.

The arithmetic underneath this is large. Alphabet is on a guided capex range of $175bn-$185bn for 2026, inside a combined Big Tech AI infrastructure spend that has crossed $650bn this year. Google has, on its own commentary, been bringing well over a gigawatt of new AI compute capacity online in 2026.

The decade-long bet on TPUs is finally producing the kind of unit-economics advantage that lets the company sell its chips, host its competitors’ models, and run its own frontier research on the same fabric. The fabric is just no longer big enough for all three uses at once.

Bloomberg’s reporting names two specific signals of the tension. Researchers including Ioannis Antonoglou, a long-tenured DeepMind contributor, have departed for startup roles in the past 18 months, a pattern that has accelerated as compute access has become harder to secure inside Google.

Oren Etzioni, the former Allen Institute for AI chief executive cited in the piece, has framed the dynamic publicly as the predictable result of an internal market in which compute is rationed by managerial seniority rather than by the unit-cost economics that govern external customer contracts.

Google has spent the past 18 months in a delicate position: it needs its TPU programme to demonstrate volume traction with named external customers to validate the technology against Nvidia, while keeping enough internal capacity for Gemini training runs and DeepMind research.

The four-partner inference-chip supply chain with 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 the demand curve requires.

Google did not dispute Bloomberg’s internal-allocation framing on the record; it pointed 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, on the cleanest reading of Q1 2026 earnings, compute-constrained relative to its own research aspirations.

What makes the Google version newsworthy is the side-by-side: the company has, at the same time, become its main competitors’ largest infrastructure supplier. If it can keep selling the asset and using it is the question the next several quarters will settle.