← Back to feed

In the space of 18 months, Google transformed its Tensor Processing Unit from an internal research advantage into the company’s most contested resource. The result, reported by Bloomberg on May 18, is an unusual internal dynamic: the researchers who built and depend on TPUs are now competing with Anthropic and Meta for access to them.

The immediate human cost is measurable. Andrew Dai, then a researcher at Google’s AI lab, discovered a meaningful blind spot in Gemini last summer while playing a board game. He photographed the board and asked Gemini who was winning. Gemini couldn’t answer. Rivals couldn’t either. Dai concluded the models needed fundamentally better spatial reasoning and discussed the idea with colleagues — then determined he could not secure enough compute inside Google to pursue it. He left.

Ioannis Antonoglou, a long-tenured DeepMind contributor known for work on game-playing agents, is among other senior researchers who have departed, partly citing the compute allocation problem.

The Deal Math

The pressure has a clear origin point. Google has committed up to one million seventh-generation Ironwood chips to Anthropic as part of a deal worth up to $40 billion, covering 5 gigawatts of TPU capacity over five years. An additional 3.5 GW from a Broadcom supply agreement comes online from 2027. Meta signed its own TPU deal earlier in 2026. Once those contracts are signed, those chips are revenue, not research budget.

Allocation inside Google is now governed by project priority rather than market economics. High-priority work — primarily Gemini product improvements and revenue-generating cloud workloads — gets resources. Exploratory research without an obvious near-term product connection waits.

“Inside Google, every TPU has three suitors,” said Oren Etzioni, a veteran AI researcher and professor emeritus at the University of Washington. “If you find yourself in the uncomfortable position where you have a pie-in-the-sky project and you are competing with a revenue-yielding customer, that’s a tough position to be in.”

DeepMind CEO Demis Hassabis acknowledged the constraint publicly, attributing the pressure to “a few suppliers of a few key components” — a reference to the high-bandwidth memory bottleneck at Samsung, Micron, and SK Hynix that limits production of any advanced accelerator. Researchers, he added, “need a lot of chips to be able to experiment on new ideas at a big enough scale.”

The Structural Irony

Google spent a decade building the TPU precisely to give its AI researchers an edge unavailable to external teams. The chip now serves the opposite function: locking research capacity to revenue-generating commitments.

The Blackstone-Google TPU joint venture, announced May 18, extends this logic further. Blackstone committed $5 billion in equity and takes majority ownership of a new TPU-as-a-service company targeting 500 MW by 2027 — the equivalent of a CoreWeave built on Google silicon. The venture is explicitly designed to sell TPU capacity to enterprises who want non-NVIDIA compute without a Google Cloud contract. More external supply, more external demand, same internal queue.

Google’s total 2026 capex guidance runs $175-185 billion, which suggests the company is investing heavily in building more capacity. Whether that expansion relieves the internal research queue or gets committed to the next wave of external contracts is the open question.

The Talent Signal

The pattern maps to a known dynamic in AI. Compute access has shifted from a cost line to a strategic advantage to a direct determinant of which researchers can run which experiments. Labs with owned silicon — or long-term locked supply — can iterate freely. Labs that depend on spot capacity or internal allocation systems cannot.

For talent, the revealed preference is clear: researchers who want to work on exploratory problems are moving to startups that have locked their own compute or can fund spot capacity without internal competition. Google is simultaneously its main competitors’ largest infrastructure supplier and a frontier research lab trying to run on the same hardware. That tension, as of May 2026, has not been resolved by spending alone.