Google's TPU Sales to Anthropic and Meta Crowd Out Its Own Researchers

Google's TPU infrastructure strategy has produced an uncomfortable side effect: the company's own researchers are now queuing for compute their employer is selling to outside competitors. The internal friction is a direct result of long-term supply commitments that were commercially attractive but have squeezed what is available inside the company.

The Next Web reported Monday that teams inside Google DeepMind are competing for tensor processing units the company has committed to Anthropic and Meta. Bloomberg's Julia Love broke the original story, framing it as a structural tension nobody fully anticipated when the agreements were signed.

The numbers make the squeeze legible. Google agreed to provide Anthropic five gigawatts of TPU capacity over five years, part of an investment arrangement worth up to $40 billion that also covers access to as many as one million of its seventh-generation Ironwood chips. A Broadcom-mediated tranche adds 3.5 gigawatts more for Anthropic from 2027, layered on top of the one gigawatt already flowing in 2026. Meta signed its own TPU supply deal earlier this year. Combined, those contracts lock up a significant portion of Google's chip output for external customers, leaving internal teams to compete for what remains.

Demis Hassabis, Google DeepMind's chief executive, has acknowledged the constraint runs in two directions: hardware supply, where a few manufacturers control high-bandwidth memory production, with Samsung, Micron and SK Hynix the most-cited bottlenecks; and research throughput, which suffers whenever teams cannot run experiments at the cadence they need.

The internal access problem

There is structural irony here. Google built its TPU program to reduce dependence on Nvidia and give its own researchers a frontier advantage in artificial intelligence. That bet paid off: Anthropic has publicly described the Google TPU stack as central to its training roadmap. That credibility is also why signing Anthropic and Meta made commercial sense for Google Cloud. What the deals did not anticipate was how fast internal demand for frontier model training would scale in parallel.

Context and implications

The hardware bottleneck is not unique to Google. High-bandwidth memory supply is constrained across the sector, and frontier model trackers like llm-stats.com show major labs shipping new artificial intelligence model versions at an accelerating pace through 2026. More training runs mean more compute demand and less slack in the supply chain.

Since late 2022, AI Release Tracker has documented over 150 frontier model releases, and cadence is not decelerating. Google's current situation is less a strategic error than an early signal that compute scarcity will become a recurring governance problem for any organization running both a chip supply business and a frontier research program at the same time.

Price Per Token data shows inference costs declining as more hardware comes online, but training-level compute is a different market, still controlled by whoever secured allocation earliest. Google locked that position for its paying customers. Whether its internal teams can negotiate equivalent priority remains the open question.

Google has three realistic paths: formalize an internal compute allocation process, build additional TPU capacity fast enough to ease the pressure, or quietly restructure some external commitments. Anthropic and Meta will not accept reduced allocations without a fight. Capacity expansion is the only palatable option, and expansion at TPU scale takes years, not quarters.

Frequently asked questions

What are Google TPUs, and why do artificial intelligence labs use them?

TPUs are custom chips Google designed for machine learning workloads. Labs adopt them as a high-throughput alternative to Nvidia GPUs, particularly for large-scale model training runs.

Why is Anthropic, which competes with Google AI products, buying Google chips?

AI infrastructure and AI product markets operate partly independently. Anthropic needed compute at a scale Google could supply, and Google needed credible large-scale cloud customers. The commercial arrangement runs alongside competitive dynamics in the model market.

How much TPU capacity has Google committed to Anthropic?

Five gigawatts over five years plus access to one million Ironwood chips, with an additional 3.5 gigawatts via a Broadcom-mediated deal from 2027, layered on top of one gigawatt already in place for 2026.

What is high-bandwidth memory and why does it constrain AI chip supply?

HBM is a critical component inside AI accelerators including TPUs and Nvidia GPUs. Only Samsung, Micron and SK Hynix produce it at significant volume, so constraints on HBM supply translate directly into finished chip shortages.