Google’s TPU Capacity: A Supply and Demand Dilemma
Google has sold so much Tensor Processing Unit (TPU) capacity that its own researchers are now facing a queue for access to the remaining resources.
(May 18, 2026 – 1:35 pm)
Image by: The Pancake of Heaven!
Alphabet, the parent company of Google, has built an enviable AI infrastructure stack over the past decade. This includes a robust cloud business, custom-designed chips, and strategic supply deals that have made its TPUs the preferred alternative to Nvidia for major external customers like Anthropic and Meta.
The success of this strategy, however, has led to an internal challenge. Bloomberg’s Julia Love reported that Google’s AI researchers, including teams from DeepMind, are competing for access to the computing resources their employer is also selling to external clients.
The Supply Chain
The primary reason for this situation is the substantial investments Google has made in Anthropic and Meta:
- Anthropic Deal: Up to $40 billion over five years, including 5 GW of TPU capacity and access to up to one million seventh-generation Ironwood chips.
- Broadcom Supply Line: An additional 3.5 GW of TPU capacity for Anthropic from 2027, building on the 1 GW already received in 2026.
Anthropic has publicly acknowledged Google’s TPU stack as central to its training and serving roadmap. Meta, another commercial-scale TPU customer, also signed a deal earlier this year, further limiting access for Google’s internal model teams.
Bottlenecks and Constraints
DeepMind CEO Demis Hassabis noted two main constraints:
- Hardware: Limited supply of key components from Samsung, Micron, and SK Hynix.
- Research Throughput: The need for ample chips to experiment with new ideas on a large scale.
While some hardware limitations are beyond Google’s control, the internal allocation constraint is a direct result of its own success.
The Numbers
Alphabet’s guided capex range for 2026 is $175 billion to $185 billion, with a significant portion dedicated to AI infrastructure. Google has brought over one gigawatt of new AI compute capacity online in 2026 alone, proving the success of its decade-long bet on TPUs.
Bloomberg’s reporting highlights two concerning trends:
- Researchers, including Ioannis Antonoglou, a DeepMind veteran, have left for startup roles in the past 18 months, accelerating in recent times.
- This internal-allocation issue signifies a potential talent drain and disruption to Google’s research capabilities.