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

"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."

Google built a strong AI infrastructure position through a healthy cloud business, custom chips, and supply agreements that make its TPUs a default alternative to Nvidia for major external customers. Third-party deals with Anthropic and Meta increased the value of internal access to computing resources. Google’s internal AI researchers, including teams inside Google DeepMind, now compete for access to the same computing capacity being sold externally. Google agreed to invest up to $40bn in Anthropic, including five gigawatts of TPU capacity over five years and access to up to one million seventh-generation Ironwood chips. Additional TPU capacity for Anthropic is supplied via a Broadcom-mediated line, and Meta has also secured TPU capacity. The locked capacity reduces availability for internal model teams, creating bottlenecks from both hardware component limits and research constraints.

Read at TNW | Google