Google has sold so much TPU capacity that its own researchers are now waiting in line for the remaining resources.

      Alphabet possesses one of the most sought-after AI infrastructure stacks in the industry. The achievements in its partnerships with Anthropic and Meta have turned internal access into a valuable competitive asset.

      Over the past decade, Google has stealthily established a leading position in AI infrastructure, featuring a robust cloud business, custom chips, and procurement agreements that designate its TPUs as the go-to alternative to Nvidia for significant external clients. However, the success of this strategy has led to an unforeseen internal challenge for the company.

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

      The underlying structural issue is clear. Google has pledged to invest up to $40 billion in Anthropic, which includes five gigawatts of TPU capacity over five years and access to nearly one million seventh-generation Ironwood chips. Additionally, a separate supply line mediated by Broadcom will provide another 3.5GW of TPU capacity to Anthropic starting in 2027, adding to the 1GW the company will already receive in 2026. Anthropic has publicly stated that the Google TPU stack is essential to its training and operational plans.

      Meta, another major customer of TPUs that Bloomberg mentions, secured a similar deal earlier this year. The capacity reserved by these agreements means that it is no longer available for Google’s internal model teams without delays.

      DeepMind’s CEO, Demis Hassabis, indicated earlier this year that the limitations are twofold. One aspect is hardware-related, with a few suppliers for crucial components. High-bandwidth memory from companies like Samsung, Micron, and SK Hynix is identified as a significant bottleneck. The second challenge lies in research throughput, as Hassabis noted that researchers "need a lot of chips to experiment with new ideas on a substantial scale." While the hardware limitations are partly beyond Google's control, the allocation of internal resources is not.

      The scale of this issue is considerable. Alphabet is targeting a capital expenditure range of $175 billion to $185 billion for 2026, amid a collective Big Tech AI infrastructure spend exceeding $650 billion this year. According to its own statements, Google plans to bring well over a gigawatt of new AI compute capacity online in 2026.

      The long-term investment in TPUs is finally yielding the kind of economic advantage that allows the company to sell its chips, support its rivals' models, and conduct its own cutting-edge research using the same resources. However, the capacity is no longer sufficient for all three purposes at the same time.

      Bloomberg has identified two specific indicators of this tension. Researchers, including Ioannis Antonoglou, a long-term contributor at DeepMind, have left for startup positions over the past 18 months, with this trend accelerating as access to computing resources has become more difficult within Google. Former Allen Institute for AI CEO Oren Etzioni has publicly described this dynamic as a predictable outcome of an internal market where compute resources are allocated based on managerial seniority rather than the unit-cost economics that drive external customer agreements.

      For the past 18 months, Google has found itself in a challenging situation: it requires its TPU program to demonstrate substantial traction with named external clients to validate its technology against Nvidia while retaining sufficient internal capacity for Gemini training runs and DeepMind research. The collaboration with Broadcom, MediaTek, and Marvell on an inference-chip supply chain is an attempt to alleviate this constraint by increasing capacity downstream of TPU training, but it has yet to reach the scale required to meet demand.

      Google did not challenge Bloomberg’s portrayal of its internal resource allocation; instead, it highlighted its broader investment approach and the reality that compute constraints are a widespread issue across the industry rather than unique to Google. This observation holds true, as every major model provider appears compute-constrained concerning its own research goals, according to the simplest interpretation of Q1 2026 earnings.

      What makes Google’s situation particularly noteworthy is the juxtaposition of its role as the largest infrastructure supplier for its main competitors while grappling with these internal challenges. Whether it can continue to sell the capacity while also using it will be determined over the next several quarters.