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

      Alphabet has established the most desirable AI infrastructure stack in the industry. The achievements of its third-party partnerships with Anthropic and Meta have turned internal access into a valuable competitive asset. Over the past decade, Google has quietly crafted a strong position in AI infrastructure, featuring a robust cloud business, custom chips, and supply agreements that position its TPUs as the primary alternative to Nvidia for significant external clients. However, this successful approach has resulted in an unexpected internal challenge for the company.

      Bloomberg’s Julia Love reported on Monday that Google’s own AI researchers, including teams within Google DeepMind, are now competing for access to the computing resources that their employer is also providing to Anthropic and Meta. The fundamental issue is clear: Google has committed to investing up to $40 billion in Anthropic, which includes an allocation of five gigawatts of TPU capacity over five years alongside access to up to one million seventh-generation Ironwood chips.

      A separate supply line managed by Broadcom will provide an additional 3.5 gigawatts of TPU capacity for Anthropic starting in 2027, building on the one gigawatt the company is set to receive in 2026. Anthropic has publicly acknowledged the importance of Google’s TPU stack to its training and operational strategies. Meta, another significant TPU user cited by Bloomberg, finalized its own agreement earlier this year. The capacity locked into these deals means there is less available for Google’s internal modeling teams, which must now wait.

      Demis Hassabis, the CEO of DeepMind, stated earlier this year that the limitations have dual aspects. Some of the issue lies in hardware availability—particularly concerning 'a few suppliers of a few key components’—with high-bandwidth memory from Samsung, Micron, and SK Hynix often mentioned as the primary bottleneck. Additionally, there is a challenge with research throughput; as Hassabis explained, researchers require substantial chip resources to test new ideas on a meaningful scale. The hardware limitations are partly beyond Google's control, while the internal allocation limits are not.

      The underlying figures are significant. Alphabet aims for a capital expenditure of between $175 billion and $185 billion for 2026 within a broader Big Tech AI infrastructure spending that has surpassed $650 billion this year. According to Google's comments, the company has been adding well over a gigawatt of new AI computing capacity in 2026. The long-term investment in TPUs is yielding the type of unit-economics advantage that allows the company to sell its chips, host its competitors’ models, and conduct its own advanced research on the same infrastructure. However, that infrastructure is no longer sufficient to support all three functions simultaneously.

      Bloomberg’s reporting highlights two clear indicators of the existing tension. Within the past 18 months, researchers such as Ioannis Antonoglou, a seasoned contributor at DeepMind, have left for startup opportunities—a trend that has intensified as securing access to computing resources within Google has become more challenging. Oren Etzioni, the former CEO of the Allen Institute for AI mentioned in the article, has publicly framed this situation as an expected outcome of an internal market in which computing resources are allocated based on managerial seniority rather than by the economic factors that dictate external customer contracts.

      Over the last 18 months, Google has found itself in a delicate situation: it requires its TPU program to show substantial traction with named external clients to validate its technology against Nvidia, while also maintaining enough internal capacity for Gemini training runs and DeepMind research. The four-partner inference-chip supply chain involving Broadcom, MediaTek, and Marvell is a strategy aimed at easing this constraint by increasing capacity downstream of TPU training, but it has yet to meet the scale needed to satisfy demand.

      Google did not contest Bloomberg’s perspective on internal allocation in its official statements; instead, it emphasized its overall infrastructure investment strategy and noted that computing constraints are an issue affecting the industry as a whole, not just Google. This assertion aligns with the evidence: every major model provider is, according to Q1 2026 earnings, constrained in computing resources relative to its research goals. What makes Google's situation particularly newsworthy is the juxtaposition: the company has concurrently become the largest infrastructure provider for its main competitors. Whether it can continue to sell this asset while using it is a question that the coming quarters will clarify.