Furthermore, the biggest beneficiary of today’s news is neither Anthropic nor OpenAI—it’s Google Cloud.
More than 20% of total TPU shipments from 3Q26 to 4Q27 are being sold directly to Anthropic.
This is not another card about Google's DeepMind leadership reshuffle. The question is where the change pushes Google's money and compute allocation, and which numbers would show that direction to Alphabet investors. SemiAnalysis makes a clear case. DeepMind is falling behind in frontier-model competition as key researchers leave, while GCP can fill the gap by selling TPUs and cloud services to outside customers. Supplying compute to rival model companies such as Anthropic can lift near-term cloud revenue, while reducing the pool of compute that Google can keep available for Gemini research.
The original post groups the departures of Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals into a signal about talent. That reshuffle is already known. The new information here is the next step: resource allocation. SemiAnalysis estimates that more than 20% of TPU shipments from the third quarter of 2026 through the fourth quarter of 2027 will go directly to Anthropic.
Google's official description shows GCP as both internal infrastructure for its own models and a platform sold to outside customers. If the platform grows by hosting rival models such as Claude, model rankings and cloud revenue can move in different directions. That protects Alphabet from turning Gemini's weakness directly into a group-wide earnings problem, while forcing investors to ask which business Google is choosing to prioritize over time.
The figures come from Google Cloud's official product description of its training and inference TPUs. They are independent infrastructure figures, separate from SemiAnalysis's shipment forecast.
2026-04-22 · Google Cloud
SemiAnalysis estimates that GCP could reach more than $73 billion in third-party AI infrastructure annualized revenue by 2027, with TPU system sales taking the external figure much higher. These are model estimates, not Alphabet guidance, so they should not be read as a reported forecast. The investable point is the path: once customers commit to buying compute, the cloud ledger can expand even while the model leaderboard moves against Gemini. The tradeoff is the opportunity cost of capacity. Selling more TPUs can speed up infrastructure payback and GCP growth. Long-term allocations to competing model labs also reduce the spare capacity Google can reclaim when Gemini needs it. Platform neutrality can create revenue while dispersing the focus of the research lab.
The next question is not whether Gemini immediately returns to number one. It is whether Gemini API usage and retention of key researchers change while GCP growth and TPU backlog keep rising. Put external GCP revenue, TPU shipments and backlog, and first-party Gemini API token growth in one table. The combination will show whether Google is rebuilding a model-first strategy or committing more deeply to an infrastructure-first one.