AI Compute Race Helps Vendors

Jefferies said this week that the AI boom is sending more power to cloud and chip suppliers than to model makers alone. (proactiveinvestors.com) After Google Cloud Next, the firm named Amazon, Microsoft, Alphabet, Oracle and CoreWeave as the main beneficiaries of continued compute shortages. It said demand for AI infrastructure still exceeds available supply. (proactiveinvestors.com) Jefferies said the tightest pinch remains in graphics processors, the chips used to train and run large AI models, across Nvidia’s Hopper and Blackwell generations. It also said spot availability for older chips remains limited. (proactiveinvestors.com) That scarcity is pushing customers toward committed-capacity contracts, where they reserve compute in advance instead of buying it only when needed. Jefferies said that shift supports higher spending and better visibility for hyperscale cloud providers. (proactiveinvestors.com) The numbers are large. Jefferies estimated capital spending across cloud providers will rise roughly 64% in 2025 and 58% in 2026 as companies keep building data centers, networking and chip capacity for AI workloads. (proactiveinvestors.com.au) A fresh AWS-Meta agreement shows what those commitments look like in practice. Amazon said on April 24 that Meta will deploy AWS Graviton processors at scale, starting with tens of millions of Graviton cores and potentially expanding further. (aboutamazon.com) Amazon said those chips will support Meta’s “agentic” AI workloads, including systems that handle real-time reasoning, code generation, search and multi-step task orchestration. Amazon also said Meta is now one of the largest Graviton customers in the world. (aboutamazon.com) Jefferies said enterprises are also trying to control rising token costs by using model-agnostic software layers that can route work across different models. That favors providers that own not just models, but the cloud, chips and tooling around them. (proactiveinvestors.com) The same logic is showing up in other capacity deals. Jefferies said Anthropic’s growth has been helped by an agreement with Google and Broadcom for about 3.5 gigawatts of next-generation TPU compute expected in 2027. (proactiveinvestors.com) The result is an AI market where access to electricity, servers, chips and reserved cloud capacity is becoming a product in its own right. For now, the vendors selling that access appear to have the strongest hand. (proactiveinvestors.com)