Hyperscaler spending reorders bargaining power

Amazon just put a number on the table that most cloud rivals and most artificial intelligence startups cannot match: Amazon Web Services said its artificial intelligence business is now running at more than $15 billion a year, and Andy Jassy said Amazon expects about $200 billion of capital spending in 2026, mostly for artificial intelligence infrastructure. (aboutamazon.com) That changes the conversation from “whose model looks smartest in a demo” to “who can actually deliver enough chips, power, and data center space on time.” Jassy wrote that Amazon can monetize capacity as soon as it comes online because demand is already there. (aboutamazon.com) (constellationr.com) A hyperscaler is just a cloud company so large it can build computer warehouses the way a railroad once laid track. Amazon Web Services, Microsoft Azure, and Google Cloud are the three names that matter here because they buy the chips, finance the buildings, and rent the computing to everyone else. (aws.amazon.com) (microsoft.com) (abc.xyz) The bottleneck is not software alone. Training and running large artificial intelligence models requires thousands of specialized chips linked together, plus substations, cooling systems, networking gear, and long construction lead times for new data centers. (nvidia.com) (s206.q4cdn.com) That is why capital spending has exploded. Alphabet said it spent $91.4 billion in 2025 and guided 2026 capital spending to $175 billion to $185 billion, while Microsoft said it planned to spend $80 billion in fiscal 2025 on artificial intelligence-ready data centers. (s206.q4cdn.com) (finance.yahoo.com) (cnbc.com) Once spending reaches that scale, bargaining power shifts toward the company writing the checks. If a cloud provider can pre-buy enough NVIDIA Blackwell systems or build enough custom chips years ahead, customers and model companies have fewer alternatives when they need capacity fast. (nvidia.com) (aws.amazon.com) Amazon’s answer is to reduce its dependence on outside suppliers where it can. Amazon Web Services says its Trainium family and Graviton processors are custom silicon designed to improve price-performance, and its Trn2 instances use 16 Trainium2 chips for large model training and inference. (aws.amazon.com 1) (aws.amazon.com 2) (aws.amazon.com 3) Custom silicon matters because the cloud provider that designs the chip also controls the supply, the pricing, and the software stack around it. Jassy said two large Amazon Web Services customers asked if they could buy all of Amazon’s Graviton capacity for 2026, which is a blunt sign that scarce compute is becoming a contract weapon. (constellationr.com) The same logic is showing up in model partnerships. Amazon and Anthropic built Project Rainier around nearly 500,000 Trainium2 chips, and Anthropic said Claude is already training and serving on almost 1 million Trainium2 chips across that relationship. (aws.amazon.com 1) (aws.amazon.com 2) Microsoft is playing the same game from the other direction by locking in giant NVIDIA clusters. Microsoft said it built the first large-scale production cluster with more than 4,600 NVIDIA GB300 NVL72 systems for OpenAI workloads and plans to scale to hundreds of thousands of Blackwell Ultra graphics processors across its data centers. (azure.microsoft.com) So the new pecking order in artificial intelligence is being set less by a flashy chatbot launch and more by who controls the industrial base underneath it. In 2026, the strongest negotiating position belongs to the company that can say yes to 100,000 chips, 1 gigawatt of power, and a signed multi-year contract on the same call. (aboutamazon.com) (azure.microsoft.com) (s206.q4cdn.com)