OpenAI and Anthropic now drive roughly half of hyperscalers' AI orders, internal tracking shows

Cloud infrastructure used to be sold like retail. Lots of customers. Lots of workloads. Steady diversification. AI is pushing it toward something much more concentrated. A handful of frontier model companies now absorb so much compute that they are starting to shape how the biggest cloud providers build, finance, and even site new capacity. That shift came into focus over the past few weeks because two deals put real numbers on it. Anthropic said on April 20 that it will spend more than $100 billion on AWS technologies over 10 years, while securing up to 5 gigawatts of capacity and taking another $5 billion from Amazon, with up to $20 billion more possible. OpenAI, meanwhile, has been scaling Stargate with Oracle and SoftBank, including a July 2025 Oracle agreement for 4.5 additional gigawatts and a broader path toward $500 billion of U.S. AI infrastructure. ### Why are these deals such a big deal? Because gigawatts are not normal cloud language. They are utility-scale language. When an AI lab books 4.5 or 5 gigawatts, it is not renting a few more server clusters — it is effectively helping determine how many data centers get built, where power gets procured, and which chips get prioritized. BloombergNEF says global data center IT capacity under construction topped 23 gigawatts, with 15.9 gigawatts in the U.S. alone. (anthropic.com) A single customer commitment can now move the industry totals. ### Why OpenAI and Anthropic? Because they are no longer just model vendors. They are infrastructure anchors. OpenAI raised $40 billion in March 2025 specifically to scale compute infrastructure, and Anthropic’s AWS deal locks in a decade-long demand stream tied to Amazon’s Trainium roadmap. These companies are becoming the tenants that justify hyperscaler capex the way a giant retailer might justify a warehouse buildout. (anthropic.com) ### What does the hyperscaler get? Revenue visibility, first. Strategic pull, second. Amazon gets Anthropic as a primary training customer and a showcase for Trainium chips. Oracle gets OpenAI as the flagship tenant for Stargate capacity. The catch is that these are not plain customer relationships. They are intertwined with equity stakes, custom silicon roadmaps, and long-dated infrastructure plans. That makes the top line look stronger, but it also makes each provider more exposed to a few very large counterparties. (openai.com) ### Why does concentration matter? Because diversification used to be the safety valve. If one enterprise customer slowed spending, the cloud kept humming. But if a meaningful slice of future AI demand sits with two labs, then delays in model launches, fundraising, regulation, or chip transitions can ripple straight into hyperscaler utilization assumptions. That is why investors have started treating OpenAI and Anthropic almost like shadow business units inside the cloud giants. (anthropic.com) ### Where do sovereign deals fit in? They make the same pattern more geopolitical. Separate from the lab-hyperscaler tie-ups, sovereign compute projects are getting huge. Deloitte expects more than $100 billion to be committed to sovereign AI compute in 2026. That means governments are joining frontier labs as anchor buyers of scarce capacity, often with national-security or industrial-policy goals attached. (cnbc.com) ### So what changes next? Capacity planning gets less generic. More power contracts. More chip-specific lock-in. More multi-cloud arrangements that are really risk-sharing deals between a few giant players. Anthropic already spans AWS, Google Cloud, and Microsoft Azure, but AWS is clearly the center of gravity. OpenAI has multiple partners too, but Stargate makes Oracle central in a way that would have seemed unlikely two years ago. (deloitte.com) ### Bottom line The important change is not just that AI demand is huge. It is that the demand is clumping. When two model companies can justify multi-gigawatt buildouts and $100 billion-plus commitments, hyperscalers stop acting like broad malls and start acting like landlords built around a few anchor tenants. (anthropic.com)