AI infra deals harden competition
CoreWeave signed a multi‑year deal to power Anthropic’s Claude at production scale while OpenAI is telling investors it has a computing advantage, signalling that raw access to GPUs and capacity remains a primary competitive axis. At the same time, moves toward specialised and hybrid inference hardware — exemplified by SK Telecom partnering with Arm and Rebellions on CPU‑NPU servers — indicate architectures will increasingly split work between local and remote execution. (thenextweb.com) (letsdatascience.com) (koreatimes.co.kr)
The fight between big artificial intelligence labs is starting to look like the fight between airlines for airport gates: whoever locks up the scarce slots first gets to move more people. On April 10, CoreWeave said Anthropic signed a multi-year deal to run the Claude model family on CoreWeave’s cloud, with new computing capacity coming online later in 2026. (coreweave.com) CoreWeave is not a consumer app company. It rents out giant clusters of graphics processing units, the chips that train and run systems like Claude and ChatGPT, and it said this Anthropic deal is its second major infrastructure announcement in 48 hours after expanding its Meta agreement by $21 billion. (cnbc.com) That tells you where the bottleneck still is. The hardest part of selling artificial intelligence in 2026 is not finding customers for a chatbot; it is finding enough electricity, data center space, and graphics processing units to answer millions of prompts without slowing down. (bloomberg.com) (coreweave.com) OpenAI made that point directly to investors this week. In a memo seen by Bloomberg and CNBC, OpenAI argued that its early push to secure computing gives it a structural edge over Anthropic and said it plans to have 30 gigawatts of compute by 2030, while it expects Anthropic to have about 7 to 8 gigawatts by the end of 2027. (bloomberg.com) (cnbc.com) A gigawatt is power-plant scale. When an artificial intelligence company starts comparing itself in gigawatts instead of servers, it is saying the race is no longer just about better code; it is about who can finance and feed entire industrial sites full of chips. (cnbc.com) But there is a second shift happening at the same time. Not every artificial intelligence task needs the same kind of machine, so companies are starting to split the work the way a restaurant splits prep work between a central kitchen and the line cooks serving dinner. (koreatimes.co.kr) On April 10, SK Telecom said it would work with Arm and Rebellions on servers that pair a central processing unit, the general-purpose chip that handles broad system tasks, with a neural processing unit, a specialized chip built to run artificial intelligence inference more efficiently. (koreatimes.co.kr) (news.sktelecom.com) Inference is the moment a trained model answers your request. Training is like teaching a chef over months; inference is the chef making tonight’s meal, and SK Telecom said the new server design will be tested in its artificial intelligence data center for telecom and sovereign artificial intelligence workloads. (rebellions.ai) (koreatimes.co.kr) That is why these two stories fit together. CoreWeave and OpenAI are showing that the biggest labs still need vast remote clusters, while SK Telecom, Arm, and Rebellions are betting that more of the serving work will move onto mixed servers tuned for specific jobs instead of relying only on giant pools of graphics processing units. (coreweave.com) (bloomberg.com) (news.sktelecom.com) The result is a market that is getting more brutal, not simpler. The winners may not be the labs with the smartest model demo on a laptop screen, but the ones that can secure industrial-scale power for training and then route everyday inference onto cheaper, more specialized hardware without users noticing the handoff. (cnbc.com) (koreatimes.co.kr)
