Anthropic leak and compute strain
Social reporting claims Anthropic accidentally leaked a full‑stack coding product and notes Opus—their model—has degraded under compute shortages, while rivals are locking long‑term deals to secure capacity. These posts portray tension between product development, operational constraints and competitive responses in the compute market. (x.com/ziwenxu_/status/2043395528004624645, x.com/haider1/status/2043318397878878373)
Anthropic is juggling two problems at once: an accidental Claude Code source leak and a fresh scramble to lock down enough computing power. (bloomberg.com, anthropic.com) On March 31, Anthropic said a packaging mistake exposed source code for Claude Code, its coding assistant, through version 2.1.88 of its npm package. Bloomberg reported the leak covered about 1,900 files and 512,000 lines of code, and Anthropic said it was caused by “human error,” not a breach. (bloomberg.com, thehackernews.com) Anthropic then sent GitHub copyright takedown notices to remove mirrors of the leaked code, but TechCrunch reported the effort briefly swept up thousands of unrelated repositories before the company reversed course. Anthropic told Bloomberg it had changed its automation process after the release error. (techcrunch.com, bloomberg.com) A coding product like Claude Code is a wrapper around a model: the model writes text, while the software around it decides what files it can read, what commands it can run, and how it loops through a task. That outer layer is where many of the practical product choices live, so a source leak can show rivals how a tool is orchestrated even when the model weights stay private. (thehackernews.com, eweek.com) At the same time, Anthropic has been describing infrastructure as a bottleneck. In a September 17, 2025 engineering post, the company said three infrastructure bugs had intermittently degraded Claude’s response quality, and in April 2026 it announced a new Google and Broadcom agreement for multiple gigawatts of next-generation TPU capacity starting in 2027. (anthropic.com, anthropic.com) Anthropic’s recent announcements show how large the capacity race has become. On October 23, 2025, it said it planned to expand use of Google Cloud technologies by up to one million tensor processing units, with well over a gigawatt expected online in 2026, and on April 10, 2026, CoreWeave said it had signed a multi-year agreement to run Anthropic workloads at production scale. (anthropic.com, coreweave.com) Rivals are making similarly long-dated bets. OpenAI said in July 2025 that its Oracle partnership would bring Stargate to more than 5 gigawatts of capacity under development, and in a later infrastructure update it said Stargate had grown to nearly 7 gigawatts of planned capacity while OpenAI also expanded work with NVIDIA and CoreWeave. (openai.com, openai.com, openai.com) Anthropic has also tied customer growth directly to the need for more chips. In its April 6 announcement, the company said run-rate revenue had surpassed $30 billion, up from about $9 billion at the end of 2025, and that the number of business customers spending more than $1 million on an annualized basis had doubled to more than 1,000 in less than two months. (anthropic.com) The immediate question is whether Anthropic can keep shipping coding tools while tightening release controls and adding capacity fast enough to stabilize performance. Its own statements point in both directions: more safeguards after the leak, and more multi-year compute contracts to keep Claude running at scale. (bloomberg.com, anthropic.com, coreweave.com)