Cloud compute getting concentrated
- Major AI compute capacity is being locked up through big bilateral cloud and chip deals. - Meta signed for millions of Amazon AI CPUs while Anthropic secured long-term AWS Trainium commitments. - Those pacts concentrate training and inference resources, raising competition for smaller robotics firms that lack hyperscaler ties ( ).
The market for AI computing is tightening around a handful of giant private deals, with Meta and Anthropic now locking in Amazon chip capacity at unusual scale. (techcrunch.com) Meta signed a deal to use millions of Amazon Web Services Graviton chips for AI workloads, Amazon announced Friday, April 24. TechCrunch reported the chips are CPUs rather than the graphics processors that have dominated recent AI spending. (techcrunch.com) Four days earlier, Amazon said Anthropic would commit $100 billion over 10 years to Amazon Web Services technologies and secure up to five gigawatts of Trainium capacity. Amazon said the package also includes current and future Trainium generations and tens of millions of Graviton cores. (aboutamazon.com) Trainium is Amazon’s in-house chip for training AI models, the step where a model learns from huge datasets, while Inferentia is built for inference, the step where a trained model answers prompts. Amazon says Trainium1 can cut training costs by up to 50% versus comparable Amazon Elastic Compute Cloud instances, and Inferentia2 can deliver up to four times higher throughput than the first Inferentia chip. (aws.amazon.com, aws.amazon.com) Those commitments come after Anthropic also expanded a separate compute partnership with Google and Broadcom this month, according to Futurum. The analyst firm said that arrangement gives Anthropic dedicated artificial intelligence infrastructure beyond its Amazon relationship. (futurumgroup.com, futurumgroup.com) The shift is moving cloud capacity away from the old model of mostly on-demand rentals and toward long-term reservations by a small set of buyers with enough scale to sign bilateral contracts. Futurum said Anthropic’s Amazon pact alone would anchor one leading model lab to a single cloud provider for a decade. (futurumgroup.com) Meta is not relying only on outside suppliers. In March, Meta said its own Meta Training and Inference Accelerator program would add four new chip generations within two years, with custom silicon central to its artificial intelligence infrastructure strategy. (about.fb.com) Amazon is also building more services around its chips instead of selling only raw capacity. In March, Amazon Web Services said a new Cerebras collaboration would pair Trainium for one part of inference with Cerebras systems for another, and make the service available through Amazon Bedrock. (press.aboutamazon.com) For smaller AI companies, including robotics startups that need both training runs and steady inference capacity, the practical choice is narrowing to whatever compute remains after the largest customers pre-book supply. The result is a cloud market where access increasingly depends less on spot demand and more on who already has a hyperscaler at the table. (techcrunch.com, futurumgroup.com)
