Meta’s massive Graviton buy
- Meta agreed on April 24 to use hundreds of thousands of Amazon Web Services Graviton chips for at least three years, adding another major compute contract to its 2026 artificial intelligence buildout. - Amazon said the deal makes Meta one of Graviton’s five biggest customers; CNBC reported Meta had already made $48 billion of recent AI infrastructure commitments with CoreWeave and Nebius. - The purchase extends Meta’s shift toward a mixed chip stack spanning CPUs, GPUs and in-house silicon as AI capacity tightens. (cnbc.com)
Meta agreed to use hundreds of thousands of Amazon Web Services Graviton chips in a deal that runs at least three years. (cnbc.com) Graviton is Amazon’s Arm-based central processor, the kind of chip that handles broad computing work rather than the graphics processors used for the heaviest AI training jobs. Amazon and Meta said the chips will help run CPU-intensive workloads tied to agentic artificial intelligence. (cnbc.com) (finance.yahoo.com) CNBC reported Meta had already committed a combined $48 billion in recent weeks to CoreWeave and Nebius, two suppliers that rent access to Nvidia graphics processing units. Amazon did not disclose the dollar value of the Graviton agreement. (cnbc.com) That mix shows how Meta is spreading its AI demand across several kinds of hardware at once. The company has been buying Nvidia capacity, working with outside cloud providers and rolling out new versions of its own Meta Training and Inference Accelerator chips. (cnbc.com) (finance.yahoo.com) Amazon’s Nafea Bshara told CNBC that Meta had used Graviton on a small scale before and that this expansion will make Meta one of the top five Graviton customers. He said Graviton is used for pre-training by many foundation model companies. (cnbc.com) Meta’s infrastructure push is landing as the company says 3.6 billion people use its apps every day and it expects to operate 32 data centers once a new Oklahoma site is complete. CNBC also reported Meta disclosed plans on April 23 to cut about 8,000 jobs, or 10% of its workforce. (cnbc.com) Amazon has pitched Graviton as a lower-energy, lower-cost option for many cloud workloads, while Nvidia-style graphics processors remain the scarce resource for the largest model training runs. Meta’s latest deal suggests it is reserving premium GPU capacity for the work only GPUs can do and pushing other tasks onto cheaper processors. (cnbc.com) The result is not a switch away from Nvidia so much as a bigger architecture split inside Meta’s AI systems. Meta is buying more compute wherever it can get the right chip for the right job. (cnbc.com 1) (cnbc.com 2)
