Snowflake has signed a $6bn deal with Amazon Web Services, deepening an existing relationship that tells you something about where enterprise AI spending is actually going.
The deal centres on AWS's Graviton CPU chips, not Nvidia GPUs. That distinction matters.
The CPU moment
The AI narrative has been dominated by GPUs for three years. Nvidia's stock price reflects the assumption that every meaningful AI workload requires its hardware. That is true for training large models. It is increasingly less true for running them.
AI agents, the workloads enterprises are actually deploying, often run more efficiently on CPUs. Inference at scale does not always need the brute-force parallel processing that GPUs provide. For many enterprise applications, CPUs are cheaper, more available and good enough.
Snowflake's decision to lock in CPU capacity rather than fight for scarce GPU allocation is pragmatic. It also suggests that the company sees its future in running AI workloads, not training models from scratch.
What Amazon gets
For AWS, the deal strengthens its position as the default infrastructure provider for enterprise data workloads. Snowflake already runs heavily on AWS. Six billion dollars of committed spend makes that relationship structural rather than transactional.
It also validates Amazon's investment in custom silicon. Graviton chips are Amazon's own design, which means every workload that runs on Graviton instead of Nvidia is margin Amazon keeps rather than margin it sends to Jensen Huang.
The rental economy wins again
The broader trend is clear. Even well-funded companies prefer renting compute to owning it. SpaceX rents. Snowflake rents. The economics of building and maintaining your own infrastructure only work at a scale that very few companies reach.
For everyone else, the cloud providers have won. The only remaining question is which cloud provider captures the most spend, and deals like this suggest AWS is not losing its grip.
