Snowflake Commits $6B to AWS Over Five Years for Graviton CPUs — Almost Its Entire Marketplace Lifetime Spend

Snowflake signed a five-year, $6 billion commitment to Amazon Web Services on May 27, with the spend pointed squarely at Amazon’s Graviton CPUs to run agentic AI and inference workloads — not Nvidia GPUs. The deal, disclosed in Snowflake’s FY2027 Q1 8-K and reported by GeekWire and Yahoo Finance, is one of the largest single CPU commitments AWS has ever booked from a software company.

The most telling number is the comparison. Snowflake has run on AWS as its primary cloud since the company was founded in 2012, and over fourteen years it has booked roughly $7 billion of revenue through AWS Marketplace — its entire commercial history on the platform. This single new contract is worth $6 billion in five years. Annual AWS spending by Snowflake’s customers had already doubled in 2025 to roughly $2 billion, driven mostly by demand for Cortex AI, the company’s natural-language query and automated-reporting layer on top of the warehouse.

What is unusual is where the money goes. The commitment is framed around Graviton — AWS’s ARM-based CPUs — rather than Nvidia GPU instances. Snowflake and AWS describe the workload as the “rest of the AI stack”: the agentic loops, retrieval, orchestration and structured-data joins that surround a model call, all of which scale on CPU rather than GPU. Andy Jassy has spent 2026 pitching Graviton as offering “better price-performance” than Nvidia at the rack level, and Snowflake’s public bet of $6 billion is the largest external endorsement of that pitch to date.

It also is not an isolated deal. In April, BitsMinds covered Meta’s multi-billion Graviton commitment — hundreds of thousands of chips over at least three years — which made Meta one of AWS’s top five Graviton customers. Snowflake is now the second multi-billion Graviton anchor inside seven weeks. In parallel, AWS continues to scale Trainium and Inferentia for the GPU side of the workload, and on the buy-side Nvidia CEO Jensen Huang has been previewing a new Vera CPU as a “brand new $200 billion market” to keep CPU dollars from quietly migrating to ARM.

For Snowflake the strategic logic is more boring and more important: lock in five years of CPU at negotiated pricing before agentic query volumes — the kind Cortex AI generates by default for every business user — explode against an inflexible budget. For AWS it is balance-sheet validation that the next leg of AI revenue is not just GPU rentals at $40-an-hour but a much larger, much steadier CPU stream sitting underneath them. The training arms race got the trillion-dollar headlines in 2025. In 2026, the inference arms race is starting to get the trillion-dollar contracts.