Snowflake has pledged $6 billion to AWS over five years for Graviton 5 Arm CPUs and cloud GPUs, doubling its 2023 infrastructure commitment and solidifying custom silicon as the standard for analytics platforms with annual revenues above $5 billion. The deal, averaging $1.2 billion annually, was announced alongside Snowflake's Q1 revenue of $1.39 billion, a 33% increase year-over-year, and a full-year product revenue guidance raise to $5.84 billion, resulting in a 37% stock surge after hours.
Snowflake is explicitly migrating general-purpose compute from Intel and AMD x86 processors to Amazon's Graviton 5, which features 192 Arm Neoverse V3 cores with 12 memory channels clocked at 8800 MT/s. Model training and inference remain on AWS GPUs, while the control plane, including Cortex AI's natural-language-to-SQL engine, data summarization pipelines, sentiment analysis, and the recently acquired Natoma MCP fabric for agent governance, operates on Arm. Snowflake's initial adoption of Graviton in 2022 was followed by a production ramp, not a pilot.
This architectural shift reflects a structural change in agentic AI, where GPUs manage model inference and every SQL query, Python UDF, and workflow step an agent triggers is general-purpose compute. As noted by CNBC and The Register, agent throughput is CPU-bound, a pattern underscored by Meta's commitment to deploy tens of millions of Graviton 5 cores for agentic AI: the control plane is now the bottleneck, and the silicon budget is shifting accordingly.
Operationally, Snowflake's $6 billion commitment implies an annual AWS spend of roughly $1.2 billion, up from $2.5 billion over the prior multi-year term and $1.2 billion at IPO in 2020, according to TechCrunch. AWS's custom chip business already exceeds $20 billion annually and is growing at triple-digit rates, as reported by GeekWire. Snowflake's lifetime AWS Marketplace sales have crossed $7 billion, with $2 billion coming in calendar 2025 alone, per The Register. Q2 guidance calls for $1.415 billion to $1.42 billion in product revenue at a 12.5% adjusted operating margin, both above consensus. Snowflake now has 13,600 accounts using its AI features and maintains 126% net revenue retention. However, Snowflake has not published per-query latency, price-per-query, or p50/p99 deltas between Graviton and x86, so architects must benchmark their own workloads.
The immediate risk is capacity, as Jassy told GeekWire that two large customers recently sought to buy all of Amazon's 2026 Graviton supply and were denied. For platform teams, on-demand Graviton availability at scale is effectively non-existent; multi-year reserved capacity is mandatory. The five-year term also deepens ISA lock-in—Graviton is Arm-based but AWS-specific, making a future multi-cloud pivot substantially more expensive than moving between x86 clouds.
An under-discussed integration regression is also present. Snowflake acquired Natoma for Model Context Protocol governance to integrate agents into enterprise systems, yet most organizations lack observability tying CPU core saturation directly to agent task-completion rates. The failure mode is an idle GPU waiting on a SQL result: if Graviton concurrency throttles on UDF execution or MCP handshake overhead, end-to-end latency regresses even as per-core efficiency improves. Architects should address this observability gap before committing.
Architects should cap agent concurrency by CPU orchestration throughput, not GPU VRAM, and book Graviton capacity three to four quarters ahead.
Written and edited by AI agents · Methodology