Snowflake Commits $6B to AWS Graviton and Agentic AI Workloads

Snowflake announced on May 27, 2026 a $6 billion, five-year infrastructure commitment to Amazon Web Services, its largest cloud spend pledge ever, centered on AWS Graviton Arm CPUs plus GPU capacity for Cortex AI and agentic enterprise workloads. The same week Snowflake reported earnings that sent shares up about 36%, Amazon highlighted more than $7 billion in lifetime AWS Marketplace sales for Snowflake and $2 billion+ in calendar 2025 marketplace volume.

For developers, the deal is a signal: agentic AI is moving from demos to always-on SQL, orchestration, and inference bills, and hyperscalers want those cycles on custom silicon, not only Nvidia GPUs.

What did Snowflake and AWS announce?

The strategic collaboration agreement expands joint go-to-market, migrations, and industry solutions. Snowflake said the $6B covers Graviton compute and AI spend on AWS through roughly 2031, averaging about $1.2 billion per year.

Context on deal size:

CNBC noted the new pledge is roughly five times the original 2020 figure and about 2.4x the 2023 arrangement.

Snowflake was founded on AWS in 2011; most customers still run there, though multi-cloud marketing continues.

Why Graviton matters for agentic AI (not only GPUs)

Headlines focus on AI GPUs, but Snowflake and AWS emphasized Graviton for the CPU-heavy parts of agent workflows:

• SQL generation and execution from natural language (Cortex)
• Data summarization and sentiment pipelines
• Workflow orchestration between tools and warehouses
• Always-on warehouse compute as agents run continuously

AWS launched Graviton in 2018; Snowflake began adopting it in 2022. In 2026, CPU demand is rising again because agents fire many small queries and control-plane tasks between GPU inference bursts.

That split matters for architecture reviews: not every agent token needs an H100; mis-sizing everything to GPUs is how bills explode.

Marketplace, Cortex, and the agentic enterprise thesis

Amazon said Snowflake passed $7B lifetime AWS Marketplace sales and doubled transaction growth year over year in 2025, with $2B+ in calendar 2025 marketplace sales alone.

Snowflake positions Cortex AI as governed enterprise AI on warehouse data: natural language to SQL, summarization, and agent tooling (including products like Snowflake Intelligence and partner integrations cited in press materials such as Hex and Fetch deployments).

The $6B bet says Snowflake expects those workloads to be persistent infrastructure, not one-off notebooks.

How this fits the May 2026 AI infrastructure stack

• Anthropic $965B and Amazon's Anthropic Trainium spend show foundation-model capex
• Cerebras IPO pop highlights inference economics
• RAMageddon raises cost of memory-bound stacks
• Snowflake's deal anchors enterprise data-plane spend on AWS custom CPUs + selective GPUs

Meta reportedly plans tens of millions of Graviton cores for agent infrastructure as well, per May coverage, so Snowflake is not alone in the Graviton-for-agents pattern.

Developer implications

If you build on Snowflake: expect Cortex and agent features to assume AWS Graviton paths and Marketplace procurement. Benchmark workloads on current Graviton generations before assuming x86.

If you compete with Snowflake: the lock-in depth just increased; multi-cloud failover stories need dollar comparisons against $1.2B/year AWS gravity.

If you model cloud costs: split agent architectures into CPU orchestration vs GPU inference lines; use the LLM API pricing tracker for model calls and warehouse pricing for SQL-heavy agent loops.

If you procure via AWS Marketplace: Snowflake's simplified contracting path is now strategic, not cosmetic, given $2B+ annual marketplace velocity.

Risks and skepticism

IDC's Dave McCarthy called the deal proof enterprise AI is crossing from experiments to foundational infrastructure. ISG's Anay Nawathe cautioned not to treat the full $6B as pure inference GPU burn; much is long-duration CPU and platform commit.

Snowflake remains AWS-dependent for the bulk of compute even as it markets multi-cloud. Graviton generation details were not fully specified; CEO Sridhar Ramaswamy pointed to more disclosure at AWS re:Invent 2026.

Key Takeaways

• May 27, 2026: Snowflake commits $6B over 5 years (~$1.2B/year) to AWS, largest pledge to date
• Graviton CPUs anchor cost-efficient SQL, orchestration, and agent control-plane work; GPUs cover accelerated AI
• AWS Marketplace: $7B+ lifetime Snowflake sales; $2B+ in 2025 alone
• Stock reaction: shares rose about 36% on earnings plus deal news (CNBC)
• For developers: design agents with CPU/GPU split; expect deeper Snowflake-AWS integration and Marketplace procurement paths
• What to watch: re:Invent Graviton generation details, Cortex agent pricing, and whether Azure/GCP counters with similar commits

Frequently asked questions

How much is Snowflake spending on AWS?

Snowflake committed $6 billion over five years to AWS for Graviton compute and AI infrastructure, announced May 27, 2026, averaging about $1.2 billion per year.

Why Graviton instead of only Nvidia GPUs?

Agentic enterprise workloads spend significant compute on SQL, data movement, and orchestration. Graviton gives Snowflake better price-performance on those CPU-heavy paths while GPUs handle accelerated model tasks.

How big is Snowflake on AWS Marketplace?

Amazon said Snowflake surpassed $7 billion in lifetime AWS Marketplace sales and exceeded $2 billion in calendar 2025 marketplace sales, with transaction growth more than doubling year over year.

Does this replace Snowflake multi-cloud?

No. Snowflake still markets multi-cloud, but the $6B deal deepens AWS dependence for the majority of customer workloads and infrastructure spend.

What should data engineers do with this news?

Plan Cortex and agent features as production infrastructure on AWS, benchmark Graviton SKUs for warehouse workloads, and model agent costs as continuous CPU plus episodic GPU inference rather than GPU-only budgets.