Snowflake signed a $6 billion five-year agreement with Amazon Web Services on Wednesday for expanded access to AWS's Graviton CPU chips, a contract that nearly matches Snowflake's entire $7 billion AWS Marketplace revenue since the company's 2012 founding. The deal underscores how rapidly AI workloads are shifting enterprise cloud spending toward CPU-heavy infrastructure as agentic AI moves from training to production deployment.
Snowflake's AWS spending doubled to $2 billion in calendar year 2025, driven by customers accelerating adoption of Cortex AI, the company's data-layer AI tool that provides natural-language database queries and automated reporting. The spike in demand reflects a broader architectural reality: while GPUs handle model training and inference, CPUs handle the orchestration, memory management, and task execution for AI agents running in production.
The contract positions AWS's homegrown ARM-based Graviton chips as a credible alternative to Nvidia's CPU offerings. Amazon CEO Andy Jassy said last month that Amazon's AI chips offer better price-performance than Nvidia's, a claim AWS reinforces by passing cost savings to customers. AWS still deploys Nvidia GPUs widely, but the Graviton bet is a deliberate play for the exploding CPU-driven segment of AI infrastructure.
“better price-performance”— Andy Jassy, Amazon CEO
Key facts
- 01Snowflake signed a $6 billion five-year agreement with AWS for Graviton chips, doubling its AWS Marketplace revenue since 2012.
- 02Snowflake's AWS spending doubled to $2 billion in 2025 as customers accelerated AI workloads through Cortex AI.
- 03Meta signed a deal for millions of Graviton chips last month, following a $10 billion Google Cloud deal months earlier.
- 04Nvidia's Vera CPU represents a $200 billion market, with $20 billion already sold according to Jensen Huang.
- 05AWS, Google, and Microsoft are all deploying homegrown AI chips to compete with Nvidia on price and availability.
AWS closed a similar multi-billion-dollar Graviton deal with Meta last month, a notable win after Meta had signed a $10 billion deal with Google Cloud months earlier. Meta's willingness to split its AI compute across multiple cloud providers signals that no single vendor yet owns the agentic-AI infrastructure market, and that price and availability matter as much as raw performance.
Google has manufactured its own AI chips for years, and Microsoft launched its Maia AI chip in January. The three hyperscalers are collectively attempting to carve out margin and lock in customers by vertically integrating chip design, a strategic shift that directly threatens Nvidia's dominance in AI hardware.
Nvidia is not ceding the market quietly. CEO Jensen Huang said last week that the company's new Vera CPU represents a $200 billion market opportunity for Nvidia, with $20 billion already sold. Vera is architected specifically for AI workloads, positioning Nvidia to defend its turf in both training and inference while now competing in the CPU layer that agents depend on.
The Snowflake deal matters less for the headline number than for what it reveals about enterprise AI spending: the bottleneck is shifting. Training large models remains capital-intensive, but the marginal cost of running millions of agents in production is increasingly CPU-bound. Whoever wins the CPU war wins the recurring revenue stream that compounds as AI agents proliferate across enterprise workflows.
AWS's bet is that price-conscious enterprises will choose Graviton over Nvidia when performance differences narrow and cost savings compound at scale. If that bet pays off, AWS captures both the cloud hosting revenue and the chip margin, a vertical integration play Amazon has executed successfully in logistics and retail. The risk is that Nvidia's software moat — models and frameworks optimized for its architecture — makes switching costs prohibitive even when alternatives are cheaper.
Snowflake's doubled AWS spend in 2025 suggests that enterprises are already treating AI infrastructure as a recurring operational expense rather than a capital experiment. The five-year commitment locks in that trajectory, and it positions AWS to replicate this playbook with every other AI-native platform still running multi-cloud or evaluating chip options.
The cloud providers are extracting value from the AI boom regardless of which model lab wins the frontier race. AWS, Google Cloud, and Azure host the training runs, serve the inference workloads, and now manufacture the chips those workloads run on. That vertical integration gives them leverage over both the model providers who rent their infrastructure and the enterprises who consume the models as services.
Nvidia still holds the high ground in GPU-accelerated training and inference, but the CPU war is a new front where AWS has home-field advantage. The company that built its cloud empire on commodity x86 servers is now betting that ARM-based Graviton chips can do to Nvidia's CPU ambitions what AWS did to on-premise data centers: make them economically obsolete through scale and relentless cost optimization.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
