Snowflake Bets $6 Billion on AWS Graviton as AI Agent Demand Scales
Table of Contents
A massive bet on the ‘plumbing’ of AI
Snowflake has entered into a five-year, $6 billion agreement with Amazon Web Services (AWS), a move that underscores the massive infrastructure shift required as generative AI moves from theoretical chatbots to active enterprise agents. While the headline figure is staggering, the real story lies in what Snowflake is actually buying: increased access to AWS’s homegrown Graviton ARM-based CPUs.
To put the scale of this deal into perspective, AWS notes that Snowflake has generated roughly $7 billion in total sales via the AWS Marketplace since its inception in 2012. This single contract essentially matches the entirety of that historical revenue stream, signaling an aggressive scaling phase for Snowflake’s compute needs. The company reports that customer spending on AWS is accelerating rapidly, with projections for 2025 hitting $2 billion for the calendar year alone.
The shift from GPUs to CPUs in the AI lifecycle
For the last two years, the AI narrative has been dominated by GPUs—specifically Nvidia’s H100s and B200s—which are essential for the heavy lifting of training large language models (LLMs) and complex reasoning. However, as enterprises deploy tools like Cortex AI, the compute requirements are shifting.
Cortex AI allows businesses to run LLMs directly on their data, enabling natural language database queries and automated summary reports. As these tools evolve into ‘AI agents’—software that can independently execute tasks rather than just answering questions—the burden of processing shifts toward the CPU. While GPUs handle the ‘thinking,’ CPUs handle the orchestration, data movement, and systemic execution that allow an agent to actually function within a corporate workflow.
By locking in a massive supply of Graviton chips, Snowflake is positioning itself to handle this increased ‘inference’ load more efficiently. Amazon’s ARM-based architecture is designed specifically for cloud efficiency, offering a price-performance ratio that is often superior to traditional x86 processors when handling the specific types of workloads associated with AI orchestration.
The cloud giants’ war against the Nvidia monopoly
This deal is part of a broader, quiet rebellion by cloud service providers (CSPs) against their dependence on Nvidia. Amazon CEO Andy Jassy has repeatedly emphasized that homegrown silicon provides better price-performance, a claim that is gaining traction with hyperscalers. Just last month, AWS secured a deal to provide millions of Graviton chips to Meta, a significant win considering Meta had previously signed a $10 billion deal with Google Cloud.
Google and Microsoft are following the same playbook. Google has leveraged its TPU (Tensor Processing Unit) for years, and Microsoft recently debuted its Maia AI chip. The goal for these companies is vertical integration: by designing the chip, the hypervisor, and the cloud environment, they can slash margins and offer lower costs to enterprises like Snowflake.
Nvidia, however, is not standing still. CEO Jensen Huang recently pointed to the launch of ‘Vera,’ a new AI-specific CPU designed to capture the very market AWS Graviton is targeting. Huang claims this represents a $200 billion market opportunity, with $20 billion in early sales already recorded.
For Snowflake, the choice of Graviton is a pragmatic one. By leveraging the efficiency of ARM architecture, they can scale their AI offerings without the prohibitive costs associated with GPU-only environments. As the industry moves toward an era of autonomous agents, the battle for the CPU may prove just as critical as the fight for the GPU.
