Snowflake just signed a five-year, $6 billion agreement with Amazon Web Services. Let that sink in for a second. Snowflake has generated roughly $7 billion in total revenue through AWS Marketplace since its founding back in 2012. So this single contract is essentially a "do it all again, but faster" bet—compressed into five years instead of thirteen.

That's not a routine vendor renewal. That's a signal.

How Did We Get Here?

Snowflake has always called AWS home. Yes, it eventually spread to Azure and Google Cloud—that's just what enterprise software does when it wants to look platform-agnostic in sales meetings—but AWS is where its roots are. And apparently, its customers have been spending aggressively there. Snowflake says its AWS-routed customer spend doubled in 2025, hitting $2 billion for that calendar year alone. That kind of growth trajectory makes a $6 billion forward commitment look less like a gamble and more like arithmetic.

The engine behind all of this? AI, obviously. Snowflake has been pushing its Cortex AI platform for a couple of years now, and the value proposition is actually coherent: enterprises already store enormous amounts of structured data in Snowflake, so layering AI features on top—natural language querying, automated summaries, anomaly detection—makes sense in a way that a lot of bolt-on AI products don't. You're not moving data to the AI. The AI comes to the data. That's the right architecture.

The Graviton Angle Is the Interesting Part

Here's where this deal gets technically juicy. A significant piece of this $6 billion commitment is earmarked for access to AWS's homegrown ARM-based CPU, Graviton—not GPUs. And that distinction matters more than most headlines are letting on.

The industry's collective imagination has been captured by GPU scaling and training runs, but inference workloads and—increasingly—agentic AI systems are a different beast. Agents aren't just running a single forward pass through a neural network. They're orchestrating tools, managing state, parsing outputs, calling APIs, looping through decision trees. That's CPU-heavy work. GPU utilization in agent pipelines is often surprisingly low; you're spending more cycles on coordination logic than on matrix multiplication.

So when Snowflake locks in a multi-billion dollar commitment to Graviton, it's making a practical architectural bet: that the workloads its customers actually run day-to-day are CPU-bound more than GPU-bound. And given that Cortex AI is largely about querying, summarizing, and automating data workflows rather than training foundation models, that bet is probably right.

Amazon's Quiet Chip Strategy Is Working

AWS CEO Andy Jassy has been vocal about Graviton's price-performance advantages over Nvidia's offerings. Take that claim with appropriate skepticism—"better price-performance" is a phrase that requires a very specific workload definition to mean anything—but the underlying economics are real. Amazon builds Graviton for Amazon's infrastructure. It doesn't pay Nvidia's margins. Those savings, passed on to customers, become a legitimate competitive lever.

The evidence that this lever is working? Last month, Meta signed a deal to take millions of Graviton chips from AWS for its AI compute needs. Meta, which had just inked a $10 billion agreement with Google Cloud a few months prior. That's not a company fleeing Google—that's a company diversifying its compute supply chain and taking a cheaper option where it makes sense.

Snowflake doing the same thing, at $6 billion, reinforces the pattern. AWS's homegrown silicon is converting into real enterprise commitments, not just benchmark press releases.

Where Nvidia Fits Into All of This

Before anyone writes Nvidia's obituary—and some people on the internet are very eager to do this—let's be precise about what's actually happening. Graviton is a CPU. Nvidia dominates the GPU market, and GPUs remain essential for training and for the most compute-intensive inference workloads. These aren't the same product competing for the same job.

What is threatening to Nvidia is the new Vera CPU that Jensen Huang announced last week, positioning it as a "$200 billion market opportunity." Nvidia is explicitly entering the CPU space to capture the agentic AI workflow that AWS's Graviton is currently eating. Huang claims $20 billion in Vera commitments already. That's a lot of money, but it also tells you that Nvidia recognizes what's coming: the next big compute battle isn't over who has the fastest GPU for training—it's over who handles the sprawling, CPU-heavy infrastructure of AI agents at scale.

Google has its TPUs. Microsoft quietly launched its Maia AI accelerator in January. Every major cloud provider is now in the custom silicon business. Nvidia isn't being displaced—it's being forced to fight on more fronts.

The Actual Takeaway for Engineers

If you're building on Snowflake or designing AI data pipelines, the Cortex AI expansion is worth watching closely. Not because AI-powered natural language queries are going to replace your data engineers (they won't, at least not the good ones), but because the infrastructure economics are shifting in ways that affect your architecture decisions.

Graviton-backed inference is cheaper. If your workload is mostly CPU-bound—and more AI applications than you'd expect are—routing through ARM-based infrastructure rather than GPU clusters can meaningfully cut your costs. That's not a theoretical benefit; it's what Snowflake is apparently betting $6 billion on.

The broader lesson here is one that gets lost in the GPU hype cycle: most of what AI actually does in production isn't training, it's everything else. Orchestration, retrieval, formatting, routing, calling tools. That's where the real compute spend is going, and that's where the chip war is quietly being fought.

AWS is winning some of those battles. Whether that chips away at Nvidia's longer-term dominance or just carves out a parallel market is the more interesting question—and we won't have a real answer until agentic AI workloads mature enough to audit properly.