Snowflake’s $6B AWS deal says a lot about where AI compute is going

Snowflake has signed a new five-year, $6 billion agreement with Amazon Web Services. The size of the deal is the point.

AWS says Snowflake has sold about $7 billion worth of services through AWS Marketplace since Snowflake was founded in 2012. This new contract nearly matches that lifetime total in a single five-year commitment.

AI spending is driving it. Snowflake says its customers doubled their AWS-related spending in 2025 to $2 billion for the year. A large share of that demand is tied to Snowflake’s Cortex AI services, which bring natural-language querying, summarization, and other AI features closer to the enterprise data already inside Snowflake.

The technical detail worth watching is the hardware. Snowflake is signing up for expanded access to AWS’s Arm-based Graviton CPUs.

That matters because AI infrastructure coverage has been obsessed with GPUs. GPUs still dominate model training and high-end inference, especially in Nvidia-heavy environments. But once AI moves into production, much of the load shifts back to CPUs: orchestration, retrieval, data filtering, API calls, agent loops, permission checks, embedding pipelines, query planning, and the dull glue code that keeps AI apps running.

AI systems still need a lot of general-purpose compute.

Why Snowflake wants more Graviton capacity

Snowflake is a database and data platform company first. Its AI products sit on top of that.

Cortex AI lets customers run AI-assisted operations where their structured and semi-structured data already lives. That includes asking questions in plain English, generating summaries, classifying content, extracting information, and building model-driven workflows without moving everything into a separate AI stack.

For enterprise teams, the appeal is obvious. Data movement is expensive, slow, and risky. Copying sensitive business data into another system for AI processing creates governance problems and security review pain. Keeping AI close to the warehouse simplifies some of that, at least in theory.

It also changes the compute profile.

A natural-language query tool does a lot before and after a model call. A real implementation may involve:

• Authenticating the user and checking row-level or role-based permissions
• Translating intent into SQL or another query plan
• Fetching metadata and schema context
• Running retrieval over relevant tables, documents, or embeddings
• Calling one or more models
• Validating output against data access rules
• Generating explanations, summaries, or follow-up actions
• Logging everything for audit and billing

Some of that hits GPUs or AI accelerators. A lot of it lands on CPUs.

That’s where Graviton fits. AWS’s Graviton processors are Arm-based server CPUs designed for cloud workloads where price-performance and energy efficiency matter. They’re widely used for web services, containers, data processing, and, increasingly, the support layer around AI systems.

For Snowflake, CPU economics matter at enormous scale. If Cortex AI usage keeps rising, shaving cost from the non-GPU parts of the pipeline can improve margins or give Snowflake room to price AI features more aggressively.

CPUs are back in the AI infrastructure conversation

The industry spent the past few years treating Nvidia GPUs as the center of AI infrastructure, for good reason. Training frontier models and serving large models at low latency still depends heavily on Nvidia’s hardware and software stack. CUDA remains a serious moat. Many AI applications, frameworks, and deployment patterns assume Nvidia compatibility.

Production AI workloads are messier than benchmark charts.

Agentic systems, in particular, can burn a surprising amount of general-purpose compute. An agent that uses tools, queries databases, calls APIs, updates records, and plans multi-step tasks may issue many small operations around each model call. It may spend more time waiting on services, parsing data, and coordinating state than doing pure matrix math.

That doesn’t reduce the need for GPUs. It makes CPU capacity and cost harder to ignore.

AWS clearly sees the opening. Amazon CEO Andy Jassy recently argued that Amazon’s homegrown AI chips offer better price-performance than Nvidia’s alternatives, though AWS still offers and heavily depends on Nvidia GPUs. The company isn’t pretending customers can drop Nvidia across the board. It’s trying to own more of the stack where it has pricing power.

Graviton is one part of that. AWS also has Trainium and Inferentia for model training and inference. The Snowflake deal specifically highlights Graviton, which says something about the workloads Snowflake expects to scale.

The issue isn’t only serving giant language models. It’s running the surrounding machinery cheaply enough that AI features can become default product behavior rather than expensive demos.

Cloud providers want the AI margin

AWS has had a strong month on the custom silicon front. In April, Meta reportedly signed a deal for millions of Graviton chips to support its AI compute needs. That followed Meta’s earlier $10 billion agreement with Google Cloud, which made the AWS win look even sharper.

Snowflake’s $6 billion commitment gives Amazon another high-profile customer for its argument that major AI consumers want alternatives to buying every unit of compute through Nvidia-backed infrastructure.

Google has been building its own AI chips for years with TPUs. Microsoft launched its Maia AI accelerator in January. AWS has Graviton, Trainium, and Inferentia. Every major cloud provider wants to reduce its dependence on Nvidia while still selling Nvidia capacity at high prices where customers demand it.

Nvidia sees the same shift. Jensen Huang recently positioned Vera, Nvidia’s new AI-specific CPU, as a $200 billion market opportunity and said the company has already sold $20 billion worth. That’s a blunt signal that Nvidia doesn’t plan to leave the CPU side of AI infrastructure to the cloud vendors.

For developers and platform teams, the practical takeaway is that the AI compute stack is fragmenting.

A few years ago, the default enterprise answer was simple enough: run AI workloads on Nvidia GPUs, probably in a hyperscaler cloud. Now teams need to think harder about workload placement. Some tasks belong on GPUs. Some can run on CPUs. Some may fit cloud-specific accelerators. Some should stay near the data store to avoid movement, latency, and governance problems.

That complexity is annoying. It can also save serious money.

What Snowflake customers should watch

Snowflake customers probably won’t care which CPU runs under the hood unless it affects price, latency, region availability, or feature reliability. Infrastructure details matter when they show up in the bill or the incident report.

There are a few implications worth tracking.

First, Cortex AI is likely to become a bigger part of Snowflake’s platform strategy. The company has every incentive to turn its data warehouse footprint into an AI application layer. If customers already trust Snowflake with governed enterprise data, Snowflake can pitch AI features with less friction than a standalone AI vendor asking for fresh data access.

Second, performance claims should be evaluated at the workflow level, not the chip level. Graviton may offer strong price-performance for many services, but AI workloads vary wildly. A natural-language analytics feature that performs well on one schema, data volume, and concurrency pattern may behave very differently under heavy multi-tenant usage.

Third, lock-in risk grows. Snowflake already abstracts infrastructure away from most users, which is part of its appeal. But deeper alignment between Snowflake AI services and AWS hardware could make certain workloads cheaper or faster on AWS than on Azure or Google Cloud. Snowflake is available across all three major clouds, but economics often shape architecture more than slide decks do.

If your organization runs Snowflake as part of a multi-cloud strategy, track this closely. Feature parity is one thing. Cost parity is another.

The developer angle: AI apps need boring systems engineering

The Snowflake-AWS deal is also a reminder that production AI is largely a systems problem.

Developers building AI features inside enterprise applications tend to focus on model choice first: GPT, Claude, Gemini, Llama, Mistral, or a domain-specific model. That choice matters. After the prototype, the harder questions are usually about data access, latency budgets, observability, authorization, and cost control.

A text-to-SQL assistant needs a capable language model, but that’s only the start. It needs schema grounding, safe query generation, guardrails against expensive scans, permission-aware responses, and clear failure modes. If the assistant hallucinates a column name, leaks restricted data, or runs a monster query against a production warehouse, the model benchmark doesn’t matter.

Snowflake has an advantage because it controls a lot of the data execution environment. It can integrate AI features with governance, metadata, and query execution in ways that external tools have to bolt on. That doesn’t make the problem easy. It gives Snowflake better starting conditions.

For technical leaders, the right questions are concrete:

• Can AI-generated queries respect existing access controls?
• How are prompts, completions, and retrieved context logged?
• What data leaves the Snowflake boundary, if any?
• Are workloads isolated enough for sensitive use cases?
• How predictable is pricing under high concurrency?
• Can teams choose models or accelerators based on workload needs?
• What happens when the AI feature gets something wrong?

The last point matters. AI assistants inside data platforms can create false confidence quickly. A polished summary over stale, partial, or misinterpreted data is worse than a slow dashboard in many business settings. Snowflake and AWS can improve the infrastructure economics, but application teams still need evaluation pipelines, human review paths, and auditability.

AWS gets the kind of AI win it needed

For Amazon, this deal lands well.

AWS remains the largest cloud provider, but Microsoft has benefited heavily from its OpenAI relationship, and Google has strong AI credibility through DeepMind, Gemini, and TPUs. Amazon has needed visible proof that its AI infrastructure strategy works beyond renting Nvidia GPUs.

A $6 billion Snowflake agreement helps. So does the Meta Graviton deal. These are not small experimental deployments. They suggest AWS’s custom silicon is becoming part of mainstream AI infrastructure planning, especially for customers that care about cloud economics at scale.

There’s a caveat. Custom cloud chips can reduce costs, but they can also tighten dependence on a specific provider’s architecture, tooling, and pricing model. Nvidia’s ecosystem is expensive, but it’s broadly supported. Cloud-native chips often look best when you’re already deep inside that cloud.

Snowflake can absorb much of that complexity for customers. Individual engineering teams may not have to port applications to Arm or tune services directly for Graviton. But someone still pays attention to the trade-off, and at this scale, that someone is probably staring hard at gross margins.

The AI boom is turning CPUs back into strategic hardware. Snowflake’s $6 billion AWS deal makes that visible. GPUs still get the attention, but the day-to-day economics of enterprise AI may be decided by the less flashy compute sitting around the model.