Sridhar Ramaswamy is turning Snowflake into an artificial-intelligence platform at the same time that its infrastructure economics are becoming more concentrated. On May 27, 2026, Snowflake expanded its relationship with Amazon Web Services and committed to spend $6 billion on AWS over six years. The companies also agreed to deepen product integration and joint selling.
The agreement is rational. AWS provides global capacity, specialised chips, enterprise distribution and a large base of customers already using Snowflake. A long commitment can improve commercial terms and planning. Joint engineering can remove friction between data, models and applications. For a consumption business, better infrastructure economics can support both growth and margins.
It also creates a leadership test. Snowflake’s appeal has long included the ability to bring a common data experience to several clouds. Customers use that option for resilience, regional requirements, acquisitions and bargaining power. Ramaswamy must show that a large supplier commitment does not convert multi-cloud from an operating principle into marketing language.
Growth gives Snowflake leverage
Snowflake entered the agreement from a position of strength. In the first quarter of fiscal 2027, ended April 30, 2026, product revenue rose 34 per cent to $1.33 billion. Total revenue reached $1.39 billion, remaining performance obligations increased 38 per cent to $9.21 billion and 779 customers were spending more than $1 million a year. Net revenue retention was 126 per cent, indicating that existing customers continued to expand.
Those numbers reflect more than warehouse migrations. Snowflake has added application development, data sharing, machine learning and agentic services around a governed data layer. Ramaswamy, a computer scientist who previously led Google’s advertising products and founded the search company Neeva, has pushed the company to move faster on products while maintaining enterprise controls.
AI increases infrastructure consumption in unpredictable ways. Training, retrieval and agent workloads can create bursts of demand. Customers want access to models and chips without moving sensitive datasets through complicated pipelines. Closer AWS integration can reduce latency and simplify procurement. It can also help Snowflake reach buyers through the AWS marketplace and field organisation.
A six-year commitment signals confidence that demand will persist. It may let Snowflake negotiate lower unit costs and pass some savings to customers. The company can plan capacity and product road maps with greater certainty. Yet the commitment creates a fixed economic target even if customers’ preferred clouds change or new computing architectures emerge.
Portability is a product, not a contract clause
Customers will evaluate multi-cloud credibility through practical differences. New Snowflake features should reach AWS, Microsoft Azure and Google Cloud on a transparent schedule. Performance and pricing should remain comparable enough that workload placement reflects customer needs. Data-sharing and disaster-recovery functions must work across providers without excessive egress fees or administrative complexity.
Snowflake does not need identical capability everywhere on the same day. Cloud providers release different hardware and services, and engineering teams must prioritise. It does need clear explanations when one platform receives preferential access. A pattern of AWS-first launches could influence customers long before any formal retreat from other clouds.
The company should publish measures of cross-cloud replication, regional coverage and feature availability. It can maintain strong engineering teams for Azure and Google Cloud and align executive compensation with overall platform adoption rather than AWS consumption alone. Those signals would demonstrate that the $6 billion is an economic commitment, not a strategic lock-in.
Ramaswamy must also protect Snowflake from becoming a reseller of one provider’s AI stack. Its value lies in governing data and workloads across a heterogeneous environment. Customers may use models from Anthropic, OpenAI, Google or internal teams. Snowflake should make those choices easier to compare and control, even when AWS has a preferred partner or model marketplace.
Agents raise the value of governed context
Snowflake’s May acquisition of Natoma, an enterprise platform built around the Model Context Protocol, illustrates the direction. Agents need controlled connections to data, applications and tools. Enterprises need to know which systems an agent can reach, what information it can transmit and how an action can be reversed. Natoma is intended to help discover, manage and govern those connections.
The strategic logic is compelling. Snowflake already holds governed enterprise data. If it becomes the control plane for agent context and action, it can move from a place where analysis occurs to infrastructure through which work is executed. Consumption could expand beyond queries into continuous operational workflows.
That expansion creates new responsibilities. An agent with incorrect context can make a poor recommendation; an agent with excessive permissions can change a production system. Snowflake must provide identity controls, policy enforcement, logging and human approval that extend beyond its own database. It should allow customers to use competing agent frameworks and clouds without weakening those safeguards.
AWS can accelerate this road map through model access, compute and its own agent services. The partnership should nevertheless preserve a clean boundary between Snowflake’s governance and Amazon’s infrastructure. Customers must understand which company processes data, where logs reside and who is responsible for an incident. Shared selling cannot substitute for shared accountability.
The consumption model requires cost discipline
Snowflake charges primarily according to usage. AI can increase that usage, but it can also make costs harder to predict. Agents may run queries continuously, repeat failed tasks or pull more context than necessary. Customers will not tolerate surprise bills simply because a workload is labelled intelligent.
Ramaswamy should treat efficiency as a product feature. Workload controls can set budgets, route jobs to appropriate compute and show the cost of each model or agent. Better query optimisation can let customers do more work for each credit while supporting long-term retention. Snowflake’s revenue may grow more slowly per task when efficiency improves, but durable adoption is more valuable than waste.
The AWS commitment adds another layer. Snowflake needs enough consumption to meet its negotiated spending path without encouraging uneconomic use. Finance teams should separate supplier utilisation from customer value and resist internal incentives to fill committed capacity. If demand falls short, the answer should be revised planning rather than opaque pricing.
Infrastructure concentration can also affect margins. Preferential terms may improve gross profit, while rapid adoption of expensive accelerators may pressure it. Snowflake should disclose enough about AI workload economics for investors to distinguish productive growth from subsidised consumption. The balance between product revenue, remaining obligations and cash generation will show whether the platform is scaling efficiently.
Joint selling must not narrow customer choice
AWS has enormous enterprise reach. Coordinated account planning can shorten sales cycles and help customers consolidate procurement. It can also create channel tension with Microsoft, Google and independent consultancies. Snowflake needs those relationships to serve global customers whose infrastructure decisions vary by region and workload.
Ramaswamy can keep the ecosystem balanced by continuing co-selling and technical certifications across providers. Marketplace incentives should not make equivalent deployments materially harder outside AWS. Partners need confidence that Snowflake will not route every opportunity through one channel.
Competition makes this discipline essential. Databricks, hyperscale cloud vendors and open data technologies all challenge Snowflake’s position. If customers believe the company has become dependent on AWS, they may prefer a native service with tighter integration or an open architecture with more portability. Snowflake wins by offering both convenience and independence.
The six-year horizon is long in cloud computing. Chips, models, regulation and enterprise architecture can change quickly. Ramaswamy has bought attractive economics and a powerful distribution alliance, but he has also reduced some optionality. His job is to rebuild that optionality at the product layer.
Asia will expose the limits of a single-cloud bias
Snowflake’s Asia-Pacific customers operate across very different regulatory and infrastructure environments. Financial institutions in Singapore may prioritise resilience and auditability. Japanese manufacturers can have long-standing relationships with several technology suppliers. Indian digital businesses may optimise aggressively for cost, while governments and state-linked enterprises often impose local residency requirements.
No single cloud has identical strength across those markets. Snowflake’s ability to maintain consistent governance across regions is therefore commercially important. It should add regions where demand exists, support sovereign and sector-specific controls and make cross-border replication policies easy to understand. AWS can be the largest partner without becoming the only practical choice.
Local partners are equally important. Systems integrators and managed-service providers translate a global platform into domestic compliance and operating practice. Snowflake should give them comparable training and incentives regardless of their preferred hyperscaler. A neutral partner network can offset the gravitational pull of the AWS agreement.
Ramaswamy can use Asia as an early-warning system. If customers begin moving workloads because features or commercial terms diverge, the company should respond before the pattern becomes structural. Regional adoption data, customer councils and transparent road maps will be more useful than broad assurances about choice.
Sridhar Ramaswamy does not need to apologise for choosing a close infrastructure partner. Scale requires commitments. He does need to prove that Snowflake can concentrate purchasing without concentrating customer choice. If AWS helps the platform grow while Azure and Google Cloud deployments remain first-class, the agreement will look like disciplined leverage. If product road maps and sales incentives tilt steadily toward Amazon, the $6 billion saving could cost Snowflake the neutrality that made it valuable.
