Snowflake’s $6bn AWS chip deal signals shift from Nvidia dominance

Snowflake has committed $6 billion to Amazon Web Services for custom AI and CPU chips, according to TechCrunch AI, in what represents the largest known enterprise agreement for non-Nvidia AI compute infrastructure. The multi-year deal positions AWS’s proprietary silicon—including its Trainium AI accelerators and Graviton processors—as a viable alternative for large-scale AI workloads.

The agreement marks a strategic inflection point for enterprise AI infrastructure procurement. Rather than relying exclusively on Nvidia’s H100 and forthcoming Blackwell GPUs, Snowflake will leverage AWS’s vertically integrated chip designs to power its data cloud platform, which serves over 10,000 enterprise customers including Capital One, Sony, and Nielsen.

The deal’s structure reflects growing enterprise confidence in custom silicon. AWS developed Trainium specifically for training large language models and deep learning workloads, whilst Graviton processors target inference and general compute tasks. Both chip families promise cost advantages over equivalent Nvidia-based instances, though AWS has not disclosed specific performance benchmarks for Trainium against Nvidia’s latest offerings.

For AWS, the agreement validates years of chip development investment and provides a marquee customer reference as it competes with Microsoft Azure and Google Cloud Platform for AI infrastructure dominance. Amazon has reportedly invested over $10 billion in custom chip development since acquiring Annapurna Labs in 2015, and the Snowflake commitment helps amortise those research costs whilst demonstrating production readiness at scale.

Snowflake gains predictable capacity allocation—a critical advantage given persistent Nvidia GPU shortages that have constrained AI deployment timelines across the industry. The company’s CEO has previously cited compute availability as a growth constraint, and the AWS agreement presumably includes guaranteed chip supply alongside volume pricing discounts.

The implications for Nvidia are nuanced rather than immediately threatening. The chipmaker still commands approximately 80% market share in AI accelerators and faces no near-term risk of displacement in cutting-edge model training. However, the Snowflake deal demonstrates that hyperscaler custom silicon has matured sufficiently for production enterprise workloads, particularly for inference and fine-tuning tasks that represent the majority of AI compute demand by volume.

Microsoft and Google have pursued parallel strategies with their Maia and TPU chip families respectively, but neither has announced enterprise commitments approaching the Snowflake-AWS scale. The deal may accelerate similar agreements as enterprises seek to diversify chip suppliers and reduce dependency on single vendors—a strategic priority that has intensified following recent GPU allocation challenges.

Financial analysts note the agreement’s revenue recognition will likely span five to seven years, providing AWS with predictable high-margin infrastructure income whilst Snowflake locks in compute costs during a period of AI workload expansion. Both companies’ stock prices rose modestly following the announcement, suggesting investors view the arrangement as mutually beneficial rather than zero-sum.

The broader market impact centres on compute cost trajectories. If AWS custom silicon delivers comparable performance at materially lower costs—some estimates suggest 30-40% savings versus equivalent Nvidia instances—enterprise AI economics improve substantially. This could accelerate AI adoption amongst cost-sensitive organisations that have delayed deployments due to GPU expense.

Industry observers will watch whether other major cloud software providers follow Snowflake’s lead. Databricks, Confluent, and MongoDB all operate substantial AWS workloads and face similar compute scaling challenges. Additional large-scale commitments to AWS silicon would further validate the custom chip approach and potentially pressure Nvidia’s pricing power in the enterprise segment.

The deal underscores a fundamental shift in AI infrastructure strategy: hyperscalers are no longer merely reselling third-party chips but offering vertically integrated compute stacks optimised for specific workloads. For enterprises, this creates both opportunity—through improved cost-performance—and complexity in evaluating chip architectures that lack the standardisation of Nvidia’s CUDA ecosystem.