Nvidia will deliver 1 million graphics processing units to Amazon Web Services between 2026 and 2027, an Nvidia executive confirmed. The company says the shipments will start this year and continue through 2027, establishing a clear timeline for a major cloud infrastructure partnership.
Scope beyond GPUs
According to Nvidia, the agreement extends beyond GPUs to include several other chip families and networking products. Amazon Web Services will buy Nvidia's Spectrum networking chips and newly added Groq processors, which Nvidia obtained through a $17 billion licensing transaction with the AI chip company Groq late in 2025. AWS plans to combine Groq chips with six other Nvidia chip types to address inference workloads, which are the compute tasks that generate AI system outputs.
Nvidia's Ian Buck, vice president of hyperscale and high-performance computing, summarized the technical rationale succinctly: "Inference is hard. It's wickedly hard." He added that a single chip is not sufficient for leading inference performance, saying, "To be the best at inference, it is not a one chip pony. We actually use all seven chips."
Networking gear and a shift in deployment
The agreement also contemplates deploying Nvidia's Connect X and Spectrum X networking equipment inside AWS data centers. That represents a notable operational change because AWS has traditionally relied on custom-built networking hardware developed internally over many years. Buck acknowledged AWS will continue to use its proprietary systems, saying, "They're still going to do that, of course. But we are collaborating now on deploying Connect X and Spectrum X for those important workloads and biggest customers across AI with AWS." Financial terms for the arrangement were not disclosed by the companies.
Market opportunity as framed by Nvidia leadership
The timeline of the AWS initiative coincides with Nvidia CEO Jensen Huang's projection that the company faces a $1 trillion sales opportunity for its Rubin and Blackwell chip families through 2027. Huang's estimate excludes several categories - central processing units, networking chips, Groq-based products, and a Rubin Ultra variant - indicating the total addressable market could be materially larger if those elements are included.
Huang has also outlined expectations for Groq integration, suggesting it could unlock $300 billion in annual revenue per gigawatt, and projecting that roughly 25% of GPU workloads will be paired with Groq chips. Nvidia and Groq's combined system, referred to as LPX, is positioned as an optional integration with Nvidia's Vera Rubin platform, but Nvidia notes the LPX is not yet in use at scale.
What investors and industry participants should watch
Several metrics and milestones will be important in evaluating the commercial and operational impact of the partnership:
- GPU delivery pace - whether Nvidia fulfils the one million unit commitment and how deliveries break down by chip generation.
- Groq deployment timeline - the timing and extent to which AWS moves Groq chips beyond pilot stages for inference workloads.
- Networking equipment adoption - the degree to which AWS incorporates Connect X and Spectrum X versus continuing to rely on custom networking solutions.
- Fiscal 2027 results - anticipated in early 2027, these results will provide the first public financial read on the partnership's contribution.
- Competitive response - how other chipmakers, including AMD and peers, react to Nvidia's expanded presence in AWS infrastructure.
Broader implications
The arrangement touches multiple segments of the technology supply chain and capital markets. It directly affects semiconductor suppliers, cloud infrastructure providers, and firms supplying data-center networking gear. The commitments also bear on AI service providers and enterprises that rely on AWS for inference workloads, as changes in hardware mix and performance characteristics can affect cost, latency, and available model deployment options.
Disclosure
No financial details of the transaction have been made public, and the timeline and technology adoption remain subject to execution over the coming years.