An Interview with Nvidia’s Ian Buck: Why AWS Is Buying a Million GPUs
Tae Kim
· Key Context by Tae Kim
· March 20, 2026 at 09:00
· ⏱ 7 min read
| Read on Substack ↗
Summary
The article reports that AWS is deploying over one million Nvidia GPUs through 2027 as part of a major partnership expansion driven by surging demand for agentic AI. It validates Nvidia's product roadmap (Blackwell, Vera Rubin, LPUs) and networking solutions, reinforcing Nvidia's dominant position in AI infrastructure and implying sustained strong demand for its hardware.
•AWS will deploy more than one million Nvidia GPUs through 2027, starting with Blackwell and scaling into Rubin.
•Amazon raised its 2026 capex outlook to $200 billion, far exceeding the $146.6 billion Wall Street estimate.
•The partnership includes adoption of Nvidia's ConnectX networking for East-West traffic and Spectrum-X AI Ethernet.
•Nvidia is working closely with AWS to scale capacity for OpenAI, including joint monitoring of fleet health and debugging across stacks.
•Data center planning and GPU orders are placed years in advance; the first Vera Rubin ramp is expected in the second half of 2026.
•Networking gear is typically installed before GPUs to optimize infrastructure readiness and minimize downtime.
Read time7 min
Length7,036 chars
Categoryfinance
Trade Ideas
Tae KimSenior writer, Barron's; author of The Nvidia Way
Amazon's commitment to over one million Nvidia GPUs through 2027, alongside adoption of Nvidia's networking (ConnectX, Spectrum-X) and joint work with OpenAI, confirms sustained demand for Nvidia's AI
Amazon's commitment to over one million Nvidia GPUs through 2027, alongside adoption of Nvidia's networking (ConnectX, Spectrum-X) and joint work with OpenAI, confirms sustained demand for Nvidia's AI hardware and software, including Blackwell and Vera Rubin.
Risk: Execution risks in manufacturing ramp (six-month lead time) and potential competition from custom ASICs or alternative networking solutions.