- Amazon CEO Andy Jassy’s 2026 shareholder letter defends a $200 billion capital expenditure commitment, citing an OpenAI pledge to spend $100 billion on AWS as evidence of concrete demand.
- Amazon’s Trainium AI chip line has reached a $20 billion annual revenue run rate in internal consumption; Jassy states capacity for Trainium4—still roughly 18 months from availability—is already nearly sold out on reservations.
- Amazon’s Graviton CPU is now used by 98% of the top 1,000 EC2 customers, with two unnamed companies requesting to purchase all available Graviton instance capacity in 2026.
- Amazon’s satellite broadband service, scheduled to launch in mid-2026, has already signed contracts with Delta Airlines, AT&T, Vodafone, Australia’s National Broadband Network, and NASA.
What Happened
Amazon CEO Andy Jassy published his 2026 annual shareholder letter on April 9, using the document to defend the company’s $200 billion capital expenditure plan while positioning AWS’s homegrown Trainium and Graviton chips as direct alternatives to Nvidia and Intel offerings. The letter arrives as Amazon’s stock has remained below $200 per share, increasing pressure on Jassy to demonstrate that the infrastructure spending is supported by real customer commitments.
Why It Matters
Amazon’s $200 billion capex figure for 2026 exceeds announced infrastructure spending by any other major technology company, including Microsoft, Google, and Meta, all of which have also committed to significant data center buildouts. Jassy’s letter is his most direct public argument yet that AI infrastructure demand justifies the investment, at a moment when analysts are actively debating whether the sector’s capital commitments have outpaced actual revenue. “I’ve followed the public debate on whether this technology is over-hyped, whether we’re in ‘a bubble,'” Jassy wrote, framing Amazon’s position as grounded in signed customer agreements rather than projections.
Technical Details
On AI chips, Jassy wrote that “virtually all AI thus far has been done on NVIDIA chips, but a new shift has started,” attributing the change to AWS customers seeking better price-performance from Amazon’s Trainium silicon. Amazon’s Trainium line has reached a $20 billion annual revenue run rate in internal consumption terms; Jassy estimated that if the chips were sold externally at market rates, they would represent a $50 billion ARR business. Capacity for Trainium3 is described as nearly exhausted, and Trainium4—approximately 18 months from general availability—is also nearly fully reserved.
On the CPU side, Jassy stated that Amazon’s Graviton processor, which competes with Intel’s x86 architecture, “is now used expansively by 98% of the top 1,000 EC2 customers.” Two unnamed enterprise customers requested to purchase all available Graviton instance capacity in 2026; Amazon declined both requests. For scale context, Nvidia reported $215.9 billion in actual revenue for its most recent fiscal year, a figure that substantially exceeds Amazon’s current Trainium internal run rate.
Who’s Affected
The letter signals continued competitive pressure on Nvidia and Intel in the cloud infrastructure market, where AWS holds the largest global share. Customers building large-scale AI training workloads on AWS are the primary target for the Trainium pitch, which Jassy frames explicitly as a lower-cost alternative to Nvidia GPUs. SpaceX’s Starlink faces a new entrant in the satellite broadband market: Amazon’s low-earth-orbit service, scheduled to launch in mid-2026, has already secured commercial agreements with Delta Airlines, AT&T, Vodafone, Australia’s National Broadband Network, and NASA.
What’s Next
Jassy disclosed that beyond OpenAI’s commitment to spend $100 billion on AWS, “there are several other customer agreements completed (and unannounced), or deep in process” for the new data center capacity, though he named no additional counterparties. Trainium4 is expected to reach general availability in approximately 18 months. Jassy also indicated Amazon may eventually commercialize robotics solutions derived from operational data gathered by its fleet of one million warehouse robots, though he provided no product specifics or timeline.
