Amazon's AI Revenue Hits $15B Run Rate as Jassy Defends $200B Capex Bet

Every year, Amazon CEO Andy Jassy writes a shareholder letter that functions less as a financial summary and more as a state-of-the-company manifesto. The 2026 edition, released this week, had one number at its center: $15 billion.

That is the annualized revenue run rate that AWS’s AI services reached in the first quarter of 2026. To put it in the context Jassy himself chose: at the same point in AWS’s lifecycle, three years into the cloud computing wave, AWS’s total run rate was roughly $58 million. AI’s equivalent figure is 260 times larger.

“We have never seen a technology get adopted this fast at this scale,” Jassy wrote. “The hard part isn’t selling. The hard part is keeping up with demand.”

The $200 Billion Defense

The $15 billion figure is not just a milestone — it is an argument. Amazon announced earlier this year that it would invest approximately $200 billion in capital expenditure in 2026, the largest single-year infrastructure spending commitment ever made by a private company. The plan encompasses data center construction, custom silicon development, power infrastructure, and networking buildout across dozens of regions worldwide.

That number triggered skepticism from analysts who questioned whether the return on investment could justify the outlay, particularly as hyperscaler spending has repeatedly drawn criticism for outpacing near-term revenue visibility. Jassy’s shareholder letter is in large part a rebuttal.

The core argument: AWS AI demand is growing faster than Amazon can build capacity to serve it. Several enterprise customers that signed multi-year AI infrastructure commitments have had to queue for compute allocation. Others have been told that expanded capacity won’t be available until late 2026 or early 2027. In this framing, $200 billion in capex is not a bet on future demand — it is an attempt to catch up with demand that is already there.

“We are not doing this on a hunch,” Jassy said in remarks accompanying the letter. “Every dollar of that spend has a customer need behind it.”

Trainium Quietly Becomes a $20 Billion Business

Buried beneath the AWS AI headline is a figure that may prove just as significant over time. Amazon’s custom chip business — encompassing Graviton server processors, Trainium AI training accelerators, and Nitro networking cards — now generates an annualized revenue run rate of over $20 billion, doubling from the $10 billion disclosed with the company’s Q4 2025 results.

The Trainium number is the most strategically important. When Amazon began developing its own AI training silicon, the conventional wisdom was that NVIDIA’s CUDA ecosystem was an insurmountable moat — customers simply would not accept the engineering overhead of porting training workloads to a new chip architecture. That assumption has been eroding steadily.

Trainium 3, released in late 2025, closed much of the performance gap with NVIDIA’s H200 on transformer training workloads, and Amazon has used its position as both chip maker and cloud provider to bundle Trainium capacity at prices that undercut the NVIDIA-based alternatives by a meaningful margin. Enterprise customers training large proprietary models — the highest-value segment of the AI infrastructure market — have taken notice.

NVIDIA’s share of AI training revenue on AWS has declined for four consecutive quarters. It is still dominant, but the trajectory has changed.

AWS Re-Accelerates Above $2.5 Trillion Market Cap

The combination of AI revenue growth and infrastructure credibility has propelled Amazon’s market capitalization above $2.5 trillion, making it the third company in history (after Apple and Microsoft) to sustain that threshold. AWS, which generates the vast majority of Amazon’s operating income, is now being valued by some analysts as a standalone business worth more than $1.5 trillion — roughly equivalent to the entire market cap of Microsoft a decade ago.

The re-acceleration is notable because AWS growth had slowed meaningfully through 2023 and early 2024 as enterprises “optimized” their cloud spend following the pandemic-era buildout. That optimization phase is over. AI workloads, which tend to be additive to rather than substitutive of traditional cloud usage, have created a second growth curve.

The Infrastructure Question Everyone Is Asking

The elephant in every hyperscaler earnings discussion right now is power. A $200 billion capex plan implies not just data centers but the electricity to run them. Industry estimates project that the combined AI infrastructure buildout across Amazon, Microsoft, Google, and Meta will require adding the equivalent of several new countries’ worth of power generation capacity over the next five years.

Amazon’s response has been to move aggressively into nuclear power procurement. The company has signed purchase agreements with three small modular reactor developers and is funding construction of a large-scale nuclear project intended to come online in 2030. It has also struck agreements with renewable energy providers in regions where grid capacity allows.

Whether those power commitments can keep pace with data center construction timelines is the central uncertainty hanging over not just Amazon’s plan but the entire AI infrastructure supercycle. For now, the demand side of the equation is not in doubt. Jassy’s $15 billion number made sure of that.