Pomegra
---
---
AWS AI Revenue Hits $15B Run Rate — and Is Still Capacity-Constrained
Amazon Web Services posted an AI revenue run rate exceeding $15 billion as of Q1 2026, with Jassy acknowledging the business would be growing even faster if the company weren't still supply-constrained. AWS added 3.9 gigawatts of compute capacity throughout 2025 and has committed to doubling that figure by 2027, underscoring the scale of demand Amazon is racing to meet.
The CEO disclosed that two large AWS customers have already approached the company asking to purchase all of its custom Graviton CPU capacity for 2026 — a request Amazon declined in order to preserve service availability across its broader customer base. The anecdote highlights both the extraordinary appetite for Amazon's in-house silicon and the strategic leverage it now holds in the AI infrastructure landscape.
---
Trainium2 Sold Out, Trainium3 Nearly Fully Subscribed
Jassy's letter offered the clearest public signal yet of Trainium's commercial momentum. Trainium2 is completely sold out, while Trainium3 — which began shipping in early 2026 — is described as "nearly fully subscribed," a remarkable uptake for a chip that entered general availability just months ago.
The Trainium3, built on a 3-nanometer process manufactured by TSMC, runs on next-generation Trn3 UltraServers that Amazon claims deliver comparable AI performance at up to 50% lower operating cost than traditional cloud servers built around third-party GPUs. New custom Neuron switches allow every Trainium3 chip within a cluster to communicate with every other chip in a mesh configuration, dramatically reducing inference latency — the most critical performance bottleneck in modern AI deployments.
Over 1.4 million Trainium chips are currently deployed across all three generations. Anthropic's Claude models run on more than 1 million Trainium2 chips, with Project Rainier — one of the world's largest AI compute clusters, featuring 500,000 chips — coming online in late 2025 to support Anthropic's workloads.
---
A $50B Revenue Opportunity Hidden Inside AWS
Jassy offered a striking financial reframing of Amazon's chip business. While the segment currently carries an annual revenue run rate of $20 billion — growing at triple-digit year-over-year percentages — the CEO argued that figure dramatically understates its true value, because Amazon only monetizes its chips today through its AWS EC2 cloud compute service.
Valued as a standalone chip business, Jassy said Amazon's silicon operations would carry a run rate of approximately $50 billion. That figure positions the Trainium franchise as a potential direct competitor to Nvidia's data center GPU segment, which generated roughly $115 billion in fiscal year 2025 revenue but faces mounting pressure from hyperscaler-developed alternatives.
"There's so much demand for our chips that it's quite possible we'll sell racks of them to third parties in the future," Jassy wrote, a line that immediately drew analyst attention and drove AMZN up 4.96% in Thursday trading to $232.23.
---
The Cost Argument Against Nvidia
The competitive framing in Jassy's letter was pointed. While Amazon continues to procure Nvidia GPUs for AWS — and has committed to receiving 1 million Nvidia chips by 2027 — Jassy characterized customer demand as increasingly driven by price-performance considerations, an area where Trainium is gaining ground.
"Having our own hotly demanded AI chip opens up many possibilities, but perhaps none larger than the ability to lower costs for customers and secure better economics for AWS," Jassy wrote. "At scale, we expect Trainium will save us tens of billions of capex dollars per year, and provide several hundred basis points of operating margin advantage versus relying on others' chips for inference."
The Trainium team has also moved aggressively on developer ecosystem compatibility, ensuring full support for PyTorch — the dominant open-source AI framework — requiring only a single-line code change to migrate workloads from Nvidia's CUDA ecosystem to Trainium hardware. This step directly targets one of Nvidia's most enduring competitive moats: developer lock-in through its proprietary software stack.
---
Big Tech Clients Signal Ecosystem Expansion
The breadth of Amazon's Trainium customer base has expanded well beyond its own internal workloads. OpenAI, as part of a landmark $50 billion AWS partnership announced in February 2026, has agreed to receive 2 gigawatts of Trainium compute capacity — making AWS the exclusive provider of OpenAI's Frontier AI agent builder platform. Apple publicly praised Amazon's Graviton and Inferentia chips in 2024, and the AWS chip team is widely understood to be in active conversations with additional hyperscale enterprise clients.
AWS has also struck a new integration deal with Cerebras Systems, embedding Cerebras inference chips alongside Trainium hardware in UltraServers designed for ultra-low-latency AI applications, further broadening the competitive threat to conventional GPU suppliers.
---
$200 Billion Capex Commitment Backs the Ambition
The chip ambitions sit atop a staggering capital commitment. Amazon disclosed in February 2026 that it plans to spend $200 billion in capital expenditures during the year alone — the largest single-year infrastructure investment in the company's history. AMZN shares declined approximately 7% following that announcement but have since recovered, with Thursday's rally reflecting renewed confidence in the company's AI monetization trajectory.
Jassy's letter also highlighted that Amazon now operates more than 1 million robots in its fulfillment network, is exploring third-party robotics sales, and is preparing to launch its Amazon Leo satellite internet service in mid-2026, with early customers including Delta Air Lines, JetBlue, and AT&T — dimensions of a diversified technology empire that increasingly competes across hardware, infrastructure, and connectivity markets simultaneously.
---
Market Reaction and Competitive Stakes
NVDA edged up 0.92% Thursday to $183.67 while AMD gained 2.14%, suggesting markets have not yet fully priced the competitive implications of a potential Amazon chip sales business. Historically, hyperscaler-designed chips have displaced meaningful GPU volumes from Nvidia primarily within the walls of their own data centers; an Amazon decision to sell Trainium racks externally would mark a qualitatively different threat — a direct assault on Nvidia's and AMD's core enterprise and cloud customer base.The AI accelerator market is projected to surpass $400 billion annually by the end of this decade, with inference workloads — where Trainium now holds a cost advantage — representing an increasing share of total spend. For Amazon, converting internal silicon strength into external chip revenue would represent the most consequential business model expansion in AWS history.
---
Mentioned tickers: AMZN, NVDA, AMD, DAL, JBLU, T
