Custom AI chips, also known as application-specific integrated circuits (ASICs), are processors designed from the ground up to handle artificial intelligence workloads like training large language models and running inference at scale. Unlike general-purpose GPUs from companies like Nvidia, these chips are built by cloud providers to optimize performance, cost, and energy efficiency within their own data centers. Amazon Web Services (AWS) has been developing its own line of custom AI silicon, called Trainium and Graviton, through its subsidiary Annapurna Labs for nearly a decade. Now, CEO Andy Jassy has signaled a major strategic pivot, Amazon may start selling those chips directly to third parties.
In this article, we’ll discuss what Jassy revealed about the scale of Amazon’s chip business, why the company is considering selling racks of its custom silicon to outside buyers, and what this means for the broader AI chip market currently dominated by Nvidia. We’ll also explore how this move fits into the larger context of Amazon’s record-breaking $200 billion capital expenditure plan for 2026, and the intensifying race among hyperscalers to control their own AI hardware supply chains.
TL;DR Snapshot
Amazon CEO Andy Jassy disclosed on April 9, 2026, that they are exploring the possibility of selling racks of their chips directly to third-party buyers. The move would represent a fundamental shift from Amazon’s current model, where its chips are only available through AWS cloud services, and could have significant implications for Nvidia, AMD, and the broader AI semiconductor market.
Key takeaways include…
- Amazon’s in-house chip unit has hit a $20 billion annual run rate, with Jassy estimating its standalone value at roughly $50 billion if it weren’t limited to AWS monetization channels.
- The Trainium2 chip is completely sold out and Trainium3 is nearly fully subscribed, creating so much demand that Jassy says it’s “quite possible” Amazon will sell chip racks to third parties.
- This potential move raises the competitive stakes for Nvidia and AMD, as Amazon positions its custom silicon as a lower-cost, higher-efficiency alternative to off-the-shelf GPUs.
Who should read this: Cloud architects, AI infrastructure engineers, semiconductor investors, and technology strategists tracking the AI chip market.
The Scale of Amazon’s Custom Silicon Empire
For years, Amazon’s chip efforts have been quietly growing behind the scenes. That era of quiet growth ended on April 9, 2026. According to a Yahoo Finance report, Jassy disclosed that Amazon’s chip business has reached an annual revenue run rate of $20 billion and is growing at triple-digit percentages year over year. Jassy went further, saying the number is actually understated because Amazon currently only monetizes its chips through its AWS EC2 service. If it were a standalone business, he estimated the revenue run rate would be closer to $50 billion.
Those numbers put Amazon’s chip operation in a league that rivals some of the largest semiconductor companies in the world. To put it in perspective, AMD’s entire data center segment generated roughly $12.6 billion in revenue in 2024. Amazon’s chip unit, which most people don’t even think of as a chipmaker, is already significantly larger by run rate.
The demand picture is equally striking. Jassy revealed that Trainium2 is sold out and Trainium3 is nearly fully subscribed, even though it only started shipping at the beginning of 2026. He also noted that two large AWS customers had asked to purchase all of Amazon’s Graviton CPU capacity for 2026, a request the company had to decline because it still needs to serve its broader customer base.
Why Amazon Might Start Selling Chips Externally
The catalyst behind this potential shift is simply overwhelming demand. Jassy stated plainly that there’s so much demand for Amazon’s chips that it’s “quite possible” the company will sell racks of them to third parties in the future. But the strategic logic goes deeper than just meeting demand.
First, there’s the cost advantage. Jassy told investors that at scale, Trainium is expected to save Amazon tens of billions of dollars in capital expenditures per year, and deliver several hundred basis points of operating margin advantage over relying on third-party chips for inference. According to pricing data compiled compiled by Uncover Alpha, Trainium chips have been available at roughly $1 per hour compared to approximately $3 per hour for Nvidia H100 chips through providers like CoreWeave, with potential long-term contract discounts bringing the effective price even lower.
Second, selling externally would help Amazon amortize its massive capital investments. The company announced plans to spend $200 billion on capex in 2026 alone, according to a TechBuzz report, up from $131.8 billion in 2025. Opening a new revenue stream through direct chip sales could help justify that spending to a skeptical Wall Street.
Third, there’s a competitive angle. By selling chips externally, Amazon could pull customers away from Nvidia’s ecosystem while simultaneously undermining competitors like Google Cloud and Oracle, who don’t offer comparable custom silicon for external purchase. A recent TechCrunch article highlighted how Uber recently expanded its AWS contract specifically to use Amazon’s Graviton and trial Trainium3, joining Anthropic, OpenAI, and Apple as major customers drawn to AWS partly because of its custom chips.
The Competitive Ripple Effects
If Amazon follows through on selling its chips outside of AWS, the implications for the AI chip market would be significant. Nvidia currently dominates the AI accelerator market, but Amazon isn’t the only hyperscaler building alternatives. Google has its Tensor Processing Units (TPUs), Microsoft is developing its Maia chip series, and Meta is moving inference workloads to its own custom silicon. What makes Amazon’s situation unique is the sheer scale. A Motley Fool analysis noted that demand for custom silicon is growing across the board, but it may be growing fastest at AWS, the world’s largest cloud computing platform.
Jassy also took a direct shot at Nvidia during his remarks, saying that while Amazon continues to use Nvidia’s chips, customers are increasingly asking for better price-to-performance ratios. That sentiment is driving the rapid adoption of Trainium and positioning Amazon’s silicon as a credible alternative, not just a cost-cutting experiment.
The partnership landscape is also evolving. In February 2026, Amazon and OpenAI announced a strategic partnership that includes OpenAI committing to deploy 2 gigawatts of Trainium chips for its enterprise platform. As CNBC reported, Jassy noted that the two largest AI labs are now both significantly betting on Trainium. Meanwhile, Amazon has also struck a deal with Cerebras Systems to combine Cerebras’ wafer-scale chips with Trainium3 for optimized inference workloads, showing that Amazon is willing to play both the “build” and “partner” strategies simultaneously.
For Nvidia, all of this doesn’t spell doom, but it does signal an erosion of the moat. As more workloads shift to custom silicon purpose-built for specific cloud environments, the premium that Nvidia commands for its general-purpose GPUs may come under sustained pressure.
What Comes Next
Amazon hasn’t committed to a timeline or specific format for external chip sales. Jassy’s comments were exploratory, not a product announcement. But the building blocks are clearly in place: a chip business growing at triple-digit rates, demand that outstrips supply, a cost structure that beats the competition, and a massive capex budget that needs to be put to work.
If Amazon does begin selling Trainium racks to third parties, it would mark a transformation of the company from a cloud services provider that happens to make chips into a full-fledged semiconductor competitor. It’s a path that could reshape the AI hardware market for years to come, and one that every company building on AI infrastructure should be watching closely.
Frequently Asked Questions
AWS is Amazon’s cloud computing division and the largest cloud infrastructure provider in the world. It offers on-demand computing resources, storage, and a wide range of services that businesses use to run applications, store data, and deploy AI models without managing their own physical servers.
Trainium is Amazon’s custom-designed AI chip family, built by its subsidiary Annapurna Labs. The chips are purpose-built ASICs optimized for training and running AI models. The lineup currently includes Trainium2 (sold out) and Trainium3 (shipping since early 2026), with Trainium4 already in development.
Graviton is Amazon’s custom ARM-based CPU designed for general-purpose cloud computing workloads. Unlike Trainium, which focuses on AI, Graviton powers a broad range of applications on AWS and is known for delivering strong price-to-performance ratios compared to traditional x86 processors.
Annapurna Labs is an Israeli chip design company that Amazon acquired in 2015 for $350 million. It serves as Amazon’s in-house semiconductor division and is responsible for designing the Trainium, Inferentia, Graviton, and Nitro chip families that power AWS infrastructure.
Nvidia is the dominant supplier of GPUs used for AI training and inference globally. Amazon’s potential entry into external chip sales matters because it introduces a well-funded competitor offering chips at significantly lower price points, which could pressure Nvidia’s margins and market share in the data center segment.
Capital expenditure refers to the money a company spends on acquiring or maintaining physical assets like data centers, servers, networking equipment, and chips. Amazon’s $200 billion capex plan for 2026 is the largest in tech history and is primarily directed at AI infrastructure.
A run rate is an estimate of a company’s annualized financial performance based on current results. When Jassy says Amazon’s chip business has a $20 billion run rate, he means that if current revenue levels continue for a full year, the business would generate $20 billion.
Training is the process of teaching an AI model by feeding it massive datasets so it learns patterns and relationships. Inference is the process of using a trained model to generate outputs, like answering questions or making predictions. Both require significant computing power, but they have different hardware optimization needs.
Other Enterprise AI Articles You May Be Interested In
Google and Intel’s New AI Chip Deal: Why CPUs Are Making a Comeback
Anthropic Launches Project Glasswing: AI That Found a 27-Year-Old Security Flaw
Why Intel Is Joining Elon Musk’s Terafab to Build the World’s Largest AI Chip Factory
Onix Bets Big on Google Cloud With Wingspan-Powered AI Transformation Strategy
Google’s TurboQuant: The Compression Breakthrough That Shook the AI Industry
