A company that rents out computing by the hour has a choice about the processors inside its servers. It can buy every one from the firms that design them and pass the price on, or it can design some of its own and keep the difference. The second path costs years and a design team before a single chip earns anything. It pays only if customers pick the in-house chip when the other kind sits on the same shelf.
A shareholder sees that decision in fragments. A results release gives a run rate for the chips, which is a management figure rather than an audited line. A letter to shareholders gives a price-performance claim against the chips the company would otherwise buy, and an expectation of what the design work will save. The annual report shows the segment margin the chips are supposed to lift, without saying how much of it they already have.
Amazon has been designing its own processors for a decade, and in 2026 the company started putting numbers on them in every release. The February results gave a run rate, and the April letter to shareholders gave a margin expectation.
On BloFin the share that carries the chips business is held as the AMZNX/USDT Spot token or as the AMZNUSDT Perpetual, and whatever the chips add to AWS reaches either only through the Amazon share price.
What are Amazon's own chips?
Amazon's own chips are four families designed by AWS. Graviton, a general-purpose processor, arrived in 2018 and Trainium, an AI chip, in 2021, after AWS bought Annapurna Labs in 2015 (source: Amazon, May 5, 2026). Inferentia is an inference chip (source: AWS, Inferentia). Nitro is the network interface card (NIC) for EC2 servers (source: Amazon 2025 shareholder letter).
Say you rent a server on AWS. The processor inside it is one of three kinds. It may be an Intel or AMD chip, the x86 processors that ran almost every cloud workload until Graviton arrived. It may be an NVIDIA graphics processor, the chip that has run almost every AI model so far. Or it may be one of Amazon's own. Graviton reached its fifth generation, announced in December 2025, and Trainium its third (source: Amazon, May 5, 2026). Trainium3 is sold on servers that went on general sale on December 2, 2025 (source: Amazon, December 2, 2025). Inferentia, an inference-only family, runs on its own Inf1 and Inf2 instances (source: AWS, Inferentia). Its first instances went on sale in December 2019, two years before Trainium (source: AWS, December 3, 2019).
The chips have no company of their own. Annapurna Labs is a subsidiary, the chips are sold only as time on AWS servers, and the revenue they earn is AWS revenue, one of the parent's three reported segments (source: Amazon 2025 Form 10-K). Every Amazon share carries the chips business at the same weight as the stores, the advertising line and the rest of AWS; what Amazon stock is sets out that structure. Can you buy AWS stock gives the size of the segment the chips sit inside.
Why Amazon designs chips instead of buying them all
Amazon's stated reason is price-performance, and its evidence is what happened with processors before AI. The 2025 letter says almost every cloud workload ran on Intel chips until Graviton arrived in 2018, that Graviton offers up to 40% better price-performance than other x86 processors, and that 98% of the top 1,000 EC2 customers use it (source: Amazon 2025 shareholder letter).
Say you are an AWS customer choosing where to run an AI model in 2026. The letter's pitch is that the same arc is repeating. Almost all AI so far has run on NVIDIA chips, in Andy Jassy's words, "but a new shift has started", because customers want better price-performance (source: Amazon 2025 shareholder letter). Trainium2 had about 30% better price-performance than comparable graphics processors and has largely sold out, and Trainium3 is 30 to 40% more price-performant than Trainium2 and is nearly fully subscribed (source: Amazon 2025 shareholder letter). A significant part of Trainium4, still about 18 months from broad availability when the letter was written in April 2026, had already been reserved (source: Amazon 2025 shareholder letter).
The second reason is that Amazon is itself the largest buyer. Bedrock, the AWS service that runs AI models for customers, runs most of its inference, the work of answering queries with a trained model, on Trainium, so every query it answers on an Amazon chip is one it did not have to answer on a chip bought from someone else (source: Amazon 2025 shareholder letter). The February 2026 results release put it in numbers: Trainium2 "powers the majority of inference on Bedrock, a service used by 100,000+ companies" (source: Amazon Q4 2025 results release).
The third reason is energy, which in a data center is a cost line. Amazon describes Trainium as designed for AI from the ground up, where a graphics processor is built to be flexible. It says Trainium3 delivers 40% better energy efficiency than the previous generation and up to 4.4 times the compute of a Trainium2 server (source: Amazon, December 2, 2025). Amazon names Anthropic among customers reporting training costs cut by up to half against alternatives (source: Amazon, December 2, 2025).
How big the chips business is
Amazon's chips business had an annual revenue run rate of more than $20 billion by April 2026, up from more than $10 billion for Trainium and Graviton alone in February, growing at triple-digit percentages year over year (source: Amazon Q1 2026 results release). A run rate is a management figure, and AWS reports the chips only inside segment sales.
The largest figure Amazon gives is a hypothetical, and the letter labels it as one. Amazon says the $20 billion run rate is understated because it monetizes chips only through EC2. A stand-alone chips business selling this year's production to AWS and to third parties would run at about $50 billion, and Amazon adds that it may sell racks of chips to third parties in future (source: Amazon 2025 shareholder letter). For a shareholder the operative figure is the $20 billion, because that is the revenue AWS books, against AWS sales of $128.7 billion in 2025 (source: Amazon 2025 Form 10-K).
Amazon has published five kinds of figure for the business since February 2026: run rates, a hypothetical stand-alone run rate, chip counts, supply by generation and delivery dates. Each carries the date of the release that gave it.
Measure | Figure | Date and source |
|---|---|---|
Run rate, Trainium and Graviton | more than $10 billion, triple-digit growth | Feb 5, 2026 (Q4 2025 release) |
Run rate, chips business incl. Nitro | more than $20 billion, triple-digit growth | Apr 29, 2026 (Q1 2026 release); repeated in the letter |
Hypothetical stand-alone run rate | about $50 billion, if this year's chips were sold to AWS and third parties | April 2026 (letter) |
Trainium2 | fully subscribed; 1.4 million chips landed; Project Rainier's 500,000-plus chips | Feb 5, 2026 (Q4 2025 release) |
AI chips landed in 12 months | 2.1 million-plus, more than half Trainium | Apr 29, 2026 (Q1 2026 release) |
Trainium3 | nearly all supply expected to be committed by mid-2026 | Feb 5, 2026 (Q4 2025 release) |
Trainium4 | delivery expected to start in 2027 | Feb 5, 2026 (Q4 2025 release) |
The unit figures show the scale of the build. AWS landed more than 2.1 million AI chips in the twelve months before its April 2026 release, more than half of them Trainium (source: Amazon Q1 2026 results release). Project Rainier, the cluster Anthropic uses to train Claude, connects more than 500,000 Trainium2 chips, and AWS says its newest cluster design can link up to one million Trainium chips (source: Amazon, December 2, 2025). Andy Jassy told the first-quarter 2026 earnings call that the chips business grew nearly 40% quarter over quarter (source: Amazon, April 29, 2026).
Who has committed to Trainium and Graviton
The two largest AI laboratories have each committed to gigawatts of Trainium, the unit Amazon uses for the size of a chip commitment. OpenAI committed to about two gigawatts of Trainium capacity, ramping from 2027, and Anthropic will secure up to five gigawatts of current and future Trainium generations (source: Amazon Q1 2026 results release).
Both laboratory contracts carry a clause that ties part of the money to the chips. The OpenAI expansion of $100 billion over eight years and the Anthropic expansion of more than $100 billion over ten years each include, in the filing's words, contractual obligations related to the performance of AWS chips (source: Amazon Q2 2026 Form 10-Q). Amazon's OpenAI stake explained sets out the $50 billion investment that came with the OpenAI contract, and Amazon's Anthropic stake explained does the same for the Anthropic investment.
Amazon's releases name four groups of customers for the chips: the two laboratories, Meta on Graviton, a list of start-ups and larger companies on Trainium, and the Graviton base as a whole.
Customer | Commitment | Date and source |
|---|---|---|
OpenAI | about 2 GW of Trainium, ramping from 2027 | Apr 29, 2026 (Q1 2026 release); the $100 billion contract expansion of Feb 27, 2026 |
Anthropic | up to 5 GW of current and future Trainium; Project Rainier on Trainium2 | Apr 29, 2026 (Q1 2026 release); contract expansion of more than $100 billion over 10 years (Q2 2026 10-Q) |
Meta | tens of millions of Graviton cores for agent workloads | Apr 29, 2026 (Q1 2026 release) |
Start-ups | NEURA Robotics, Odyssey, TwelveLabs, Decart, Poolside and others on Trainium; Uber and Pinterest | Jul 30, 2026 (Q2 2026 release) |
Graviton base | 98% of the top 1,000 EC2 customers; more than 100,000 customers | Jul 30, 2026 (Q2 2026 release); May 5, 2026 (Amazon) |
Demand for the general-purpose chip is easier to read than demand for the AI chip, because it is older. Two large AWS customers asked to buy all of the Graviton capacity for 2026, requests Amazon declined because of other customers' needs (source: Amazon 2025 shareholder letter). Graviton revenue commitments rose nearly three times quarter over quarter in the second quarter of 2026, and Graviton5, released that quarter, delivers up to 25% more compute than Graviton4 (source: Amazon Q2 2026 results release).
NVIDIA beside Trainium: What Amazon still buys
Amazon still buys NVIDIA chips by the million. The letter states that AWS "will always have customers who choose to run NVIDIA", in the paragraph announcing the shift toward Trainium (source: Amazon 2025 shareholder letter). The April 2026 release added a plan for a million more NVIDIA processors (source: Amazon Q1 2026 results release).
Say you list what OpenAI's two AWS contracts buy. The original $38 billion agreement of November 2025 was for servers built around NVIDIA GB200 and GB300 processors, hundreds of thousands of them, all targeted for deployment before the end of 2026 (source: Amazon, November 3, 2025). The Trainium commitment came only with the $100 billion expansion of February 2026 (source: Amazon Q2 2026 Form 10-Q), and it starts ramping in 2027 (source: Amazon Q1 2026 results release). One customer, two contracts, and the NVIDIA one comes first.
The two chip lines are being designed to share a rack. Trainium4 is being designed to support NVIDIA's NVLink Fusion interconnect so that Trainium servers and graphics-processor servers can sit in the same MGX rack design, alongside Graviton (source: Amazon, December 2, 2025). The relationship also includes a patent case. In May 2026 a patent holder named Xockets filed a complaint at the US International Trade Commission against Amazon, AWS, Annapurna Labs, NVIDIA and Microsoft over servers built on the GB200. The commission opened an investigation in June 2026, and Amazon disputes the allegations (source: Amazon Q2 2026 Form 10-Q).
Amazon is one of three cloud companies designing its own AI chip against the same supplier. Why Google builds its own AI chips covers the TPU, and NVIDIA's moat and the custom-silicon threat reads the same shift from NVIDIA's side.
What the chips do to AWS capex and operating margin
Amazon has put two figures on what its chips are worth to AWS, and both are expectations. The letter says that at scale Trainium will save "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" (source: Amazon 2025 shareholder letter).
Neither figure sits in a financial statement, and the letter gives no date for reaching scale. Say you take the margin figure against the segment it applies to. AWS earned $45.6 billion of operating income on $128.7 billion of sales in 2025, a margin of 35.4% (source: Amazon 2025 Form 10-K). In the second quarter of 2026 the segment earned $16.6 billion on $42.2 billion, or 39.4% (source: Amazon Q2 2026 results release). One hundred basis points, or one percentage point, of margin on 2025 sales is about $1.3 billion of operating income a year, so a few hundred is a few billion dollars against a segment already earning more than $45 billion. The figure is Amazon's expectation for the inference part of AWS. The segment's margin moves for other reasons too, from the price of power to the useful life of servers, which Amazon cut from six years to five for part of its fleet from January 2025 (source: Amazon 2025 Form 10-K).
The capex figure sits against a larger number. Amazon expects to spend about $200 billion on capital expenditure in 2026, and says customer commitments make that spending predictable (source: Amazon 2025 shareholder letter). Tens of billions a year of chip savings would be a large share of that bill, which is why the letter presents the chips as changing AWS's economics rather than as a product line. Amazon's AI capex and cash flow follows the whole bill through depreciation to free cash flow, and the chips are one input to it.
The AWS margin is the line the market reads first on a results day, and the AMZNX/USDT Spot pair reprices on it the same evening; the pair's live price and order book are on its page.
What a holder of AMZNX or the AMZNUSDT Perpetual owns of the chips business
A chip dollar is an AWS dollar, and it reaches a holder of AMZNX only inside an Amazon share, because the chips have no stock of their own. Each AMZNX token is a certificate on one Amazon share held with a custodian, as what tokenized Amazon (AMZNX) is explains, and BloFin lists no instrument for Trainium, AWS or Annapurna Labs.
Say you hold 40 AMZNX, bought on September 21, 2026, when the AMZNX/USDT Spot pair traded at 255.78, which made those tokens about 10,231 USDT of Amazon. The chips business is a thin slice of that: a $20 billion run rate is about 15% of AWS's $128.7 billion of 2025 sales and under 3% of the $716.9 billion the whole company sold (source: Amazon 2025 Form 10-K). Whatever the chips add to AWS's margin reaches the token only through Amazon's reported operating income and the market's reaction to it.
The AMZNUSDT Perpetual gives a trader leverage on an index that follows the Amazon share price. Funding changes hands between longs and shorts every eight hours, and the contract holds no Amazon share and no claim on anything Amazon owns. On the same day it traded at 256.24 against an index price of 256.08, with a funding rate of +0.0089%. A results day that moves the AWS growth rate moves the index.
What moves Amazon's stock price ranks that growth rate first among the lines a release can move. Buying the token starts with the xStocks eligibility confirmation and ends with a spot order for a fractional quantity, the six steps in how to buy tokenized Amazon on BloFin.
Looking to gain exposure to Amazon? To get started, you'll need to first create a BloFin account, fund your account with cryptocurrency, and navigate to the AMZNX/USDT Spot trading page or AMZNUSDT Perpetual page.
Frequently asked questions
What is Amazon Trainium?
Trainium is the AI chip AWS designs for training and running AI models on its own servers. AWS describes it as a purpose-built AI chip designed for one goal, the best economics for high-performance AI training and inference at scale (source: AWS, Trainium). Amazon monetizes the chip only through EC2, which includes services such as Bedrock that run on it (source: Amazon 2025 shareholder letter). The current generation, Trainium3, packs up to 144 chips into one UltraServer, and AWS says its clusters can connect up to one million of them (source: Amazon, December 2, 2025).
Is Trainium a GPU?
Trainium is an AI accelerator, which Amazon's own explainer groups with GPUs as the chips that excel at the parallel arithmetic behind training and running AI models. A CPU, by contrast, handles the general tasks of running software and operating systems (source: Amazon, May 5, 2026). The difference is scope. A GPU stays flexible across a range of tasks, while Trainium is designed from the ground up for AI workloads alone, which Amazon says gives it greater performance and efficiency than general-purpose GPUs for large language models (source: Amazon, May 5, 2026).
Who makes Trainium chips for Amazon?
Trainium is designed by AWS, whose chip work began with its 2015 purchase of Annapurna Labs and produced Graviton the same way (source: Amazon, May 5, 2026). Inferentia is designed by AWS as well (source: AWS, Inferentia), and Nitro is Amazon's own EC2 network card (source: Amazon 2025 shareholder letter). Amazon says Trainium3 is built on a 3-nanometer process and that it engineered the chip, the server, the network and the software together; its pages do not name the foundry that manufactures the chips (source: Amazon, December 2, 2025). Annapurna Labs is an Amazon subsidiary that appears in the filings as a named defendant in the 2026 patent case over GB200 servers (source: Amazon Q2 2026 Form 10-Q).
Does Amazon sell its chips to other companies?
Only as computing time. A customer who wants Trainium rents it by the EC2 instance, on Trn3 UltraServers that became generally available on December 2, 2025, or through services such as Bedrock that run on the chips (source: Amazon, December 2, 2025). Inferentia is rented the same way, on Inf1 and Inf2 instances, with Inferentia2 delivering up to four times the throughput of the first generation (source: AWS, Inferentia). Amazon publishes no price for the chip itself; what a customer pays is the EC2 instance rate by the hour, and the only comparison Amazon gives is price-performance against comparable GPUs. The letter says Amazon may sell racks of chips to third parties in future, without giving a date (source: Amazon 2025 shareholder letter).
Is Trainium better than NVIDIA's chips?
Amazon's own answer is that it depends on the workload, and the comparison it publishes is price-performance against "comparable GPUs", with no chip named (source: Amazon 2025 shareholder letter). Its explainer says the choice between processor types is made per task, with GPUs the more flexible option and Trainium built for one kind of work (source: Amazon, May 5, 2026). AWS says it will keep making itself the best place to run NVIDIA while selling Trainium beside it (source: Amazon 2025 shareholder letter). For a customer the question is the price per unit of work; for a shareholder it is how much of that work lands on Amazon's own chip.
Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include Amazon's 2025 letter to shareholders, its results releases for the fourth quarter of 2025 and the first and second quarters of 2026, its 2025 Form 10-K and second-quarter 2026 Form 10-Q filed with the Securities and Exchange Commission, Amazon's and AWS's own chip pages, and BloFin's AMZNX/USDT and AMZNUSDT pages, read on September 21, 2026.
Nothing in this article constitutes financial advice, and nothing in it predicts AWS's margins, Amazon's share price or the outcome of any contract or case, or recommends buying, selling or holding Amazon in any form. The run rates, price-performance claims and margin expectations quoted here are Amazon's own management figures, not audited lines, and the chips' contribution to AWS is not reported separately. AMZNX tracks the share's price and carries the issuer's and custodians' risk in addition, and a leveraged position in AMZNUSDT can be liquidated by a single session's move. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.