AWS Trainium Revenue: $225 Billion Committed, and No Trainium-Only Number

Amazon has never published a Trainium revenue figure. What it has published is a $225 billion pile of Trainium revenue commitments, a chips run rate above $25 billion that also counts two other chip lines, and a long list of chip counts and gigawatts. AWS Trainium revenue is therefore something you have to infer, and the inference depends on which of three kinds of number you are reading.

The short answer. AWS Trainium revenue is not disclosed as a standalone number. The closest figures are a chips run rate above $25 billion at Q2 2026, which includes Graviton and Nitro, and over $225 billion of Trainium revenue commitments stated in April 2026. Commitments are future contracted spend. Neither figure tells you what Trainium earned in any quarter.

This post is written for a founder or director who buys compute, sells into AI infrastructure, or holds a view on Amazon. That reader needs to know which Trainium numbers can carry a decision and which cannot, before the next call on 29 October 2026.

Every Trainium statement Amazon has made, in order

The record below runs from the Q3 2025 call to a podcast Jassy recorded four days before this post. Each row is classified by what the number actually measures. I opened every source listed. Where I relied on a transcript service or secondary report, the row says so.

Only one row in ten carries a Trainium-specific dollar figure, and it is a commitment. The Q3 2025 call quotes are from the Motley Fool transcript. The October 2026 podcast quote is as reported by 24/7 Wall St and has not been checked against the recording.

Revenue, capacity and commitments are three different numbers

Every Trainium statement falls into one of three buckets. Mixing them is the most common error in coverage of AWS Trainium revenue, and it is easy to make because Amazon often presents them in the same paragraph.

Revenue: what was earned

A revenue run rate annualises a recent period, usually the latest quarter or month. It is the only kind of number that tells you what customers are paying now. Amazon has given three run rates for its chips business, and every one of them combines Trainium with something else.

The single Trainium-only revenue statement is qualitative. On the Q3 2025 call Jassy said Trainium2 was "a multibillion-dollar business that grew 150% quarter-over-quarter". Multibillion is a floor of $2 billion and has no ceiling, so it does not size anything.

Capacity: what was installed

Capacity numbers count chips landed, power connected or clusters built. Amazon has been generous here: 1.4 million Trainium2 chips landed by February 2026, and more than half of 2.1 million AI chips landed in the 12 months to Q1 2026 were Trainium.

A chip that is installed is not a chip that is billed. Capacity tells you the size of what can be sold. It says nothing about price, utilisation or which customer pays.

Commitments: what was promised

Commitments are contracted future spend. The $225 billion Trainium figure sits here, alongside the gigawatt pledges from Anthropic and OpenAI. They indicate demand and give visibility, but they convert into revenue only as capacity is delivered and used, over a period Amazon has not stated.

My view is that only the first bucket should ever be called "Trainium revenue". The other two are inputs to a forecast, and treating them as results is how a capacity story gets priced as an earnings story.

The AWS chips run rate changed scope between Q4 2025 and Q1 2026

The neatest-looking series in this story is the chips run rate: over $10 billion, then over $20 billion, then over $25 billion. It looks like a business that more than doubled in two quarters. The definitions say otherwise.

The Q4 2025 release, dated 5 February 2026, says "Trainium and Graviton now have a combined annual revenue run rate of over $10 billion". The shareholder letter and Q1 2026 release describe the over $20 billion figure as inclusive of Graviton, Trainium, and Nitro. Nitro is the network interface card inside every EC2 server, which Jassy's letter calls "our EC2 NIC".

Amazon's own Q1 2026 call excerpt cites nearly 40% quarter-over-quarter growth in the chips business. If the Q4 base had been exactly $10 billion, 40% growth gives about $14 billion, not $20 billion. So either the Q4 figure was well above its "over $10 billion" floor, or adding Nitro accounts for much of the gap. Both are possible, and the disclosures do not let you tell which.

Amazon has not published a restated Q4 2025 figure on the three-chip basis. Until it does, the honest growth series runs from April 2026 only: over $20 billion to over $25 billion in one quarter, on a consistent definition.

Graviton is a general-purpose CPU used by 98% of the top 1,000 EC2 customers, and its revenue commitments nearly tripled in Q2 2026. Nitro ships in every instance. Both are large, mature businesses that have nothing to do with training a frontier model.

So a reader who wants Trainium in dollars has to subtract two unknowns from a threshold. I would not put any Trainium-only revenue estimate in a model on that basis, and I am wary of anyone who does without showing the subtraction.

The $225 billion Trainium commitment figure, read carefully

On the Q1 2026 call, Jassy said Amazon now has "over $225 billion in revenue commitments for Trainium". It is the only Trainium-specific dollar figure Amazon has given. It is also nine times the combined chips run rate disclosed a quarter later.

What the figure leaves out

Four things are missing, and each one changes what the number is worth to you.

• Term. No period is given. $225 billion over 3 years and $225 billion over 10 years describe different businesses.
• Customer split. The two named Trainium commitments are Anthropic, at up to 5 gigawatts, and OpenAI, at about 2 gigawatts ramping in 2027. Amazon has not said how much of the $225 billion they represent.
• Firmness. No cancellation or minimum-take terms are disclosed. "Up to" in the Anthropic wording is a ceiling, not a floor.
• Relationship to backlog. Amazon has not said whether the $225 billion sits inside the AWS backlog it reports, or how much of it is.

The Anthropic loop

Anthropic is the anchor Trainium customer and also an Amazon investment. Amazon's releases record pre-tax gains of $16.8 billion in Q1 2026 and $53.4 billion of other income in Q2, primarily from Anthropic. That is about $70 billion in two quarters, nearly three times the chips run rate.

None of that is improper. It does mean the customer behind a large share of Trainium demand is also the source of Amazon's largest non-operating gains, which is the pattern examined in the breakdown of how the AI circular deals actually work. A commitment from a company you partly own deserves a different discount rate than one from an arm's-length buyer.

How Broadcom discloses custom ASIC revenue instead

Broadcom designs custom accelerators, which it calls XPUs, for other companies. Google's TPU is the best known. That makes it the closest public comparison for Trainium, and its disclosures are structured very differently.

One revenue line, a stated mix, named customers

Broadcom's fiscal Q3 2026 release reports AI semiconductor revenue of $16.7 billion, up 221% year over year and 54% on the prior quarter. On the call, Hock Tan said XPUs were 73% of AI revenue in the quarter, which works out to roughly $12.2 billion of custom accelerator revenue in 13 weeks. Broadcom guided $21.7 billion of AI revenue for fiscal Q4 and $58 billion for fiscal 2026.

Tan also named customers by gigawatt. Anthropic is deploying 1 gigawatt of Ironwood TPUs in 2026, is expected to take another 5 gigawatts of TPU version 8i in 2027, and has "an incremental 10 gigawatts" in line of sight for 2028. He said Anthropic "is on track to become our largest XPU customer in 2027".

Broadcom's forward figures come from the call. CFO Amie Thuener said the company does not intend to update that AI revenue guidance quarterly, so treat it as a dated statement of intent, not a running forecast.

Why the comparison is not like for like

Broadcom sells silicon. Amazon rents capacity built on its own silicon. Jassy says as much in his 2025 letter: the run rate is "somewhat understated by our currently only monetizing our chips through EC2", and a standalone chip business would run at about $50 billion. That $50 billion is a hypothetical. It is useful only as Amazon's own estimate of the gap between rental revenue and silicon value.

Putting Trainium and Broadcom on one scale: dollars per gigawatt

Both companies now talk in gigawatts, which gives a rough common unit. The calculation below is mine. It is illustrative, it rests on assumptions Amazon has not confirmed, and I show the working so you can reject any step.

Step 1: the two named Trainium commitments are up to 5 gigawatts for Anthropic and about 2 gigawatts for OpenAI, so up to 7 gigawatts. Step 2: assume, generously, that those two account for the full $225 billion. Step 3: $225 billion divided by 7 gigawatts is about $32 billion of committed spend per gigawatt.

For comparison, Tan said on the Broadcom call that Broadcom's content per gigawatt runs between $20 billion and $30 billion. Broadcom's number is chip and networking value only. Amazon's commitments should cover more than chips, because a cloud contract also pays for power, buildings and margin.

If the $225 billion covered only those 7 gigawatts, Amazon would be collecting barely more per gigawatt than Broadcom collects for silicon alone. Either the commitments include other customers, which lowers the per-gigawatt figure further, or the term runs long. Or Trainium's whole price advantage shows up here as a lower bill per gigawatt.

The third reading is consistent with what Amazon says it is trying to do. Jassy wrote that Trainium will "save us tens of billions of capex dollars per year" and give several hundred basis points of margin advantage. That is a cost claim, not a revenue claim, and it is the strongest argument for why Amazon may never break out Trainium revenue at all.

Where this analysis is weakest

The case above leans hard on what Amazon has not said. There are good reasons to discount that criticism, and you should weigh them.

Amazon may simply be reporting the business it runs

Trainium is not sold as a chip. It is consumed inside EC2 instances and Bedrock, whose inference mostly runs on Trainium. Splitting a cloud bill into the chip and everything else is an allocation choice, and any number Amazon picked would be arguable. Broadcom has a cleaner line to report because its product is the chip.

Seen that way, the combined chips run rate and the $225 billion figure are more than most clouds give. Microsoft and Google do not publish comparable run rates for Maia or TPU capacity. Amazon is disclosing more than its peers, just not the number you want.

The comparison figures have their own gaps

Broadcom's 73% XPU share comes from a call remark, not the filed release, and its 2027 and 2028 figures will not be updated each quarter. The per-gigawatt calculation stacks three assumptions. The October podcast quote comes through a secondary outlet. Any one of these could move the conclusions above, and none of them is audited.

Finally, the disclosures are dated. Amazon reports Q3 2026 on 29 October, two days after this post. A Trainium-only line could appear there and make most of this piece out of date, which would be the right outcome.

What to listen for on Amazon's 29 October call

Four questions decide whether AWS Trainium revenue becomes a number you can check or stays an inference. I would listen for each, in this order.

1. Is there a Trainium-only figure? Even a run-rate threshold, separate from Graviton and Nitro, would end the guesswork.
2. Has the $225 billion moved, and is there a term? A larger number with no period attached adds little. A smaller number with a duration adds a lot.
3. Is the chips definition unchanged? If the run rate is still "inclusive of Graviton, Trainium, and Nitro", the Q2 to Q3 change is a real growth rate for the first time across three quarters.
4. How do commitments relate to AWS backlog? Saying what share of backlog is Trainium would connect the commitment figure to reported numbers.

The backlog question matters beyond Amazon. How each cloud reports contracted revenue, and how far those figures can be trusted, is set out in the comparison of cloud backlog and RPO disclosures. The equivalent question for the other big lab customer is worked through in the analysis of OpenAI's compute commitments against its revenue.

It also helps to know where Trainium sits in the wider spending picture. The comparison of hyperscaler capex against operating cash flow shows how much of Amazon's build is funded from operations, and the ranking of AI revenue disclosure metrics scores run rates, commitments and segment lines by how much each can tell you.

Frequently asked questions

Where to start with Trainium numbers

If you only do one thing this week, build a three-column sheet: revenue, capacity, commitments. File every Trainium figure from the 29 October call into one column before you read any coverage of it. Most of the confusion in this story disappears once the columns are separate.

If you are pricing compute, ask your AWS account team one specific question: what is the committed term and minimum take on any Trainium reservation you are offered? The answer tells you how firm the $225 billion is likely to be across the customer base, and it is the one data point Amazon has not published.

Related reading