AWS Marketplace fee on a private offer of $10M or more
1.5%
Listing fee schedule effective Jan 5, 2024; renewals also priced at 1.5%
Anthropic revenue booked through cloud marketplaces
~44%
Nine months ended Sep 30, 2026, per its Form S-1/A, versus ~11% in 2023
AWS net sales
$45,984M
Quarter ended Sep 30, 2026, up 39% year over year
Azure revenue growth
+48%
Microsoft's September 2026 quarter, to about $33.2B
Amazon's holding of OpenAI Series C preferred stock
$50.0B
Nine months ended Sep 30, 2026, including $21.3B in the September quarter
Anthropic's net loss
$104.0B
Nine months ended Sep 30, 2026, against $33.2B of revenue
The event
Microsoft Lost Exclusivity in April. November 9 Changed Something Else.
On November 9, 2026, OpenAI announced that ChatGPT Enterprise is available through AWS Marketplace. Enterprise buyers discover, buy and manage seats through the purchasing relationship they already have with Amazon, and eligible spend can be applied against existing AWS commitments; the same route extends to ChatGPT for Healthcare and ChatGPT for Financial Services. Pricing is set through custom private offers, so no list price is published and account teams confirm eligibility before a customer can transact. Nothing about the model, the seat price or the contract changed. OpenAI routes seat purchases through the AWS procurement rails enterprises already run on.
- Moves the sale inside a vendor the buyer already pays, so procurement cycles shorten rather than lengthen.
- Lets eligible AI spend count against AWS commitments, which reclassifies unbudgeted software spend as billable cloud consumption.
- Extends the same private-offer rail to the healthcare and financial services editions, not just the flagship enterprise seat product.
- Leaves the compute footprint untouched for now: OpenAI products still ship first on Azure unless Microsoft declines a capability.
The framing that matters here is about where a decision gets made, not who owns a model. When Microsoft's license to OpenAI's intellectual property became non-exclusive on April 27, 2026, the exclusivity argument ended — but the enterprise seat budget did not move, because most buyers still procure through a cloud contract they already hold. A marketplace listing changes that: it makes the seat legible to a procurement system that already knows the buyer, the billing code and the discount agreement. That is a repricing of distribution, and it is why the fee Amazon collects is the least interesting number in the transaction.
Channel economics
Amazon's Fee Is 1.5% to 3%. The Prize Is the Commitment Drawdown.
AWS Marketplace listing fees are a rounding error next to the seat economics. A public SaaS offer pays 3%; a private offer pays 3% below $1 million of contract value, 2% from $1 million, and 1.5% at $10 million and above, with renewals also at 1.5% and a further 0.5% uplift when a channel partner is involved. The consequential line is not the fee. A marketplace bills the customer, collects the money and remits it to the seller net of platform and other fees — the exact mechanic Anthropic's IPO filing describes for its own marketplace sales — so the channel a lab picks now sets the gross margin on every seat it sells.
| Offer structure | Listing fee |
|---|---|
| SaaS standard (public) offer | 3% |
| Private offer, contract value under $1M | 3% |
| Private offer, $1M to under $10M | 2% |
| Private offer, $10M or more | 1.5% |
| Private offer renewals | 1.5% |
| Channel partner private offer (CPPO) | +0.5% on the listing fee |
Evidence from a filed prospectus
Anthropic's IPO Filing Shows Where Frontier Revenue Actually Lands
Anthropic's Form S-1/A, filed November 2, 2026, is the closest thing to filed evidence on how frontier-lab enterprise revenue actually gets collected. Revenue booked through cloud service provider marketplaces rose from about 11% of the total in 2023 to 32% in 2024 and 47% in 2025, and stood at roughly 44% for the nine months ended September 30, 2026, with those marketplaces also representing about 52% of net receivables. For scale, press reporting put OpenAI's annualized revenue run rate near $70 billion in late September 2026, with its enterprise segment growing 107% on a run-rate basis in the September quarter. OpenAI's move is the same trade run in a different direction: Anthropic put its enterprise product on Amazon's marketplace in July 2025, and OpenAI is now putting its seat product on the same rail a year later.
Anthropic revenue booked through cloud marketplaces
Share of total revenue, per Anthropic's Form S-1/A filed Nov 2, 2026
Unit: % of total revenue
2023
11%
2024
32%
2025
47%
9M 2026
44%
| Period | Revenue | Net loss | Cloud marketplaces as % of net receivables |
|---|---|---|---|
| 2023 | $41M | $1.5B | — |
| 2024 | $386M | $8.3B | ~42% |
| 2025 | $4.6B | $42.0B | ~60% |
| 9M ended Sep 30, 2026 | $33.2B | $104.0B | ~52% |
- Remits to the lab net of platform and other fees, which makes the channel a gross-margin decision rather than a marketing decision.
- Concentration is rising at the customer level — the top five customers were about 25% of revenue in the first nine months of 2026, with no single customer above 10%.
- Growth is outrunning the economics — nine-month revenue of $33.2B came with a $104.0B net loss, so distribution reach is not a substitute for unit economics.
The counterparty
Microsoft Loses Position, Not Contract
Microsoft's April 27, 2026 amendment kept the economics and dropped the exclusivity. The license to OpenAI's intellectual property runs through 2032 and is now non-exclusive; OpenAI products still ship first on Azure unless Microsoft declines a capability; OpenAI can serve products on any cloud; and revenue-share payments from OpenAI to Microsoft continue through 2030 at the same percentage, now subject to a total cap. OpenAI's commitment to use at least $250 billion of Azure services by 2032 also survives. Microsoft trades a legal moat for a contractual annuity — and the procurement channel was never part of either contract.
| Metric | Change | Year over year |
|---|---|---|
| Azure revenue | +$10.8B | +48% |
| Agents and Infra segment revenue | $76,795M | +25% |
| Microsoft 365 commercial cloud revenue | +$4.1B | +17% |
| Microsoft 365 commercial seats | — | +5% |
| Frontier and support services revenue | +$86M | +4% |
| Microsoft Cloud gross margin | 66% | down on Azure mix |
The line to watch is \"Frontier and support services\" — Microsoft's separately disclosed bucket that carries OpenAI's Azure-hosted Frontier product alongside support — which rose $86 million, or 4%, while Azure grew 48%. Microsoft attributes the increase to growth in Customer Success Support, so this is not proof that OpenAI seat revenue stalled; it is proof that Azure's acceleration is not coming from OpenAI's own seat product. The structure underneath is worth naming: OpenAI's first-party products, including Frontier, stay hosted on Azure, while Amazon holds the exclusive third-party cloud distribution rights to Frontier. The split already existed; the marketplace listing simply formalizes it on the buyer's invoice.
Competing rails
The Labs Are Becoming Marketplaces Themselves
Two months before OpenAI put ChatGPT Enterprise on AWS Marketplace, it launched a marketplace of its own at DevDay on September 29, 2026, with 32 launch partners. The mechanic is the mirror image: eligible enterprise customers apply part of their OpenAI commitment toward qualifying partner products, the customer contracts with and is invoiced by the partner, and OpenAI confirms eligibility and reconciles the amount against that customer's commitment. There is no self-service checkout, and eligibility is decided customer by customer, product by product, purchase by purchase. Anthropic runs the same play with its own storefront.
- Hands the lab control of the enterprise AI budget line — commitment-eligible spend now clears through the lab's own reconciliation process.
- Makes the rails interchangeable: if the binding constraint is which commitment the invoice hits, AWS Marketplace, OpenAI Marketplace and a lab's own storefront become competing paper trails.
- Opens a third route for the same seat — Microsoft Foundry, AWS Marketplace and OpenAI Marketplace can each be the invoice trail for one enterprise AI budget.
Downstream read-through
Seat-Priced Software Gets Repriced
The bill lands first on software sold per seat. ServiceNow's third-quarter results, published October 28, 2026, showed total revenue up 20.5% to $4.11 billion, subscription revenue up 21% to $3.99 billion and cRPO up 20.5% to $13.65 billion — and GAAP operating margin of 9.5%, against 16.8% a year earlier. The 2026 AI annual contract value target was raised to $1.7 billion, 70% above where the year started, while fourth-quarter cRPO growth was guided to 18.5% in constant currency, below what had just been delivered. Growth and margin are now moving in opposite directions for the seat-priced layer.
| Metric | Q3 2026 | Comparison |
|---|---|---|
| Total revenue | $4,106M | +20.5% YoY |
| Subscription revenue | $3,988M | +21% YoY |
| cRPO | $13.65B | +20.5% YoY |
| GAAP operating margin | 9.5% | 16.8% a year earlier |
| 2026 AI annual contract value target | $1.7B | +70% from start of year |
| Q4 cRPO growth guidance (constant currency) | 18.5% | vs 20.5% delivered |
Upstream read-through
Every Seat Sold Is Inference Bought
Every seat sold through a marketplace is inference that has to be paid for in silicon. OpenAI's February 27, 2026 announcement with Amazon put roughly 2 gigawatts of Trainium capacity behind OpenAI's workloads, spanning Trainium3 and Trainium4 parts with deliveries starting in 2027 and more high-bandwidth memory per accelerator. Amazon put $50.0 billion into OpenAI's Series C preferred stock in the nine months to September 30, 2026, and NVIDIA put $30 billion into the same February round on top of 3 gigawatts of dedicated inference capacity. The bill is already visible: Amazon spent $75.9 billion of cash capital expenditure in the September quarter, against $34.2 billion a year earlier, and added a further $57.1 billion of property and equipment under finance leases. Seat growth is a capacity commitment in disguise.
The memory layer is where the pricing risk sits. OpenAI says more than 9 million paying business users rely on ChatGPT for work, against more than 900 million weekly active users overall, and the Trainium4 step-up in memory per accelerator ties that demand directly to high-bandwidth memory supply. SK hynix is the clearest listed expression of it: trailing revenue of ₩268.9 trillion at a 74.5% operating margin. Yet the shares closed at ₩1,136,000 on November 9, 2026, about 62% below their 52-week high of ₩2,987,000 — the market has already started discounting the end of the memory cycle while the income statement still shows extraordinary margins. That asymmetry is the upstream risk to the whole seat story: if per-seat pricing falls faster than usage grows, the memory layer's margin, not the lab's revenue, absorbs the shock.
| Company | Metric | Period | Value |
|---|---|---|---|
| Amazon | AWS net sales | Q3 2026 | $45,984M, +39% |
| Amazon | Cash capital expenditure | Q3 2026 | $75.9B (vs $34.2B a year earlier) |
| Amazon | OpenAI Series C preferred stock | 9M ended Sep 30, 2026 | $50.0B |
| Microsoft | Azure revenue | Q1 FY2027 | ~$33.2B, +48% |
| Alphabet | Total revenue | Q3 2026 | $127,577M |
| NVIDIA | Revenue | FQ2 FY2027, ended Jul 26, 2026 | $96,221M |
| Broadcom | Revenue | FQ3 FY2026, ended Aug 2, 2026 | $29,591M |
- Turns software renewals into capacity commitments — the Trainium deal alone is about 2 gigawatts of dedicated accelerators.
- Links seat growth directly to memory content — the Trainium4 step-up in high-bandwidth memory per part is the physical form of inference demand.
- Holds four positions at once — $50.0 billion of OpenAI equity, 2 gigawatts of chips, the marketplace fee and a competing model family on Bedrock.
What to watch
Two Clocks, One Test
The channel question resolves on two clocks. In the next two quarters the evidence is price and mix: what a seat costs, whether a buyer applies it to a cloud commitment, and what OpenAI's next enterprise disclosure shows about revenue invoiced by someone else. Over one to three years the evidence is capacity and contract: Trainium4 deliveries from 2027, memory content per accelerator, and the expiry of the arrangements that currently protect both Microsoft and the labs. Watch the commitment calendars, not the launch announcements.
- Days to quarters: OpenAI's next enterprise disclosure shows how much seat revenue is now invoiced by Amazon rather than sold direct — Anthropic's IPO prospectus is the template for what gets reported.
- Days to quarters: Microsoft Foundry and bundled Microsoft 365 Copilot offers face a rival that can be bought against an AWS commitment; paid Copilot seats were about 15 million against more than 450 million commercial seats.
- One to three years: Trainium4 deliveries from 2027 and memory per accelerator decide whether this is a capacity event or a pricing event for the entire seat layer.
- One to three years: revenue share from OpenAI to Microsoft runs to 2030 under a cap and the Azure commitment to 2032, after which Foundry and bundling carry the position.
- Risk to the thesis: if procurement committees treat a marketplace listing as a compliance formality and keep buying seats through the lab's sales team, the channel shift stays cosmetic and the fee stays a rounding error.
Where the channel shift actually lands
- AWS Marketplace private-offer fees of 1.5% at $10M-plus contract value make the channel cheap for OpenAI and sticky for buyers.
- Turns seat renewals into AWS commitment drawdown, defending the $75.9B quarterly capex ramp with committed AI workload.
- Holds four positions at once — $50.0B of OpenAI equity, 2GW of Trainium, the listing fee and Nova on Bedrock — which caps how hard it can push.
- Near term: AWS net sales of $45,984M in Q3 2026 grew 39%, and the listing gives that growth a non-infrastructure source.
- Keeps revenue share through 2030 under a cap and a $250B Azure commitment to 2032, so cash economics survive April 2026's loss of exclusivity.
- Loses default seat placement: buyers with AWS commitments can invoice OpenAI seats inside Amazon procurement instead of Azure.
- Azure grew 48% to about $33.2B in the September 2026 quarter, but the separately disclosed Frontier and support services line rose only $86M, or 4%.
- One to three years: once the revenue-share cap binds and the 2032 Azure commitment ends, Foundry and bundling carry the position.
- Google Cloud is one of the three hyperscaler marketplaces Anthropic's IPO filing shows collecting 44% of revenue in the first nine months of 2026.
- Captures the same commitment-drawdown economics through Google Cloud Marketplace while supplying multi-gigawatt TPUs with Broadcom from 2027.
- One to three years: $127,577M of Q3 2026 revenue lets Alphabet price channel terms hard if labs defect to hyperscaler marketplaces.
- Owns the compute both channels must run on: $30B into OpenAI's February round plus 3GW of dedicated inference capacity.
- FQ2 FY2027 revenue of $96,221M with $59.7B of net income funds capacity for seat growth regardless of which marketplace wins.
- Longer term: seat expansion raises tokens per user, the durable driver of accelerator demand even if per-seat pricing falls.
- Halved GAAP operating margin to 9.5% from 16.8% in the September quarter while subscription revenue still grew 21%.
- Raised its 2026 AI annual contract value target to $1.7B, up 70% from the start of the year, while guiding Q4 cRPO growth to 18.5% in constant currency.
- Bearish on seat pricing: when a general-purpose AI seat can be bought through a hyperscaler marketplace, workflow-seat upsell pricing weakens.
- Near term: FY2026 subscription revenue guidance of $15.78B to $15.79B is intact; the risk is the next guidance cut, not this quarter.
- Captures custom-silicon content from marketplace-led AI build-outs, including the multi-gigawatt TPU agreement between Alphabet and Anthropic running from 2027.
- FQ3 FY2026 revenue of $29,591M and $13.1B of net income show the custom-silicon franchise compounding on hyperscaler capacity deals.
- Longer term: seats routed through hyperscalers require hyperscaler-designed silicon — Broadcom's revenue line, not the lab's.
- Earns the memory margin behind every inference token a seat generates — trailing operating margin of 74.5% on ₩268.9 trillion of revenue.
- Trailing operating margin of 74.5% means small changes in inference volume swing the line hard; at ₩1,136,000 the shares sit about 62% below the 52-week high.
- Near term: OpenAI's Trainium4 step-up in high-bandwidth memory per part from 2027 is the demand pull; the risk is memory pricing normalising first.
