The filing that closed a February promise
Amazon told the United States Securities and Exchange Commission on 31 July that it has paid OpenAI the full 50 billion dollars it agreed to in February. The filing sets out the schedule plainly: 15 billion dollars in the first quarter, 13.7 billion in the second, and the remaining 21.3 billion after 30 June. The structure had been an initial 15 billion followed by a 35 billion tranche that was described at the time as conditional.
What the filing does not do is say why the condition was met. Earlier reporting had tied the remaining money to events such as a public listing or a declared step change in capability, and OpenAI has not gone public. Amazon simply recorded that the money moved. Reporting since the filing puts the resulting holding at close to 5 percent, a figure that comes from journalists rather than from the filing itself, and it is worth keeping those two things apart when you repeat the number.
Exclusive is doing more work here than fifty billion
The sentence in Amazon's own announcement that changes a purchasing decision is not about the investment. AWS is described as the exclusive third-party cloud distribution provider for OpenAI Frontier, the product through which an organisation builds, deploys and manages teams of AI agents. Read that literally. OpenAI can still sell to you directly. Every other cloud on the market, including every European sovereign provider a company might have spent two years qualifying, cannot carry that product.
For a business in Frankfurt or Manchester that has deliberately built a multi-cloud position, this is the first time the agent layer of a major model vendor has been given a single commercial route through a third party, and that route belongs to a United States hyperscaler. The models remain available by other means. The packaged agent platform does not.
The money contracted back is larger than the money invested
Alongside the investment, Amazon says the existing multi-year compute agreement worth 38 billion dollars grows by a further 100 billion over eight years. Add those and OpenAI is contracted to spend roughly 138 billion dollars with AWS. Amazon put in 50 billion. On the published terms the contracted spend running back to the investor is about 2.8 times the investment, and that ratio is ours, derived from the two figures Amazon disclosed rather than something either company presented.
This is a familiar shape to anyone who has financed a customer, and it is not by itself a criticism. It does tell you what the transaction is for. Amazon has bought a claim on the compute budget of the company most likely to set the price of agent software in Europe, and it has bought the right to be the only third party selling that software through a cloud.
Two gigawatts of silicon that is not Nvidia
The compute is specified as roughly two gigawatts of Trainium capacity, covering Trainium3 and the next generation Trainium4, with delivery of the latter expected in 2027. Amazon's chief executive Andy Jassy said the company was excited about OpenAI choosing to go big on its custom silicon. Sam Altman framed it as putting powerful AI into the hands of businesses at real scale.
The portability consequence sits underneath the marketing. Work tuned for a vendor's own accelerators does not lift cleanly onto a different fleet, and the more of OpenAI's marginal capacity that runs on Trainium rather than on merchant GPUs, the more the practical cost of serving you is tied to one supplier's hardware roadmap. A model you reach through an interface is portable. A serving stack optimised around one accelerator family is a commitment, and it is being made on your behalf.
The clause to read before your next agent pilot
The useful action this month is narrow and it is contractual. If you are scoping an agent deployment for the autumn, establish in writing which components you are buying directly from the model vendor and which arrive through a cloud marketplace, because those two routes now have different competitive futures. A direct contract leaves your cloud position intact. A marketplace contract for OpenAI Frontier does not, because there is presently one marketplace.
The wider discipline is to stop treating a model vendor as a single supplier. It is a stack, and the layers are being sold under different terms, on different exclusivity, and increasingly on different silicon. Your exit options differ by layer, and the only way to know yours is to ask before the pilot rather than at renewal.
Read next: AWS Marked Nova Legacy in March, Not This Week | Nothing Crashed, and It Cost $1.8 Million
