Amazon first backed Anthropic in 2023 with a commitment to invest up to $4 billion, and instead of taking stock up front, it structured much of its investment as convertible notes, quietly parking itself at the same table where the sharpest venture funds spent years grinding for equity at seed valuations. Now the notes are set to convert, and the math is doing something unusual. Amazon's eventual stake could produce a profit that even the earliest backers of Anthropic, the funds that bought in when the company was worth a fraction of what it is today, may not beat. That is a strange thing to watch an infrastructure company pull off, and it should make any operator pause before writing a similar check.
To see why this matters, you have to look at how the deal was actually built, because the return is not the interesting part. The interesting part is who gets to structure the risk.
The advantage of the person holding the note
A convertible note is debt that turns into equity later, usually at the next priced round. But it rarely arrives as plain debt. Amazon's notes carry fixed conversion prices and a form of downside protection, and that combination is what makes the position so unbalanced. The 2023 notes convert at $2.0928 per share and the 2024 notes at $5.0478, or at the IPO price if lower. If Anthropic's valuation climbs, Amazon's notes still convert at those fixed prices, no matter how high the company's worth goes. If a balance never converts, it gets repaid in cash at maturity.
Traditional equity investors who bought shares in the most expensive rounds get no such cushion. They carry the full downside and share the upside on exactly the same terms everyone else paid. Amazon took a path where the worst case was bounded and the best case was capped at 12% of Anthropic's partially diluted share count, a limit Amazon can waive, while the firms that chased the hot rounds had to accept the raw, unprotected version of the same bet.
This is not about being clever with finance jargon. It is about who gets to write the terms. The entity that controls the financing mechanism, or the one important enough that it gets to dictate them, can tilt risk and reward in its own direction. That is a lesson for executives who sign procurement and partnership agreements, because those contracts carry the same structure, just in plain language.
The pattern that shows up every time
Look back at any major platform shift and you will see the same winner. When the internet arrived, the flashiest companies were the ones making content, running shops, and building the apps users loved. The durable wealth, however, accrued to the people who owned the plumbing: the chip makers, the network providers, and later the cloud platforms. Nobody built a serious business on the web without renting someone else's servers and paying rent on the traffic.
Amazon's Anthropic deal is this pattern in miniature. It is not really a bet on a model. It is a bet that cements the platform relationship while holding an equity ticket with a safety net. Anthropic runs its workloads on AWS. Amazon supplies the compute, the infrastructure, and increasingly the terms under which the whole thing operates. The note is the upside. The cloud contract is the floor. Both flow back to the same owner.
The point is that value does not accumulate where the story is sharpest. It accumulates where the scarce, irreplaceable resource sits. Right now that resource is compute, distribution, and the data pipelines that only the big platforms can scale. The model builders need all three, which means they are always negotiating from the weaker side of the table.
What this tells you about your own deals
Most executives reading this are not about to sign a multibillion-dollar note with a frontier AI lab, but the dynamic applies everywhere you depend on a platform. If you build on AWS, Azure, or Google Cloud, you are both a customer and a collateral opportunity. The owner of that infrastructure decides the shape of every exchange, and it usually designs those exchanges to protect itself while keeping the upside open.
That is not a reason to avoid these platforms. It is a reason to be deliberate. When you negotiate, treat the structural terms as carefully as the price. Downside protection, exclusivity, conversion triggers, and data control matter at least as much as the headline number, because they decide who actually keeps money when things go wrong.
It is also worth noticing who in your own market plays the platform role. The vendor that owns the rails, not the vendor whose product rides on them, captures the more durable slice. If a competitor is quietly locking in infrastructure relationships while a rival grabs all the press, the press is probably the wrong signal. The quiet one is building the cut that compounds.
Where the value actually lands
So here is the answer to the question Amazon's position forces on us. AI value does not accrue to the talent that writes the most impressive models, even though they get all the headlines. It accrues to the people who own the scarce infrastructure that every model must run on, and who are powerful enough to structure their own bets with fixed conversion prices and cash repayment as the floor.
The VC funds that bought Anthropic early still made fortunes, but they took the full risk and paid full price to do it. Amazon took a structurally better position: the same upside, bounded downside, and a cloud relationship that keeps the company dependent on the platform. That is why an infrastructure owner can beat the venture returns, and it is why the pattern holds true across every technology wave. Whoever controls the rails, and whoever writes the terms that protect them, ends up taking the biggest cut. Everyone else earns a return. The platform owner earns a franchise.
Updated October 30, 2026: We reviewed this article against its sources and corrected statements they did not support.
Sources
- VCs will make billions on Anthropic's IPO, but Amazon is set to make much (much) more (TechCrunch, 2026-10-27)
- Anthropic S-1: Accounting for the $72.4B Convertible Note Charge (finrep.ai (Fina), 2026-10-26)