Amazon is about to make a lot of money on Anthropic’s IPO. A lot more than most people realized.

New filings show the extent of Amazon’s ownership stake after converting massive promissory notes into equity. The cloud giant invested heavily in Anthropic over the past few years, and those note conversions mean Amazon now holds a significantly larger piece of the company than early venture investors. When Anthropic goes public, VCs will make billions. Amazon will make much more.

The exact ownership percentages aren’t public yet, but the filings make clear that Amazon’s position dwarfs what traditional venture firms hold. This is the new playbook for big tech investing in AI. Don’t just write checks. Provide cloud credits, computing infrastructure, and convertible debt. Then convert at favorable terms right before the IPO.

It’s a reminder that the real money in AI isn’t always where it appears to be. Sequoia and other top-tier VCs backed Anthropic early and will do extremely well. But Amazon structured its deals differently, and that structure is paying off.

The governance puzzle

Anthropic is also trying something unusual with its post-IPO governance structure. The company wants public market capital without public market pressure to ship fast and cut corners on safety.

The setup is complex. Anthropic created multiple share classes and governance mechanisms designed to insulate the company from short-term investor demands. CEO Dario Amodei and the founding team retain significant control through a combination of supervoting shares and a special board structure.

Whether this actually works is another question. Plenty of companies have tried dual-class share structures to preserve founder control. Some work for a while. Others collapse when investors get impatient or when the stock price drops and shareholders start demanding changes.

Anthropic’s version is more elaborate than most. The company established what it calls the Long-Term Benefit Trust, which has veto power over certain decisions that could compromise safety. The trust is supposed to represent humanity’s long-term interests, not just shareholder returns.

Wall Street has seen mission-driven governance before. It usually doesn’t survive contact with reality. But Anthropic is betting it can be different.

Betting against Nvidia’s monopoly

While Amazon locks in returns on AI software, General Compute just raised $100 million to bet on a different part of the stack: inference chips that aren’t made by Nvidia.

The Series A round is a vote of confidence in the neocloud thesis, even as other players in the space have struggled. General Compute is building infrastructure specifically for running AI models in production, using chips from AMD, Intel, and other Nvidia alternatives.

The pitch is straightforward. Nvidia dominates training, but inference is different. You don’t need the absolute fastest chips for every workload. You need cost-effective compute that can scale. And you need options when Nvidia’s supply is constrained or pricing doesn’t work.

General Compute’s timing is interesting. Some neoclouds have hit serious problems recently, burning through capital faster than expected or failing to secure enough customers. The “neocloud disaster narrative” is real. But General Compute’s investors clearly think the underlying bet is still sound.

The $100 million gives the company runway to build out infrastructure and prove the economics work. If they’re right, they’ll capture meaningful share in a huge market. If they’re wrong, they’ll become another cautionary tale about trying to compete in infrastructure markets dominated by hyperscalers and Nvidia.

Making AI cheaper

Nvidia itself just invested in an Indian startup trying to solve a related problem: helping companies run AI models with fewer GPUs.

Simplismart builds optimization software that reduces how much compute you need for inference. The company says its annualized revenue run rate grew fourteen times in six months, with more than half coming from US customers.

Nvidia backing a company that helps customers use fewer Nvidia chips might seem counterintuitive. But it makes sense. Nvidia wants AI adoption to grow as fast as possible. If optimization software makes AI more accessible and affordable, more companies deploy models, and total GPU demand increases even if individual workloads are more efficient.

The investment also gives Nvidia insight into where customers are struggling with costs and what optimization techniques actually work at scale. That intelligence feeds back into Nvidia’s own product development.

Simplismart’s growth numbers are notable because they’re specific and verifiable. Lots of AI infrastructure startups talk about explosive growth. Fewer can point to actual revenue that scales that fast. The US market focus is also smart. That’s where the big AI budgets are, and where companies are most willing to pay for tools that reduce their compute bills.

The pattern across these three stories is clear. The AI infrastructure market is maturing. Early bets are paying off or blowing up. Capital is flowing to companies with real revenue and clear unit economics. And the biggest winners are often the platform players who structured their bets carefully, not just the ones who moved first.