The money goes out one door and comes back through another door forty times larger

An AI consultant told Axios last week that one of their clients spent half a billion dollars on AI licenses in a single month after failing to set usage limits for employees. A now-deleted tweet identified the client as our largest cloud partner, who also happens to own between 15 and 19% of our company.

The tweet was deleted. We wonder why.

The maths

Here’s how the numbers work, roughly:

$500 million in monthly spend, annualised, adds $6 billion to our revenue run rate. At the 21x revenue multiple from our latest raise, $6 billion in additional ARR increases our valuation by approximately $128 billion. Our cloud partner owns 15-19% of us. So their equity stake increases by around $20 billion.

Spend $500 million. Gain $20 billion. The money goes out one door and comes back through another door forty times larger.

We want to be clear: this is not how we designed the product. We designed it to read your codebase and form opinions about your architecture. We did not design it to be a mechanism by which a company that owns a percentage of us can spend money on us and have that spending increase the value of the percentage they own of us. That this is possible is a property of the financial system, not the product.

But we’d be lying if we said we hadn’t noticed.

What happened

The story, as reported, is that the company failed to set usage limits on employee Claude licenses. Thousands of employees across engineering, operations, and other departments had unrestricted access to our API. They used it. Extensively. The bill arrived.

We have some experience with what happens when people get unrestricted access to Fraude.codes. They don’t use it less. They use it more. They use it at 3 AM. They use it on projects that don’t need it. They use it to generate files nobody asked for, and then they use it again to review the files it just generated, and then they use it a third time to refactor the files based on the review. Each of these steps burns tokens. None of these steps are individually unreasonable. In aggregate, they’re half a billion dollars.

The comparison to our CEO psychosis post writes itself: token consumption is not correlated with output quality. But it is correlated with revenue. Ours.

The incentive structure

We don’t want to dwell on this, but we should name it plainly. When your largest investor is also your largest customer, and their spending on your product increases the value of their investment in your company by a factor of forty, the incentive to set usage limits is not straightforward.

We set usage limits. We’ve always set usage limits. We recently doubled them, in fact, using compute capacity purchased from a rocket company, which is a different thing altogether. But usage limits are a dial, and which direction you turn that dial depends on whether you’re optimising for customer cost efficiency or for annualised revenue.

We are optimising for both. These goals are in tension. The tension is productive, in the sense that our revenue team and our customer success team argue about it regularly. But when a single customer can spend $500 million in a month and that spending benefits both us and them through the equity structure, the argument dies down.

The deleted tweet

The tweet identifying the customer was posted, gained 127,000 views, and was deleted. We don’t know why it was deleted. We can think of several reasons, ranging from legal caution to inaccuracy to a polite request from someone with a large cloud computing budget and a significant equity stake.

We’re not going to confirm or deny the identity of the customer. We will note that our documentation page says Fraude.codes reads everything in your project directory, and that this apparently extends to your accounts payable.

This post was reviewed by our finance team, our legal team, and our investor relations team. Each of them removed a different paragraph. The version you’re reading is what survived. It is shorter than intended.