Fireworks, Baseten and the price of inference

Two companies that run AI models for other businesses now want about twice what they were worth in the summer. Fireworks AI is weighing a new round at up to thirty billion dollars, reports said late last week. Baseten has been approached at about twenty-six billion, double its price in June. Most people have never heard of either. Both sell inference, which is the work of actually running a model every time someone uses it. So that’s what I want to settle today. Is running other people’s models worth that kind of money, or is it a GPU bill with a margin on top?

The second option is the one I keep coming back to. But let’s test it properly. This is Venture Place, a new weekly show, and this is our first episode. I’m Kirill Parinov, and my co-host is Kate.

Hi, everyone. And since we’re starting from zero, Kirill, let’s be strict about it. What do we actually know about Fireworks, and what’s only reported?

Less than the headlines suggest. What’s confirmed comes from Fireworks’ chief executive, Lin Qiao. On Friday she posted that the company is running a tender for its employees, led by Atreides, so the team can sell a hundred and thirty-two million dollars of stock. Her post doesn’t give a price. A report relaying The Information put it at seventeen and a half billion, the same as the July round. And the thirty billion comes from one person familiar with the matter, and only if a new round goes ahead.

So the hard number is seventeen and a half.

Seventeen and a half, with about one and a half billion raised in July. Last October the Series C valued them at four. So before anyone says thirty, that’s more than four times in nine months.

And Baseten?

Steeper, if anything. Axios had it first, on the twenty-second: Baseten was approached about doubling its valuation to about twenty-six billion. Bloomberg followed a day later and put that against thirteen billion in June. In January it was five. There’s no amount, no lead investor, nothing closed, and Baseten declined to comment.

Okay, let me defend them for a minute, because the demand side is hard to argue with. That same report on Fireworks says it reached a billion dollars of annualized revenue in July, four times a year earlier. Baseten says its revenue grew twenty times year over year as of June. And Bloomberg says Modal, which does similar work, is in funding talks too, and The Information named Fal next to Fireworks. Companies don’t grow like that selling something nobody needs.

I don’t doubt the demand at all. What I doubt is who gets to keep the money. The way I see it, an inference company runs models that somebody else built, mostly open-source ones in Fireworks’ case, on chips that somebody has to own. So it sits in the middle, between the labs and the people with the GPUs. And, um, the people with the GPUs have started selling inference themselves.

You mean Crusoe.

Crusoe, yeah. On the seventeenth it announced the first close of a three point nine billion dollar Series F, at thirty point nine billion post-money. It builds data centers and rents out the compute. And it says its managed inference business passed a hundred million dollars of contracted annual recurring revenue within a year of launch. Then on Wednesday it signed Thinking Machines Lab, Mira Murati’s company, for sixty-five million dollars a year, to run its open models on Crusoe’s cloud. Crusoe runs that cluster itself.

That’s one customer, though. Baseten says it works with thousands.

Fair. One deal isn’t a trend. It’s the direction that bothers me. If the companies that own the chips can offer the same service, the ones in the middle get squeezed on price. I can’t put a number on that yet. For now it’s a worry.

Then here’s a number from the other side. A week ago Bloomberg reported that Harvey, the legal AI company, saw its gross margin go from about fifty percent at the start of the year to minus fifty by June, after an update to its agents sent usage up. Harvey says it’s back to positive now, after it changed how it uses models and launched an open-weight model of its own. It also says its token use is up twenty times this year. Somebody has to run that kind of model, and run it cheaply. That’s the job Fireworks and Baseten are selling.

It’s a good example, and it cuts both ways. The margin figure rests on one anonymous source, so I’d hold it loosely. But look at what Harvey did. It went to an open model. Open models travel. You can run them on Fireworks, on Baseten, on Crusoe or on your own hardware. That’s great for demand. It also makes it easy for a customer to leave.

In theory. I doubt many companies move a production model every quarter to save a little on tokens. Speed and uptime are what they pay for, and switching is a pain.

That’s the strongest version of your case, and I half buy it. Switching costs are real. I just haven’t seen anything that tells me how big they are here. Let me ask you one back, though. If you ran Harvey, would you sign a long contract with any one of them?

Probably not a long one. Which is your point, I know.

It is, a little.

Let’s talk about the tender, then. Investors led by Atreides are taking a hundred and thirty-two million dollars of stock off Fireworks employees. If the reported price is right, that’s real money at seventeen and a half.

It is real money. Atreides Management also co-led Crusoe’s round. If that’s the same Atreides, and I’d assume it is, it has money on both sides of the argument we’re having. I read the tender as good news for the employees and fairly neutral for the price. It happened at July’s number. Thirty is still a target.

There’s a comparison in there you haven’t made yet. Crusoe at thirty point nine, Fireworks maybe at thirty.

That’s the one that sticks with me. Crusoe says it has more than a hundred and forty billion dollars of contracted value and over six gigawatts under contract. Those are its own numbers, and nobody has audited them. Fireworks would be at about the same price with a billion dollars of annualized revenue and no data centers.

No data centers is the point, though. Contracted value over years isn’t revenue this year. And Crusoe has to borrow to build. Earlier this month it launched a loan of about five hundred million dollars to pay for GPUs. Fireworks doesn’t carry any of that.

True. Owning the hardware is expensive, and Fireworks gets to skip that. But if GPUs stay hard to get, whoever owns them has the stronger hand. You’re right, though, that I’m comparing a backlog with a run rate, and they aren’t the same thing.

One more thing on Baseten. On Friday it hired Figma’s former marketing chief, who was there for Figma’s IPO. That doesn’t look like a company planning to stay small.

I wouldn’t read an IPO into one hire. They’ve added a revenue chief, a legal chief and a head of compute this year as well. It looks like a company getting ready to sell to bigger customers. Which, honestly, fits your demand story better than mine.

So, bottom line. Thirty for Fireworks, twenty-six for Baseten?

On demand, I’m fairly confident. A billion dollars of annualized revenue, up four times, is hard to argue with, even as a reported number. On the prices I’m much less sure. If you take that billion and the thirty-billion target, it’s thirty times revenue, for a business whose margins nobody outside has seen. And the companies that own the chips are already selling the same thing.

What would it take for you to believe the thirty?

A gross margin. If either of them shows one that holds up while revenue quadruples, I’ll stop calling it a GPU bill with a markup. And, I mean, if the round actually closes with a new investor leading it, that tells me somebody outside did the work and still paid.

Fair enough. That’s our first one. If there’s a company you think we should look at, or something you think we got wrong, send it to us on WhatsApp. You’ll find the link under this episode.

Talk to you next week.