Anthropic洽谈以60亿美元收购AI初创公司Decart

Yeah. This is a startup called Descartes, an Israeli startup that does two things. They work on world models, wild models, literally AI models that try to understand the world, uh, more rooted in in real world physics. They're trained on things like video data. It's beyond text. But the second thing that they do is have a software platform that makes chips better running AI, you know, optimizing basically. And it seems like based on our reporting, that's the bit anthropic is interested in.

So it's a $6 billion deal we're reporting citing sources. We don't know the structure of the deal, but in the context of what you just said, it would be anthropic, the biggest acquisition, acquisition to date. And with an IPO right around the corner, you know, it's interesting timing, but it kind of speaks to the the operational or technical challenges that anthropic so has is this out of the ordinary though? Uh, for anthropic though because they really make these big acquisitions.

This is a big one. I mean, but they've been really spending heavily, you know, computing power. They want to develop new products. Yeah, I don't know, $6 billion doesn't seem very big to me. Uh, when you think about the balance sheet, I mean, the way that I look at it is anthropic raised even this year, you know, north of $100 billion. And most of that sits on its balance sheet. You know, in, in, you know, even the money it raised prior to this year probably still sits on its balance sheet.

And we are seeing valuations of very specific, uh, neat niche niches that some of these startups occupy. BB at that level be high. You know, there is a lot of activity. And the frontier Labs, the hyperscalers have used M&A to acquire those specialisms. Um, in this case, you know, Descartes says that that software to optimize the the stack and make the chip run better. And do we know anything more about a potential IPO of anthropic, the timing thereof?

I mean, we're just we were mentioning space is now, uh, back above its IPO price. So presumably that that bodes well for the tech IPO space. But anything on timing. Yeah I mean the Bloomberg reporting is still that it's this full. I mean, with respect, as a British person who's lived in the States for almost a decade, I still have no idea what the full actually means. It seems an unspecified and an infinite period of time, but that's what we've reported, I think.

You know, late September or early October is what we had previously reported. And so it's kind of full steam ahead. And you make a really interesting point, which is, you know, I think I reported this at the time, right. Once the space prospectors hit, OpenAI and anthropic had a good chance to read it and decide what to do. And, you know, whatever their psychology was behind closed, it was ultimately they both filed confidentially with the SEC and space, you know, was it was, uh, led the way in that respect.

And I'm going to stick with space since you mentioned it. There was talk about actually Descartes could be acquired by SpaceX, but Elon Musk. That's right. That's not true. Fake news not happening. So why would that company be of appeal to space? Well, uh. You. Perfect summary. There were reports that SpaceX had offered wanted to offer the similarly $6 billion. Um, I think it was the report was that it was an all stock proposal, and Elon Musk said that it was fake news, as you said.

Same logic as anthropic. You know, with the greatest respect to de Carr. And again, this is all Bloomberg's reporting, right. You know anthropic has not confirmed this acquisition. Nordica. But their work in world models is interesting is not unique. There are others working world models. You know the property that that folks are interested in is that software that that just makes the chips better. Uh, from an economics perspective and operation that running AI models.

And in a world where, you know, the economics of leasing out compute capacity, running your own models on your own compute capacity is so important. And there is a supply constraint, there is not enough compute to meet the demand for running AI models that would make that that software particularly interesting. And it kind of goes back to your question on the valuation right of and in the $6 billion scale of the deal, at cost of using AI, the token cost users here, you're seeing stories that CFOs and managers are saying, whoa, whoa, whoa, whoa, what are you using I for?

What am I paying for? What do we know about just the user cost experience. Yes, there are two different, uh, parts to that. The the cost of running a model is measured on a dollar per, uh, million input tokens and a dollar per million output tokens basis. And so when we get releases from, say, like Kimmy, K3, um, from, from China, one of the reasons that was such a big news story is that it was a high performance model, highly capable, but it was priced quite reasonably, um, against how the absolute top US models were priced.

A token spend is a budgeting issue that basically technology companies around the country, some of them have got wrong. So if you're the CFO, you have to assign a budget, right? How much money you're going to allow your engineers to spend to run AI models. And because some of these companies have consumption basis arrangements with the Frontier Labs, a customer of entropic, OpenAI is consumption based. They just burn through them.

And then you have to go to the CFO and go, oh, I actually, I, you know, I said I'd spend ten, I spent 100, and then you're in trouble.