Anthropic is reportedly paying $6 billion for a company that makes chips work harder. Read which number it's really buying.

The number everyone will repeat is $6 billion, and the number underneath it is 50 percent. Anthropic is in talks to buy Decart, an Israeli startup that was worth about $4 billion in a May funding round and roughly $3.1 billion the summer before that, Bloomberg and Reuters reported this week. Nothing is signed, and the talks could still fall apart. If they don't, this would be the largest acquisition Anthropic has ever made, and it would be priced at half again what the last set of investors paid three months ago.

A 50 percent markup in a quarter is the kind of figure that gets read as momentum. It is worth reading as something plainer: a price set by a buyer who needs the asset before a specific date. Anthropic filed a confidential draft prospectus with the SEC on June 1 and is reported to be aiming for a Nasdaq listing as soon as October, at a valuation last marked around $965 billion. The interesting question about this deal is not whether Decart is worth $6 billion. It is which line of Anthropic's own income statement the purchase is meant to change before that statement becomes a public document.

Two companies inside one startup

Decart, founded in 2023 by the brothers Dean and Orian Leitersdorf and Moshe Shalev, is really two businesses wearing one logo. The one that made its name is generative: a model called Oasis that produces a playable, Minecraft-like world on the fly, and a newer one called Lucy that rewrites a live video feed in real time — dropping clothing onto a person for e-commerce, restyling a stream. eBay, which is also an investor, is a customer. It is the eye-catching half, and it is not, on the evidence, the half Anthropic is buying.

The other half is unglamorous and is the point. Decart also builds software that squeezes more work out of the same silicon — better utilization of the GPUs a model already runs on. The company claims its stack delivers roughly eight times the throughput on the same hardware, at north of 80 percent utilization against an industry norm closer to 40 to 50 percent. Treat those as vendor figures until someone independent runs them. But the direction is the entire rationale. Anthropic does not need another video model. It needs to serve more tokens per dollar of compute it has already contracted for, because the cost of serving those tokens is the number a frontier lab can least afford to have looking wrong on the day it opens its books.

The line the prospectus can't hide

Here is the mechanism, because it is where the money actually moves. A software company is valued the way it is — on high multiples of revenue — because each additional customer costs it almost nothing to serve. Gross margin is the proof of that claim; it is the share of every revenue dollar left after the direct cost of delivering the product. For a conventional software firm it sits in the eighties. For a company whose product is generated by GPUs it rents or runs, the direct cost is compute, and the margin is whatever survives the electricity, the depreciation, and the chips. That single percentage is the difference between a software business and a compute reseller wearing software's valuation.

An IPO prospectus has to disclose it. Private rounds do not: a lead investor negotiates a headline number and the gross margin stays inside the building. A registered offering forces the number into daylight, audited, next to a growth rate. Anthropic is about to become the first pure frontier lab to walk a public market through exactly that line. Buying a company whose only job is to lift utilization is, read as a balance-sheet move, an attempt to walk in with a better number than it would otherwise be reporting.

You do not pay a 50 percent markup for a technology you have time to build. You pay it for a number you need before a date you've already set.

That is what the premium is buying: not eighteen months of internal engineering, but the margin now, ahead of the filing. It is a defensible thing to want. It is also the tell. You do not overpay for efficiency you consider optional. You overpay for efficiency you have realized is load-bearing — the hinge on which the whole software-versus-reseller argument turns.

The currency is the story

Now the part the headline number obscures. Anthropic is not a company with $6 billion of idle cash to spend; deals like this are done substantially in stock. And Anthropic's stock is marked at roughly $965 billion — a figure set in a private round that closed in late May, negotiated between the company and a handful of investors who wanted in. It is not a price discovered by anyone free to sell. When the payment is equity, the acquirer is spending a currency whose value is its own most optimistic private mark. The $6 billion is real to Decart's holders only to the extent that $965 billion is real to everyone else.

This matters because the mark and the listing are about to be tested against each other. If the public market prices Anthropic below its last private round — as it has recently priced other richly-marked names on their debut — then the shares handed to Decart were worth less than the sticker, and Anthropic will have paid for its margin improvement in a currency that was itself overvalued at the moment of sale. A company can buy a better income statement with an inflated balance sheet. Whether that was a good trade depends entirely on a price that has not been discovered yet.

The same names on every side

Follow the cap table and the picture gets more concentrated, not less. Nvidia was itself reported to be in advanced talks to buy Decart, and to have been outbid when higher offers arrived. Nvidia is also an investor in Decart, through the May round. And Nvidia is, separately, one of the chip suppliers on which Anthropic's own compute depends. So the same company was a bidder for this asset, is a shareholder in it, and sells the picks-and-shovels to the firm that looks set to win it. None of that is improper. All of it is the same small group of names appearing on every side of the same transaction — the structural signature of this entire cycle, where the customer, the supplier, and the investor keep turning out to be one another.

The valuation ladder tells the same story from a different angle. Decart went from about $3.1 billion last August to roughly $4 billion in May to a reported $6 billion now — each step set inside the same tight circle of AI-infrastructure capital, each one the input to the next. A price that only ever gets marked by parties who benefit from it going up is a measurement of enthusiasm, not of worth. It holds while the enthusiasm does.

Separate the company from the trade

Both companies here are real, and that is not the argument. Decart ships products people use. Anthropic is one of the two most important AI labs in the world and, by its own projections reported this spring, turned its first operating profit in the second quarter — on the order of $559 million on about $10.9 billion of quarterly revenue, with a run-rate that third-party trackers put well north of that by late July. This is not a weak company reaching for a crutch. It is a strong one, weeks from a public listing, choosing to spend richly-priced paper to improve the one metric its new investors will scrutinize first.

The trade is the thing to interrogate, and it rests on assumptions worth naming:

• That a utilization edge is durable. Efficiency is the most perishable advantage in this field. The tricks that lift tokens-per-GPU — better batching, quantization, scheduling — diffuse fast, and the cheapest open-weight labs have made a sport of publishing them. A margin bought today is a floor competitors reach tomorrow.
• That the $965 billion mark holds through the listing. If it compresses, the equity spent on Decart was worth less than booked, and the effective price of the margin rises accordingly.
• That the people are the asset, and that they stay. A great deal of what Anthropic would be buying is a few dozen engineers. Acquired talent has a way of vesting and leaving; the throughput numbers walk out with them.
• That buying efficiency doesn't quietly concede the larger point — that after all the talk of models as the product, the margin turned out to live in the plumbing, not the model.

The downside no one is pricing

The optimistic read is clean: a profitable, fast-growing lab bolts on a genuine efficiency layer, walks into its IPO with a healthier gross margin, and the whole thing looks in hindsight like good housekeeping before company. That may well be how it plays. But the risk sits exactly where the enthusiasm is thickest. This is a company using an untested public-market valuation as currency to buy a private asset priced by the same investors who set that valuation, in order to dress a number it is about to be forced to disclose. Each of those moves is individually reasonable. Stacked, they describe a market pricing its own confidence and calling the result a fact.

The number that will settle it is not the $6 billion and not the 50 percent. It is the gross-margin line in the prospectus, and then the opening print that tells you what Anthropic's currency was actually worth on the morning it had to pay. Until both are on the record, this is a very good company making a defensible move at a price only the buyer can currently see. That combination — a real business and a self-referential price — is the one worth watching, because it is the one the whole trade is built on.