Etched is worth $20 billion. It is also worth $10 billion. Read the gap, not the number.

The interesting thing about Etched this week is not that a four-year-old chip company is reportedly being valued at twenty billion dollars before it has sold a single chip. That happens now. It happens often enough that it has stopped being news on its own. The interesting thing is that Etched is also, in the same week, being valued at ten billion dollars, by a different set of buyers who are no less sophisticated than the first. Two syndicates, one company, one product still in a lab, and a factor of two between them.

A valuation is supposed to be an estimate. It is a guess about future cash flows, discounted back to a number you are willing to write on a wire transfer today. When two estimates of the same asset, made by serious people in the same seven days, land twice apart, the number has stopped estimating anything. It has started voting. And the thing it is voting on is not Etched's cash flows, because Etched does not have any yet. It is voting on a story.

Here is what is reportedly on the table, per the Wall Street Journal's account and the coverage that followed it. Etched is raising two concurrent rounds. The larger one, at roughly a twenty-billion-dollar valuation, is led by Jane Street, an existing backer, and would roughly quadruple the company's last mark. The smaller one, at roughly ten billion, is led by Sequoia. Neither has closed. Terms could still move. Both sets of investors are the kind of people who are paid, professionally, not to be wrong about this. That is the part worth sitting with. This is not a smart-money-versus-dumb-money story. It is smart money disagreeing with smart money by a factor of two, out loud, at the same table.

The real thing, first

I try to start these columns by separating the real thing from the trade built on it, because conflating the two is how people lose money in both directions, and the AI boom has given them more chances to do it than any market I have covered. So: Etched is a real company with a genuinely interesting idea, and I want to say that plainly before I spend the rest of this column being difficult about the price.

The company was founded in 2022 by three Harvard dropouts. Its chip, called Sohu, is an ASIC — an application-specific integrated circuit — built to do one thing: run transformer models, the architecture behind ChatGPT, Claude, Gemini, and very nearly everything else you have typed into a text box in the last three years. Sohu does not train models and it does not run the older or stranger kinds. It runs transformers, and the pitch is that by hardwiring that one architecture into silicon instead of leaving it to a general-purpose GPU, you get a large efficiency win. The company's own claim, made back in 2024, is that a single eight-chip Sohu server replaces a hundred and sixty Nvidia H100s on transformer inference. It says it has more than a billion dollars in signed customer contracts waiting on a chip it is still validating, with the first racks due to ship around now.

That is not a fraud. It is not vaporware. Specialization is a real lever, and the history of computing is full of moments where a fixed-function chip crushed a general-purpose one at the exact job it was built for. If you believe, as Etched's backers evidently do, that transformer inference is about to become one of the largest and most durable workloads in the economy, then building the best possible chip for precisely that workload is a defensible thing to attempt. The question was never whether Sohu is clever. Sohu is clever. The question is what you pay for clever, and whether either of the two numbers on the table is a price or a wish.

A good idea and a sane price are different claims

A specialized chip is, by construction, a bet that the thing it specializes in does not change. That is the whole trade. You give up flexibility, you get efficiency, and the efficiency is only worth anything for as long as the world keeps running the function you froze into the wafer. Which means an ASIC's real risk is not competitive and it is not financial. It is architectural. It is the risk that the workload moves.

A specialized chip is a bet that the thing it specializes in does not change. In a field that reinvents its own architecture on roughly an eighteen-month clock, that is not a small bet to make with someone else's twenty billion dollars.

And here is the part the twenty-billion-dollar mark quietly assumes away: the field is visibly, actively trying to move off the exact design Etched has committed to silicon. The transformer is a 2017 idea. It has been extraordinarily durable, and it may stay dominant for years. But in just the last several months the frontier has been full of people building around its limits — state-space models, linear-attention variants, diffusion-style language models, the custom attention schemes that Chinese labs have started shipping under their own names. Some of these will go nowhere. That is how research works. But you do not need all of them to succeed. You need only one of them to become the thing the world runs in 2028, and a chip that can run nothing but the classic transformer becomes a very expensive paperweight with a heat sink. I am not predicting that. I am telling you the price is.

The precedent everyone has agreed to forget

The reason I distrust this particular round number is that I have the longer memory here, and the memory is unkind. The AI-chip challenger — the startup that was going to dethrone Nvidia with a better-designed piece of silicon — is one of the most reliably disappointing trades of the last decade. Not because the chips were bad. Often the chips were good. The trade disappointed because the moat kept moving, and the moat was never really the transistor count. It was the software, the ecosystem, and the workload, all of which Nvidia either owned or could chase.

Run the tape:

• Nervana, an AI-chip startup acquired by Intel in 2016 to great fanfare, its architecture effectively shelved within a few years.
• Habana Labs, bought by Intel in 2019 for about two billion dollars to be its inference answer, and largely sidelined since.
• Graphcore, once a fourteen-billion-dollar British hope with rapturous press, sold to SoftBank in 2024 for a reported fraction of that.
• Wave Computing, a 'dataflow' challenger that raised heavily and filed for bankruptcy.
• And the current generation — Groq, Cerebras, SambaNova, Tenstorrent — each with a real and defensible engineering story, each of which has spent the AI boom explaining why this time the incumbent's moat finally cracks.

The point is not that these were bad companies. Several were excellent. The point is that being an excellent chip company that is not Nvidia has, so far, been a way to lose money slowly while being right about the technology. That is the precedent the consensus has quietly agreed to forget in order to underwrite a twenty-billion-dollar mark for a company whose first product has not yet earned a dollar. The chips were real. The question was always what you paid.

What two valuations are actually telling you

Now, the dual round. Raising two rounds at once, at two different marks, has become a small fashion in this boom, and it is worth understanding as a mechanism rather than a curiosity, because the mechanism is the tell. When a company can clear a twenty-billion-dollar round and a ten-billion-dollar round simultaneously, it is not confused about its own worth. It is price-discriminating. It is letting different buyers self-select by conviction. The believers pay the believers' price. The people who want the exposure but cannot stomach the top number get let in lower. The company keeps its options open and its cap table full.

That is shrewd corporate finance. It is also, if you are the one being asked to buy at the higher mark, a piece of information the company is handing you for free. The existence of a round at half the price, for economically similar ownership, in the same week, means the twenty-billion-dollar figure is not a floor. It is the top of a range that the company itself is quietly disclosing. When the seller shows you two prices, the honest one to plan around is rarely the larger. It is the one below it that also cleared.

There is a reflexive quality to all of this that ought to make a careful buyer nervous. The higher a private mark goes, the more it functions as marketing — for the next round, for recruiting, for the customer who wants to bet on a winner. A twenty-billion-dollar headline helps close the very billion dollars of contracts that are then cited to justify the twenty-billion-dollar headline. That is not fraud either. It is just a loop, and loops feel like momentum right up until the moment they are asked to settle in cash.

What is actually priced in

Strip the story down and the higher number asks you to underwrite three things at once. First, that the transformer stays the dominant workload long enough for a single-architecture chip to earn back its silicon — architectural permanence, in a field whose entire culture is the opposite of permanent. Second, that Nvidia stands still, which is a strange thing to assume about a company that reprices its own lineup every year; the hundred-and-sixty-H100s comparison is already being measured against a part Nvidia is busy replacing, and the denominator in that ratio does not sit still for anyone. Third, that a billion dollars of contracts signed against an unshipped chip convert into durable, repeat revenue once the chip meets the messy reality of a customer's data center, which is a thing that has humbled better-funded companies than this one.

None of those three is impossible. All three at once, capitalized today at twenty billion dollars, is a wager that everything breaks Etched's way and nothing in a famously fast-moving field moves against it. The ten-billion-dollar syndicate is making a gentler version of the same wager, which is why I keep saying the two numbers are not a contradiction to be resolved. They are a range the company is telling you. Believe the lower end until the chip ships and does in a customer's rack what it does on a slide.

I will do the thing I always do and admit the risk in my own position, because being early is a way of being wrong and I refuse to launder it into being right. Etched may well grow into twenty billion dollars, and then past it, and this column will read as the usual skeptic missing the usual rocket. I have been early before. But early is not the same as wrong about the price, and the discipline that has kept me solvent is simple: do not confuse a real technology with a sane valuation, and when the seller shows you two numbers twice apart in the same week, do not flatter yourself that you know which one is right. The company just told you it does not either. Read the gap. It is more honest than either figure inside it.