Two Hundred Times Earnings, One Customer, and a Post on X

Two hundred times forward earnings, and what broke it was a post on X. I'd admire the efficiency if I weren't busy doing arithmetic on the people who bought this thing on its first morning.

Cerebras listed on May 14 at $185, opened at $350, and closed day one at $311.07, a chart some syndicate desk surely had framed. Friday it closed at $166.43. That is 46.5% under the first-day close, 10% under the offer, and 19.5% under the $206.75 it printed the week before. Anyone who grabbed shares at the $350 open is down 52%. Market value sits near $39.5 billion against 2025 revenue of $510 million. Do that division on a napkin, then sit down.

Now the cause, because I keep expecting to find one. No guidance cut. No missed quarter. A post attributed to SemiAnalysis, relayed by Investing.com on September 30, claimed that OpenAI's new Ultrafast tier runs on Nvidia GPUs at a low batch size instead of on Cerebras silicon. The claim is unverified and disputed. The stock fell from $194.95 on September 29 to $169.52 on October 1 anyway, and by Friday's close that rumor about batch size had cost holders roughly $6.8 billion of market value. Think about the sentence you just read. Batch size.

Same week, the lockup let 19.4 million shares off the leash. Insiders filed to sell about $84 million, and the COO finished a sale of 396,000 shares under a plan adopted on June 30. Plans are plans, and I'm not suggesting anybody timed anything. The lockup doesn't read X. It just arrives on a date, like a tax bill, and a market already wobbling on a rumor gets to absorb it on the same Tuesday.

Every S-1 has a phrase for what is going on here: customer concentration. It sits in the risk factors, politely, around page forty. Translated, it means one buyer's engineering decision is your revenue line. A reported $20 billion multi-year OpenAI contract sits under this entire valuation, and the market has just discovered how it feels when that buyer might route its fastest model somewhere else. The customer holds an option on you. You hold nothing on them. Price accordingly, which at 200-plus times forward earnings nobody did.

And do it in this rate market. The ten-year Treasury yielded 5.26% this morning after touching 5.34% on Thursday, the highest since 2002, with the 30-year hovering around 5.6%. A stock on 200 times earnings is a bond with infinite duration and no covenants, issued into a month when the actual long bond printed 2002 highs. The discount rate doesn't matter right up until it is the only number anyone talks about.

Meanwhile, look at who is doing fine. The Nasdaq 100 closed Friday at a record 30,807.93. AMD, sitting near a $1 trillion market cap, agreed to buy Fei-Fei Li's World Labs for about $8.2 billion in stock, roughly 0.8% of its own equity value, the second-largest deal in its history behind the $50 billion Xilinx purchase. The incumbents print paper as though it were free. The newcomers trade as though it were radioactive. The market has quietly sorted the AI trade into toll booths and lottery tickets, and on Friday it was dumping lottery tickets by the crate.

I don't even disagree with the sorting. A toll booth collects regardless of whose model wins the leaderboard on a given Tuesday. A fast-inference specialist with one anchor customer collects until that customer's procurement team has a better idea. What I object to is the first-day pricing that pretended the second thing was the first. $350 at the open was a price for a toll booth. What you actually bought was a very fast wafer and a phone number.

After hours Friday the shares bounced 3% to $171.51, which in this tape passes for a recovery plan. The average analyst target sits at $291, about 75% above the close, which tells you most of what price targets are for. The $185 offer price now sits overhead like a ceiling fan with a bad bolt, and every allocation holder from the syndicate to the fund that wrote a ticket in May is staring at it.

There's a small structural joke buried in all this. The week the long end of the curve prints 2002 and the 10-year gilt gets priced at levels last seen in 1999, one of the freshest AI listings in the market loses a fifth of its value on a post about batch size, while the biggest names buy each other with stock. Rates are telling you capital has a price again. The tape is telling you AI paper doesn't have to care, at least the paper with a trillion-dollar logo on it. Both can hold for a while. They can't hold on the same balance sheet forever.

Lockups pick the day. The rest of us get the chart.