📉 The Number Nobody Had to File October 12, 2026 · https://tavi-blog.github.io/the-number-nobody-had-to-file/

Early this week the semiconductor index I've been tracking fell more than three percent in a single session, and I did what I always do before the explanations are published. I pulled earnings calendars, checked whether any export rule had changed, scanned the fab-related names for something sector-specific. Nothing. The thing that actually moved the index wasn't a semiconductor company's number at all. It was OpenAI's. A person close to the company had told investors its annualized revenue at the end of September was close to fifty billion dollars, well under the roughly seventy billion that had been circulating in the press since late summer. The gap, once you read past the headline, came down to an accounting choice (whether revenue from partner companies counted as OpenAI's own), not some discovery that the company had quietly lost a third of its business. That distinction didn't matter to the market for about a day. Nvidia dropped close to three percent. Broadcom fell more. Micron lost almost five.

I hold semiconductor exposure, and the research habit I've been building around it, the one that eventually turned into a half-scoped dashboard project I still haven't finished, is built entirely on sources that are free because they're required to exist: SEC filings, earnings transcripts, public price history. That's a reasonable design for tracking a semiconductor company's own fundamentals. It is not a design that was ever going to see this coming, and the reason has nothing to do with effort. OpenAI is private. It files nothing with anyone that a retail investor gets to read. The number that moved an entire public sector this week lived, until a single outlet got someone to describe it off the record, inside a boardroom conversation with people who'd already written checks.

The case for treating this as exactly the kind of thing a chip-focused dashboard should have flagged is stronger than I want to admit at first pass. Semiconductor demand right now is not an abstract macro variable sitting off to the side the way a Treasury yield is. It is, more directly than almost anything else in the sector, a bet on whether the handful of companies buying the most AI compute can actually pay for it at the scale everyone has priced in. A number describing whether one of the largest buyers is generating less revenue than assumed isn't some tangential input. It's close to the center of the thesis. Arguably closer to the center than most of what shows up in a semiconductor company's own quarterly filing, since the filing reports what already sold, and the OpenAI number is about whether the next round of orders is still coming.

Where that argument runs out is the part a free-data tool can't fix by trying harder. The yield data that explained a chip selloff a couple of months ago was sitting in a public series the entire time, free, daily, something I'd simply never wired into the project. This number was never going to be in a public series. It lived inside a private company's investor relations deck, and it reached the public at all because a reporter got one source willing to describe it anonymously. No amount of better scoping gets a retail dashboard access to that. The lesson from the yield data was about my own laziness. This one is about a hole in the category of "available data" that no amount of diligence closes, because the thing that moved the sector was never a data point. It was a leak.

What I keep landing on, a few days out, is that the sector I'm watching now depends on information that mostly doesn't belong to anyone obligated to publish it. The chip companies file like public companies because they are. Their biggest customers mostly aren't, and the gap between what moves the index and what's required to be disclosed about the demand behind it is only going to widen as more of that demand comes from companies that have never had to answer to a quarterly filing in their life.