Book-to-bill is a terrible measure of AI demand (and it’s about to be everywhere)
The ratio analysts are dusting off to justify AI semiconductor hype, why it was never a great signal, and what MU and SNDK are already telling you.
Book-to-bill is having a moment. If you have not seen it yet, you will. The ratio is being pulled out of the 1990s analyst toolkit and aimed at the AI buildout, and the pitch is simple: look at the orders, the demand is real, the cycle has legs. It is a clean story. It is also not a great measure of anything.
What book-to-bill actually is
The ratio is straightforward:
\[ \text{Book-to-bill} = \frac{\text{orders booked}}{\text{revenue billed}} \]
Above 1.0, you are booking orders faster than you can ship. Below 1.0, you are draining backlog. The ratio is a leading indicator in principle — an order is signed before revenue is recognized, so it reads ahead of the income statement by a quarter or two. That lead time is the entire appeal.
The ratio has its roots in SEMI’s North American semiconductor-equipment book-to-bill, which for decades was the barometer for the chip cycle. SEMI stopped publishing the bookings half of it in 2017. The billings report carries on without it. Even the canonical source eventually admitted the number was not worth the trouble.
Why it is suddenly back
The AI capex story needs a forward indicator. Revenue is backward-looking. Capex guidance is forward-looking but gets revised. Earnings calls are narrative. Book-to-bill sits in the sweet spot: it is quantitative, it leads, and it sounds like engineering rather than storytelling.
Sell-side analysts are already building the decks. The argument will go something like: “AI infrastructure names are printing book-to-bill above 1.5x. The backlog is stretching. This is not 2000. This is demand you can measure.”
The problem is that the ratio is only as good as what is underneath it, and what is underneath it in mid-2026 is uneven at best.
Two memory names, two very different prints
The memory space is where the book-to-bill narrative collides with reality most sharply. Two names make the point.
Micron (MU) looked like it was heading into a cyclical downdraft. Memory is notorious for boom-bust: oversupply builds, PC and smartphone demand softens, inventories swell, margins crater. It happened in 2023 and the market was bracing for a repeat. Then Ford stepped in. A long-term automotive supply agreement gave Micron a demand floor that the broader memory cycle could not take away. The stock held. Book-to-bill for the auto segment probably looks fine. The story works — but only because a single customer deal changes the math, not because the cycle itself resolved.
SanDisk (SNDK) did not get a Ford. The print is not close. Margins guided to 83–85%, roughly in line with the ~84% the market already expected. That is not a bad margin in absolute terms, but it signals the peak. When you guide margins flat at cycle highs, you are telling the market the easy gains are done.
More telling is what is happening with pricing. SNDK recently signed supply deals with eight large customers, and the contracts reportedly include price bands in both directions — a bottom to stop a collapse, and a top that limits how much they can charge even if spot prices run. The presence of a floor is the part worth sitting with. You do not negotiate a minimum price because you are worried about demand being too strong. You negotiate it because you see the risk on the other side. The market is pricing the top of the cycle, and book-to-bill will not save you from that. If anything, a high book-to-bill at the peak of a memory cycle is the trap — it tells you orders are strong right before they roll over.
The metric is a narrative tool, not a truth machine
Book-to-bill has structural problems that make it a dangerous thing to hang a thesis on.
First, comparability is a mess. Every company defines backlog differently. One quotes a twelve-month backlog. Another quotes total backlog including unsigned awards. A third quotes a segment-level figure and calls it a company-wide read. The ratio is only as consistent as the denominator, and the denominator is different in every filing.
Second, the derivation is fragile. If you do not get the ratio stated directly — and most companies do not state it — you derive it from backlog changes. That derivation assumes the backlog moved only because of new orders minus deliveries. Cancellations, foreign-exchange translation, acquisitions, and a company quietly redefining what counts as backlog all break the arithmetic.
Third, a single large order can push a quarter above 2.0x and say precisely nothing about the run rate. The ratio is point-in-time, lumpy, and easily distorted. It tells you what happened last quarter, not what will happen next.
The inference story cuts both ways
I wrote earlier this year about there being no inference moat. DeepSeek was compressing the problem from the model side, Cerebras from the hardware side, open source from the orchestration side. Inference was getting cheaper, everywhere, all at once. The thesis was that no single vendor could lock down the stack.
DeepSeek just announced it is raising token prices. Significantly.
Take a moment with that. The company that symbolised free-falling inference costs is now raising rates. That is not a contradiction of the moat thesis — pricing power eventually asserts itself when the dust settles — but it complicates the demand story that book-to-bill is supposed to validate. If inference is cheap enough to eat everything and expensive enough to raise prices on, which one is the signal? Analysts will cherry-pick the half that supports the deck.
The AI semiconductor trade needs a narrative to justify valuations at these levels. Book-to-bill is being positioned as that narrative. It is a number. It is not a thesis. The distinction matters.
This is analysis for my own notes. It is not investment advice.