The chips repriced. The numbers didn't.

The Read · 90 secondsThe Read

Samsung reported the best quarter in its history this month, operating profit up roughly nineteenfold from a year earlier. The stock fell about seven percent on the news. Across three weeks the chip complex gave back more than a fifth of its value while almost every number underneath it got better.

That gap, between what the businesses reported and what the market paid for them, is the month's story, and it is the kind our instruments are built to separate. When prices and fundamentals move in opposite directions, the useful question is which one is doing the moving.

FNC-1, our index of the seven companies actually pouring the concrete of the AI buildout (NCB-003), reads against the SOX, the chip index most people watch, and the S&P 500. Beside it sits the Belief Index, a read on how convinced the betting markets are that the boom keeps going.

The decision this Reading informs belongs to a utility executive weighing whether to commit interconnection capacity to a data-center contract running past 2030. A twenty percent drawdown in chip stocks is either the first sign the demand curve was fiction, or it is noise in a market that reprices faster than transformers get built. Those readings imply opposite answers, and the instruments separate them more cleanly than the tape does.

A note on the gap. This is the first Reading since R-012 on June 29. Four weeks of data in one issue, and every comparison to June is dated as such.

Capital

The builders kept falling less than the chip trade around them, for the second month running. That is the half of the sentence that matters.

The events were concentrated and they were about price, not production. The chip index peaked around June 22 at an all-time high after a twelve-month run of roughly 130 percent, then shed more than twenty percent in three weeks. Intel fell 21 percent across seven sessions. Micron came off its peak by about 17 percent. The proximate triggers were a report that SK Hynix would slow its high-bandwidth memory expansion in favor of higher-margin DDR5, profit-taking around that company's Nasdaq debut on July 13, and a hawkish turn at the Federal Reserve under Chair Kevin Warsh, where nine of eighteen policymakers now project a 2026 rate hike against zero in March. Goldman Sachs data showed hedge funds trimming technology-hardware exposure for a fourth straight week.

Against that, the operating numbers ran the other way. Samsung's record quarter. Micron reporting $41.46 billion in quarterly revenue at an 84.9 percent gross margin, with high-bandwidth memory booked out through fiscal 2026. Hyperscalers still guiding to roughly $725 billion of combined 2026 AI capital spending, up about 77 percent on 2025.

The instrument confirms the shape. At the July 20 close, FNC-1 reads 128.5, the SOX 209.3, the S&P 500 116.0, all rebased to 100 at the start of the trailing twelve months. The gap to the chip index stands at 80.9 points, in from roughly 100 in June, and it closed the same way: the chips fell toward a basket that did not move as far.

The mechanism is what the chip index holds that our basket does not: the merchant and memory names, Intel and Micron among them, the parts that trade on sentiment and led the drawdown. FNC-1 holds the seven whose capital expenditure builds data centers, power, and frontier models. They fell less because they are priced as construction rather than momentum.

Inside the basket the composition has gone extreme, and this is the number to carry out of the issue. Decomposed by substrate on the same twelve-month window: Compute 184.2, Energy 121.0, Frontier 120.4, Biological 45.6. Nearly 139 points between the strongest and weakest substrates of one index. The buildout, measured this way, is one substrate carrying three.

Belief

The Belief Index reads 49.7, up from 44.3 at the June 22 reading. That comparison spans five weeks, not one. Read the direction, not a weekly slope.

Read

R-012 closed with a call: watch whether the builders keep outholding the chips on the way down the way they underperformed them on the way up. Five weeks on, they did. The chip index gave back more than twenty percent from its June peak, the basket gave back a fraction of that, and the gap closed by roughly nineteen points for the second consecutive reading, the same way both times. The call held, through a larger test than the one that produced it.

What the four-substrate decomposition shows, and no chip ticker can, is that the basket's resilience is not evenly owned. Compute at 184.2 against Biological at 45.6 is one substrate carrying the composite while a second sits at less than half its base. Note what that does to the June frame. R-012 read the builders as the low-beta core the chip index swings around. That holds for the composite, but the part doing the holding is the compute-linked part, which is the part most exposed to the same complex that just fell twenty percent. The distinction between builders and chips is thinner than it looked.

Belief moved up, and its composition says what it said in June from the other direction. In June the index fell eleven points, entirely on the timing bets, while the crash bet held. This month it recovered five points almost entirely on one timing bet, GPT-6, while the crash bet again barely moved. Two readings, opposite directions, same decomposition: the market keeps repricing when the acceleration arrives and keeps declining to reprice whether the floor holds. A drawdown this size that leaves implied burst odds at 16 percent is not a market rehearsing a collapse.

Put the two together and the month reads as a valuation event that left the option space almost intact. For the utility executive, very little closed: capex guidance did not move, memory is booked out through fiscal 2026, and a price correction in listed equities does not un-commit a contracted load. What did narrow is more specific. SK Hynix slowing high-bandwidth memory expansion in favor of DDR5 is a capacity decision with a fab-scale lead time, and capacity not started is capacity that cannot be summoned inside the window a 2027 decision would need it. The drawdown is reversible the way prices are reversible. A deferred expansion is reversible in principle and slow in practice, and the second kind of move quietly removes options while the first kind gets the headlines.

Nothing broke this month. What it demonstrated is that this complex will reprice by a fifth on positioning and rate expectations while every operating number improves. That capacity was always latent and is now observed, which is a different thing from a buildout that has stopped.

The Register

We opened our threshold register this week: what we track, what counts as movement, and the date we grade ourselves. Two dimensions to start.

The FNC-1 / SOX spread, 80.9 points. Threshold: closes to 40 or widens past 120. The Belief Index, 49.7. Threshold: two readings above 55, or one below 40. Both graded August 31.

We also published what we declined to register. The Biological substrate rests on one listed company, which is too thin to carry a threshold. Full entries and limitations at fp1.ai/register.

Radar

The delta since the June 29 sweep on SA-001, covering four weeks rather than one.

Substrate (substrate gates capability) — held, strengthened. July's binding constraint was not capability or demand. It was a memory producer's decision about which capacity to build, and a rate environment pricing the financing of physical plant.

The Seam (US–China) — moving, inside the band. Reporting that Beijing is consulting its own AI and chip firms on export controls, including tiered review that could keep frontier systems at home. In June, Washington offered chips and Beijing declined them. Same substrate-sovereignty logic, other side.

The board holds at five of five, one strengthening, no falsifications. Full board at fp1.ai/radar.

What this measures, what it does not

The Belief Index is one venue's positioning, the Polymarket panel, weighted by FP1's view of which markets matter. Individual components will be wrong. Three of the four resolve on December 31, 2026, and as they approach resolution their prices compress toward certainty and the index carries less information. The panel is scheduled for expansion before that binds.

FNC-1 is a measurement instrument, not an investment vehicle. It is rebased to a rolling twelve-month window, so the basket level is not strictly comparable across issues and the spread to the chip index is the number to watch. Substrate readings here are comparable to each other on the same window, not to the substrate numbers printed in June.

Both fail openly. Every number traces to the chart pipeline or a named source, every forward claim carries a falsifier and a grade date, and when a read breaks we log it. The method is in NCB-003 and NCB-004.

Cadence

The Reading runs every Monday, returning to weekly with this issue. The first Reading of each month carries the Radar Delta. Standing calls are tracked at fp1.ai/radar, and the register is at fp1.ai/register.