Cross-reference three sources before believing one. That is the rule I run every file through, and it is the reason the latest round of semiconductor earnings stops me — because for once, the sources line up in the same direction, and a story that has been told for quarters is starting to carry a ledger.
Here is what the numbers say, in order. A major domestic memory maker reported first-half revenue of 150.31 billion yuan, up 873.64% year on year, and swung to a net profit of 77.605 billion yuan — a turnaround from a loss. A leading foundry reported net profit of 4.467 billion yuan, up 94.20%. Those are two independent companies, two different parts of the chip chain, both reporting the same direction in the same half-year. That is not one source telling a flattering story. That is triangulation.
Separate the signal from the noise first
The first thing to do with numbers this large is to be suspicious of them. A revenue line up eightfold could be a base effect — a depressed prior year making any recovery look heroic. So let me check that before going anywhere else. The memory maker’s prior-year half was indeed weak; that is what makes the percentage look extreme. But the swing to a positive 77.6 billion profit is not a base-effect artefact. Profitability of that size is the actual signal, and it says the pricing environment for memory is not merely recovering — it is in a phase where pricing power has returned to the sellers. Call it what the data shows: a supercycle in memory pricing, not a polite rebound.
No, let me be more precise, because ‘supercycle’ is a word people use loosely. A rebound is a price recovery back toward a normal range. A supercycle is a period where supply discipline meets structurally elevated demand and prices stay high long enough to re-rate the whole industry. The 873.64% revenue growth alone does not distinguish the two. The combination does: record-scale revenue, a swing into deep profitability, and a research view — from a brokerage house, dated this month — that AI-driven demand pushes global semiconductor equipment spending to a record high this year. Equipment spending is the forward indicator. Ordering machines to add capacity is what sellers do when they believe the current pricing is not a blip. That is the part I trust most.
What the foundry number adds
The foundry’s 94.20% profit growth matters for a different reason. Memory and logic are cousins, not twins, and when both print strong numbers in the same quarter it reduces the chance that we are looking at one product-specific spike. The foundry’s gain is broad — it reflects utilisation, pricing, and mix across a wide customer base, not one commodity line. Two different supply chains telling the same story is precisely the cross-validation pattern I look for in a file: if it were one company, I would file it as a single-source claim. With two, in different sub-sectors, the hypothesis moves from unverified to leaning-confirmed.
The part analysts keep conflating
Now the discipline part, because this is where I have seen otherwise careful people drift. There are two claims on the table. Claim one: the memory cycle is strong — this is now well-supported by the earnings data. Claim two: the domestic chip supply chain has ‘delivered’ — this is a different statement entirely, and the earnings do not prove it by themselves. A memory maker swinging to profit proves the market is good. It does not automatically prove that the domestic production base is cost-competitive or that the advanced-process bottlenecks have been cleared. Profitability in a seller’s market can mask a lot of structural cost that will reappear the moment pricing normalises.
This is the hedged position, and I want to state it plainly so it is not mistaken for doubt about the cycle. The cycle is real. The durability of the domestic supply chain at the next downturn is unverified. Those two sentences can both be true, and an honest analyst keeps them in separate drawers. The moment they are merged into a single ‘semiconductor is winning’ narrative, you have a story, not a dossier.
A concrete check from my own file
Let me ground this in a specific scene from the desk. When the first earnings release crossed my screen, I flagged the 873.64% figure with a red tag — base effect, single source, treat with caution. It sat there for two days while I waited for the foundry numbers and a second independent write-up. When both arrived and agreed, I moved the file from ‘unverified’ to ‘leaning-confirmed’. That is the whole process in miniature, and it is deliberately dull: nothing about the method is clever, and everything about it is about refusing to let one loud number finish a sentence.
The same discipline applies to the equipment-spending view. A brokerage forecast is a forecast, and forecasts are hedged instruments by design — useful for direction, unreliable for magnitude. What makes it more than a guess is the corroborating detail in the same report: that the record is driven by AI demand specifically, which ties it to the same structural driver behind the earnings. When the demand story, the earnings story, and the forward-spending story all point at the same engine, the probability has genuinely shifted. I still would not call it confirmed. I would call it the strongest semiconductor-cycle signal of the year so far.
What the ledger means for decision-makers
For someone making capital allocation or procurement decisions, the practical reading is straightforward. The cycle being real changes the default assumptions: memory prices stay firm, capacity comes back online, and anyone building a plan on ‘cheap memory is around the corner’ should re-check their arithmetic. The counter-cyclical buyers who waited for a pullback have now missed the turn, which is how cycles punish prudence when prudence was mis-timed. The honest note is that the entry point has passed; the defensible move now is to size exposure to the cycle’s remaining runway, and to keep the structural-risk drawer open.
I will close with the discipline, because it is the deliverable. Two independent earnings reports, in memory and in logic, agree with each other and agree with a forward equipment-spending view. That is the strongest corroboration I can construct for the supercycle claim, and I am still holding one question open: how the domestic supply chain behaves when the seller’s market ends. The cycle has a receipt now. The receipt does not cover the whole bill.
Walk the chain before you trust the headline
The discipline I run on any market file is to walk the chain, not the headline. A memory maker reports a record quarter; a foundry reports accelerating profit; an investment bank forecasts record equipment spending. Three different links in the same industrial chain, three different institutions, three different kinds of documents — and they agree on direction. That is the triangulation pattern that moves a claim from unverified to leaning-confirmed. But note what the chain does not tell you. It tells you demand is real; it does not tell you how durable the pricing is. It tells you capacity is being bought; it does not tell you what utilization will look like in a year. The chain establishes the signal. The durability is still a question the market will answer one quarter at a time, and an honest file keeps that distinction drawn.
Two bets, one ledger: separate them
The most common error in reading this kind of report is to treat two different bets as one. Bet one: the memory supercycle is real — that is a cyclical claim about pricing, supply and demand. Bet two: domestic capacity is being substituted in — that is a structural claim about supply chains, policy and self-reliance. The two bets can move together, as they are now, but they falsify on different evidence and they deserve different position sizes. A hedged reader does not blur them. The cycle can turn while the structural trend keeps going; the structural trend can stall while the cycle still glows. Asking which bet you are actually making is the first step of not confusing signal with noise, and it is the step most commentary skips entirely.
The falsification list, filed with the view
Here is the falsification list I keep on the file, so the view stays honest. I will change my reading if any of three things happens: first, if memory prices stop rising and begin to roll over for two consecutive quarters; second, if reported utilization rates fall while new capacity keeps being added; third, if the equipment-spend forecast for the year is revised down materially. Until one of those lands, the direction of the data is consistent, and I treat the magnitude as unverified — the profit numbers are real, but how much of them is pricing windfall versus operating efficiency is a question the income statement alone cannot separate. That is the honest file: the signal is triangulated, the size of the bet is hedged, and the view stays open to the three pages that would close it.
The equipment-spend line, read with care
The third line of the chain — the forecast that global semiconductor equipment spending this year will hit a record — deserves a careful read, because it is a forecast and not a fact. Forecasts are made by institutions with models, and models have assumptions. The value of the line is not its precision; it is its direction, and its agreement with the two hard numbers already in the file. A record profit, a record revenue, and a record equipment-spend forecast, from three different parts of the chain, pointing the same way, is the pattern an analyst trusts. The magnitude is unverified by definition — it is a forecast. The direction is triangulated, and the direction is the part to hold.
The one thing I would not do
Let me close the analyst’s file with the one thing I would not do, because discipline is as much about restraint as about reading. I would not size a position as if the supercycle and the domestic-substitution trend were one bet with one risk. They are two bets with two falsification conditions, and a hedged book treats them that way. I would not treat the forecast as a fact. And I would not let a single record quarter convert a leaning-confirmed signal into a settled one — the file stays open until the falsification list is either triggered or disproven. Signal versus noise is a discipline, and the discipline is what survives the next quarter regardless of which way it goes.
No single source holds — but two sources, from different parts of the chain, agreeing on the same signal, is how a conclusion earns its way into a file. This one has earned it, with one drawer left open.