For two years the question under the AI trade was whether the demand was real — whether anyone outside the model labs would pay for the compute, or whether the whole structure was capex chasing a revenue line that did not yet exist. This week, with the Asian memory names falling double digits in a single session and the complex de-levered at once, the question returned in its hardest form: is this the moment the demand assumption breaks?

This week answered it, from two directions at once.

From the bottom up, Micron reported a gross margin of 84.9% on a product that has spent its entire history as a low margin commodity — a number a commodity cannot produce unless it has stopped behaving like one, because supply can no longer expand to meet demand. Scarcity, quantified from the inside of a single company.

From the top down, Exponential View published the most rigorous independent decomposition of AI demand yet attempted. Where most analysis documents the supply side — the chips, the data centres, the spend — their work does the harder thing and takes the demand apart, firm by firm, and finds it real: an annualised run rate near $175bn, growing roughly 3 times faster than any prior computing wave, built on external customers paying real money. The same finding from the outside — demand, measured across the whole ecosystem.

Two independent methods running in opposite directions, and the same verdict: the demand is real, and it is accelerating. That would seem to settle the matter — except the more careful of the two does not stop there. Its read on the capex is that the buildout is, for the moment, earning its keep, and the hedge carries the weight: the verdict holds only while the buildout continues, and the report's own pages show that buildout running into order backlogs supply cannot fill and a power demand the grid was never built to carry. The work that most rigorously proves the demand is real also documents, in its own footnotes, that what the demand is buying may not be physically deliverable at the pace the money assumes.

That is the question this edition takes up — the one their work raises and leaves on the table. Not whether the demand justifies the buildout, which is now well evidenced from both ends, but whether the physical layer can deliver it.

It is the right question, because demand is not what will stop the AI trade. If anything stops it, it is the physical layer the trade runs on — the gases a fab cannot operate without, the packaging step that turns silicon and memory into a working accelerator, the electricity that feeds the entire chain. And the physical layer has a property the demand does not: it does not bind evenly, but by geography, by sourcing concentration, and by which fab sits on which grid drawing which gas through which strait. The tape this week has sensed this much — it sold the names closest to the physical layer hardest — but it could not, in a session, do the sorting: which constraints are real and which are already priced, which ease as the conflict de-escalates and which are indifferent to it, which threaten the whole memory complex and which threaten only one side of it.

That sorting is the work of what follows. It begins where the demand proof began — with Micron's margin, and the careful distinction between what it proves and what it does not — and from there walks the physical layer one constraint at a time: the input register, and what the strait's reopening actually reaches; the packaging ceiling almost no one is pricing; the single input no inventory can cover; and, at the end, the asymmetry that means the names the market sold as one trade are, underneath, not the same trade at all.

On 24 June, Micron printed the cleanest demand reading this cycle has produced. Revenue of $41.5bn — up 346% YoY and 74% sequentially, the largest sequential dollar gain in the company's history — against a consensus near $35.8bn; adjusted EPS of $25.11 where the street had looked for roughly $20; and a gross margin of 84.9%, a company record, guided higher still to ~86% for the current quarter.

That margin is the part that matters, and the most likely to be misread.

Memory has always been a commodity business: supply expands to meet demand, makers undercut one another on price, and margin gets competed away — which is why memory gross margins have historically sat in a 30-45% band and turned negative in the worst troughs. A year ago Micron's gross margin was ~38%; two quarters ago, 74.9%; now 84.9%, guided to 86%. A commodity does not earn a near-software margin; 84.9% guided higher is a contradiction in terms — unless the thing has stopped behaving like a commodity because supply can no longer expand to meet demand. Pricing power that severe exists only where the buyer has no alternative and the seller cannot make more, and the margin is the price of scarcity, quantified.

The composition rules out an accounting quirk. DRAM revenue of $31.3bn came on pricing up in the low 60s% against bit growth in the low single digits; NAND of $9.9bn on pricing up in the mid 80s% against mid single digit bits. The revenue is almost entirely price and almost none of it volume — the precise fingerprint of a supply constrained market. And it is contracted, not spot: ~$100bn of minimum revenue now locked across sixteen customer agreements, HBM sold out, HBM4 already ramping for NVIDIA's Vera Rubin platform. Management concedes it can fill only 50-66% of the demand in front of it. This is the cyclical-to-structural shift written into the order book, not a quarter's spot spike.

Every word of that runs in favour of my thesis, which is exactly why the next step has to be the careful one. It is tempting to read 84.9% as the input crisis quantified — the helium and the tungsten and the bromine surfacing at last in a single number. It is not, and resisting that read is the whole credibility of what follows, a record margin proves scarcity-pricing-power however, it does not tell you which scarcity. And this one is overdetermined: AI pull, supply discipline among the three makers left standing, the wafer intensity of HBM, and fab capacity lead times that run to fiscal 2028 all produce the same number — and all of them predate the conflict and run independent of any strait.

Three facts point the cause away from inputs. Micron's capex is rising hard, to ~$27bn, much of it construction to pull cleanroom capacity forward — the spending profile of a company expanding, not one starved of feedstock. The constraint management names is its own build clock: new fabs not delivering meaningful output until FY2028, and Micron excluded geopolitical effects from its guidance outright. Underneath all three sits a mechanical tell: input disruption is, first-order, a cost event — a rationed feedstock lifts the cost of goods and compresses the margin, a margin moving from 84.9% to a guided 86% is the opposite signature. This is a pricing event, not a cost event; if severed inputs were already biting Micron's own line, 84.9% is not the number you would be looking at.

So be exact about what the print settles and what it leaves open, it settles that the demand is real, that it is accelerating, and that supply cannot currently meet it — scarcity, quantified, and demand led. It does not settle that the physical input layer is the binding constraint. That is a separate claim, resting on separate evidence, and it is the one this edition exists to make.

What the print does is sharpen the question rather than answer it. Micron has told us, in the only language a balance sheet has, that the buyer has no alternative and the seller cannot make more, and the seller's reason is a fab capacity clock that resolves, in theory, by 2028. The rest of this edition asks whether "cannot make more" runs deeper than that clock — whether, beneath the capacity story every desk can model, there is an input story almost no one is pricing.

If the demand is real, the question becomes whether the physical layer can deliver it — and the instinct, faced with that question, is to reach for a list: Helium, tungsten, bromine, neon and sulphuric acid, the inputs a modern fab cannot run without, each with a number attached and a red bar beside it. I had published that list before, the trouble is that a list is not an analysis. It tells you how many things are disrupted but stops short of telling you which disruptions matter, and it flattens inputs with entirely different mechanisms into a single descending ranking that implies they are the same kind of problem. They are not.

The discipline that separates them is the one event the entire market is currently pricing as the all clear, de-escalation. The Strait of Hormuz is reopening to transit, Brent has fallen back below $80, and the conflict premium that built through the spring is unwinding across every commodity that moves by sea. The reading on most desks is that the supply scare is therefore over, it is not — and the reason is that the reopening discriminates. It reverses the premium and the transit cost; it does not repair physical damage, it does not reach supply that never moved through the Strait in the first place, and it does nothing to a bottleneck that was never a conflict story at all. Sort the inputs by how each one behaves against the good news, and the register resolves into three tracks, the sorting is the analysis.

These are the inputs the war actually touched, and even here the reopening helps only one of them.

Helium is the anchor, and the model for the whole exercise. Roughly a third of world supply runs through Qatar's Ras Laffan complex, one of only two facilities on earth producing semiconductor-grade helium at scale, and it has been offline since 28 February with two LNG trains destroyed, the repair is turbine-gated and runs 3 to 5 years; no ceasefire shortens it, which is why helium is reopening-proof out to roughly 2029. Korea sources about 64% of its helium from Qatar, spot has run +40–100%, and something like 200 cryogenic containers are stranded in the disrupted logistics. The point to hold is mechanical: this is physical destruction, not a blocked sea lane, so the de-escalation does precisely nothing for it. Helium is the durable floor under the entire register — and notably, TSMC is largely insulated, carrying inventory, heavy recycling, and a new Air Liquide plant in Taiwan, which is the first hint of an asymmetry we will return to.

Bromine is the memory specific one, and the one the consensus is getting wrong in both directions. It is the feedstock for the hydrogen bromide that etches every DRAM and NAND die; Israel and Jordan together hold roughly two-thirds of global supply, and Korea draws about 97% of its bromine from Israel. The Neot Hovav industrial zone was struck on 29 March — the Adama plant specifically — and the easy version of the story is that the line has been cut, it hasn't. ICL's Dead Sea bromine and HBr output continues; what the strike produced is a risk premium of roughly 2× and a live exposure to the missile corridor and to climbing Israeli port war-risk, not a severed supply. Armed and proximate is a different risk from severed, and it resolves differently: it eases if the exchange stops and detonates if a single facility takes a hit, calling it "cut" overstates today and understates the tail.

Gulf sulphur is the one Track A input that genuinely eases, because its constraint was always transit rather than damage. FOB pricing ran to roughly $695–700/t, some $200 above pre-conflict, on the closure of the seaborne route; as the Strait reopens, that premium comes off. For our purposes it is also the least important of the three — electronic grade acid is carved out of the disruption and fabs take priority allocation, so the heavy impact routes to mining and fertiliser, not to wafers. It earns its place here mainly as the control case: the input that proves the reopening does work, on the one input whose problem the reopening can actually solve.

This is the track the market keeps folding into the Hormuz story, and it is the most important error to avoid, because none of it runs through the Strait and none of it eases when the Strait reopens. It is a separate regime — the US–China tech war, currently in a fragile truce — sitting on an intact and expanding architecture of export controls.

Tungsten is the live one. China holds north of 80% of tungsten powder, and the semiconductor specific product is tungsten hexafluoride — WF6 — the interconnect gas deposited into contacts and NAND structures, WF6 supply contracts have locked in at +70–90% for 2026, the Japanese specialty gas houses that supply roughly a quarter of the world have warned their inventory could run dry around mid 2026, and the upstream signal is unambiguous: APT, the benchmark tungsten intermediate, has risen more than 550% since early 2025 against a long run average near $300/mtu, with Chinese exports cut to a trickle and only fifteen firms licensed to export through 2027. This is the cleanest dated trigger in the entire register — a named inventory cliff, mid 2026, with Chinese capacity unable to qualify replacement supply in under 12 to 18 months, and it falls harder on Samsung than on SK Hynix, which matters for where the pressure lands.

The rest of the basket is structural leverage of varying temperature. Gallium, where China holds about 99% of supply, and germanium have both been driven effectively to zero for Japan. Antimony is the instructive one: it spiked roughly 2,600% from a depressed base after the 2024 controls, peaked near $59,750/t in mid 2025, and has since retreated — held down by the suspension of the US specific ban agreed in November, which expires on 27 November 2026. That expiry is the tell for the whole track, the truce suspends but does not dismantle, the whitelists, the zeroing of exports to Japan, the licensing architecture all remain in place, ready to re-arm. Track B doesn't ease with the Strait; it eases, partially and conditionally, only for as long as a separate clock keeps running.

Neon belongs here too, and it is worth correcting the record, because the line you will see repeated — neon as a live, single sourced casualty of the current conflict, sometimes pinned to Iran's South Pars — is wrong on every count. Neon's chokepoint was Ukraine, which supplied around 90% of US semiconductor grade gas; the shock was 2022, not 2026. The industry has since de-risked it hard — 90% plus recycling, new Western air separation capacity, and China now a dominant producer — and EUV lithography, the leading edge, does not use neon at all. It is a latent, China-gated tail worth watching, not an acute disruption, and folding it into the conflict story is the kind of error that costs credibility when someone checks. Sulphuric acid rounds out the track and, as noted, is second order for fabs specifically.

The third track is neither severed nor China-gated. It is the risk that a single jurisdiction, or a single supplier, simply is the whole world's capacity — no damage required. ABF substrate and photoresist sit here, concentrated in Japan, with the 2019 Japan–Korea export episode as the live precedent for how fast a concentrated node can become a weapon. And the largest single item in this track is consequential enough that it gets the next movement to itself: CoWoS, the advanced packaging step where the entire AI accelerator complex now actually bottlenecks. For here, the point is only that Track C exists and that the reopening is irrelevant to it — you cannot de-escalate your way out of geographic concentration.

Stand back from the three tracks and the durability map is the payoff, run the whole register against the de-escalation and exactly one input genuinely clears — Gulf sulphur, the transit bound control case. Helium is physically destroyed and indifferent to 2029, the China basket never touched the Strait, the concentration nodes were never a conflict story and the single event the market is treating as the all clear reaches one line in the register and leaves the rest standing. That is what "multi sourced" actually means, stated precisely: not that there are many disruptions, but that they have independent causes, so that no single resolution — not the ceasefire, not the reopening, not the truce — clears more than a fraction of them. The good news: the de-escalation everyone is pricing is real; it just isn't the same event as the disruption clearing.

The AI trade has been argued, this whole cycle, at the wrong layer. It was a demand story when the question was whether anyone would pay for the compute, and NVIDIA became the proxy for the answer. It changed to a memory story this quarter, when Micron's margin proved the buyer had no alternative, becoming an exotic input story — helium, tungsten, the gases a fab cannot run without. All three are real and none of them is a binding constraint. The thing that actually caps how many AI accelerators can physically exist in 2026 is a back end assembly step that, until about two years ago, almost no one outside the industry could name — and it sits one layer downstream of everything the market has been pricing.

The step is CoWoS — Chip-on-Wafer-on-Substrate — and it is worth being concrete about what it does. A finished logic die, even a flawless 3nm GPU, is not an AI chip, it’s an inert tile until it is bonded onto a silicon interposer alongside its stacks of high bandwidth memory, wired to them at a density no circuit board can approach, and turned into a single package — CoWoS is that bonding step. Without it the wafer and the memory are two unrelated objects; with it they become a Blackwell or a Rubin, which means the memory supercycle from the opening of this edition and the packaging bottleneck here are not two problems — they are one problem seen from two ends. Every additional HBM stack an accelerator carries is more interposer area consumed, more CoWoS per unit. The memory boom does not relieve the packaging constraint; it tightens it. And the newest accelerators make it worse on a second axis: as reticle sizes grow — Rubin runs to several times the standard area — a single wafer yields fewer finished units, so the same packaging line produces less of what the market most wants.

Now place that step geographically, and the picture resolves into the third track from the register — concentration without disruption. No strike hit this, no export licence gates it, no ceasefire or truce is even relevant to it, however; the entire world's advanced packaging capacity for AI accelerators sits, effectively, inside one company, in one country, say it with me: TSMC, in Taiwan. Its three backend packaging fabs are sold out through 2027, with lead times of 52 to 78 weeks — meaning an order placed today is delivered, at best, a year and a quarter from now. Demand has roughly tripled in 2 years, from around 370,000 wafers in 2024 to near a million in 2026, while TSMC ramps capacity about as hard as a fab can be ramped — close to a tenfold increase since late 2023, heading toward 120,000–130,000 wafers a month by year-end — and the lines stay fully booked anyway. NVIDIA alone holds roughly 60% of it, the top 3 customers more than 85%, and NVIDIA has reportedly booked over half of the 2026–27 expansion before it has even been built. The capacity that does not yet exist is already spoken for.

The proof that this is a physical ceiling and not a financial one is Google. Alphabet cut its 2026 production target for its own flagship AI chip, the TPU, from roughly 4 million units to 3 — a quarter of the programme, gone — and it did so not because demand for the chips softened, and not because it lacked the money to build them. It cut because, as a relative latecomer to volume packaging, it sat at the back of the queue behind Apple and NVIDIA and could not secure the CoWoS slots. One of the most cash rich companies on earth absorbed a 25% reduction to its strategic silicon because capital could not buy interposer area that was already allocated. That is the Micron logic from the start of this edition, generalised to the entire complex: the buyer has the money and still cannot get the product, when money cannot clear a constraint, the constraint is not about money. The smaller custom-chip designers fare worse still — without the scale to lock multi-year allocations, they face multi-quarter delays, and some of that demand simply consolidates back to the one supplier who already holds the capacity.

The reason this matters for the question this edition is built around — can the AI trade deliver — is that it does not resolve on any of the clocks already discussed. It is not a conflict input, so the de-escalation does nothing for it, it’s also not China-gated, so the truce does nothing for it, it is pure geographic concentration, and the only thing that lifts it is more capacity, which is precisely what is sold out. TSMC's expansion is real and aggressive, and it deserves to be weighed honestly: if the 120,000–130,000 wafer target lands on schedule and reticle growth plateaus, the ceiling does begin to lift, but the offsetting facts are heavier: the expansion is pre-sold, US packaging capacity in Arizona is realistically a 2028 story, not a 2027 one — the Phoenix fab's earlier node qualification already ran behind, and qualifying greenfield lines to CoWoS tolerances is its own multi-year problem, and panel-level packaging, the structural fix that escapes the wafer's geometry, ramps in 2028–29. So the honest horizon is narrow and specific: the ceiling is binding through 2027, the relief arrives in 2028 at the earliest, and across that whole window the chips themselves keep getting larger, pushing packaging consumed per accelerator up while supply crawls.

Which returns the whole argument to a single, slightly uncomfortable place. The demand is real, the exotic inputs are real, and as the next movement will show, they point hardest at Korea rather than at the leading edge. But underneath all of it, the rate at which the AI buildout can physically happen is governed not by appetite, not by capital, and not even by the gases that nearly everyone is now learning to worry about. It is governed by how fast one company in Hsinchu can bond memory to logic.

Every constraint in this edition so far shares a hidden property, and it is worth naming because the final one breaks it. Helium can be held, but only barely and only briefly — it is the second smallest molecule in the universe and permeates straight through most containment at the molecular level, so commercial scale, long term storage is economically impractical, and the one strategic stockpile that exists, the US Federal Helium Reserve at Amarillo, holds crude grade gas in deliberate drawdown that never reaches a Korean or Japanese fab. What buffers TSMC is not a warehouse of helium but inventory in working volumes plus heavy recycling — a margin measured in weeks of operation, not a hedge against a third of world supply vanishing. WF6 can be stored, which is the only reason the mid-2026 inventory cliff is mid-2026 and not now, bromine, gallium, the photoresists: all of them can be bought ahead and held against disruption, which is what every fab procurement desk on earth has spent the past year doing, even CoWoS, the packaging ceiling, is in the end a queue — the capacity is finite, but the wafers exist and the slots can be booked years out, the entire register, the whole physical layer, is in principle hedgeable with a warehouse and a forward contract. There is exactly one input the fabs run on that cannot be put in a warehouse, and it is the one this movement is about: electricity.

Electricity is different in kind, not degree. Grid scale storage is still marginal, so the power a fab draws is consumed in the same instant it is generated, continuously, at a scale few industrial users on earth approach. And the exposure is not just to outages: a modern fab running EUV lithography is so sensitive that a momentary voltage sag — not a blackout, a dip lasting a fraction of a second — can ruin wafers mid process and force a line restart, so where a fab's exposure to its gases is buffered by inventory and measured in months, its exposure to its grid is absolute and measured in milliseconds. You cannot stockpile your way out of it, which makes the question of which grid the world's most important fabs sit on rather more than an infrastructure footnote.

They sit on Taiwan, and Taiwan's grid is among the most import dependent on earth: the island produces about 4% of its own energy and imports the other 96%. Roughly half its electricity comes from LNG that arrives by ship. Its one source of dispatchable, fuel storable baseload — nuclear — was phased out in 2025, the last reactor going offline just as this crisis began; a restart is now being debated, but it is a political fight, not a near term supply. The system's reserve margin runs chronically below the government's own 15% target, and it has produced 3 sizeable outages in 7 years on ordinary days. This is a grid running hot with thin margins before anything is added to it.

Now the storage figures, because this is where thin becomes acute. Taiwan holds about 146 days of oil in reserve and roughly 39 days of coal, its LNG reserve is about 11 days. The gap is not a planning failure; it is physics — LNG resists storage, boiling off if held too long, so an island cannot simply warehouse a season of it the way it warehouses crude, and that 11 day figure is the comfortable one, it describes winter consumption, in summer, when air conditioning load lifts demand, the same reserve covers only about 7 days. So the buffer is at its very thinnest in exactly the months the load is at its highest — and July demand on this grid runs around 40% above February's, layer the calendar onto the geography and the picture is stark: a roughly 1 week margin of the critical fuel, in the hottest part of the year, on a grid with no nuclear and no spare capacity, where a third of that fuel still comes from the Middle East through the same waters as everything else in this edition.

This is the input the entire AI buildout terminates on, and the load is not holding steady — it is exploding. TSMC alone consumed about 8% of Taiwan's electricity in 2023; on S&P's modelling that climbs toward 24% by 2030 as the nodes get smaller and more power hungry. One company, heading toward a quarter of a national grid. The packaging ceiling from the previous movement and this grid are not two separate vulnerabilities; they are the same one, stacked. The glue that bonds every AI accelerator is in Hsinchu, and Hsinchu's power arrives just in time, on ships, into a system that keeps a week of the fuel that makes half of it.

The honest accounting has to cut both ways, and here it genuinely helps the bullish case in the near term. This is one of the two places in the whole register — Gulf sulphur is the other — where the de-escalation does real work: the reopening restores the Qatari route, relieves the LNG transit premium, and takes the acute edge off. Taiwan's economic ministry has secured supply through September and ruled out a shortage, and on current trajectory it will get through this summer. What the episode did was not break the grid, but X-rayed it, and that X-ray shows the structural facts the reopening does not touch: the buffer is still a week in summer, the imports are still 96%, the nuclear is still off, the margin is still thin, and the load is still climbing toward a quarter of the grid for a single firm. The de-escalation buys time, but does not change these facts.

There is one tail the reopening reaches even less, and it is where the physical reading hands off to the strategic one, so it is worth stating plainly and then leaving for another edition to develop. The same week of gas vulnerability that the Iran shock merely grazed is precisely the pressure point a blockade would aim at first — gas is the input that energy security analysts expect to collapse soonest under a quarantine, ahead of oil and coal, for exactly the storage reasons above. Beijing has already floated, during this very crisis, the offer of "energy security" to Taiwan in exchange for its sovereignty. That is not a supply chain question any more, and this edition will not pretend to resolve it, but it is the reason the grid is not merely the thinnest link in the physical chain — it is the link where the physical chain and the geopolitical one are the same object.

So this is the un-bufferable tail, and it is a different kind of risk from everything that precedes it. The gases, the metals, the packaging — each can be modelled, priced, hedged, and waited out, because each can be held. The grid cannot. The honest answer to whether the physical layer can deliver the AI trade rests, at its very end, not on a price or a quota or a lead time, but on whether the lights stay steady through August on an island that keeps a week of fuel. There is no inventory that covers that one.

Go back to the tape that opened all of this: this week, the memory names went down together — Micron off around 12%, SK Hynix off about the same in Seoul — and the market read it as flow: a deleveraging cascade, the yen carry trade unwinding, quarter end rebalancing, a hawkish Fed raising the cost of holding risk. All of that was real, and all of it explains why everything fell at once. What it doesn't explain is why it fell in that order. A pure flow event hits the highest beta names hardest and doesn't single anything out by function — yet memory sat at the very bottom of the tape, equipment beside it, and the demand proxy, NVIDIA, held up best. That dispersion was the tell, and it was mine to read, not the market's: the names were sorting by proximity to the physical layer, but even my own first read of it sorted too coarsely. It treated memory as a single object — one trade, one exposure, one direction — and on the inputs this edition has spent four movements laying out, memory is not one object at all. The exposure inside it is concentrated, and it is concentrated by geography.

Lay the live and armed inputs from the register against the two sides of the memory market and the concentration is hard to miss. Helium: Korea draws roughly 64% of its supply from Qatar, the single source sitting offline at Ras Laffan until the back end of the decade, bromine, and the hydrogen bromide that etches every memory die: Korea sources about 97% of it from Israel, a near total single country dependence running through an active missile corridor, WF6, the tungsten interconnect gas on the mid-2026 inventory clock: the Japanese specialty gas houses that supply it feed the Korean fabs, and the exposure there falls harder on Samsung than on SK Hynix. Three of the inputs most likely to bind in 2026 describe, with some precision, the sourcing profile of the Korean memory makers. They do not describe Micron's.

Micron is the leading edge memory maker least exposed to exactly these vectors. It is US headquartered, its footprint spans the US, Taiwan, Japan and Singapore rather than concentrating in one country's import channels, and — the detail from its own quarter — it excluded geopolitical effects from guidance entirely while raising capex and expanding margin, its helium is buffered the way TSMC's is, through recycling, inventory and diversified supply, not through a 64% from one source pipe, its etch chemistry is not 97% dependent on a single country at war. So when the input register is read properly, the disrupted inputs thesis is not bearish memory in general. It is bearish Korean memory specifically — and by the same mechanism, through the supply side, it is a tailwind for Micron.

That is the inversion worth sitting with, because it runs opposite to how the tape traded it. In a market that is already sold out, already printing 84.9% margins, already unable to fill more than two-thirds of demand, a harder constraint on your two largest competitors is not a threat to you. It is pricing power and allocation share handed to you, without a cent added to your own cost line. If helium or bromine or WF6 actually bites Samsung and SK Hynix output, the memory market — the tightest it has been in its history — tightens further, and the maker that isn't bound on those inputs captures the incremental pricing and the displaced demand. For Micron, an input crisis at the Korean fabs is a competitor's problem that lifts its own economics, and the market sold Micron alongside SK Hynix as if they were the same trade. They are on the demand cycle, but on the inputs they are very nearly opposite.

Now the discipline, because asymmetry is not immunity and my credibility rests on saying so plainly. Micron is not untouched. Its own fabs in Taiwan and Japan draw on some of the same gases, so its exposure is lower concentration and better buffered, not zero — the claim is relative, a question of degree and sourcing breadth, not a clean separation. The asymmetry also only pays if the inputs actually fire, and the second section was deliberate that most of them are armed or forward rather than severed: bromine is a risk premium, not a cut; the WF6 cliff is a mid-2026 date, not a present fact; helium is the one already live. And Micron shares the demand side throttle this edition spent section three on — its HBM goes into NVIDIA accelerators that are gated by CoWoS, so if packaging caps how many accelerators can be built, it eventually caps HBM demand too. Micron sits on the favourable side of the input question and the same side as everyone else on the packaging question. This is an asymmetry inside a constrained complex, not a clean long.

Which makes the test specific and near. The asymmetry expresses, or it doesn't, in the next two quarters of Korean output guidance: if Samsung or SK Hynix flag input driven production cuts — helium rationing, bromine shortfalls, a WF6 wall — while Micron's guidance stays clean, the gun was pointed where the analysis said it was, if they don't, the input register was a real risk that never fired, and memory collapses back into one trade running on one demand cycle, the way the tape is treating it now. Watching which of those happens is worth more than any view held in advance. This is research, not a recommendation, and the honest position is that the swing factor is external to every company involved: it is whether the physical layer binds before the buyers diversify around it.

And that is where the whole edition lands. The question it set out to answer — can the AI trade deliver — does not have a single answer, because the trade is not a single thing. The demand will not be what stops it; Micron proved that from the bottom up and the independent revenue work proved it from the top down. If anything stops it, it is the physical layer — the gases, the packaging, the grid — and the physical layer does not bind evenly. It binds by geography, by sourcing concentration, by which fab sits on which grid drawing which gas through which strait. The tape this week sensed that the physical layer mattered and sold proximity to it as a bloc. The work of this edition was to do the sorting the tape did crudely: to show that proximity to the physical layer is not one exposure but many, pointing in different directions, and that the names the market moved together are, underneath, not the same trade at all.