Micron Doubles Capex: The Supply Arrives in 2028, Not 2026

Micron guided fiscal-2026 capital spending to roughly $27 billion in June, nearly double the prior year, and the market read it the way a memory boom is always read: a company investing because demand is there. The demand is real. Revenue for the fiscal quarter ended May 28 came to $41.46 billion, up 346% from a year earlier, and management guided the August quarter to $50 billion at about an 86% gross margin — figures that would have looked absurd eighteen months ago.

The investor question is what the capex tells you about the cycle, and there the two halves of the headline do not line up in time or in kind. Doubling capex in 2026 does not add supply in 2026. And the revenue the market is rewarding is not coming from more memory shipped. It is coming from the price of it.

Revenue Is Price, Not Volume

The boom's ledger is useful to look at before any of the architecture. In Micron's fiscal quarter ended February 26, average DRAM selling prices rose roughly 65% sequentially while bit shipments grew only about 5%; in the prior quarter, ASPs rose about 20% with bit growth essentially flat. TrendForce's contract data shows the same shape: conventional DRAM contract pricing surged 93%–98% quarter over quarter in Q1 2026, with another 58%–63% projected for Q2.

That pricing is the entire dollar story. Industry DRAM supply growth for 2026 is forecast near 16%, below historical norms, and Micron itself describes 2026 bit demand as constrained by supply rather than by anything a customer is refusing to buy. The consumer economy supplies the most readable evidence: a DDR5 kit roughly tripled or quadrupled in price, and a 32GB DDR4 kit went from a $60–$90 range to $150–$180 in a single quarter.

The distinction matters because price-led growth is reversible by a single supply release or demand pause, and volume-led growth is not. Everything about Micron's current margin structure — 81%–86% gross margins, operating cash flow near $51 billion on a trailing basis — depends on scarcity persisting. That is the whole wager.

Why New Construction Won't Ease 2026: The HBM Wafer Tax

The reason the shortage resists a capex fix is structural, and it is worth stating plainly because it is the driver the demand narrative gets backwards. High-bandwidth memory consumes roughly three times the wafer capacity per bit of standard DRAM — a 3:1 ratio that rises toward 4:1 with the HBM4 generation — because each HBM stack packs a dozen thinned DRAM dies connected by through-silicon vias. HBM has gone from under 2% to roughly 6% of the industry's DRAM bits, and cloud-related memory has climbed from 17% to about half of Micron's own DRAM revenue.

Every HBM stack built for a GPU is wafer capacity denied to DDR5, DDR4, and phone memory. This is a zero-sum wafer economy, and it means the shortage is a product-mix tax, not simply a demand surge: even flat end-market demand would not cure it. It also caps the value of the build-out itself, because new wafers converted to HBM yield fewer net bits than the same wafers did in the DDR era.

The constraint has already migrated past the fab. Micron has committed up to $3 billion to move upstream into raw silicon, including a $500 million financing package for GlobalWafers' 300mm wafer plant in Sherman, Texas, and a ten-year supply agreement. When a memory maker is paying to secure raw wafers, that tells you where the binding bottleneck sits: not in the demand forecast, and no longer only in the cleanroom.

What the Doubling Actually Buys

The capex ladder below is compiled from Micron's fiscal-2026 reporting and guidance, and the timing column is the part of the headline that is doing the real work.

Where capital goes and when it becomes supply

Fiscal year	Capex	Where it goes / what it delivers
FY2025 (ended Aug-25)	~$14B actual	Trough aftermath; restrained spend
FY2026 (ends Aug-26)	~$27B guided	~$10B in the August quarter alone
FY2027	>$40B implied	Construction ~$10B+ higher year over year; equipment output meaningful only 2H 2027 into 2028
Through FY2035	$250B planned	~40% of DRAM made in the U.S.; Idaho wafer output mid-2027, New York fabs still later

Management's own words set the delivery date. Production from the aggressive expansion is not expected to become material until the second half of fiscal 2027 into 2028, and more than half of the fiscal-2027 increase is construction — land, shells, cleanrooms — which produces no bits at all.

None of this is a distressed decision. The company holds roughly $20 billion more cash than debt, it brought in about $22 billion of customer deposits under strategic agreements covering roughly a fifth of DRAM and a third of NAND through 2030, and it says it will return 100% of excess cash over time. Micron is not building because it must; it is building to lock in share in a market it believes stays short for years.

Two Markets, Two Economies

The single most counterintuitive detail in this cycle is where the margin sits versus where the demand story sits. HBM is sold out under multi-year contracts with hyperscalers, and management has said HBM's total addressable market crosses $100 billion in 2027. Yet on the last call, the company noted that non-HBM margins currently run higher than HBM, and that DDR5 profitability has overtaken HBM.

The reason is inherent to how the shortage allocates pricing power. HBM is contracted and repriced slowly; commodity DDR is burdened with every marginal dollar of scarcity. So the product everyone calls the AI engine is not the product generating the fastest margin expansion — the squeezed commodity market is. Any analysis that averages the two, or reads the boom through the HBM product alone, misses the actual split.

The Risk Is Everyone Else's Doubling

Micron's build is not happening in isolation, and that is the piece of the story the single-company headline obscures. SK Hynix intends to double wafer capacity over five years. Samsung has guided to roughly $73 billion of semiconductor investment in 2026 and about $648 billion in Korea over a decade, and the Korean state has wrapped its own ~$520 billion plan around four new memory fabs.

Post-2022, the memory cycle inverted its usual shape because the suppliers collectively restrained capex, technology-migrated, and let prices do the recovery. A synchronized global doubling is, by definition, the end of that restraint. The memory industry's historical record is that oversupply arrives two to three years after the moment everyone simultaneously proves confidence — which puts fiscal 2028–29 on the calendar. The one structural defense is the same wafer tax that caused the shortage: new wafers converted to HBM4 deliver fewer net bits than the dollar size of the build suggests, which is how the industry could still absorb the capacity if AI bit demand stays constrained into 2028.

The share price has already placed this bet. Micron fell from $114 a year ago to a July high above $1,255, and trades near $935 today, around a trillion-dollar market cap — roughly 12 times sales and 21 times trailing earnings, a multiple sitting on peak-scarcity dollars that are themselves the product of prices up 80%–90% in a quarter.

The key issue is not whether Micron doubles capex, or whether the shortage is real. It is what net bit supply the synchronized build delivers, and when that arrives against the AI demand curve. If the HBM wafer tax holds and AI keeps absorbing capacity, tightness can persist into 2028 and the doubling is defensive participation in a structurally short market. If the booked demand normalizes and 2028's capacity arrives as net bits, this capex cycle is the standard precursor to the next bust. Prices wrote the gains; the volume ledger will end the cycle. Watch bit growth and utilization for the turn, not the revenue headline.