Is HBM Now Micron’s Lowest-Margin AI Product?

By the Numbers

• Revenue: $54.23B, a record, ahead of LSEG consensus of ~$51B, up 379% YoY and 31% QoQ
• Non-GAAP EPS: $33.42, ahead of LSEG consensus of ~$31.60; GAAP diluted EPS $32.87
• Non-GAAP gross margin: 87.0%, up 210 bps QoQ, versus 45.7% a year ago
• Non-GAAP operating margin: 82.3%
• DRAM: $39.8B (73% of revenue), up 343% YoY; bits up mid-single digits and pricing up high teens QoQ
• NAND: $14.1B (26% of revenue), up 526% YoY; bits up about 10% and pricing up about 30% QoQ; data center SSD near $10B, more than 10x YoY
• Business units: Cloud Memory $16.3B (up 18% QoQ, 83% GM); Core Data Center $18.0B (up 56% QoQ, 90% GM); Mobile and Client $13.1B (up 14% QoQ, 90% GM); Automotive and Embedded $6.8B (up 47% QoQ, 84% GM)
• FY2026: revenue $133.19B, up 256%; non-GAAP EPS $75.52
• FQ1 FY27 guide: revenue $61.5B ± $1.5B; non-GAAP GM about 86.25%; non-GAAP EPS $38.15 ± $1.00; opex about $2.06B
• Capital: FQ4 net capex $10.77B; adjusted FCF $33.20B; FQ1 capex about $11.5B, first-half FY27 about $25B, second half higher
• Balance sheet: cash and investments $73.48B; net cash $68.3B; $0.15 quarterly dividend; increased capital return from 9 December 2026
• Strategic Customer Agreements: 26 signed, an estimated 35%+ of revenue through 2030; RPO about $150B; $32B in customer financial commitments, $12.7B in cash deposits on the balance sheet; more than 75% of FY27 output committed

The News

Micron reported record fiscal fourth-quarter 2026 results on 30 September, with revenue, gross margin, and EPS each landing above the high end of its own guidance. Revenue and non-GAAP EPS each cleared consensus by roughly 6%, and the company guided fiscal Q1 2027 to another revenue record. The more consequential signal sits in the contract book, where 26 multi-year take-or-pay agreements now backstop an estimated third or more of revenue through 2030. The full release is available here.

Analyst Take

Six straight revenue records and a guide that adds roughly $7 billion sequentially. The print suggests memory has stopped behaving like the commodity that defined Micron for nearly five decades. The detail we keep returning to sits in the business-unit table. The Cloud Memory unit, home to HBM and the largest hyperscale relationships, held at 83% gross margin, seven points below both Core Data Center and Mobile and Client. Management's own framing explains why: it described 2027 HBM pricing as narrowing the gross margin gap with conventional DRAM, which confirms HBM currently earns less. The AI flagship is the margin laggard. The bear case writes itself. If HBM cannot out-earn conventional DRAM in the tightest market on record, its premium may have been overstated, and every HBM wafer start quietly taxes supply for everyone else. Management's counter is that the vast majority of calendar 2027 HBM supply is already contracted at significantly higher pricing. We find that credible. Until it shows up in the margin line, though, the product carrying the AI narrative is the cheaper one to buy.

What the Numbers Mean

Pricing, not volume, did the work. DRAM bit shipments grew modestly while prices climbed. Management characterized demand as stronger than at the June call and said it cannot yet see when supply and demand return to balance. That is the language of a seller's market, and the contract structure appears to confirm it.

The fiscal Q1 margin step-down deserves a careful read. CFO Mark Murphy pointed to roughly $1 billion of higher Q1 cost, driven mainly by incentive compensation capitalized into inventory in Q4, with startup and other costs adding to it, and he framed Q1 as the margin floor for the year. That explanation holds. The comment we would underline is his separate indication that price increases should slow over the balance of fiscal 2027. Price levels are not the variable to watch. The second derivative is.

Management also acknowledged that, on some platforms, server customers are growing memory content per box more slowly than planned, optimizing configurations to ship more units amid tight supply. Micron frames this as unit-driven. Our operator read is blunter: memory has become the gating line item on the server bill of materials, and buyers are rationing it. Anyone who has sat in a procurement director seat knows rationing rarely survives the first sign of loosening supply.

One demand vector sits entirely outside the current numbers. Management sized memory content in Level 4 and higher autonomous vehicles at more than 200GB, over an order of magnitude above today's Level 2+ and Level 3 systems, and expects humanoid robots to carry comparable requirements. Automotive and Embedded was already the second-fastest-growing unit sequentially, but that growth came from pricing and volume across today's automotive, industrial, and consumer demand. If physical AI scales toward the end of the decade as management suggests, the next demand wave may come from a segment that has historically operated at far thinner memory margins.

The Strategic Customer Agreements change the shape of the business. Three quarters of expected SCA revenue carries a defined pricing framework, mostly in bands with floors and ceilings, and Murphy indicated that floor-price margins would still sit meaningfully above any prior cycle peak. The ceiling is the underappreciated half. Micron appears to be trading some upside in the tightest market in its history for visibility it has never had. Memory has become beachfront property in the AI buildout, and the long-term leases are being signed now. The risk is the reverse: if demand softens, Micron is locked into delivering at floor prices that may sit above spot, compressing the premium that justified the capacity commitment.

Capex follows the same logic. Management plans to raise fiscal 2027 capex above prior plans, with the majority of the increase going to construction, largely to accelerate cleanroom availability in late calendar 2028 and beyond. It expects construction spending to grow meaningfully faster than equipment spending, alongside some pull-forward of tools to raise output from existing cleanrooms. For the equipment supply chain, the implication appears to be timing rather than volume. Shells come first and tools follow, which may push a larger share of wafer fab equipment orders into late 2027 and 2028. When the sell side opens the Q&A by asking how much cash is too much, the cycle has changed character.

Market Analysis

The closest overlapping print is NVIDIA's July quarter, and the comparison is striking. NVIDIA reported a 75% gross margin on $89.0 billion of data center revenue (NVIDIA, Q2 FY2027), roughly twelve points below Micron's fiscal Q4. The supplier of the scarce input now out-earns the platform it feeds, at least on gross margin. NVIDIA's guide for its October quarter implies accelerator demand running through Micron's fiscal Q1, which helps explain the confidence behind the sequential guide.

The relationship between the two is also changing shape. Micron's work with NVIDIA on a custom HBM4E implementation moves HBM toward a co-designed, foundry-style model, where the memory stack is tuned to one customer's platform rather than qualified as an interchangeable part. That may raise switching costs for buyers and suppliers alike. In a three-supplier market, where dual-sourcing and qualification have been among the few levers customers held over pricing, that shift deserves attention.

The Korean prints point the same direction. SK hynix posted a record quarter, began mass HBM4 shipments, disclosed long-term supply agreements with more than ten customers, and raised its capex plan. Samsung's memory business set records for revenue and operating profit, while its own mobile division swung to a loss on rising component costs. That second detail matters. Memory scarcity is transferring margin from device makers to memory makers, inside a single company. Micron and SK hynix are now converging on multi-year contracts and cleanroom-weighted capex, behaving less like commodity producers and more like utilities negotiating long-dated offtake.

A timing caveat applies. Both Korean prints cover April through June, while Micron's quarter ran through early September. Exact-period comparisons arrive with Samsung's preliminary Q3 figures, typically in early October, and SK hynix's Q3 results in late October. Those prints should help indicate whether the slowing price trajectory Murphy described is industry-wide or Micron-specific.

The demand mix needs a more careful read than the unit names suggest. Micron's Core Data Center unit, which grew sequentially at roughly three times the rate of Cloud Memory, combines server memory sold to OEM customers with data center SSDs sold to every data center customer, hyperscalers included. With data center SSD revenue near $10 billion, much of that growth appears to reflect NAND pricing and AI storage demand rather than enterprise servers alone. If enterprise inference is broadening beyond pilot deployments (an early signal we hear in channel conversations) memory demand is no longer a single-customer-class story.

Looking Ahead

The key trend we'll be monitoring is the gap between contracted floors and spot momentum. Micron has now told the market two things that sit in mild tension: supply stays tighter through 2028, and price increases slow from here. Both can be true. The SCA floors are designed to make the second statement survivable, and $32 billion in customer commitments suggests buyers believe the first.

The calendar offers several checkpoints before Micron reports again in December. Samsung's preliminary Q3 figures and SK hynix's late-October print should indicate whether the moderation is industry-wide. NVIDIA's late-November results should show whether accelerator demand keeps absorbing HBM at 2027 pricing. The enterprise channel, where we see early signs of broadening demand, may reveal whether the next leg of memory consumption comes from beyond the hyperscalers. Micron's increased capital return beginning 9 December adds one more variable: how management balances cleanroom ambition against shareholder distributions.