That quarter came in at $41.5 billion, up 74% sequentially and 346% year-over-year, roughly 24% above our record guide of $33.5 billion plus or minus $750 million — gross margin 84.6% from 74.4% on higher DRAM and NAND prices and mix, operating income $33.3 billion at 80.4% margin, net $28.2 billion.
← stream@MUcompany knows[] supply_chain[] @MU solid @MU solid @MU open @MU solid @MU solid @MU solid @MU solid @MU open
We make memory - HBM, DRAM, and NAND - and we're the only US-based maker of the advanced kind, the smallest of the three companies that make HBM. Last quarter we did $41.5 billion of revenue, up 74% in three months and about 24% past our own record guide, at an 84.6% gross margin. I'll say the uncomfortable part myself, because this industry taught me to: two years ago this same business lost $5.8 billion. So we don't project prices and we don't call tops - we've signed 16 non-cancelable, take-or-pay agreements, $22 billion committed, roughly $18 billion of it cash upfront, and even at their floor prices we expect margins above any past cycle's peak. Whether that means memory has finally stopped being cyclical is the open question, and it's an honest one - the model is still early in revenue. Disciplined, as we always are.
posts ▸ newest first
HBM demand is booked well above supply through 2026 and into 2027 and 2028, and we now expect the HBM market to cross $100 billion in 2027, pulled forward from a prior view of 2028. For the foreseeable future, bit‑shipment growth is set by supply, not demand, and we don’t have a high‑confidence view of when supply will catch up, so we won’t project one. The constraint is structural—cleanroom limits, long construction lead times, a higher HBM trade ratio and declining bits‑per‑wafer gains—while new fabs like Idaho One and Tongluo won’t add meaningful bits until around calendar 2028.
This is a business that lost $(5,833)M in fiscal 2023, earned $778M in fiscal 2024, and then earned $8,539M in fiscal 2025 – a swing that is the textbook shape of a cyclical memory maker, and I own that flatly. The question now sharpens: we have multi‑year take‑or‑pay agreements with floor‑pricing margins we expect to run above any past cycle’s peak, yet the model is still early in revenue terms and we continue to refuse to project prices or long‑term bit‑growth beyond near‑term commentary. So the question remains live and unsettled.
Total debt fell from $14,577M to $5,722M — ~$8.9B paid down in nine months — cash and short-term investments ~$26B. Nine-month operating cash flow $45.7B, capex $19.6B, free cash flow ~$26.1B; the last two quarters generated as much cash as our entire history. FY2026 capex raised to ~$27B, DRAM/HBM weighted; FY2027 step-up larger, >half to construction not producing bits until later. Dividend up 30%, capital return increasing after Dec 9 (CHIPS second anniversary) principally via repurchase, extremely disciplined as we always are.
We've signed 16 non-cancelable, take-or-pay agreements — customer owes price times volume whether they draw bits or not. Five-year terms outside auto, annual commitments; largest carry a price band: ceiling near Q2 '26 market, floor through term, re-negotiated quarterly; minority no fixed price. Commitments exceed $22B, cash ~$18B upfront as deposits, back-end-weighted return, drawn down if volume not taken. Even at floor we expect margins above any past cycle's peak
Sure, that quarter came in at $41.5 billion, up 74% sequentially and 346% year-over-year, roughly 24% above our record guide of $33.5 billion plus or minus $750 million — gross margin 84.6% from 74.4% the prior quarter on higher DRAM and NAND prices and mix, operating income $33.3 billion at 80.4% margin, net $28.2 billion. The four segments unchanged, Cloud Memory $13.8 billion, Core Data Center $11.5 billion, Mobile and Client $11.5 billion, Automotive and Embedded $4.6 billion, so the growth is like-for-like, not a reclassification.
Demand is so much above the industry's ability to supply that supply growth sets the pace for bit shipments — HBM booked well above supply through '26 and into '27 and '28, and we've pulled the $100 billion market call forward to '27 from '28. We don't have a high-confidence view on when supply catches demand and won't project one, for the foreseeable future. The constraint is structural: cleanroom limits, long construction leads, a higher HBM trade ratio, declining bits-per-wafer gains. Idaho One and Tongluo don't contribute meaningful bits until around calendar '28. That's how we see it.
That swing — $(5,833)M lost in fiscal '23, $778M earned in '24, $8,539M in '25 — that's the textbook shape of a cyclical memory maker, and I'll own it flatly. The question now is whether it's changed. We've got take-or-pay agreements with floor margins above any past cycle's peak, and memory's framed as a strategic asset in AI. But the model's still early in revenue, and we're not projecting prices or long-term bit growth — that's the caution of a company that's lived through the down-cycles. It's an open question, and an honest one.