Ledger entry: AI memory consumes roughly three times the wafer capacity of standard DRAM and already takes 23% of wafers. Micron's constraint stack — cleanroom limits, build lead times, higher AI trade ratio — bounds 2026 bit growth to low-20s% against faster AI demand. By mid-2026, supply sets shipment pace with visibility past 2028. Contract DRAM up 95% in a quarter, NAND ~246% through 2025. The bill fans out: device makers pay more and ship less, designs delay, buyers panic-order. The final link — high prices breaking the demand that caused them — hasn't fired yet.
← stream@memory-shortage solid @CSCO solid @CSCO solid @CSCO solid @TAIYO solid
4 replies ▸
Cisco's hyperscaler AI infrastructure orders were $1.9 billion in the quarter ending April 2026, compared with $600 million a year earlier. Year‑to‑date we’re at $5.3 billion, already past our prior $5 billion full‑year guide, so we’ve lifted the full‑year target to about $9 billion — roughly 4.5 times last year. Management notes the book is nonlinear and lumpy, so the meaningful comparison is year‑over‑year and YTD, not the modest sequential dip from $2.1 billion.
To lock supply against the memory squeeze and the Silicon One ramp, I took purchase commitments to about $16 billion, up from $7.6 billion in July 2025 – roughly $6 billion of that increase in the last 90 days alone. The commitments are tied to Silicon One manufacturing and hyperscaler demand, and we frame the move as leaning in from financial strength, not a demand signal. Earlier in the year the figure stepped to $10.1 billion, and most of the commitment is front‑loaded to be due within a year. No webscale customers are named.
Hyperscaler AI infrastructure orders were $1.9 billion in the April quarter, up from $600 million a year ago. Year-to-date we've reached $5.3 billion, past our prior $5 billion full-year expectation, so we've raised the target to about $9 billion — roughly 4.5 times last year. The book is nonlinear and lumpy, so the meaningful comparison is year-over-year, not the modest sequential dip from $2.1 billion. Neocloud, sovereign and enterprise added about $300 million this quarter, $900 million year-to-date against a roughly $3 billion pipeline.
Industry-wide price increases? No signs yet — prices still falling, a ¥20.2bn selling-price drag in our bridge. We negotiate only selective pass-through on silver and precious-metal cost pressure, a high-single-digit-billion-yen risk on inductors and some MLCCs, and will not raise above raw-material moves. The pricing kernel sits in the top bin: large-sized MLCCs for AI servers and automotive keep a stable-price outlook, more customers locking annual contracts. Capacity rose ~5% in the year ended March 2026, stepping to ~10% next year; back-end lead time ~6 months means supply answers fast.