Bernstein: Memory Revenue Surges 451% YoY, AI Demand Underpins Semiconductor Boom
nashnova research
Global chip sales rose 131.4% year-on-year in July, with memory alone surging 451.7%. Bernstein argues the AI-driven memory super-cycle is now reshaping how profits flow across the entire semiconductor industry.
What do the July sales numbers actually tell us?
Global semiconductor sales fell 9.5% month-on-month in July, slightly worse than the historical seasonal average of 8.5%, but still up 131.4% year-on-year. This means → the sequential dip is normal seasonality; the real story is the triple-digit annual growth rate.
Memory chips rose 451.7% YoY. Strip memory out, and the rest of global semis grew roughly 35%. In plain terms = one segment — memory — delivered the vast majority of the industry's incremental revenue.
Memory's July sequential decline was 16.3%, versus a historical average of 26.1%. This reflects demand resilience well above consensus — the off-season was milder than expected.
How large is the memory price surge?
DRAM revenue grew 427.8% YoY; bit shipments rose 47.5%. NAND revenue also grew 427.8% YoY.
DRAM average selling price per bit jumped 257.9% YoY; NAND ASP per bit jumped 344.1%. This means → shipments rose less than half, but prices tripled to quadrupled — this cycle is "volume and price together," with price elasticity far exceeding volume growth.
Year-to-date global semiconductor sales reached roughly $861 billion, up from $408 billion a year earlier. Memory contributed about $355 billion in incremental sales; price and mix shifts alone accounted for roughly $306 billion — about 68% of the industry's total revenue gain. In plain terms = nearly seven-tenths of the industry's extra earnings came from memory price increases.
Why can AI keep pulling memory demand higher?
OpenAI's GPT-6 Astra scored 98% on FrontierMath Level 4 and 99.9% on ARC-AGI-3. Nvidia CEO Jensen Huang declared "AGI has arrived"; training used over 100,000 Nvidia GPUs, with another 400,000 coming online.
OpenAI disclosed it has reached an "automated research intern" stage — for every human work-day, the team now uses roughly 3.1 agent work-days. This means → "AI developing AI" is becoming its own demand curve for compute, independent of external commercial applications.
At the hardware level: training stores weights, activations, gradients, and optimizer states → consuming HBM (high-bandwidth memory — ultra-fast memory sitting next to the GPU). Long-context inference expands the KV cache → consuming server DRAM. Automated research increases experiment checkpoints → consuming enterprise SSDs. In plain terms = the smarter AI gets, the more types of memory it devours — and the bigger its appetite.
How big is the memory super-cycle?
WSTS projects memory revenue will rise from $230 billion in 2025 to $804 billion in 2026 — up 249.5% YoY — then reach $1.06 trillion in 2027.
Memory's share of global semiconductor sales is forecast to climb from 28.9% to 53.2%, then to 55.5%. This means → by 2027 memory will no longer be a segment of the chip industry; it will be the outright majority.
TrendForce expects server DRAM contract prices to rise roughly 270% cumulatively in 2026, with enterprise SSD prices up about 235%. In 2027, HBM contract prices may still climb 70%–140%. DRAM and NAND combined are projected to account for 68% of major cloud providers' capex in 2027, up from 47% in 2026.
What are downstream customers signaling?
AMD reiterated at the Citi Tech Conference that the AI data-center accelerated-computing market has expanded to $2 trillion through 2030.
Meta and two other AI labs have given procurement forecasts above initial expectations set when their strategic partnerships were formed. AMD's server CPU business is expected to grow over 80% YoY in the second half of this year and over 70% next year.
This reflects that large downstream buyers are not just maintaining AI spending — they are actively raising procurement targets. That is the most direct demand-side validation of the memory super-cycle.
What are Bernstein's price targets — and where is the risk?
Bernstein maintains "outperform" ratings on Samsung Electronics, SK Hynix, Micron, and SanDisk, with targets of ₩440,000, ₩3.3 million, $1,300, and $3,000 respectively. Nvidia's target is $400 — among the most bullish on Wall Street.
At those targets, the implied upside is roughly 77.2% for Nvidia, 77.8% for SK Hynix, 72.6% for SanDisk, and 62.7% for Samsung.
Bernstein also flags a risk: memory price increases raise bill-of-materials costs for GPU and server makers, so profits will not expand evenly. This means → the key validation metrics going forward are memory makers' realized ASPs, shipment volumes, and free cash flow — and whether downstream customers can sustain AI-deployment returns as hardware and financing costs rise. That will determine whether the super-cycle holds.
市场有风险，内容仅供研究参考，不构成投资建议。