JPMorgan: HBM Demand CAGR of 63% from 2026-2028, Supply Remains Tight
nashnova research
J.P. Morgan's latest report argues that even if HBM specs are downgraded, bit demand will still grow at a 63% CAGR from 2026 to 2028, with the supply gap holding at -16%. This means the pricing story for HBM is shifting from cyclical to structural shortage — and upward price momentum is unlikely to reverse soon.
Specs got downgraded — why is supply still this tight?
Markets feared that HBM (High Bandwidth Memory — high-speed memory designed specifically for AI chips) spec downgrades would crush demand. J.P. Morgan's math says otherwise: cumulative three-year bit demand is essentially unchanged at 163 billion Gb.
This means → the downgrade reshuffles the product mix, not the total. Demand shifted from high-stack to economy-tier products; the pie didn't shrink.
The supply gap narrowed slightly from -20% to -16%, but cumulative shortage duration actually rose. In plain terms = the gap looks a touch smaller, but it lasts longer — manufacturers aren't getting real breathing room.
How much DRAM capacity will HBM consume?
58% of new DRAM capacity added between 2025 and 2028 will go to HBM manufacturing. HBM's share of total DRAM capacity rises from 19% to 31%.
This means → HBM is no longer a niche offshoot of DRAM — it's becoming the backbone, locking up nearly a third of all capacity.
In market-size terms, HBM is projected to reach $160–282 billion in 2027–2028, accounting for 18–24% of total DRAM revenue across the three major memory makers.
Downstream buyers: who will overtake Nvidia as the top customer?
In 2026 Nvidia remains the dominant buyer at 58% of total demand. But ASIC (custom AI chips designed in-house by Google, Amazon, and others) shipment growth hits 102% year-on-year, far outpacing Nvidia's 15%.
By 2027 ASICs claim 48% of demand; Nvidia drops to 43%. Custom chips will surpass Nvidia as HBM's single largest end market.
This reflects a deeper shift: AI compute procurement is migrating from "buy Nvidia's whole system" to "design your own chip, pick your own memory." HBM vendors now negotiate with a more fragmented customer base.
How long can prices keep rising?
HBM average selling prices rise 54% year-on-year in 2027, then another 25% in 2028, reaching $3.80 per Gb.
Operating margins hold at 60–70% — below long-term-agreement server DRAM, but a significant improvement from prior years.
In plain terms = HBM is already one of the most profitable product categories in the memory industry, and because supply stays tight, non-contract pricing still has room to climb through 2028.
Among the big three, who's in the lead?
SK Hynix leads today, but Samsung and Micron are closing in — their combined HBM revenue share reaches 59% by 2027.
The technology roadmap is splitting into multiple SKU tiers: 8Hi (eight-layer stack) extends its life as the economy option, 12Hi targets high performance, TCB (thermo-compression bonding — a packaging process that presses multiple chip layers together) is the current mainstream method, and 16Hi has been pushed back to 2029 at the earliest.
This means → no single product dominates in the near term. Customized HBM becomes the next competitive battleground — Nvidia's custom HBM product is expected by late 2028.
What's the investment takeaway?
J.P. Morgan is bullish on the memory sector. The three supporting pillars: tight supply-demand, value uplift from customization, and price resilience.
On shareholder-return potential, the report favors SK Hynix first; Samsung's investment case improves further if it steps up dividends and buybacks.
Key checkpoints ahead: whether the 2028 supply gap narrows on schedule, and whether ASIC customer demand materializes — custom chips from AI labs like OpenAI and Anthropic are flagged as an unpriced long-term upside variable.
市场有风险，内容仅供研究参考，不构成投资建议。