The hottest corner of the AI-hardware trade is a memory shortage — and some of the people who make the memory are quietly building the thing that ends it. Four forces are pressing on the “memory supercycle,” and together they argue it is a more ordinary cycle than the price of the shares implies.
Memory has been one of the biggest — and most crowded — winners of the AI build-out. The question for the next year is not whether the shortage is real — it is — but how long the market keeps paying as if it were permanent.
HBM (High-Bandwidth Memory) is fast memory stacked in a tower and bolted right next to an AI chip, so data can be fed to it quickly. AI accelerators are starving for exactly this, supply has been tight, and that scarcity has handed the three makers — SK Hynix, Samsung and Micron — unusually fat margins. When people say “the memory supercycle,” HBM is what they mean.
That the shortage is real is not in dispute. Nvidia’s own commitments to suppliers ballooned to $279bn last quarter, primarily for memory, and it told investors its gross margin would slip from ~74% to about 71.5% the following quarter because of memory prices. When the most powerful buyer in the industry is paying up and taking a margin hit to secure supply, the scarcity is not imagined.
The temptation is to extend that line straight upward. A commodity earning extraordinary margins does one thing reliably: it pulls in capital, capacity and substitutes until the margin is competed away. Below are the four forces already doing that to memory — the first two visible in the supply and the software, the third and fourth the more telling, because they come from inside the industry itself.
China treats dependence on foreign memory as a strategic problem, and it is spending accordingly. YMTC, its flash-memory champion, is raising roughly $4.9bn in a Shanghai listing and has signalled it intends to become the world’s largest NAND maker by the end of 2027 — from low-teens share today, an ambition that implies adding close to another Samsung’s worth of output in under two years. Its DRAM sibling, CXMT, is chasing the same path. Goldman Sachs estimates China’s advanced-chip supply will grow around 46% a year through 2035, narrowing its self-sufficiency gap to roughly a third.
The two main kinds of memory. DRAM is fast, expensive “working memory” (HBM is a stacked, premium form of DRAM). NAND (also called flash) is slower but far cheaper and holds far more — the kind of storage that keeps photos and files in a phone, laptop or USB stick. YMTC makes NAND; CXMT makes DRAM. Remember the difference — it is the whole point of Force 3.
Even setting China aside, the incumbents are finally expanding into these margins, and new entrants are drawn by them. Whether or not YMTC hits its 2027 target, the direction of travel for supply is unmistakable — up.
Scarcity does not only pull in new supply; it also rewards anyone who can need less of the scarce thing. Across AI, engineers are now doing exactly that — treating memory as something to optimise around rather than a fixed cost to accept.
Two very different jobs, with very different memory needs. Training is building the model — feeding it oceans of data while it repeatedly rewrites its own billions of parameters until it is capable. It is write-heavy, done relatively rarely, and demands the fastest, most expensive memory. Inference is using the finished model — it mostly reads a fixed set of already-learned weights to answer each prompt. Think of it as writing the textbook once, then reading it a billion times: the writing needs the best equipment, the reading can be done from a cheaper copy. Inference is now the larger and faster-growing share of AI compute — which is why, as the balance tips from training toward inference, the memory work-arounds below start to bite.
When a chatbot writes a reply, it keeps a running “short-term memory” of the conversation so far, called the KV-cache. It sits in expensive HBM and grows with every word. If you can shrink or compress it, you serve the same users with far less memory — or far more users with the same memory.
This is the force that should give the supercycle thesis most pause, because it is not coming from a Chinese rival or a software team. It is coming from the incumbents.
HBF is a simple idea with a big consequence: stack cheap NAND the way HBM stacks expensive DRAM, to get HBM-class read speed at several times the capacity for the money. The catch is that flash is slower to write — but inference mostly reads (see above), so the trade is often worth it: park the model’s fixed weights in cheap, roomy HBF and keep only the hot, fast-changing data in costly HBM.
None of this replaces HBM tomorrow — HBM stays essential for training and for the most demanding work. But it caps how far, and how permanently, HBM can price, precisely because the alternative has the memory industry’s own backing.
The memory makers’ best hope for keeping premium margins was to move up the value chain — to sell not a commodity chip but a custom, integrated HBM stack that only they could design. Nvidia’s latest move is aimed straight at that hope.
An HBM stack sits on a small controller chip called the base die — the brain that manages the memory. With NVHBM, Nvidia pulls that controller into its own design and sets a common specification that any of the three makers can build to. The memory becomes a standardised part Nvidia designs and they manufacture — which keeps them competing on price.
The question that decides who keeps the margin is simply “who controls the design?” By answering “Nvidia,” NVHBM denies the memory vendors the custom-HBM escape route and keeps all three in a commodity contest for Nvidia’s business. It is the same pattern Touchstone has tracked in AI models and elsewhere: the platform commoditises its suppliers and keeps the design rent for itself.
None of this is lost on the memory makers, and they are not standing still. Through the up-cycle they have converted a spot commodity into contracted revenue: all three — SK Hynix, Samsung and Micron — have reportedly sold out their 2026 and 2027 output, and are signing multi-year supply contracts (LTAs) that fix volumes and terms years ahead. The terms are being fought over — SK Hynix has reportedly removed the price cap in its contracts, letting prices float up, while Micron kept caps — which itself tells you both sides are already pricing in a turn.
This deserves real weight. Contracted, pre-sold revenue means the makers’ realised prices lag the spot market, so any down-cycle arrives delayed and smoothed — a glide rather than a cliff — and they have bought perhaps two years of visibility and a demand floor while China ramps. It is the strongest argument that this cycle is gentler than the last.
But a defence that delays is not one that repeals. Contracts reset at renewal; in a genuine glut, committed volumes get renegotiated (the fight over price caps shows buyers already have leverage); the share price re-rates on expectations well before the contracted revenue rolls off; and two of the four forces bite straight through the structure — Nvidia’s standardisation (Force 4) attacks the very custom, differentiated HBM the makers are trying to lock in, and the whole edifice leans on a handful of large buyers who are themselves funding the cheaper alternatives.
Our read is that the memory shortage is real today and priced as if it were permanent — and that four forces are already working against permanence: new supply (China), a falling appetite for memory (software), a cheaper substitute tier (HBF), and a platform standardising the vendors into a commodity contest (Nvidia). None of these breaks the cycle this quarter. Together they argue it is a cycle, with a top, not a new plateau.
The most persuasive signal is not the Chinese newcomer or the clever software — it is that the memory makers are building and standardising the cheaper alternative themselves. People do not hedge a shortage they believe will last — and they do not pre-sell two years of output at fixed terms either. Those sold-out order books buy the incumbents a year or two of visibility, so this is not a call on next quarter’s price: contracted revenue can hold even as the cycle rolls. It is a caution that the multiple tends to turn before the earnings do — one well-followed analyst puts the capital-markets turn within roughly six months; we would not date it that precisely, but we would treat today’s pricing as a peak-cycle input, not a permanent one.
What we are watching: whether HBM spot prices and lead-times start to ease; whether HBF and HBM-light designs win real inference deployments (the proof the substitute is more than a spec); YMTC’s and CXMT’s actual output ramps versus their announcements; and how memory-maker managements themselves guide on pricing durability into 2027.
What we are not doing. This is general market commentary on an industry trend, not advice or a recommendation on any security. The companies named are described to illustrate the dynamic, not as a trade. What any of this means for a particular portfolio is a suitability conversation — one that needs a person, not a page. We are always glad to have it.
General commentary only — not investment advice. By contacting us you consent to Touchstone using your personal data to respond to your enquiry, in accordance with the PDPA. We do not share it with third parties.
AI 硬件交易中最炙手可热的一角，是一场存储短缺——而部分存储厂商，正悄悄打造终结这场短缺的东西。所谓“存储超级周期”，指的是 HBM（高带宽存储）紧缺赋予 SK 海力士、三星、美光三家异常丰厚的利润。短缺是真实的：英伟达上季度对供应商的承诺激增至 2,790 亿美元、主要用于存储，并因存储涨价而指引毛利率由约 74% 降至约 71.5%。但把这条线一路向上延伸是危险的——四股力量已在侵蚀“永久短缺”的假设。
力量一 · 竞争（新供给）
中国将存储自给视为战略问题并据此投入：长江存储（YMTC）正于上海募资约 49 亿美元，宣称要在 2027 年底成为全球最大 NAND 厂商（意味着不到两年内新增近“一个三星”的产量）；长鑫（CXMT）在 DRAM 上追赶同一路径。高盛估计中国先进芯片供给将以约 每年 46% 增长至 2035 年。无论 YMTC 能否达标，供给方向明确——向上。
力量二 · 替代（软件）
短缺同样奖励“少用”的一方。工程上正把存储当作可优化的对象而非固定成本：一篇研究显示，压缩推理时的 KV 缓存比“加更多 GPU／存储”便宜约 1.2–2.0 倍（按该论文所述，单位美元容量最高提升约 16 倍）；微软自研 Maia 200 芯片被描述为围绕“更高效地搬运更少数据”设计。每单位 AI 的存储强度在下降。
力量三 · 替代（架构 · HBF）——最值得警惕，因为来自厂商自身
2026 年 8 月，SanDisk 与 SK 海力士通过开放计算项目（OCP）发布了 HBF（高带宽闪存）的首个行业标准，目标带宽最高约 3 TB/s：以廉价 NAND 堆叠出接近 HBM 的读取速度，而单位美元容量高出数倍，适合以“读”为主的推理（把模型权重放入廉价 HBF、热数据留在昂贵 HBM）。高通则推出刻意“少用 HBM”的推理加速器（AI200／AI250）。当最接近短缺的人开始打造更便宜的替代品，他们是在告诉你他们如何看待这场短缺的价值。
力量四 · 去中介化（英伟达）
存储厂商保住溢价的最大指望，是把 HBM 做成只有它们能设计的“定制”整合堆叠。NVHBM 正对准这一点：英伟达把 HBM 的控制芯片（“基片”）纳入自家设计，并设定三家均可代工的统一规格——“谁掌握设计？”答案是“英伟达”，从而堵死存储厂商借“定制 HBM”逃离商品化的出口。竞争压低价格，替代封顶价格，去中介化则确保厂商无法爬出商品化的盒子。
拓石研判
存储短缺当下真实、却被按“永久”定价；四股力量（中国新供给、软件层“少用”、更便宜的 HBF 替代、英伟达把厂商标准化为商品竞争）虽不会在本季度终结周期，却共同表明这是一个有顶的周期，而非新的高原。最具说服力的信号，是存储厂商自己在打造并标准化更廉价的替代品——没有人会为一场自认会长久的短缺去对冲。某位广受关注的分析师认为“资本市场层面”的拐点或在约六个月内到来；我们不会如此精确地择时，但会把当前定价视为周期高位的输入、而非永久输入。本摘要为节选，非全文翻译；如与英文版存在歧义，概以英文版本为准。