---
格式版本: 2
标题: "Why High Bandwidth Memory Stopped Behaving Like a Commodity"
原文链接: "https://sanieinstitute.substack.com/p/why-high-bandwidth-memory-stopped"
发布日期: "2026-08-31"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:scrape:strict_html_body"
发布时间证据: "div class=pencraft pc-display-flex pc-gap-12 pc-alignItems-center pc-reset byline-wrapper: Aug 31, 2026"
发布时间校准原因: "规则确认唯一严格发布时间，来源 scrape:strict_html_body"
发布时间校准置信度: "high"
发布时间候选数量: 11
发布时间严格候选数量: 2
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-09-03T23:54:02+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-09-03T23:53:49+08:00"
入库时间: "2026-09-03T15:54:03.019Z"
来源平台: "Substack 数据中心相关博客搜索"
搜索渠道: "source_template"
搜索词: "site:substack.com HBM"
匹配关键词:
  - "HBM"
  - "bandwidth"
  - "AI"
  - "roadmap"
相关厂家:
  - "NVIDIA"
  - "Broadcom"
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 33
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "文章聚焦HBM定价与AI加速器路线图，但未涉及超节点/AI Rack机柜级系统、架构、供电散热互连等核心主题，属供应链市场分析，与项目关注范围偏离。来源为个人博客，权威性低，技术细节少。"
AI质检模型: "zj-deepseek-v4-flash"
AI质检时间: "2026-09-03T23:54:13+08:00"
AI主题相关性: 5
AI来源权威性: 2
AI新颖性: 5
AI技术细节: 5
AI商业部署信号: 8
AI完整性: 8
AI摘要: "HBM已脱离传统DRAM周期：主导买家由PC/手机OEM转为AI加速器厂商，他们为多年硅片路线图提前锁定产能，使需求不再对价格敏感。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-04T00:18:28.454Z"
采集批次: "2026年9月3日22点43分34秒"
采集批次ID: "20260903-224334-406"
去重键: "https://sanieinstitute.substack.com/p/why-high-bandwidth-memory-stopped"
---

[institute.sanie.com](http://institute.sanie.com/)

---

##### Executive Summary Audio (free)

0:00

\-2:30

Prefer listening? Paid subscribers can access a full narration of this article through the Substack app. Listen while commuting, traveling, walking, or exercising.

---

High-bandwidth memory has traded like every other memory product for the past two years: tight supply, rising prices, and a narrative that treats the shortage as an unusually severe turn in the same cycle DRAM has run for four decades. That framing is no longer accurate. The buyers reshaping HBM demand are not OEMs reacting to end market sales; they are AI accelerator vendors executing multi-year silicon roadmaps they have already committed capital against. That difference does not just explain why HBM prices are high. It also explains why the mechanism setting those prices has permanently changed, and why the risk beneath the current boom is not the risk most coverage is pricing.

The reading is direct. HBM has decoupled from the historic DRAM cycle because its buyers behave nothing like the OEMs who set the terms of every prior memory cycle, and the same concentration handing memory makers unprecedented pricing power removes the demand elasticity that used to cushion the downturn.

This post examines why HBM buyers behave differently than the memory industry’s traditional customers, what that difference has already done to pricing power inside SK Hynix, Samsung, and Micron, and the fragility that structural story is quietly building in. It closes with what this means for how memory exposure should actually be modeled. (🔒 for paid subscribers)

- The Memory Cycle Investors Think They Are Still Watching
- Why HBM Buyers Do Not Behave Like Memory Buyers
- 🔒 The Pricing Power This Handed to Memory Makers
- 🔒 The Fragility Hiding Inside a Structural Story
- 🔒 Investor Implications
- 🔒 Bottomline: HBM Now Prices to Roadmaps, Not to …

**The Memory Cycle Investors Think They Are Still Watching**

The memory industry has run the same basic pattern for forty years. Demand from PC or phone OEMs outpaces supply, prices spike, and every DRAM maker on earth adds capacity to capture the upside. That capacity eventually lands, the OEMs pull back the moment cost pressure builds into consumer products, and the same suppliers that raced to expand a year earlier are left holding a glut. The 1993 Windows PC cycle drove DRAM demand up roughly fourfold per device and then prices fell more than 60% by 1995 and 1996, as close to 50 new fab construction plans landed at once ([Memory Mania: How a Once-in-Four-Decades Shortage Is Fueling a Memory Boom](https://newsletter.semianalysis.com/p/memory-mania-how-a-once-in-four-decades)). The 2010 mobile and cloud cycle peaked and then fell nearly 46% within months. Every one of those cycles ended the same way, because every one of those cycles was set in motion by a buyer who could walk away the moment price stopped making sense for its product.

Most coverage of the current HBM shortage assumes a variation on that same story, a sharper version of an old cycle rather than a genuinely new one. That assumption treats HBM as an intense instance of a familiar mechanism, just bigger, just tighter, just longer. The mechanism itself is what actually changed this time, and it did not change because AI made memory more valuable. It changed because AI changed who is buying it.

**Why HBM Buyers Do Not Behave Like Memory Buyers**

The OEMs who set the terms of every prior cycle bought memory the way any component buyer sources a commodity, compare price across a handful of qualified suppliers, commit a quarter or two ahead, and shift volume when a cheaper part clears qualification. That buyer no longer sets the price of HBM. AI accelerator vendors do, and they do not source memory the way a phone OEM ever did.

Samsung signed a five year memory and advanced packaging agreement with Broadcom worth more than $200B running through 2030, and SK Group’s HBM supply commitments to Nvidia over the same horizon run past $500B, with every major supplier’s entire 2026 output already committed before either agreement was even announced ([Samsung and SK Hynix Lock In AI Chip Supply Through 2030 With $950B in US Deals](https://www.techtimes.com/articles/321554/20260725/samsung-sk-hynix-lock-ai-chip-supply-through-2030-950b-us-deals.htm)). Those are not purchase orders. They are capacity reservations tied to chip roadmaps that were locked in years before the silicon ships.

A buyer signing a five-year roadmap commitment does not behave like a buyer comparing quotes each quarter. It cannot walk away when price rises, because the accelerator generation it is building was designed around a specific memory configuration years in advance, and switching suppliers mid-roadmap means requalifying a part inside a chip that has already taped out. HBM demand is now roadmap-locked rather than price-elastic, and a roadmap-locked buyer cannot discipline a supplier’s price the way a price-elastic one always could. That is the actual mechanism, and it has nothing to do with how tight the current shortage happens to be.

What follows moves to the consequences of that mechanism, the pricing power it has already handed to memory makers, the fragility building underneath a story most coverage is still reading as purely structural, and what this means for how memory exposure should actually be modeled. Access is reserved for paid subscribers.

**The Pricing Power This Handed to Memory Makers**

A roadmap-locked buyer base does more than remove price discipline. It actively redirects supply. HBM wafer input is on pace to reach approximately 30% of total DRAM wafer input by the end of 2027, up from about 18% in 2025, and every point of that shift comes directly out of capacity that would otherwise have produced conventional DRAM ([Tight DRAM Supply Gives Suppliers Greater Pricing Power in HBM, with HBM Contract Prices Expected to Surge Multiples Higher in 2027](https://www.trendforce.com/presscenter/news/20260602-13074.html)). That crowding-out effect is deliberate. HBM’s own profitability recently fell below standard DDR5 64GB RDIMM, which gave suppliers a direct incentive to reallocate wafers toward HBM and push through the price increases needed to restore its position as the higher-margin product.

This is a genuinely new position for memory makers. SK Hynix, Samsung, and Micron have never before been able to set price across a multi-year horizon rather than a spot market, because no prior buyer base would have tolerated it. AI accelerator vendors can tolerate it, because the cost of memory is a rounding error against the cost of the compute it feeds, and because switching now would mean redesigning a chip that has already shipped to hyperscaler customers on a fixed schedule. Pricing power that used to evaporate the moment supply caught up now persists for as long as the roadmap commitments behind it hold, which for the current generation of agreements runs to 2030.

**The Fragility Hiding Inside a Structural Story**

The concentration that gave memory makers this pricing power is the same concentration that removes the buffer every prior cycle relied on to soften its landing. A price-elastic OEM base is, in effect, a shock absorber: it pulls back gradually as cost rises, spreading the correction across millions of independent purchasing decisions. A handful of roadmap-locked accelerator vendors is not a shock absorber. They are a small number of correlated decisions, and if any one of them slows its accelerator rhythm, delays a generation, or hits a demand air pocket of its own, the resulting drop in HBM demand arrives all at once rather than gradually.

Morningstar’s own view of the broader semiconductor cycle is a useful check on how consensus is still framing this risk. The firm argues a typical semiconductor cycle runs about four years and that AI is stretching the limits of the current upcycle rather than replacing the pattern outright, and it explicitly notes that memory makers remain exposed to the intense cyclicality of the sector regardless of how structural the current story looks ([Even the AI Gold Rush Can't Stop the Chip Industry's Boom-Bust Cycle](https://finance.yahoo.com/news/even-ai-gold-rush-cant-154213354.html)). That framing is directionally right and still understates the mechanism. It is not that HBM will eventually behave like a normal cycle again. It is that when this cycle turns, it will turn through three or four correlated customers rather than millions of independent ones, and that concentration is likely to make the drop sharper and harder to see coming, not softer.

Nothing about that risk shows up in how the current shortage is being priced. Coverage treats sold-out capacity and multi-year contracts as evidence of durability, when a five-year commitment from a small number of buyers is exactly the structure that produces a step change rather than a gradual correction once one of those buyers changes course.

**Investor Implications**

The practical consequence is that memory makers should not be modeled the way they were modeled through the last three decades of DRAM cycles, and they should not be modeled as if AI has simply made that old framework obsolete either. The right model tracks a different set of inputs entirely. Traditional cycle analysis watched DRAM channel inventory, OEM order patterns, and end market unit sales, because those were the levers that actually moved price. None of those levers explain HBM pricing today. What matters now is accelerator roadmap cadence at three or four named companies, the pace at which each renews or extends its capacity commitments, and any signal that a generation is slipping, being redesigned around a lighter memory configuration, or facing a demand pause of its own.

That also changes what a memory maker’s backlog is actually worth. A backlog built from diversified OEM orders has always carried a reasonable assumption of renewal, because losing any single customer barely moves the aggregate. A backlog concentrated in a handful of multi-year AI accelerator contracts carries a different kind of value, higher visibility as long as those specific customers keep executing their roadmaps, and materially higher downside if any one of them does not. Investors pricing SK Hynix, Samsung, or Micron on a simple earnings multiple against current bookings are pricing the upside of that concentration without pricing what it costs on the way down. The names to watch closely are not the memory makers themselves so much as the handful of accelerator vendors whose roadmaps those makers are now priced against.

**Bottomline: HBM Now Prices to Roadmaps, Not Elasticity, and That Cuts Both Ways**

HBM’s pricing power is real, and it is not going away on the timeline the last three DRAM cycles would suggest. That is precisely why it is being misread. Investors have correctly identified that this cycle looks different from 1993, 2010, or 2018, and have concluded from that difference that the usual downside case no longer applies. The opposite is closer to true. The usual downside case does not apply because a sharper one has taken its place, one where the trigger is not a broad OEM pullback but a single accelerator vendor’s roadmap slipping, and where the correlated concentration behind today’s pricing power removes the gradual cushion that made every prior memory downturn survivable rather than abrupt.

We expect the coming quarters to reward investors who are tracking accelerator roadmap execution as closely as they are tracking memory maker earnings, because by the time a slowdown shows up in DRAM channel data the way it always used to, the concentrated structure behind this cycle will have already made the correction sharper than the data suggested it could be.

*[institute.sanie.com](http://institute.sanie.com/)*

*Note: This post is for informational purposes and does not constitute investment advice.*