HBM4 is sold out, CXMT just IPO'd: the new memory bottleneck in the AI buildout

The new choke point is memory, not compute

In February 2026, Samsung became the first memory maker to put HBM4 — built on its sixth-generation 10-nanometre-class (1c) DRAM and a 4-nanometre logic base die — into mass production. The chip runs at 11.7 gigabits per second steady, up to 13 Gbps, and delivers up to 3.3 terabytes per second of bandwidth. By April, Samsung's chief financial officer Park Soon-cheol confirmed the obvious on an earnings call: the entire 2026 HBM4 allocation was already sold out. TrendForce has since raised Samsung's 2026 HBM4 shipment outlook from 3.5 billion to roughly 4 billion gigabits.

This matters because the next generation of AI accelerators is unusually hungry for stacked memory. Nvidia's Rubin GPU ships with up to 288 GB of HBM4 and 22 TB/s of bandwidth. AMD's Instinct MI455X, supplied primarily by Samsung under a new memorandum of understanding, scales to 432 GB of HBM4 and up to 23.3 TB/s. The accelerators are only as fast as the memory underneath them.

Samsung is leading, but not alone

Samsung shipped the industry's first HBM4E samples to customers in May 2026, and at FMS 2026 (August, Santa Clara) the company previewed its next concept: zHBM, a vertically stacked architecture that places HBM directly above the AI accelerator rather than beside it. Samsung claims zHBM can deliver roughly eight times the performance of HBM5, more than ten times the memory density, three times the energy efficiency and more than a 50% drop in thermal resistance. It also unveiled V10 BV-NAND, an industry-first 400-plus-layer bonded V-NAND with 58% higher density than the previous generation.

Micron is not standing still. In March 2026 it began volume shipments of a 36 GB, 12-high HBM4 product designed for Nvidia Vera Rubin, with more than 2.8 TB/s of bandwidth and a 20% power-efficiency improvement over HBM3E. In July, Micron raised its planned US fab and technology investment to more than $250 billion through 2035. Shares are up more than 200% year to date.

SK Hynix, the HBM3E cycle leader, is bringing HBM4 to commercial launch in 2026, with samples already out since March. It is, however, redirecting some capacity to DDR5 and LPDDR5, where pricing has become more attractive, and its HBM4 mass production slipped into the third quarter because of interface-synchronisation delays.

CXMT has arrived — but is one to two generations behind

The most consequential new entrant is Changxin Memory Technologies (CXMT), China's leading DRAM maker. On 27 July 2026, CXMT listed on the Shanghai stock exchange in a debut that surged almost 466%, raising up to $9.8 billion in what is likely to be the largest semiconductor IPO in Chinese history.

According to Counterpoint Research, the global DRAM market today is led by Samsung at 38%, SK Hynix at 29% and Micron at 22%. CXMT holds around 8%. SemiAnalysis estimates CXMT will finish 2026 with roughly 350,000 wafer starts per month of DRAM capacity — about 25,000 short of Micron — on its way to becoming the third-largest memory supplier in the world.

The product story, however, is more measured:

• CXMT unveiled its latest DDR5 chips in November 2025 and is preparing LPDDR6 mass production in the second half of 2026.
• The company's new G4 DRAM technology features memory cells roughly 20% smaller than its G3 generation.
• CXMT is targeting HBM production from the end of 2026. According to Counterpoint's MS Hwang, the first HBM products are likely to be HBM3 or HBM3E, leaving CXMT one to two generations behind the HBM4 and HBM4E leaders.
• The hard ceiling is lithography. Without access to extreme ultraviolet (EUV) machines — restricted under US export controls — CXMT needs around 30% more wafers than its rivals to produce the same amount of memory. The shortfall is structural, not a roadmap problem.

China's domestic demand is large enough to absorb a great deal of that output. Counterpoint and MST Financial both expect CXMT's share gains to come primarily from Chinese customers, including Tencent, ByteDance, Alibaba and domestic phone manufacturers. Western hyperscalers are unlikely to qualify CXMT HBM at the front of the AI accelerator stack in 2026.

The market is pricing the squeeze

UBS analyst Timothy Arcuri expects stronger-than-expected HBM4 and HBM4E pricing through 2027. The reason is simple: TrendForce says tight conventional DRAM supply has pushed some server memory returns above HBM returns in early 2026, giving suppliers an incentive to be careful about how quickly they move capacity between product lines. Micron's share price and Samsung's sold-out order book are the visible market signals; the underlying dynamic is that AI's most physical input is now rationed.

What this means for UK organisations

UK buyers do not need to memorise the part numbers. They do need to absorb three practical consequences:

1. AI infrastructure timelines are memory-bound. Any 2026 procurement plan for Nvidia Rubin, AMD Instinct MI455X or successor accelerators will be paced by HBM allocation, not GPU availability. If your roadmap assumes a generous 2026 supply, that assumption is now wrong.
2. Pricing pressure is moving up the stack. HBM4 and HBM4E contract prices are expected to climb sharply into 2027. Cloud inference pricing, hosted model costs and downstream AI product margins will absorb some of that increase.
3. Supplier concentration is increasing. Three companies control roughly 89% of the global DRAM market and 100% of leading-edge HBM. CXMT is closing the capacity gap but is constrained from closing the technology gap in the near term. Buyers should plan for limited substitution at the AI accelerator tier.

The short version

In 2026, the question for AI infrastructure is no longer who has the most GPUs. It is who can feed those GPUs. Samsung has the lead, Micron is investing at a scale that suggests it intends to keep that lead contested, SK Hynix is the legacy HBM3E champion now transitioning, and CXMT is the new entrant that has just bought itself a public balance sheet to catch up — without, for now, the lithography to reach the front.

---

Sources

• Samsung Global Newsroom, Samsung Unveils Next-Gen 3D-Memory Vision at FMS 2026, 5 August 2026 — https://news.samsung.com/global/samsung-unveils-next-gen-3d-memory-vision-at-fms-2026-charting-the-future-of-ai-infrastructure
• Samsung Global Newsroom, Samsung and AMD Expand Strategic Collaboration on Next-Generation AI Memory Solutions, 2026 — https://news.samsung.com/global/samsung-and-amd-expand-strategic-collaboration-on-next-generation-ai-memory-solutions
• CNBC, CXMT's blockbuster debut in Shanghai sets stage for next test against global memory giants, 31 July 2026 — https://www.cnbc.com/2026/07/31/cxmts-sk-hynix-samsung-micron-memory-chip.html
• SemiAnalysis, China's CXMT Is Set to Challenge DRAM Incumbents, 23 June 2026 — https://newsletter.semianalysis.com/p/chinas-cxmt-is-set-to-challenge-dram
• Tom's Hardware, CXMT close to matching Micron's memory capacity in 2026, August 2026 — https://www.tomshardware.com/pc-components/dram/cxmt-close-to-matching-microns-memory-capacity-in-2026-research-claims-would-put-china-on-track-to-become-worlds-second-largest-dram-producer
• StartupFortune, Samsung Sold Out Every Gigabit of Its 2026 HBM4 Memory Chips, 13 August 2026 — https://startupfortune.com/samsung-sold-out-every-gigabit-of-its-2026-hbm4-memory-chips/
• Wikipedia, ChangXin Memory Technologies — https://en.wikipedia.org/wiki/ChangXinMemoryTechnologies
• Counterpoint Research, Global DRAM and HBM market share, 2026 — https://counterpointresearch.com/en/insights/global-dram-and-hbm-market-share

Reported figures attributed to company earnings calls, Counterpoint, TrendForce, SemiAnalysis, Yonhap and Seoul Economic Daily are company-reported or analyst-estimated. Specific HBM yield, allocation and price data is not independently audited.