A memory shortage that started with laptops and smartphones has now reached the chips that power the AI boom itself. HBM4, the sixth generation of high-bandwidth memory and the component NVIDIA, AMD, and every major AI accelerator maker need to ship next-generation hardware, is running short just as mass production ramps up. Contract pricing for HBM4 is projected to jump 121% year over year in 2027, according to a report from the Korea Economic Daily. That spike is no longer a rounding error for chipmakers. It is reshaping how AMD, NVIDIA, and three memory suppliers plan the next two years of AI hardware.
AMD CEO Lisa Su flew to Seoul twice in seven months to secure more supply directly from Samsung Electronics and SK hynix, underscoring how memory, not logic chips, has become the tightest link in the AI supply chain. The scramble touches everyone from hyperscale data center operators buying NVIDIA’s Vera Rubin platform to PC buyers who will eventually feel the squeeze through higher prices on consumer DRAM, which shares the same fabs and the same engineering talent as HBM4.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What HBM4 Is and Why the Timing Matters
HBM4 stands for sixth-generation high-bandwidth memory, the stacked DRAM architecture that sits directly next to AI accelerator dies and feeds them data at speeds ordinary DDR5 cannot match. Mass production for major customers began in the second quarter of 2026, moving the technology from engineering samples into real shipments inside a matter of months. That timeline matters because it lines up almost exactly with NVIDIA’s rollout of its Vera Rubin platform, the GPU architecture succeeding Blackwell, which depends on HBM4 to hit its target memory bandwidth.
The jump from HBM3E to HBM4 is not incremental. HBM4 doubles the per-stack interface width from 1,024 I/O connections to 2,048, according to specifications detailed by VLSIresearch Korea. Samsung’s HBM4 implementation is rated at up to 3.3 TB/s of bandwidth per stack, roughly 2.7 times the HBM3E figure, running at pin speeds between 11.7 and 13.0 Gbps. SK hynix has separately demonstrated a 16-layer HBM4 stack packing 48 GB of capacity at around 2 TB/s per stack. Those numbers explain why NVIDIA, AMD, and every hyperscaler racing to stand up new AI clusters see HBM4 as the gating factor, not the logic die.
SK hynix says its HBM4 production yields are now approaching the maturity level of HBM3E, a sign the technology has moved past the early, expensive ramp phase that typically drags down supply in a new memory generation’s first year. The company is already planning the next step, HBM4E, with full-scale mass production targeted for 2027. That overlapping roadmap, HBM4 scaling up while HBM4E is already being planned, shows how compressed the memory industry’s product cycles have become under AI demand.
AMD’s Lisa Su Goes to Seoul, Twice
The clearest sign of how tight HBM4 supply has gotten came on October 7, 2026, when AMD chief executive Lisa Su met with Samsung Electronics executives in Seoul to negotiate expanded memory supply and discuss broader semiconductor cooperation. It was her second trip to South Korea in seven months, according to Tech Times, a frequency that signals just how central Korean memory makers have become to AMD’s AI accelerator roadmap.
Su was direct about the pressure facing the industry. “Memory supply is very tight,” she said, describing the state of HBM4 and broader memory availability as AMD plans its AI-chip supply chain, according to Chosun Biz. She added that “AI demand is very high, so we are planning long-term supply chains with our partners,” framing the Samsung and SK hynix relationships as multi-year commitments rather than spot purchases. Su called the two Korean suppliers “core supply chain partners for AMD,” a characterization that puts memory sourcing on the same strategic tier as AMD’s foundry relationship with TSMC.
Speaking separately about the depth of those ties, Su said: “We have gotten great support from both Samsung and SK Hynix, and we view the partnerships on HBM not just one generation but really multiple generations.” That comment, reported by Tech Times, points to a structural shift. AMD is no longer treating HBM sourcing as a single-generation negotiation. It is locking in supply commitments that stretch across HBM4, the coming HBM4E, and whatever comes after, a hedge against exactly the kind of price spike the market saw in early October.
The 117% Price Spike, Explained
The number driving most of this week’s coverage is the 121% year-over-year jump in HBM4 contract pricing projected for 2027, a figure that reflects buyers bidding up scarce allocation rather than a simple cost-of-goods increase. Longer-term forecasts cited by industry trackers suggest a 12-high, 36 GB HBM4 stack could rise from roughly $600 in 2026 to about $1,300 in 2027, more than doubling if current supply-chain projections hold, according to the Korea Economic Daily report.
Pricing on that scale changes the economics of every AI accelerator built around HBM4. A single high-end AI GPU can carry multiple HBM4 stacks, so a doubling in per-stack cost does not just squeeze memory-maker margins. It flows directly into the bill of materials for NVIDIA’s Vera Rubin boards, AMD’s Instinct MI455 accelerators, and every custom AI chip hyperscalers are designing in-house. Data center operators who modeled 2027 capital budgets around 2026 memory pricing are now working with a materially different number.
| Metric | HBM3E (prior generation) | HBM4 (current generation) |
|---|---|---|
| Per-stack I/O connections | 1,024 | 2,048 |
| Samsung peak bandwidth per stack | ~1.2 TB/s (baseline) | Up to 3.3 TB/s |
| Pin speed | Lower than HBM4 | 11.7–13.0 Gbps |
| SK hynix 16-layer stack capacity | Not applicable at this density | 48 GB per stack, ~2 TB/s |
| 12-high 36 GB stack price (estimate) | Lower baseline pricing | ~$600 in 2026, ~$1,300 projected for 2027 |
| Year-over-year contract price move (2026→2027 proj.) | N/A | +121% |
This is the same dynamic that has already hit consumer electronics. Tech Insider covered how smartphone shipments crashed 20% to a 14-year low as memory shortages rippled through handset bills of materials, and how PC shipments dropped 20.1% in the third quarter as DRAM prices surged. HBM4’s price spike is the AI-data-center version of the same story, playing out in server rooms instead of retail stores.
Samsung’s Bet: More Than Half of All HBM
Samsung Electronics has set an explicit internal target for its share of the HBM market going forward. Hwang Sang-jun, an executive vice president at Samsung Electronics, said the company’s goal “is for HBM4 to account for more than half of total HBM,” according to Chosun Biz. That statement predates the current price spike but explains why Samsung is positioning itself so aggressively in the AMD supply talks and in NVIDIA’s qualification process.
Samsung, SK hynix, and Micron have all reportedly passed NVIDIA’s HBM4 qualification process, clearing the way for all three to ship into NVIDIA’s Vera Rubin platform. Industry supply-chain estimates put SK hynix at roughly 60–70% of HBM4 volume for Vera Rubin, with Samsung at 25–30% and Micron supplying the remainder. Those shares are not fixed. Samsung’s explicit ambition to pass 50% of total HBM output, combined with its new multi-year memorandum with AMD covering HBM4, advanced DDR5, joint memory development, and advanced packaging, suggests the company intends to close the gap with SK hynix rather than settle for a secondary role.
The AMD-Samsung memorandum, signed around Su’s October 7 visit, runs three years and goes beyond a simple supply contract. It covers potential future foundry manufacturing as well, meaning Samsung is angling to turn its memory relationship with AMD into logic-chip foundry orders too, according to Tom’s Hardware’s reporting on the talks. That is a notable strategic trade: Samsung uses scarce HBM4 capacity as leverage to win manufacturing business that would otherwise go to TSMC.
SK hynix and Micron’s Positions in the Three-Way Race
SK hynix currently holds the largest share of HBM4 shipments into NVIDIA’s platform, consistent with the lead it has held since HBM3E. Its yields nearing HBM3E maturity levels mean it can scale output faster than a company still working through early-generation defect rates. The company’s 16-layer, 48 GB stack demonstration also positions it at the high-capacity end of the market, useful for AI accelerators that need maximum memory per package rather than maximum per-stack speed alone.
Micron, the smallest of the three qualified HBM4 suppliers by reported volume, has less public detail available on its specific HBM4 shipment commitments but has cleared NVIDIA’s qualification bar alongside its larger rivals. For AMD and NVIDIA, having three qualified suppliers instead of two is itself a hedge against the kind of single-source risk that can turn a shortage into a production stoppage. It does not, however, solve the aggregate capacity problem. All three companies are constrained by the same wafer capacity, advanced packaging bottlenecks, and skilled-labor limits that defined the DRAM shortage already hitting PCs and phones.
| Supplier | Reported HBM4 share for NVIDIA Vera Rubin | Notable HBM4 milestone |
|---|---|---|
| SK hynix | ~60–70% | 16-layer, 48 GB stack demoed; yields near HBM3E maturity |
| Samsung Electronics | ~25–30% | Up to 3.3 TB/s per stack; target of 50%+ of total HBM market |
| Micron | Remainder (smallest of the three) | Passed NVIDIA HBM4 qualification alongside Samsung and SK hynix |
AMD’s Instinct MI455 Puts HBM4 to Work
AMD has already begun shipping its latest Instinct MI455 AI accelerator with 432 GB of HBM4 memory onboard, a configuration deployed as part of the company’s rack-scale Helios AI systems. That is a meaningful jump in onboard memory capacity for a single accelerator, and it is the clearest evidence that HBM4 is not a future promise for AMD. It is already shipping in production hardware, which is exactly why Su’s supply negotiations with Samsung carry urgency. Every MI455 unit AMD wants to ship competes for the same constrained HBM4 pool that NVIDIA’s Vera Rubin platform is drawing from.
AMD’s own AI infrastructure ambitions compound the squeeze. The company’s Helion-class rack designs, built for agentic AI workloads, pack tens of thousands of cores per rack and depend on dense memory bandwidth to keep those cores fed. Readers can see how that scale plays out in Tech Insider’s coverage of AMD’s Helion rack packing 32,768 cores for AI agents, a design philosophy that only works if HBM4 supply keeps pace with AMD’s rack shipment targets.
NVIDIA’s Vera Rubin Platform Raises the Stakes
NVIDIA’s Vera Rubin platform, the architecture succeeding Blackwell, is reported to already be in mass production and shipping rather than sitting at the sampling stage. That timing puts NVIDIA, the single largest buyer of AI accelerator memory in the world, drawing from the exact same HBM4 pool that AMD is fighting for. With SK hynix, Samsung, and Micron all qualified to supply Vera Rubin, NVIDIA has more sourcing flexibility than AMD does today, which may partly explain why AMD’s leadership is pursuing direct, multi-year commitments with Samsung rather than relying on spot-market allocation.
The broader implication is that HBM4 scarcity is now a shared industry problem rather than a company-specific one. NVIDIA, AMD, and every hyperscaler building custom silicon are effectively bidding against each other for the same finite wafer and packaging capacity at Samsung, SK hynix, and Micron. That dynamic mirrors what ASML has said about AI demand outpacing chip-tool supply through 2028, a supply-versus-demand mismatch Tech Insider covered in its report on ASML warning that AI demand will top chip-tool supply through 2028. HBM4 is simply the memory-side expression of the same structural shortage.
How This Echoes the DRAM Shortage Already Hitting Consumers
HBM4 and the DDR5 memory used in ordinary laptops and desktops are not the same product, but they come from the same fabs, the same silicon wafers, and the same limited pool of advanced packaging engineers. When Samsung, SK hynix, and Micron prioritize HBM4 production for AMD and NVIDIA because it carries far higher margins per wafer, that capacity gets pulled away from consumer DRAM lines. Tech Insider has tracked that spillover directly, from Elon Musk’s public remarks on Micron memory constraints, covered in Musk’s Micron memory remarks spreading across six outlets, to the Intel and AMD CPU price increases tied to the same shortage, detailed in Intel’s CPU prices jumping up to 20% as AMD preps its own hikes.
That same capacity competition is now showing up in gaming hardware sales figures, which fell to a 13-year low as DRAM prices climbed, a trend Tech Insider broke down in its report on game hardware sales hitting a 13-year low as DRAM soars. The HBM4 shortage and the consumer DRAM shortage are, in practical terms, the same shortage wearing two different price tags.
Historical Context: How HBM Got Here
High-bandwidth memory has gone through six generations since its original JEDEC standardization, moving from a niche technology for high-end graphics cards into the single most important memory format in AI computing. Each generation has roughly doubled interface width and bandwidth while adding more stacked layers, a trajectory HBM4 continues with its jump to 2,048 I/O connections per stack. What has changed is demand. Previous HBM generations served a relatively small market of high-end GPUs and supercomputers. HBM4 is being bought by every major cloud provider and AI lab simultaneously, at a volume the memory industry has never had to plan for on this short a timeline.
SK hynix’s push toward HBM4E mass production in 2027, announced even as HBM4 itself is still ramping, shows how compressed that planning cycle has become. Memory makers used to have years between generations to amortize capital spending on new fabs and packaging lines. With AI demand accelerating, Samsung, SK hynix, and Micron are now running HBM4 production and HBM4E development in parallel, a capital-intensive approach that only makes financial sense if AI accelerator demand keeps growing at its current pace.
Market Impact: What the Shortage Means for AI Infrastructure Spending
For hyperscalers and AI labs already committing hundreds of billions of dollars to data center buildouts, a doubling of HBM4 stack pricing between 2026 and 2027 is a direct hit to the cost per AI accelerator deployed. That pressure lands at a moment when companies like Microsoft, Meta, and Google have already disclosed combined AI capital spending in the hundreds of billions of dollars for the year. Memory cost inflation on this scale does not cancel those buildouts, but it changes the math on how many accelerators a given budget can actually purchase, and it strengthens the case for AMD and NVIDIA locking in supply contracts years in advance rather than buying on the spot market.
It also raises the stakes for AMD’s three-year memorandum with Samsung. If HBM4 pricing keeps climbing through 2027 as the Korea Economic Daily’s estimates suggest, locking in volume and pricing now looks less like a routine supply agreement and more like a hedge against a memory market that could get considerably more expensive before new capacity comes online.
Competitive Comparison: AMD’s Approach vs. NVIDIA’s Approach
AMD and NVIDIA are responding to the same shortage with different strategies. AMD, with Lisa Su personally negotiating in Seoul and a formal multi-year memorandum now in place with Samsung, is pursuing direct, bilateral supply security. That approach concentrates risk and reward: if Samsung’s HBM4 output scales as planned, AMD gets priority access and favorable terms. If Samsung falls short, AMD’s Instinct MI455 roadmap is more exposed than a company sourcing from all three qualified suppliers.
NVIDIA, by contrast, has already qualified Samsung, SK hynix, and Micron for Vera Rubin, spreading its sourcing across all three. That diversification gives NVIDIA more flexibility to absorb a shortfall from any single supplier, but it also means NVIDIA is competing with AMD and every hyperscaler’s custom silicon program for the same aggregate HBM4 supply, rather than locking in a dedicated channel the way AMD’s Samsung memorandum does. Both approaches are rational responses to scarcity. Which one proves more resilient will depend on how fast Samsung, SK hynix, and Micron can each scale HBM4 output through 2027.
What Comes Next: Five Predictions
- HBM4 contract pricing will likely stay elevated through at least the first half of 2027, given that SK hynix, Samsung, and Micron are all directing capital toward HBM4E development rather than pure HBM4 capacity expansion.
- AMD will probably pursue additional supply memoranda, potentially with SK hynix or Micron, to avoid overreliance on the Samsung agreement alone, mirroring the multi-supplier approach NVIDIA already uses for Vera Rubin.
- Samsung’s push to exceed 50% of total HBM output will intensify competitive pressure on SK hynix, which has held the largest HBM4 share for NVIDIA to date, potentially triggering further price negotiations between the two Korean suppliers’ customers.
- Consumer DRAM and NAND pricing will likely remain under upward pressure into 2027 as fabs continue prioritizing higher-margin HBM4 production, extending the trend already visible in PC and smartphone shipment data.
- Expect more AI accelerator vendors to disclose specific HBM4 capacity onboard their chips, following AMD’s lead with the Instinct MI455’s 432 GB configuration, as memory capacity becomes as much a marketing differentiator as raw compute throughput.
Why Gamers and PC Buyers Should Still Pay Attention
HBM4 itself never reaches a gaming PC. It is reserved for data-center AI accelerators. But the fabs producing it are the same ones producing GDDR and DDR memory for consumer graphics cards and systems, which is why the HBM4 shortage and the consumer memory shortage move in tandem. Readers building a new system in this environment can see how tight margins have gotten for component selection in Tech Insider’s guide to building a Mini-ITX gaming PC with a 310mm GPU fit, where memory and GPU availability now factor into build planning in ways they did not two years ago.
The takeaway for consumer hardware buyers is simple. As long as AMD, NVIDIA, and hyperscalers keep bidding up HBM4 allocation at Samsung, SK hynix, and Micron, the same three companies have less incentive to expand the consumer DRAM and NAND lines that determine RAM kit and SSD pricing. That link, not any direct technical overlap, is why an AI-accelerator memory shortage in Seoul factories is already showing up in PC component price tags.
Frequently Asked Questions
What is HBM4?
HBM4 is the sixth generation of high-bandwidth memory, a stacked DRAM format used in AI accelerators. It doubles the per-stack I/O interface to 2,048 connections compared with HBM3E and reaches bandwidth of up to 3.3 TB/s per stack in Samsung’s implementation.
Why is HBM4 in short supply in October 2026?
Mass production only began ramping for major customers in the second quarter of 2026, just as NVIDIA’s Vera Rubin platform and AMD’s Instinct MI455 accelerators both began drawing heavily on the same supplier base, driving contract pricing up 117% week over week in early October, according to the Korea Economic Daily.
Who makes HBM4?
Three companies currently produce qualified HBM4: SK hynix, Samsung Electronics, and Micron. All three have reportedly passed NVIDIA’s qualification process for the Vera Rubin platform, with SK hynix holding the largest reported share.
What did AMD and Samsung agree to?
AMD and Samsung signed a three-year memorandum covering HBM4 supply, advanced DDR5, joint memory development, advanced packaging, and potential future foundry manufacturing, following CEO Lisa Su’s October 7, 2026 meetings with Samsung executives in Seoul.
Will the HBM4 shortage affect gaming PC prices?
Not directly, since HBM4 is used only in data-center AI accelerators. But the same fabs produce consumer DDR and GDDR memory, and prioritizing higher-margin HBM4 production has already contributed to DRAM price increases affecting PCs, smartphones, and gaming hardware.
What is HBM4E and when does it arrive?
HBM4E is the next planned iteration after HBM4. SK hynix has targeted full-scale mass production for 2027, even as HBM4 itself is still scaling up, reflecting how compressed memory product cycles have become under AI demand.
How much onboard HBM4 does AMD’s latest AI chip have?
AMD’s Instinct MI455 accelerator ships with 432 GB of HBM4 memory onboard, deployed as part of the company’s rack-scale Helios AI systems.
Is NVIDIA also affected by the HBM4 shortage?
Yes. NVIDIA’s Vera Rubin platform draws from the same pool of qualified HBM4 suppliers as AMD, though NVIDIA has diversified its sourcing across SK hynix, Samsung, and Micron rather than concentrating on a single bilateral agreement.
Related Coverage
- RX 9070 XT vs RTX 5070 Ti: 8% FPS, $340 Price Gap [2026]
- Wooting 80HE vs Keychron Q6 Ultra 8K: $60 Price Gap [2026]
- Google Fitbit Edge Debuts at $179.99, Ships Oct. 27 [2026]
- Nasdaq Slips 0.2%, Oil Jumps 2% on Ukraine Strikes [2026]
- Tesla Cuts AI5 Chip Memory, Reverses Course to 96GB [2026]
