TL;DR
A late-June Thorsten Meyer AI report says High Bandwidth Memory has become the main pressure point in the 2026 memory squeeze. The report says HBM demand from AI accelerators is taking fab capacity from DDR5 RAM and GDDR7 graphics memory, with supply sold out through 2026.
A late-June report from Thorsten Meyer AI says High Bandwidth Memory has become the component driving the 2026 memory squeeze, as AI accelerator demand pulls fab capacity away from DDR5 RAM and GDDR7 graphics memory.
The report describes HBM as a vertical stack of DRAM dies, linked by through-silicon vias and mounted near a GPU on an interposer. That layout gives AI accelerators roughly five to ten times the bandwidth of standard graphics memory, helping chips such as Nvidia’s H100, H200 and B200 feed data fast enough to keep compute units busy.
The same design makes HBM costly in fab capacity. Thorsten Meyer AI says one bit of HBM consumes roughly three to four times the wafer area of one bit of DDR5, because stacked dies are larger, yields are harder and a single defect can spoil a full tower. The report estimates per-stack prices at about $200 for HBM3, $300 for HBM3E and around $500 for HBM4.
The report says SK Hynix leads the HBM market, with Samsung attempting a 2026 comeback and Micron sold out for 2026. It also says all three major suppliers had qualified for HBM4 by June 2026, shifting the race from basic qualification to volume, yield and execution.
HBM ate the fab
The thing the factories make instead of your RAM is a tower of stacked memory bolted to every AI chip. In three years it went from niche part to the component that sets the price of nearly all the world’s memory — and now a chunk of its GPUs.
A tower, not a sheet
HBM stacks DRAM dies vertically, links them with thousands of through-silicon vias, and sits beside the GPU to deliver 5–10× the bandwidth of normal graphics memory. AI is bandwidth-bound — without it, the world’s most expensive silicon sits starved for data. But stacking is inefficient: one HBM bit eats 3–4× the wafer area of DDR5, and one defect can ruin a whole tower.
≈ 8 HBM stacks wrap every AI GPUThis isn’t artificial scarcity — AI really is bandwidth-bound, HBM really is the fix, and it really does eat 3–4× its weight in fab capacity. The discomfort is structural: one component, coupled to one customer’s demand, now sets the price of nearly all memory and a slice of GPUs. The market is now $35B → ~$100B by 2028, ~41% of all DRAM revenue (was 8% in 2023), and sold out through 2026. The one hope: with all three suppliers finally racing on HBM4, competition can add supply. The matching risk: if AI demand corrects, HBM is where it breaks first. Next: DDR5 now, DDR6 soon.
AI Demand Reprices Memory
The shift matters because HBM competes for the same broad manufacturing base that supplies ordinary RAM and parts of the consumer GPU market. If memory makers can earn far more from AI-bound HBM stacks than from commodity modules, the report says fab allocation naturally moves toward higher-margin HBM.
That pressure is no longer limited to servers. Thorsten Meyer AI says shortages of GDDR7 memory, used in consumer graphics cards, contributed to Nvidia reportedly cutting RTX 50-series production by a third or more in the first half of 2026. The report frames the issue as structural supply pressure, not a simple inventory hiccup.
High Bandwidth Memory HBM3 module
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Niche Stack to Bottleneck
HBM moved from a specialty part to a central AI component within three years, according to the report. HBM3 powered the H100 era at roughly 819 GB/s per stack, while HBM3E reached about 1.18 TB/s and became the 2026 workhorse for chips such as the H200 and B200.
The next generation, HBM4, is expected to reach roughly 2.8 TB/s per stack and use a new logic base die, with Nvidia’s Rubin platform cited as a major target. Thorsten Meyer AI, citing Silicon Analysts, Introl, TrendForce, DigiTimes, Unibetter, Astute Group and Reuters, says the HBM market could grow from $35 billion to about $100 billion by 2028.
“HBM Ate the Fab”
— Thorsten Meyer AI report
Nvidia Tesla P100 900-2H400-0000-000 GPU Computing Processor – 16 GB – HBM2 – PCIE 3.0 X16 (Certified Refurbished)
GPU Computing Processor
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Pricing and Output Remain Fluid
Several figures remain estimates or reported ranges, including HBM4 per-stack pricing, supplier market shares and the scale of any RTX 50-series production cuts. The report says its pricing data is point-in-time and that market conditions were fast-moving in late June 2026.
It is not yet clear how quickly HBM4 qualification will turn into stable volume, how much yield will improve, or whether AI accelerator demand will keep rising at the same pace. A demand pullback would likely hit HBM suppliers before the rest of the memory market.
GIGABYTE GeForce RTX 5080 Gaming OC 16G Graphics Card, WINDFORCE Cooling System, 16GB 256-bit GDDR7, GV-N5080GAMING OC-16GD Video Card
Powered by the NVIDIA Blackwell architecture and DLSS 4
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
HBM4 Shipments Set the Pace
The next milestone is the HBM4 ramp for Nvidia Rubin and other AI platforms. Memory buyers will be watching whether SK Hynix, Samsung and Micron can turn qualification into enough supply to ease pressure on DDR5, GDDR7 and next-generation memory pricing.
Patriot Viper Venom DDR5 RAM 16GB (2X8GB) 6000MT/s CL36 1.35v UDIMM Desktop Gaming Memory Kit Compatible with Intel XMP/AMD EXPO – PVV516G600C36K
Capacity: 16GB(2 x 8GB)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is HBM?
High Bandwidth Memory is stacked DRAM placed very close to an accelerator, giving AI chips much higher data bandwidth than standard graphics memory. It uses vertical die stacks and through-silicon vias to move data quickly.
Why does HBM affect DDR5 supply?
The report says HBM uses far more wafer area than DDR5 for the same bit capacity. When manufacturers assign more wafers to AI memory, fewer are available for commodity RAM.
Are gaming GPUs affected?
According to the report, GDDR7 shortages have affected consumer graphics cards as suppliers prioritize HBM production. Nvidia reportedly reduced RTX 50-series output in the first half of 2026, though the exact scale remains reported rather than confirmed by the company in the source material.
Which companies make HBM?
The report names SK Hynix, Samsung and Micron as the three main suppliers. It says SK Hynix leads, Samsung is trying to regain share in 2026, and Micron is sold out for 2026.
What could ease the memory squeeze?
More HBM4 capacity, better yields and stronger supplier competition could ease pressure. The report says the main risk is that AI demand keeps absorbing capacity faster than memory makers can add it.
Source: Thorsten Meyer AI
