TL;DR
High Bandwidth Memory has become the pressure point in the 2026 memory crunch, according to Thorsten Meyer AI’s latest series installment. The report says HBM now consumes far more wafer capacity than standard DDR5 while AI accelerator demand keeps supplies sold out through 2026.
High Bandwidth Memory has become the component driving much of the 2026 memory supply squeeze, according to a new Thorsten Meyer AI report, because AI chip demand is pulling wafer capacity away from standard DDR5 and some graphics memory used in consumer GPUs.
The report says HBM, once a niche specialty memory product, now sits beside leading AI accelerators from Nvidia and AMD and supplies the bandwidth those chips need for training and inference workloads. A modern AI GPU uses multiple HBM stacks, often around eight per accelerator, rather than a single memory package.
Unlike standard DRAM modules, HBM is built as a vertical stack of eight, twelve, or sixteen DRAM dies connected through through-silicon vias. That design delivers far more bandwidth than ordinary graphics memory, but the report says it is far less efficient to manufacture: one bit of HBM can consume roughly three to four times the wafer area of one bit of DDR5.
Thorsten Meyer AI, citing Silicon Analysts, Introl, TrendForce, DigiTimes, Unibetter, Astute Group and Reuters, says the economics have pushed suppliers to favor HBM. The report estimates per-stack pricing at about $200 for HBM3, around $300 for HBM3E, and roughly $500 for HBM4, while noting that pricing is point-in-time and subject to change.
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.
HBM Now Sets Memory Prices
The shift matters because HBM competes for the same advanced DRAM production resources that would otherwise support commodity memory. If suppliers commit more wafer starts to HBM, the report argues, fewer wafers remain available for DDR5 and related products used in PCs, servers and other systems.
The impact is not limited to system memory. The report says HBM demand has also contributed to pressure on GDDR7, the graphics memory used in newer consumer cards. It cites reports that Nvidia cut RTX 50-series production by a third or more in the first half of 2026, though that figure remains attributed to external reporting rather than confirmed directly by Nvidia in the source material.
For readers, the practical effect is that AI infrastructure spending may be influencing RAM prices, GPU availability and the cost structure of hardware far beyond data centers. The report frames this as a structural supply issue, not simply short-term scarcity.
Firepro S9300x2 Standard Airflow
Model: FirePro S9300 x2 – Detailed description of the product's model name
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
AI Accelerators Need Bandwidth
HBM’s rise follows the bandwidth demands of modern AI accelerators. The report says AI chips can perform calculations faster than ordinary memory can feed them data, making memory bandwidth a limiting factor for training and inference performance.
That is why chips such as Nvidia’s H100, H200 and B200, the expected Rubin platform, and AMD’s MI300-series rely on HBM. The report lists HBM3 at roughly 819 GB/s per stack, HBM3E at about 1.18 TB/s, and HBM4 at an estimated 2.8 TB/s, based on JEDEC and vendor specifications cited in the source material.
The supplier race is concentrated among SK Hynix, Samsung and Micron. Thorsten Meyer AI says SK Hynix leads with about 50% to 62% share, Samsung holds roughly 28% to 40%, and Micron has about 5% to 10%. The report says all three had qualified for HBM4 by June 2026, shifting the question from qualification to manufacturing scale and quality.
“The thing the factories make instead of your RAM is a tower of stacked memory bolted to every AI chip.”
— Thorsten Meyer AI
HBM3 memory modules
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Supply Gains Remain Unproven
Several points remain unsettled. The report says HBM is sold out through 2026, but the scale and timing of added capacity from SK Hynix, Samsung and Micron are still developing. It is also not yet clear how quickly HBM4 yields will improve or how much usable supply new qualification wins will produce.
Some figures in the report are explicitly described as estimates, including per-stack pricing and some market-share ranges. The reported RTX 50-series production cut is also attributed to outside reporting, and Nvidia’s direct confirmation is not provided in the source material.
ASUS ROG Strix GeForce RTX 4090 OC Edition Gaming Graphics Card (PCIe 4.0, 24GB GDDR6X, HDMI 2.1a, DisplayPort 1.4a), 3 Year Warranty
NVIDIA Ada Lovelace Streaming Multiprocessors: Up to 2x performance and power efficiency
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
HBM4 Race Moves To Volume
The next phase is production execution. With all three major suppliers reported as qualified for HBM4, buyers will watch which company can ship high-volume stacks with strong yields, stable pricing and enough supply for the next wave of AI accelerators.
The report says the upside is that broader supplier competition could add supply. The risk is that if AI demand weakens, HBM may be the first part of the memory market to face a correction because so much capacity and pricing power has gathered around one fast-growing product category.
Biwin RS200 DDR5 RAM 16GB 5600MHz, SODIMM 262-pin CL46, Laptop Computer Memory, Mini PC, 1.1V / Independent PMIC/On-die ECC, High Performance RAM for AI or Gaming Computer, Single
Next-Gen DDR5 Performance: BIWIN DDR5 SODIMM 16GB experiences double the bandwidth of DDR4, with speeds up to 5600MT/s—perfect…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is High Bandwidth Memory?
High Bandwidth Memory is a stacked form of DRAM placed close to an AI accelerator or GPU. It uses vertical connections called through-silicon vias to move data far faster than standard memory designs.
Why does HBM affect ordinary RAM prices?
The report says HBM uses far more wafer area per bit than DDR5 and is more profitable. When memory makers allocate more capacity to HBM stacks, less capacity is available for standard DRAM products.
Which companies make HBM?
The main suppliers are SK Hynix, Samsung and Micron. Thorsten Meyer AI says SK Hynix currently leads the market, while Samsung and Micron are pushing for stronger positions in HBM4.
Is the shortage confirmed to affect consumer GPUs?
The source material says supplier focus on HBM has contributed to tighter GDDR7 supply and cites reports of reduced RTX 50-series production. That production-cut claim is attributed to reporting cited by Thorsten Meyer AI and is not presented as directly confirmed by Nvidia in the provided material.
When could the pressure ease?
The report points to HBM4 production as the next possible relief valve, but timing depends on supplier yields, capacity additions and continuing AI chip demand. As described, the market remains sold out through 2026.
Source: Thorsten Meyer AI
