TL;DR
Thorsten Meyer AI’s latest Memory Squeeze report says high-bandwidth memory has become the key pressure point in the 2026 memory crunch. The report attributes higher RAM prices and tighter GPU memory supply to AI accelerator demand, HBM’s heavy wafer use, and limited capacity among SK Hynix, Samsung and Micron.
High-bandwidth memory has become the central pressure point in the 2026 memory crunch, according to a new Thorsten Meyer AI report that says AI accelerator demand is pulling fab capacity away from ordinary RAM and tightening parts of the graphics card market.
The report identifies HBM, a stacked DRAM technology used beside AI GPUs, as the component now shaping availability and pricing across much of the memory market. It says HBM moved in roughly three years from a specialty part to a product that influences the price of DRAM and, increasingly, the availability of consumer GPU memory.
HBM differs from standard DDR5 because it is built as a vertical stack of eight, twelve or sixteen DRAM dies connected by through-silicon vias and mounted close to the GPU. The report says this design gives AI accelerators far more bandwidth than ordinary graphics memory, but also uses far more manufacturing capacity. It estimates that one HBM bit consumes roughly three to four times the wafer area of a DDR5 bit.
Thorsten Meyer AI says the supply race is concentrated among SK Hynix, Samsung and Micron. The report estimates SK Hynix holds about 50% to 62% of the HBM market, Samsung about 28% to 40%, and Micron about 5% to 10%. It also says all three suppliers had qualified for HBM4 by June 2026, shifting the market question from basic availability to execution, yield and volume.
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 report matters because it links higher memory prices to a specific manufacturing tradeoff: wafers used for AI-focused HBM are wafers not used for ordinary DDR5. If HBM uses several times the wafer area per bit, even a strong increase in HBM output can remove a much larger amount of standard memory capacity from the market.
The economics described in the report help explain why suppliers are prioritizing HBM. Thorsten Meyer AI estimates HBM3 stacks at about $200, HBM3E at about $300, and HBM4 at roughly $500 per stack. The report says Samsung and SK Hynix raised HBM3E prices by about 20% for 2026 while demand still exceeded supply.
The pressure is not limited to server memory. The report says supplier focus on HBM also tightened GDDR7, the memory used in consumer graphics cards, and cites reports that Nvidia cut RTX 50-series production by a third or more in the first half of 2026. That production claim remains attributed to reporting cited by Thorsten Meyer AI.
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.
How HBM Became Central
HBM is used because modern AI accelerators are often memory-bandwidth-bound: the processor can perform calculations faster than standard memory can feed it data. By placing multiple HBM stacks close to the compute die, chips such as Nvidia H100, H200, B200 and future Rubin parts can access data at far higher rates.
The report describes an annual product race. HBM3 delivered around 819 GB/s per stack in the H100 generation. HBM3E moved to about 1.18 TB/s per stack and is described as the 2026 workhorse for parts such as H200 and B200. HBM4, tied in the report to Nvidia’s Rubin generation, is listed at about 2.8 TB/s per stack with a new logic base die.
Thorsten Meyer AI cites Silicon Analysts, Introl, TrendForce, DigiTimes, Unibetter, Astute Group and Reuters as sources and says pricing figures are estimated or point-in-time. The report says the HBM market is estimated at about $35 billion now and could approach $100 billion by 2028, representing about 41% of DRAM revenue, up from 8% in 2023.
“The thing the factories make instead of your RAM is a tower of stacked memory bolted to every AI chip.”
— Thorsten Meyer AI
CORSAIR Vengeance RGB DDR5 RAM 32GB (2x16GB) Up to 6000MHz CL36-44-44-96 1.35V Intel XMP 3.0 Computer Memory – Black (CMH32GX5M2E6000C36)
Disclaimer: Maximum Speed requires overclocking/PC BIOS adjustments. Maximum speed and performance depend on system components, including motherboard and…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Capacity Risks Still Unsettled
Several details remain uncertain. The report’s per-stack prices for HBM3, HBM3E and HBM4 are identified as estimates, not fixed market prices. The cited claim that Nvidia cut RTX 50-series production by at least a third is also reported, not independently confirmed in the source material.
It is also not clear how quickly new HBM4 capacity can reach stable yields, or whether added supply from Samsung and Micron will meaningfully relieve pressure before the end of 2026. The report says the market is sold out through 2026, but supply conditions can change if AI orders are delayed, reduced or reallocated.
Waveshare Jetson Orin NX AI Development Kit for Embedded and Edge Systems 8GB Memory Memory Jetson Orin NX Module (5 Items)
This package include a Orin NX development kit with 8GB Jetson Orin NX Module, and other accessories, 5…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
HBM4 Race Sets 2027
The next milestone is the ramp of HBM4 for upcoming AI accelerators, including parts tied to Nvidia Rubin. If all three major suppliers can increase output with strong yields, the report says competition could add supply and ease some pressure.
The main risk runs in the opposite direction: if AI demand weakens after suppliers commit more capacity to HBM, the market segment that rose fastest could be the first to show stress. Thorsten Meyer AI says the next article in the series will examine DDR5 and DDR6, where the effects of the HBM shift are expected to show up for mainstream buyers.
ASUS Dual NVIDIA GeForce RTX 3050 6GB GDDR6 OC Edition Gaming Graphics Card – PCIe 4.0, HDMI 2.1, DisplayPort 1.4a, 2-Slot Design, Axial-tech Fan Design, Steel Bracket, 3 Year Warranty
NVIDIA Ampere Streaming Multiprocessors: The all-new Ampere SM brings 2X the FP32 throughput and improved power efficiency.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the news development in this report?
Thorsten Meyer AI published a new installment in its 2026 memory crunch series arguing that HBM demand for AI chips is now a major driver of memory prices, supply limits and some GPU availability issues.
What is HBM used for?
High-bandwidth memory is stacked DRAM placed close to an AI accelerator or GPU. It gives chips much faster access to data, which matters because AI workloads often wait on memory bandwidth rather than compute power.
Why would HBM affect normal RAM prices?
The report says HBM consumes more wafer area than DDR5 and has harder yield requirements. When memory makers allocate more fab capacity to higher-priced HBM, less capacity is available for ordinary DRAM.
Who makes most of the world’s HBM?
The report identifies SK Hynix, Samsung and Micron as the three main suppliers. It estimates SK Hynix remains the leader, with Samsung and Micron competing to expand share in HBM4.
What remains unknown for buyers?
It is still unclear how fast HBM4 supply can ramp, whether DDR5 prices will ease in 2026, and how much consumer GPU availability has been affected by HBM-related capacity choices.
Source: Thorsten Meyer AI
