TL;DR
Thorsten Meyer AI’s latest report says high bandwidth memory, or HBM, has become the main pressure point in the 2026 memory shortage. The report attributes the squeeze to AI accelerator demand, HBM’s heavy wafer use, and supplier focus on higher-priced stacks over commodity memory.
High bandwidth memory has become the central component behind the 2026 memory squeeze, according to a new Thorsten Meyer AI report, as AI accelerator demand pulls wafer capacity away from DDR5 and GDDR7 and affects both memory prices and graphics card supply.
The report says HBM, once a specialty memory product, now helps determine the price and availability of much of the memory market. It cites AI chips from Nvidia and AMD as the main demand driver because modern accelerators need stacked memory positioned close to the compute die to feed data at very high bandwidth.
HBM is built as a vertical stack of eight, twelve or sixteen DRAM dies connected by through-silicon vias and mounted near a GPU or AI accelerator. Thorsten Meyer AI says this structure gives HBM roughly five to ten times the bandwidth of conventional graphics memory, but at a much higher manufacturing cost.
The report attributes the shortage to wafer economics. It says one bit of HBM can consume about three to four times the wafer area of one bit of DDR5, while a defect in one layer can spoil an entire stack. Pricing estimates cited in the report put HBM3 near $200 per stack, HBM3E near $300, and HBM4 around $500, making HBM a more attractive use of fab capacity than ordinary memory.
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 Sets Memory Prices
The report’s central point is that the shortage is not limited to data-center buyers. If suppliers allocate more wafers to HBM for AI systems, fewer wafers remain for DDR5 memory, servers, PCs and consumer devices. That can push prices higher across products that do not use HBM directly.
The effect has also reached graphics cards, according to the report. It says HBM prioritization helped tighten GDDR7 supply, the memory used in newer consumer GPUs, and says Nvidia reportedly cut RTX 50-series production by at least one-third in the first half of 2026. That specific production claim remains attributed to the report and its cited industry sources.
For readers, the practical impact is price and availability. More expensive AI memory stacks can crowd out cheaper commodity memory, while limited GPU memory supply can affect gaming cards, workstations and small AI developers that rely on consumer hardware.
Firepro S9300x2 Standard Airflow
- Model: FirePro S9300 x2
- Memory: 8 GB High Bandwidth Memory
- Interface: PCI Express x16 3.0
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Niche Part to Bottleneck
HBM’s rise tracks the rapid growth of AI accelerators. The report names Nvidia’s H100, H200, B200 and planned Rubin platform, along with AMD’s MI300 series, as chips built around high-bandwidth memory rather than ordinary DIMMs or standard graphics memory alone.
The generational cadence has moved quickly. The report lists HBM3 at about 819 GB/s per stack, HBM3E at about 1.18 TB/s, and HBM4 at about 2.8 TB/s with a new base logic die. It says HBM3E is the 2026 workhorse for platforms such as H200 and B200, while HBM4 is tied to next-generation AI systems.
Supplier concentration adds pressure. Thorsten Meyer AI cites SK Hynix as the leader with roughly 50% to 62% share, Samsung at about 28% to 40%, and Micron at around 5% to 10%. The report says all three have qualified for HBM4 as of June 2026, shifting attention from basic qualification to volume, yield and customer allocation.
“The thing the factories make instead of your RAM is a tower of stacked memory bolted to every AI chip.”
— Thorsten Meyer AI report
Supply Gains Are Still Unproven
Several points remain uncertain. Per-stack pricing for HBM3, HBM3E and HBM4 is described as estimated and point-in-time, and the report says conditions were moving quickly in late June 2026. Actual contract prices can vary by buyer, volume, generation and packaging needs.
It is also not yet clear how much added HBM4 supply Samsung, SK Hynix and Micron can bring online, or how quickly improved yields could ease pressure on DDR5 and GDDR7. The report frames supplier competition as the main relief valve, but does not confirm that added supply will arrive fast enough to change 2026 availability.
The demand side is another open question. The report says HBM is sold out through 2026, but also warns that if AI demand weakens, HBM could be the first segment to feel the correction because so much capacity and revenue growth is tied to one use case.
HBM4 Shipments Become the Test
The next milestone is the move from qualification to large-scale HBM4 shipments. If Samsung, SK Hynix and Micron can raise output and yields, the report suggests competition could add supply and reduce some pressure on the wider memory market.
Readers should watch HBM4 ramp updates, 2026 DRAM contract pricing, and any further signs of GPU production limits tied to memory availability. Thorsten Meyer AI says the next installment in the series will examine DDR5 now and DDR6 soon, shifting the focus from the AI memory stack to the mainstream memory products affected by it.
Key Questions
What is the actual development in this story?
Thorsten Meyer AI published a new analysis arguing that HBM is the main component driving the 2026 memory crunch, because AI demand is pulling wafer capacity away from ordinary memory.
Is HBM shortage confirmed or just a market claim?
The report states that HBM is sold out through 2026 and attributes that view to industry sources including TrendForce, DigiTimes and Reuters. Specific pricing and market-share figures should be read as reported estimates, not audited company results.
Why does HBM affect normal RAM prices?
HBM uses more wafer area per bit than DDR5 and is more difficult to manufacture. When memory makers use more capacity for higher-priced HBM stacks, less capacity is available for commodity DRAM.
How does this connect to GPUs?
The report says supplier focus on HBM for AI chips has also tightened GDDR7, the memory used by consumer graphics cards, contributing to reported RTX 50-series production cuts in early 2026.
Could the shortage ease soon?
It depends on HBM4 output, yields and AI demand. The report says all three major suppliers have qualified for HBM4, but it remains unclear how quickly added production can reach the market.
Source: Thorsten Meyer AI
