TL;DR
Thorsten Meyer AI reports that High Bandwidth Memory has become the main pressure point in the 2026 memory squeeze, as AI accelerators pull wafer capacity away from DDR5 and GDDR7. The report says HBM is sold out through 2026, with HBM4 supplier competition now shaping future supply and pricing.
High Bandwidth Memory has become the pressure point in the 2026 memory squeeze, with a new Thorsten Meyer AI dispatch reporting that AI-chip demand is pulling wafer capacity away from DDR5 RAM and contributing to short supplies of GDDR7 graphics memory.
HBM is stacked DRAM mounted next to advanced AI processors, not ordinary memory placed on a motherboard stick. The dispatch describes stacks of eight, twelve or sixteen DRAM dies linked by through-silicon vias, allowing the memory to feed GPUs at roughly five to ten times the bandwidth of conventional graphics memory.
The manufacturing tradeoff is the core of the report. According to the dispatch, one bit of HBM can consume roughly three to four times the wafer area of one bit of DDR5, because the dies are larger, stacking is harder and one defect can spoil a full stack. The report says the economics favor HBM: estimated pricing runs from about $200 per HBM3 stack to about $300 for HBM3E and an estimated $500 for HBM4.
The report says SK Hynix, Samsung and Micron are the three main suppliers in the race, with SK Hynix leading current HBM share and Micron already sold out for 2026. It also says all three have qualified for HBM4 as of June 2026. A separate point remains attributed to supplier and industry reporting: Nvidia reportedly cut RTX 50-series production by a third or more in the first half of 2026 as GDDR7 supply tightened.
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 Orders Reprice Memory Supply
The report matters because HBM demand no longer affects only data-center buyers. If memory makers move more wafers into AI accelerator stacks, fewer wafers are available for PC RAM, server DIMMs and consumer graphics memory, which can raise prices or limit availability across markets.
Thorsten Meyer AI frames the squeeze as structural rather than a temporary ordering mistake. The dispatch says the HBM market has moved from about $35 billion toward a projected $100 billion by 2028, equal to roughly 41% of DRAM revenue, up from about 8% in 2023. Those figures are attributed to the cited analyst and industry sources and remain projections, not guaranteed outcomes.
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.
HBM4 Widens Supplier Race
The shift has accelerated with each AI accelerator generation. The dispatch places HBM3 in the Nvidia H100 era at about 819 GB/s per stack, HBM3E around the H200 and B200 generation at about 1.18 TB/s, and HBM4 near the coming Rubin platform at an estimated 2.8 TB/s.
That speed is needed because AI training and inference are constrained by how quickly data reaches the compute die. Nvidia’s H100, H200 and B200, AMD’s MI300-series and future high-end accelerators rely on multiple HBM stacks around the main chip. The dispatch says a modern AI GPU can use about eight HBM stacks, which magnifies the pressure on wafer starts.
“The thing the factories make instead of your RAM is a tower of stacked memory bolted to every AI chip.”
— Thorsten Meyer AI dispatch
TEAMGROUP T-Create Expert CL30 Overclocking 10L DDR5 32GB Kit (2 x 16GB) 6000MHz (PC5-48000) Intel XMP 3.0 & AMD EXPO Compatible Desktop Memory Module Ram Black – CTCED532G6000HC30DC01
TEAMGROUP T-CREATE EXPERT 32GB KIT 2 X 16GB DDR5-6000 PC5-48000 CL30 DUAL CH
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Capacity Gains Still Unproven
Several details remain unsettled. The exact supplier share ranges, HBM4 yields, customer allocations and the scale of RTX 50-series production cuts have not been confirmed directly by every company named in the dispatch.
It is also unclear how much new HBM4 supply will relieve the broader market. If AI accelerator demand keeps rising, added capacity could be absorbed quickly. If demand weakens, HBM pricing and memory-maker capital plans could face pressure first.
PNY NVIDIA GeForce RTX™ 5080 Epic-X™ ARGB OC Triple Fan, Graphics Card (16GB GDDR7, 256-bit, Boost Speed: 2775 MHz, PCIe® 5.0, HDMI®/DP 2.1, 2.99-Slot, NVIDIA Blackwell Architecture, DLSS 4)
NVIDIA DLSS 4 – Supreme Speed. Superior Visuals. Powered by AI. DLSS is a revolutionary suite of neural…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
HBM4 Shipments Set The Pace
The next markers are HBM4 ramp quality, supplier yield updates and contract pricing for DDR5 and GDDR7 through the second half of 2026. Buyers should also watch Nvidia and AMD accelerator schedules, because AI GPU volume will shape how much memory capacity remains for PCs, servers and graphics cards.
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.
Key Questions
What does ‘HBM ate the fab’ mean?
It means memory manufacturers are directing more wafer capacity toward High Bandwidth Memory because AI-chip customers need it and pay higher prices for it.
Is HBM the same as normal RAM?
No. DDR5 RAM is typically laid out on modules for computers and servers. HBM is a vertical memory stack mounted close to a GPU or AI accelerator to provide far higher bandwidth.
Why would AI chips affect PC memory prices?
The report says HBM uses far more wafer area per bit than DDR5. When fabs prioritize AI memory stacks, less capacity may be available for ordinary DRAM, which can lift prices.
Did HBM cause RTX 50-series shortages?
The dispatch says GDDR7 shortages contributed to a reported RTX 50-series production cut in early 2026. The exact scale remains attributed to industry reports rather than direct company confirmation.
Could HBM4 ease the memory squeeze?
Possibly, if SK Hynix, Samsung and Micron ramp HBM4 with strong yields. The risk is that AI demand absorbs the added supply before it eases pressure on other memory products.
Source: Thorsten Meyer AI
