HBM Memory: The Quiet Bottleneck Slowing AI Development

What is HBM and why can't standard DRAM replace it?

High Bandwidth Memory (HBM) is a stacked DRAM architecture that delivers dramatically higher memory bandwidth than conventional DRAM by connecting memory dies directly to the processor die through thousands of micro-scale copper pillars — a technique called through-silicon vias (TSVs) and Chip-on-Wafer (CoW) bonding. A single HBM3e stack can transfer over 1.2 terabytes of data per second. Standard DDR5 DRAM tops out at roughly 80 GB/s per module. That is a 15x bandwidth difference.

This gap matters profoundly for AI. Training a large language model requires moving enormous amounts of weight data between memory and compute at every forward and backward pass. The GPU's compute units are so fast that they are almost always waiting for data — the bottleneck is memory bandwidth, not arithmetic throughput. HBM eliminates most of that wait. Without HBM, training a 70-billion parameter model at acceptable speed and cost is not feasible on current GPU architectures.

HBM also has a unique physical integration: it is bonded directly to the GPU die on a silicon interposer, forming a single module (the GPU package). This means HBM cannot be swapped post-sale. Each GPU package requires a fixed allocation of HBM stacks determined at fabrication. NVIDIA's H100 requires 6 HBM3 stacks. The H200 requires 8 HBM3e stacks. The B200 (Blackwell) requires 8 HBM3e stacks delivering 8 TB/s aggregate bandwidth. No HBM, no GPU — the dependency is structural.

Who makes HBM — and why only three companies can?

As of June 2026, three companies manufacture HBM at commercial scale: SK Hynix, Samsung Memory, and Micron Technology. Each took a different path to the technology, and each holds a different market share.

SK Hynix (000660.KS: ₩2,160,000, 52-week high ₩2,407,000 on June 2) was the first mover and remains the market leader, commanding approximately 50–60% of global HBM supply. SK Hynix developed the CoW bonding technology first, built the tooling, and accumulated the yield-improvement knowledge that makes mass production viable. Its HBM3e product is the standard that NVIDIA qualifies GPUs against. The company's stock hit a record high on June 2, 2026, reflecting market recognition that HBM supply will remain constrained relative to AI demand for years.

Samsung Memory (005930.KS: ₩351,500, all-time high ₩370,000 on June 2) holds approximately 25–30% of HBM supply and is aggressively ramping HBM4 production to close the gap with SK Hynix. Samsung's HBM ramp has been slower than expected partly because of yield challenges in CoW bonding — a process that bonds dies with sub-micron precision and is extraordinarily sensitive to thermal and mechanical variation.

Micron Technology (US-listed) holds approximately 10–15% and is expanding HBM capacity at its US and Japanese fabs. Micron's geographic diversification gives it supply chain resilience that SK Hynix and Samsung — both concentrated in South Korea — do not have.

The reason only these three can manufacture HBM commercially: CoW packaging requires $10–20 billion in specialized tooling investment, 5–10 years of process learning to achieve acceptable yields, a supply chain of micro-scale copper pillar materials and precision substrates that itself takes years to qualify, and the institutional knowledge accumulated through thousands of failed production runs. There is no shortcut. Memory companies that attempted to enter the HBM market — including some Chinese DRAM producers — have not achieved commercial yields.

How the Iran war made Korea's HBM supply chain more fragile overnight

Korea's HBM duopoly — SK Hynix and Samsung Memory — sits at the center of global AI hardware supply. The two companies together control 80% of HBM and nearly 70% of global DRAM. This makes Korea's supply chain stability a matter of global AI concern.

The Iran war exposed three simultaneous vulnerabilities.

Energy: South Korea imports approximately 70% of its crude oil through the Strait of Hormuz. The effective closure of the strait beginning March 4, 2026 threatened the energy supply for Korean semiconductor fabs, which run 24/7 and require extremely stable power. While Korea has LNG reserves and a mixed energy grid, prolonged closure creates both cost pressure (energy prices spiked) and availability risk. Samsung and SK Hynix saw their combined market cap decline by over $60 billion in two days before partially recovering.

Bromine: Korea sources approximately 90% of its bromine from Israel. Bromine enters the semiconductor supply chain through specialty photoresist chemicals and flame retardants used in printed circuit boards and packaging materials. Israel and Jordan together supply two-thirds of global bromine. The Iran war — which drew Israel into active military operations and disrupted Middle Eastern shipping — placed this supply under direct geopolitical stress. While a full bromine shortage has not yet materialized, the supply chain is operating on reduced buffer inventory and elevated prices.

Helium: HBM's CoW packaging process uses helium for both leak detection in vacuum bonding equipment and as a carrier gas during the ultra-precise die attachment steps. Qatar's Ras Laffan — offline since early March — supplied roughly 35% of global semiconductor-grade helium. SK Hynix and Samsung each hold weeks, not months, of helium buffer inventory. If the disruption persists through Q3 2026, HBM production schedules will face direct capacity pressure.

Why HBM supply continues to lag AI demand — and when it might close

The HBM shortage is structural, not cyclical. It is driven by two compounding forces that are both accelerating.

Demand is exponential. Every major AI model generation requires more parameters, which require more GPUs, which require more HBM. NVIDIA's B200 (Blackwell) requires 8 HBM3e stacks per GPU. The next-generation Rubin architecture is expected to require even more. Cloud providers — AWS, Azure, Google Cloud, and challengers like CoreWeave and xAI — are all ordering at the limits of NVIDIA's production capacity, which is itself limited by HBM availability. TSMC CEO C.C. Wei stated on June 4, 2026 that chip supply will continue to lag AI demand for years. The HBM tier is a primary reason why.

Supply expansion is slow. A new HBM production line requires a 2–4 year investment cycle from capital commitment to first commercial output. SK Hynix has announced major capacity expansions, and Samsung is ramping HBM4. Micron is building HBM capacity at its US fab in Boise and in Japan. But even aggressive investment cannot close the gap in the near term.

The Iran war has added a near-term capacity risk. If helium disruption extends through Q3 2026, SK Hynix and Samsung may need to allocate available helium selectively across product lines — potentially prioritizing leading-edge HBM4 over HBM3e maintenance production. Bromine shortages add friction to the photoresist supply chain that feeds into wafer preparation.

Best-case scenario: HBM supply catches up to demand sometime in 2027–2028 as Micron's US capacity and Samsung's HBM4 ramp contribute meaningfully. The Iran war and helium shock may push this to 2028–2029. For investors in SK Hynix and Samsung: the shortage is a structural tailwind, but the geopolitical exposure is a structural risk that cannot be diversified away without geographic relocation of production.

Track HBM supply chain exposure with AIChipMap

AIChipMap maps SK Hynix, Samsung Memory, and Micron's positions in the full supply chain — showing their upstream material dependencies (silicon wafers from Shin-Etsu and SUMCO, specialty chemicals from JSR and Sumitomo Chemical, packaging substrates from Ibiden and Shinko Electric) and their downstream customers (NVIDIA, AMD, cloud providers).

The trace view for SK Hynix shows exactly which companies depend on its HBM output — and by extension, which are exposed if HBM supply tightens further. The company page shows SK Hynix's export control tier, supply risk rating, and its position as a chokepoint in the HBM memory layer.

For Korean investors following SK Hynix (₩2,160,000) and Samsung (₩351,500): use the trace view to see which of their Japanese material suppliers — Shin-Etsu (¥7,641), SUMCO, JSR — are themselves subject to supply pressures from China's material export controls or the Iran war's helium and bromine disruptions. The compound exposure visible in the graph is not captured by any single company's earnings report.

For researchers and analysts tracking the HBM shortage: AIChipMap provides a visual model of the dependency chain from wafer substrate to deployed AI model — so you can see which tier, in which country, is the current binding constraint on GPU production volumes.