Who makes HBM3E for AI GPUs?

Only SK Hynix and Samsung produce HBM3E at scale. SK Hynix holds roughly 60% of the global HBM market. Micron is developing HBM but has not reached comparable volume.

Why is HBM a bottleneck for AI?

Every Nvidia H100 and H200 depends on HBM for memory bandwidth. With only two volume suppliers, any disruption or capacity limit in HBM production directly limits how many AI GPUs can be built and shipped.

How does HBM relate to the 2026 RAM and DRAM crisis?

HBM and standard DRAM compete for the same fab capacity at Samsung and SK Hynix. More HBM production means less DDR5 and similar memory for PCs and servers, which is one driver of the memory price spikes and shortage documented in the broader RAM crisis.

Can other companies start making HBM soon?

Micron is the main candidate to become a third HBM supplier but has not yet reached the volume of Samsung and SK Hynix. A meaningful change in the duopoly is unlikely in the next 12 to 24 months.

Hardware AI Tech Industry

SK Hynix and Samsung Are the Only HBM3E Suppliers. That Is the AI Bottleneck.

Abhishek Gautam·March 7, 2026·6 min read

Quick summary

SK Hynix holds ~60% of HBM and has sold out through 2026. Only two firms make HBM3E at scale. HBM market to hit $54-58B in 2026. What developers should do.

Every Nvidia H100 and H200 runs on High Bandwidth Memory (HBM). Only two companies in the world make HBM3E at scale: SK Hynix and Samsung. SK Hynix holds roughly 60% or more of the global HBM market and has sold out its DRAM, NAND, and HBM supply to Nvidia and other major customers through 2026. That duopoly is the AI memory bottleneck, and it is not going away soon.

What HBM Is and Why Only Two Companies Make It

HBM stacks DRAM dies vertically and connects them to the GPU over a wide, fast interface. It is the only memory type that can feed today's large AI chips without limiting throughput. HBM3 and HBM3E are the current workhorse generations; HBM4 is in development for the next Nvidia platform (Vera Rubin). The process is difficult, yields are tight, and the number of firms that can do it at volume is exactly two: SK Hynix and Samsung, both in South Korea. Micron is ramping HBM but has not reached the same volume; TrendForce and others still put the bulk of HBM3E supply with the two Korean suppliers.

So for the bulk of the world's AI training and inference, the memory supply depends on two Korean suppliers. Nvidia does not have a second source in the way it has multiple logic fabs; for HBM, it is SK Hynix and Samsung or nothing at scale.

The Numbers: Market Share, Sold-Out Capacity, and Growth

Roughly 60%: SK Hynix share of global HBM market (sources such as TrendForce; some estimates put it higher in HBM3E specifically)
70%: SK Hynix share of Nvidia initial HBM4 orders for the Vera Rubin platform (2026 to 2027)
Sold out: SK Hynix has committed its DRAM, NAND, and HBM capacity through 2026 to Nvidia and other major customers; it cannot take new 2026 orders in the way it used to
About two-thirds: share of total HBM shipments that will still be HBM3E in 2026 (HBM4 ramps later)
USD 38 billion to USD 54 to 58 billion: projected global HBM market growth from 2025 to 2026 (around 50% growth)
Samsung and Micron: Samsung is chasing HBM4 verification; Micron is the only plausible third supplier but at lower volume than the Korean duopoly

The HBM market is in a supercycle. AI cluster buildouts are driving demand far beyond what the two leaders can supply in the short term. That is why lead times for H100 and H200 have stayed long and why data center demand has helped push memory prices up and tightened supply for consumer and enterprise RAM. The HBM bottleneck and the broader RAM crisis are two sides of the same story: the same fabs make both HBM and DDR5, and HBM wins on margin.

Why There Is No Quick Fix

You cannot spin up HBM capacity in a year. The technology is hard, and qualification with Nvidia and AMD is strict. SK Hynix has locked in the lion's share of Nvidia HBM4 for Vera Rubin; Samsung is investing to close the gap. Micron will matter more in 2027 and beyond, but for 2026 the practical reality is two volume suppliers. Any disruption in Korea (natural disaster, geopolitical shock, or labour) would immediately tighten GPU availability worldwide.

What Developers and Infrastructure Teams Should Do

You cannot buy HBM directly. You buy GPUs. So the practical impact is on lead times, allocation, and pricing for H100, H200, and future Nvidia and AMD datacenter GPUs. If you are planning a large cluster, your procurement timeline should assume that HBM supply stays constrained through at least 2026. Lock in orders early, and do not assume spot availability or quick delivery. Diversification would require Micron to reach much higher volume or a shift to different memory architectures; neither is likely in the next 12 to 24 months.

Key Takeaways

Two: only SK Hynix and Samsung make HBM3E at volume; Micron is ramping but not at the same scale
Roughly 60%: SK Hynix share of global HBM; it has sold out capacity through 2026 to Nvidia and others
70%: SK Hynix share of Nvidia HBM4 orders for Vera Rubin (2026 to 2027)
HBM market: projected to grow from about USD 38B (2025) to USD 54 to 58B (2026)
For developers: Plan for continued HBM-driven constraints on GPU availability and lead times; the bottleneck is structural; lock orders early and do not count on spot supply
What to watch: Micron HBM3E ramp, Samsung HBM4 verification, and Nvidia Vera Rubin HBM4 rollout in 2026 to 2027