The HBM4 Supply Race: Who Gets to Feed NVIDIA

The most important contest in semiconductors right now isn’t about logic chips — it’s about memory. HBM4, the sixth generation of high-bandwidth memory, is entering mass production in 2026, and the three companies that can make it are racing to lock in the customer that matters most: NVIDIA.

Why HBM4 is the prize

Every NVIDIA AI accelerator is paired with a stack of HBM, and the next-generation Rubin platform is built around HBM4. Winning a slice of NVIDIA’s orders means years of high-margin revenue. That’s why this round of the memory supercycle has turned into a head-to-head supply race rather than a quiet commodity business.

HBM4 isn’t a minor bump. SK Hynix says its design delivers roughly 40% better power efficiency, and Samsung has shown modules pushing past the 10 Gbps mark to around 11.7 Gbps. At data-center scale, those gains compound into real savings on power and performance.

The three players

Just three firms make HBM, and they entered 2026 pulling their schedules forward to early in the year:

• SK Hynix is the incumbent leader, holding roughly 62% of the HBM market, and is reported to supply about two-thirds of NVIDIA’s HBM4.
• Samsung is pushing hard to catch up — accelerating mass production at its Pyeongtaek campus, targeting early delivery, and reportedly planning a 50% expansion of HBM capacity this year.
• Micron has emerged as the third force, overtaking Samsung on some metrics and parlaying its momentum into deals like its strategic partnership with Anthropic.

Together SK Hynix and Samsung control about 90% of HBM supply; add Micron and you have essentially the entire market.

What’s at stake

For investors, HBM4 share is the cleanest read on who captures the AI memory windfall. Allocation to NVIDIA’s 16-high HBM4 stacks is, in effect, a multi-year revenue guarantee — and the reason memory names have been so volatile as the market handicaps the winners. It’s a big part of why Micron’s stock keeps swinging.

It also raises the stakes for NVIDIA’s rivals. AMD’s MI400 accelerators lean on HBM4 too, which means the memory makers are selling into both sides of the AI-chip war — a comfortable position to be in.

The takeaway

HBM4 is the bottleneck inside the bottleneck. SK Hynix leads, Samsung is spending to close the gap, and Micron has turned itself into a credible third supplier. Whoever wins the largest, most reliable share of NVIDIA’s HBM4 orders captures the richest vein of the AI buildout — which is why a memory generation most consumers will never see is one of the defining competitions of 2026.