Every top AI accelerator needs HBM, and only three companies make it at scale. Here's where each stands on HBM4, using what they've said themselves, plus one big analyst estimate.
New to HBM? Start with HBM explained or GDDR7 vs GDDR6 vs HBM.
The standard
JEDEC's HBM4 standard (Apr 16, 2025) specifies a 2048-bit interface at up to 8Gb/s, for up to 2TB/s per stack. Vendors can and do go faster.
Samsung
On Feb 12, 2026, Samsung said it had begun HBM4 mass production and shipped commercial products, calling it an industry first. Its claims:
- 11.7Gbps per pin consistently, up to 13Gbps
- Up to 3.3TB/s per stack
- 24GB to 36GB with 12-layer stacks; up to 48GB with 16 layers later
- Built on its 1c DRAM process with a 4nm logic base die
- Expected HBM sales to more than triple in 2026 vs 2025
- HBM4E sampling expected in 2H26
SK hynix
In its Q2 2026 results (Jul 29), SK hynix said:
- It began mass shipments of HBM4 in Q2 and would ramp production in 2H26.
- HBM4 hit customer-required speeds with what it called industry-leading power efficiency.
- HBM4E sample shipments were completed in the first half.
- It has long-term agreements with around 10 customers.
Q2 revenue was 79.3187 trillion won, with 60.5426 trillion won of operating profit. Its Q3 report is set for Oct 28; see our SK hynix Q3 preview.
Micron
On Micron's Sept 30, 2026 earnings call, per the Motley Fool transcript:
- CTO Scott DeBoer said Micron has worked with NVIDIA for over a year on a co-designed HBM4E it called "NV HBM," describing it as "the first major custom HBM product out in the market."
- Management said HBM pricing was increased significantly for calendar 2027.
- On share, management said Micron expects HBM share to be around its broader DRAM share, without a fixed target.
- It's preparing to ramp an HBM facility in Singapore next year.
Micron's full results are in Micron's record FQ4.
The Bernstein estimate (reported, not confirmed)
As reported by Investing.com on Oct 22, 2026, Bernstein used Korean regional export data to estimate Q3 2026 HBM revenue:
- Samsung: about $8.7B, up 33% QoQ (10% below Bernstein's earlier forecast)
- SK hynix: about $7.6B, up 11% QoQ
Bernstein reads that as Samsung passing SK hynix, and its "value per weight" metric suggests a Samsung lead in HBM4 pricing. This is a model based on export data, not company-reported numbers. SK hynix reports Q3 on Oct 28, which will be the first real check.
Why it matters
HBM supply sets how many AI accelerators can ship. TrendForce expects HBM to stay supply-constrained in 2027, with blended prices projected up 121% YoY. Some chip designers are weighing 8-high stacks to stretch supply. That's the backdrop for why RAM is so expensive for everyone else.
