Former Intel CEO Pat Gelsinger dropped a sharp line at Hot Chips 2026 this week: "HBM is lousy." We think this is worth taking seriously — HBM accounts for nearly half of NVIDIA's high-end GPU bill of materials, making it the most expensive block in any AI chip, and SK Hynix is its top supplier.
What this is
HBM (High Bandwidth Memory) is the "stacked memory tower" sitting next to an AI chip — multiple DRAM (Dynamic Random-Access Memory) dies stacked and packaged together, letting GPUs read and write massive datasets at high speed. At the same event, an SK Hynix vice president added: HBM is not the final answer to the "memory wall" AI compute is running into.
Industry view
The skeptics argue HBM4 plans to stack 20 layers high, and every added layer dilutes bandwidth further — it is essentially trading height for throughput, a losing exchange. They prefer HBF (High Bandwidth Flash, a bandwidth solution built on NAND) — larger capacity, lower cost, potentially better long-term.
The pushback is just as firm: SK Hynix and Samsung are still aggressively expanding HBM3E and HBM4 lines, with record capex in 2026. One semiconductor analyst told us: "Take a retired CEO's comments as a reference — Hynix is voting against skepticism with real money poured into HBM." Also worth flagging: HBF is still at the concept stage, and whether it can match HBM's high-frequency characteristics remains unverified.
Impact on regular people
For enterprise IT: if the HBF path works out, AI compute rental prices per unit could drop meaningfully in 2–3 years, shortening payback for capital-heavy AI projects (like self-built data centers).
For individual careers: no direct impact on your work today, but enterprise AI tool subscription fees and internal procurement budgets have room to loosen over time.
For the consumer market: on-device AI (running on phones and PCs) is currently bottlenecked by memory capacity and bandwidth. If HBF matures, the prospect of running large models locally on laptops rises — though this is a 3–5 year story.