Pat Gelsinger, Intel's former chief executive, has called high-bandwidth memory (HBM — stacked DRAM modules placed next to a processor to feed it data at very high speed) "terrible memory" in a public remark now circulating across semiconductor trade coverage.

The sting comes with a caveat. Gelsinger made clear that despite his harsh assessment, HBM remains the best memory option available for artificial intelligence workloads today, according to the report carrying his comments.

What did Gelsinger actually say?

Two claims, and they sit in tension. First: HBM is, in his words, "terrible memory." Second: it's still the best option for AI. That pairing is the story. A veteran semiconductor executive — a man who ran the company that spent years trying to break into the AI accelerator market — doesn't mince words about the technology, yet concedes the industry has nothing better to deploy.

Why would an engineer call a memory technology terrible while simultaneously endorsing it? The trade-press answer is well understood in the field, and Gelsinger's framing maps onto it. HBM is expensive per bit compared with conventional DRAM. Its manufacturing complexity is high. Yield challenges have constrained supply. Yet AI accelerators are ravenous for bandwidth — the rate at which data moves between memory and compute — and HBM delivers bandwidth that standard DDR memory (the type in conventional servers) cannot approach.

So the verdict lands as a statement about constraints, not preferences. Terrible on cost and engineering terms. Unbeatable on the metric that AI hardware actually needs.

What does this change for the memory market?

In the near term, nothing structurally. Gelsinger left Intel in December 2024, and his remarks carry no policy weight. But they sharpen a debate that vendors and buyers are already having.

For buyers of AI hardware, the comment is a reminder that HBM's dominance reflects a lack of alternatives rather than technical satisfaction. For memory suppliers, it signals that the door remains open for a successor — whether that's a new packaging approach, a different memory architecture, or incremental HBM generations that attack the cost side of the complaint.

Gelsinger's position also carries a particular weight given his history. Under his leadership, Intel pursued its own AI accelerator strategy and confronted exactly the memory-bandwidth bottleneck he's describing. A critic who spent years inside the problem is harder to dismiss than an outside analyst.

Why the quote travels

The remark is quotable because it compresses an industry-wide unease into five words. Every AI accelerator shipped today — from every major vendor — depends on HBM sitting beside the compute silicon. If the former head of the world's one-time semiconductor leader calls that foundation "terrible," procurement teams and architects will repeat it.

The second half of his sentence, though, is the practical takeaway: it's still the best option. Builders who need AI systems now have no decision to reconsider. The quote is a flag planted for whoever eventually builds the memory that replaces it.