# HBM demand — X 热门讨论 (2026-09-25 05:49 UTC)
## @imnotharsh (ImNotHarsh | 📈💸) · 09-24 21:57 · ♥31 ↻3 💬1 Compute is worthless if you can’t fab enough chips. Compute without silicon is nothing.
The bottleneck is in the building.
Demand is exponential for energy and memory. But, the supply - for either of those to matter - is constrained. For the US, there is one process owner already running leading-edge logic and High-NA in volume. This is not disregarding TSMC, but it is the difference between a US company with IP that owns the node and a Taiwanese company building capacity on US soil.
There is a reason TSMC is finally putting some respect on Intel’s name. Go listen to TSM’s latest earnings call with C.C. Wei’s answer to Intel as competition. Then listen to what he said 2 quarters ago. He laughed Intel off. Now, not so much.
Why do you think yields matter? Yields are the part everyone treats as a rounding error.
The binding constraint on an accelerator includes advanced packaging first and foremost, alongside the front-end wafer.
Cannot build quality chips fast enough. Only one company can build its own designs and also cater to the ASICs. It is the reason $GOOGL Google (for TPU) tapped Intel. It is the reason $AMZN AWS (for Trainium) has tapped Intel. It is the reason $MSFT Microsoft has tapped Intel. It is the reason there are talks of Anthropic tapping Intel.
The building is the long-cycle constraint, yields are the constraint inside the building, and packaging and HBM are why a finished wafer still is not a shipped accelerator.
$INTC. https://x.com/imnotharsh/status/2103242563922821495