# HBM demand — X 热门讨论 (2026-09-25 09:06 UTC)
## @Taog_1575 (aog T) · 09-24 23:38 · ♥33 ↻4 💬2 Wow, I am not expected that.
From OIP conference interview, they said:
1. Customer demand is still 2-3x to the capacity TSMC can offer even on 2030.
They share a story: last year an AI giant CEO (not NV) told them to raise the capacity, and TSMC ask “okay, how much point?” and that CEO reply “not %, the unit they ask is fab number”.
TSMC do not believe it, but this changed by last year end. So this year, total 19 fab is constructed…. And more to come in next few years.
But even with this, they expected even those 19 fab and more fab to finish and joint the production in 2029-2030, they still far behind the demand…
They also point out that their fab construction speed is higher than Korean competitor by 1-2 year and US competitor by 2-4 year, this is another advantage factor.
2. Power will be the key factor for PPA
They said the electrical power may be the bottleneck and pricing issue for AI, so power in PPA will be higher priority for recent customer feedback.
Yield / capacity / power reduction will be the key.
3. Memory / chip bottleneck
Memory is the bottleneck, even worse than chip in this and next year.
So a smart customer redesign their product with lower HBM, so they can ship more product.
And they believe 3D DRAM will be important in future year.
But they also mentioned that start with 2028, the largest bottleneck will come back to foundry chip.
———————————
Takeaways:
We will see the demand still outpacing TSMC capacity even with this year’s crazy expansion.
So although we do not know Terafab will be success or not, but concern for chip shortage is true…
And other foundry like Samsung and Intel may gain the customer and order, it is just because even TSMC capacity is still far behind the market request…. https://x.com/Taog_1575/status/2103267824542286160