By: Mark Roulo (nothanks.delete@this.xxx.com), September 22, 2021 12:37 pm
Room: Moderated Discussions
Moritz (better.delete@this.not.tell) on September 22, 2021 6:40 am wrote:
> For 10 months now, Nvidia GPUs (GTX16x0 Turing) have not been available
...
> I understand that Covid could not have been planed for and caused supply chain
> disturbance, but this is not disturbance but market and industry failure.
> Looking at the (mm)² sold, I remember a consumer die to be 150(mm)² 20 years ago. Today
> a mid-range GPU is three times as big at 450(mm)². Could no one predict that it would take
> more of the same 300mm wafers to make the same number of three times bigger dies?
> Because Nvidia is effectively a monopolist (No other product is sufficiently identical) they would
> set the price rather high and would not sell at maximum sourcable capacity. Even without Nvidia being
> pushy about production, they must be at a point where selling more would increase their profit.
TSMC (and other foundries) are capacity constrained. If you want a new fab today, you needed to start building it (and placing orders for equipment) two or three years ago.
But (leading edge) fabs have two very bad properties:
The upshot is that no one wants to build a $10B+ fab without a lot of confidence that it will be mostly full. It demand exceeds the anticipated supply then one gets shortages. Which is what we are seeing now.
And Nvidia isn't the only customer bidding for TSMC wafer starts. AMD is in the same situation (both for GPUs and CPUs). As are lots of other customers.
But very few folks are willing to commit to billions of dollars of wafer purchases a few years out (Apple may be the only exception here).
> For 10 months now, Nvidia GPUs (GTX16x0 Turing) have not been available
...
> I understand that Covid could not have been planed for and caused supply chain
> disturbance, but this is not disturbance but market and industry failure.
> Looking at the (mm)² sold, I remember a consumer die to be 150(mm)² 20 years ago. Today
> a mid-range GPU is three times as big at 450(mm)². Could no one predict that it would take
> more of the same 300mm wafers to make the same number of three times bigger dies?
> Because Nvidia is effectively a monopolist (No other product is sufficiently identical) they would
> set the price rather high and would not sell at maximum sourcable capacity. Even without Nvidia being
> pushy about production, they must be at a point where selling more would increase their profit.
TSMC (and other foundries) are capacity constrained. If you want a new fab today, you needed to start building it (and placing orders for equipment) two or three years ago.
But (leading edge) fabs have two very bad properties:
- They are expensive. $10B to $20B expensive.
- Their value depreciates quickly.
The upshot is that no one wants to build a $10B+ fab without a lot of confidence that it will be mostly full. It demand exceeds the anticipated supply then one gets shortages. Which is what we are seeing now.
And Nvidia isn't the only customer bidding for TSMC wafer starts. AMD is in the same situation (both for GPUs and CPUs). As are lots of other customers.
But very few folks are willing to commit to billions of dollars of wafer purchases a few years out (Apple may be the only exception here).