I think ASIC isn’t what many people think—dedicated circuitry built just for a particular model architecture. It should be circuitry built for specific Model Layers and Operators, and the inside is programmable.

ASIC isn’t so stupid that it can only run one specific model. Their own model team also needs to iterate—how could they not communicate?
A very simple design pattern: profile your own model to see which operators consume the most time, then have the designers optimize those operators and turn them into an ASIC. This makes it easy to improve your workload efficiency.
Such an ASIC is actually a collection of operators. Of course, it might drop some less commonly used operators for efficiency reasons, but it can still run general models.
At least there’s no fundamental difference from what NVDA is doing now. The only difference is that NVDA needs to accommodate every company, while an ASIC can be dedicated to handling your own specific batch of traffic—so it’s valuable.
As companies differentiate more from each other, the value of ASIC should keep increasing, not going the other way. There’s no need to pay a premium for features your own use case doesn’t need.
So, my view on ASIC is that in the future it can at least capture 50% of the inference market. NVDA’s crisis is real. I don’t think it will go to $300, and I won’t buy it.
Thank you, everyone.