We have been writing about custom ASICs for many years. The idea that some non-chip companies had gotten large enough that designing their own chips made more sense than buying merchant silicon. For a time it looked like “everyone” was taking this path, we have a list of over 100 companies working on their own silicon, from smartphones to cars to Wi-Fi access routers. That being said, technology moves in cycles, and we may be reaching the apex of this arc of the pendulum.
We think it is important to remember that the case for custom silicon is an economic choice, not a technical one. The primary driver of the move to custom silicon is the consolidation wave that has gripped the semis industry over the past 20 years. When every socket had a dozen vendors vying for business, large customers could wrangle control of someone’s roadmap to suit their needs. Now that the industry has gone from 2,000 suppliers to something closer to 200, even the largest customers started to lose that negotiating leverage.
That does not mean that companies should rush to self-design in order to save a few dollars of vendor margin, but when a chip can convey a strategic advantage then it makes sense, with Apple as a leading example, as is Google.
This matters because it is starting to look like companies are bumping up against a countervailing economic force. Designing chips, especially at the leading edge, has become very expensive. The costs of talent, IP and manufacture have all risen steadily in recent years.
So it seems reasonable to start to question the drive towards custom designs. We do not have to look very far to find examples. As much as the hyperscalers are all talking about their own designs, our sense is that many of them are starting to rethink their efforts. Microsoft and Meta, in particular, seem to be having a harder time getting their designs out the door. We do not think these companies will abandon their efforts entirely, but we would not be surprised to see them narrow the scope of their ambitions. Does everyone really need their own CPU? We think they will continue to pursue their own AI accelerators, but even Google (who has their own semis design tools) seems to be slowing down the pace of their new chip introductions. We see something similar in smartphones – where even Samsung has cut back the use of its in-house chip, as is the case with the PRC handset makers.
And this leads to a key determinant in all this – software. Custom chips only make sense when they convey strategic advantage, but that is really only viable when the company designing their own chip controls all the software running on it. Absent that, the benefits of custom design diminish fairly quickly.
In the end, we do not think the industry is entirely finished with custom silicon. The big AI houses (i.e. OpenAI and maybe Anthropic) will probably make a stab at it. And then the world’s auto makers still have to make up their minds on this topic too. So we have not reached the high water mark of custom silicon yet, but we are getting close.
Leave a Reply