Hacker News Reader

Comment by hedgehog | Hacker News Reader

Tough situation to navigate out of. They lost the process lead, now TSMC in particular does so much more volume that Intel is at a significant scale disadvantage (makes it harder to catch up). The margins in the historically lucrative server business are under threat as Xeon has to compete with hyperscalers' in-house chips as well as credible alternatives from AMD. Moore's law is over so the big performance jumps that justified rapid PC refresh cycles are gone. They have trouble entering new markets themselves and their M&A track record is poor. It seems very possible that Intel will never return to the strong position they enjoyed ten years ago.

View on HN

I think that's largely true, but it's also worth pointing out that "Moore's law is over" means that they don't have much "catching up" to do. The truth is that all semiconductor vendors are heading for a future where they're selling an undifferentiated commodity product.

Sure, Intel isn't going to be able to sell Xeons for $1k a pop in a future where anyone can go get their chips fabbed at TSMC. But neither is TSMC going to be able to command the margins they do right now if Apple and NVIDIA know they can trot right over to Intel to get a better deal.

To me, the story here is much more "maturing industry is losing its margins" than it is a two-way horse race.

>The truth is that all semiconductor vendors are heading for a future where they're selling an undifferentiated commodity product.

This is what I'm hoping for as well. But there are huge upfront costs (>20b). This will limit the number of potential players.

The upfront cost may plummet if you don't need to redevelop and replace your entire tooling every five or so years.

This is an excelent point I hadn't thought of. The industry is currently chasing itself upward, which prevents the supply chain from maturing.

I found that the cost to build a 65nm plant was $1.2B in 2021[0] which is a pretty mature node and that's still a large upfront cost, and smaller nodes would be even more expensive. So it's hard to say how low prices could fall.

[0]https://finance.yahoo.com/news/bosch-inaugurates-1-2b-factor...

The differentiator in the server space comes in power management and thermals, all things that AMD does better than intel. AMD also has better yields and better margins because of their chiplet design philosophy.

ARM also has a significant thermal and power edge.

Intel is chasing the top, but actually they should be thinking about efficiency.

I mostly agree, and the compaction of value accrued at the lower layers of the stack are a real thing, but there are two wrinkles to the industry that I find interesting. First, like advanced turbine engine technology semi development continues to get more expensive as it matures and that drives consolidation rather than proliferation. Second, for now the end of Moore's law for companies with high pricing leverage looks like a move from "more for less" to "more for more" where they use performance and vertical integration to create a moat around their part of the market. You can see this in Apple shipping chips that are multiples larger than previous laptop processors, or NVIDIA pushing performance up but jacking prices up a the same time. Intel is going to have a hard time putting a 400mm2 chip into a Dell laptop at pricing and margins everyone can live with. At the same time they're going to have a hard time making margin on Xeons and still being competitive in TCO terms with whatever the Annapurna team can cook up for AWS using ARM IP on TSMC. This is all a roundabout way of saying unlike say the early PC industry I don't think maturing is going to level the field or lower prices, actually the opposite.

The problem is that Intel competes with Apple, Nvidia and AMD. They have a conflict of interest, just like Netflix has with content producers, and that's going to cause issues for their growth.

Roughly what are TSMC's volume compared with Intel's on their leading logic nodes?

Hard to tell, TSMC's faster cadence means you'd have to get a snapshot of total output on N7 and newer but I don't think they break that out. They report numbers here and there in the press but usually only for the new process they're promoting and for example a lot of N7 from two years ago probably got cannibalized by N5 etc. For N5 and newer Intel doesn't have any equivalent process, and it sounds like they will be doing some products on TSMC at N3? If you did a histogram of wafers by effective feature size I think the shapes for Intel and TSMC would be very different but we know TSMC has at least two big customers selling N4 product today (Apple and NVIDIA).

Reply on news.ycombinator.com