I’m surprised this article had nothing to say about Intel, TSMC, and GloFo all getting down toward the absolute bottom of what is possible in silicon. We are at 14nm right now, folks. Intel has plans all the way down to 7nm, but transmigration is a huge issue and it is getting worse. Their last node jump only seems to have given them a 10–15% IPC gain over the previous 22nm node.
Chips are cheaper than ever, and “AI” (a less hype term would be machine learning) is better than ever, but you have to combine the two, and frankly cheap chips cannot compute big AI in reasonable time frames. All the above AI examples are extremely domain-specific; what disrupts all of that? big, generalist AI: take a very good chip that can do all its work LOCALLY, without asking for help from big datacenter chips in the cloud, and then you have something that can do every single thing the domain-specific AI can do, plus it can integrate those things and do new things as they come up. That’s what really makes the future look different. But to do that, we basically need to move off silicon, which is going to be a huge issue.