The Real End of Moore's Law
What will happen when TSMC raises prices to $40,000/wafer? $50,000? And what could stop them?
In which we take Intel’s Board to task for the last decade, and what it will mean when TSMC has unconstrained pricing power.
Just a reminder that we operate our newsletters on a paid model. Paid subscribers will get three newsletters a month including our China Deep Tech notes. Paid subscribers will also get early access to the newsletter as well other benefits coming soon. Please subscribe and support our work.
Highlights from our Blog
It now looks likely that Intel will fall off the Moore’s Law curve and thus the foundry business. With Samsung’s Foundry business looking almost as dismal, we need to start thinking about what will happen when TSMC loosens its self restraint and exercises true monopoly pricing power. Next year, we could start seeing leading edge process prices touch $50,000 per wafer. At those levels, many companies will start to rethink their need for moving to those advanced nodes. More than anything else, this will drive home the end of Moore’s Law as our smartphones and PCs stop improving at the rate we have taken for granted for the past forty years. This is the reason we think the many large companies (Qualcomm, Marvell, Apple and the hyperscalers) who are not TSMC’s top customers should step in to support Intel Foundry.
After months of mounting pressure, Intel’s Board fired the CEO. So, now what? Seriously, what are they going to do next? We know the many criticisms of Pat Gelsinger, and some may be valid, but Intel now seems to be moving away from ambitious plans to become a viable foundry. It was always going to take a decade to achieve that, and the Board did not even give Gelsinger four years to get those five nodes. Short-sighted thinking seems to rule these days.
Nvidia reported normal earnings, with results any other company would have been ecstatic to deliver, but for a company that has had a run like Nvidia’s they were nothing special. Our sense is that they are just taking a breath ahead of what they think will be a very strong ramp for Blackwell next year, or maybe they have just gotten so large that it is time to revert to more normal metrics.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
A review of the latest AmpereONE server with 192 cores of power. Fairly good performance, with the major caveat that to achieve that performance the software running on the server needs to be optimized for Arm.
Construction Physics has a great profile on the Influence of Bell Labs. Definitely worth a subscription.
Smartphones killed the camera, right? Camera maker Leica just reported record revenue.
Networking and Wireless
Last month ‘someone’ cut two fiber connections across the Baltic Sea. The Internet did a good job of routing around those cuts so that no one really saw a service disruption. Good news - the Internet works. Bad news - this will not be the last time the Internet gets tested like this.
NATO posted a profile and some photos of the new Secretary General. So of course someone deconstructed the security of his telecom set-up.
Software and the Cloud
This is a good overview of the reasoning behind multi-modal AI models, or models that look at images, sound and text. This is pretty dense, and still only scratches the surface, but it is an important topic that is going to become even more prominent next year.
Car & Driver published an overview of “Every Hands-Free Driving System Available in 2024”. This is a good overview of the state of autonomous driving, but there is a mistake the in the title. It should read “Available in the US”, because they are missing about a dozen vendors in China.
A review of the Industry Structure of the “LLM market”. We contend that no one really knows what the product for LLMs will ultimately be, and so the industry structure of today is likely to change significantly.
A big topic next year will be the cost of AI inference. This AI company did a profile of all the optimizations they did to optimize the utilization of their hardware. The key takeaway is just the very large number of ways they found to accomplish this. There is ample room for innovation, experimentation and alternative approaches to improving inference.
Climate and Science
Solar is the fastest growing energy technology in history.
The world now has 2 Terawatts of solar energy generating capacity. The total electricity generation capacity of the world is around 8.5 TW.
Diversions
For anyone looking for that perfect gift - the Lego Twinscan EXE:5000, available on the ASML website now.
Image by Lucas Films
Thank you for reading D2D Newsletter. This post is public so feel free to share it.
I agree with the thesis but isn't Apple TSMC's biggest customer and Qualcomm and the hyperscalers are dependent on TSMC? And isn't that one of the biggest issues for Intel: Who can afford to upset TSMC?