D2D: CES, TSMC and More
In which we recount our visit to CES, the futility of tariffs, Nvidia’s immense software momentum and a quick look at TSMC’s results.
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
CES was its usual carousel of delights. The biggest attention getter was Jensen Huang’s Day One keynote. He brought Nvidia’s usual onslaught of bets - many of which may even play out. In particular, we thought Nvidia’s biggest message was “We have all the software”. Maybe the hyperscalers will not use it, but we think many others will.
Overall, we thought there was not enough AI at CES. Which, if you were there, sounds crazy. Everything seemed to carry an “AI” label, but the vast majority of that “AI” was just software and clever algorithms, pre-2022 technology. After the flame outs of last year’s chat gadgets (i.e. Humane Pin and Rabbit), we think all the other product companies hit pause on their efforts to bring Chat GPT features to devices. This is probably good news as the industry is being more deliberate in its search for the best ways to productize all the new technology.
We spent a whole day in the South Halls visiting hundreds of companies form China. We had many thoughts on that which you can see in our China Newsletter. However, one thing is very clear, PRC companies have retooled their geographic footprint to prepare for more tariffs. They now have alternative sites all over South East Asia explicitly built to avoid US tariffs on PRC goods.
TSMC reported good earnings which brightened everyone’s mood. They saw “AI” revenue triple last year and expect it to more than double this year. We remind everyone that TSMC is at the tail end of the supply chain bullwhip and will find out when the party ends after many others, but for the moment there is no sign of that happening any time soon. Also, we stopped by their Arizona plant on Christmas Day, and the parking lot was full.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
Cerebras posted a piece on how they deal with defects on their wafers. Big chips, and Cerebras’ is as big as they come, suffer more from fab defects than smaller chips. Cerebras did some smart architecting in anticipation of that, but ultimately that involved design trafe-offs and redundant functionality taking up real estate. Fun technology, but we are not convinced that it really alters the landscape or solves the problems customers are actually facing.
All About Circuits has a list of predictions from a dozen semis companies. Of course there is a lot of AI, but also more about chiplets, new materials and some fairly nuanced discussion of big trends.
An academic review of on-chip interconnect in GPUs. The key takeaway from this is that they found theoretical vulnerabilities which now seem likely to find their way into real world attacks.
Networking and Wireless
This company introduced a Wi-Fi router with a 10 mile range. We missed this at CES, and while the idea sounds appealing, this is far from the first attempt to do something like this. The world record for Wi-Fi signal is over 200km. Ubiquiti Networks was built on their founder’s attempt to do something similar. And if we go even further back, there were a host of cellular standards-based attempts to do something similar.
Starlink is now one of the cheapest Internet service providers in many countries in Africa. It is tempting to think of Starlink as an end-run around wireless operators, but even with all the vehicles they are putting in space, there is a limit to their capacity. Cheap for now, but ultimately users will face limits to bandwidth.
Why is Ashburn Virginia the data center capital of the US? Ready access to power, Internet and population centers.
Nokia’s internal presentation on their first reaction to the iPhone. They seem to miss the point of software and apps (in their defense, the App Store was still a year away). Instead, they wanted to find a way to make a cheaper version. We think both Nokia and RIM entirely missed the power of software and apps. Keep in mind for future technology transitions.
Akamai is exiting China. They seem to be saying that they are leaving for commercial reasons, not political ones. Which is a fair argument, but then we have to ask why are their commercial prospects there so poor?
Software and the Cloud
A very useful reading list for state-of-the-art AI papers and trends.
A paper from Stanford introduces another level of reasoning to AI inference, meta chain-of-thought. Interesting extension of current trends, and can also be read as the demand for semis is just going to grow more.
On a similar note, one more iteration of the transformer model - Transformers^2. The viability of this approach is beyond our scope, our point is just that we are going to keep seeing more of these for the foreseeable future.
Interesting overview of one company’s move from public cloud to self-hosted cloud. This is perfectly viable for any company above a few dozen employees, but requires some internal expertise and careful planning, but is much cheaper over the long-term than paying Amazon’s margins.
Climate and Science
Diversions
A primer on US equity market structure. Handy for the aspiring day traders in the readership.
We really enjoy visiting San Antonio’s Riverwalk, so a post about all the civil engineering it takes to operate it makes for good reading.
Image by D2D
Thank you for reading D2D Newsletter. This post is public so feel free to share it.