The Bubble Must Eat
AMD makes progress, but still disappoints?; Intel seems to have lost the plot; an overabundance of AI accelerators; and more.
In which we review AMD’s analyst day, question why the US has 63 AI chip start-ups, and Intel’s questionable choices around gaining new business from Sony.
Please note we will be on medical leave next week and will delay publication of our next China note by one week.
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
By our count there are now 63 AI accelerator start-ups in the US. This is too many. Then factor in: Nvidia’s 80+% market share; AMD’s heroic struggle to carve out 5% share; the top six customers consume 70% of the market, with another ~15% going to “captive” Nvidia neo-clouds; and those six customers are all working on their own chips. It takes heroic assumptions to imagine that many of those 63 are going to get any return, let alone venture-scale returns.
Companies with large fixed costs, such as manufacturing operations, build cost accounting systems around those assets. This makes it very hard for them to adapt when the industry changes. This sank Nokia in the 2010’s, the US auto makers in the 80’s and 90’s, and now seems to be part of the problem at Intel, which reportedly lost Sony’s chip business over a minor difference in pricing. Of course there are a lot of factors that lead to corporate strategic blind spots, but when the CEO gets financial models that have hidden dependencies based on those old models it reinforces bad past practices.
AMD held an AI event that was great for a semiconductor company, but disappointing for an AI company. The company announced a refresh of its Mi Series AI accelerator, some interesting networking products and general progress on other important areas like its software. Their stock fell noticeably on all of this. The Bubble Must Eat, and punishes those who do not deliver excitement.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
The IEEE interviewed the leaders of Amazon’s chip design team. No news here, but a lot of insight into how the hyperscalers are approaching chip design and the ways in which that differs from what merchant chip vendors do. For instance, they want their chip designers to have a good grasp of software coding, while a standalone chip company would very much treat those as two distinct skill sets.
More from Google’s DeepMind team on how they trained an AI model to design chips.
European researchers set out to test various forms of LLM hardware acceleration, testing 49 all together. They found that just calibrating the results was too hard to do because the differences between all the forms of hardware, models, semis production processes and the rest delivered results that were all over the map. As we note above, there are 63 AI accelerator start-ups in the US alone, and given all this fragmentation and noise, buying Nvidia just looks like the safest solution.
The world’s best audiophile amplifiers now come from no-name Shenzhen based companies, powered by the Texas Instruments’ TPA 3255 audio amplifier chip. You will still need good speakers, but the wonders of semiconductors have now come fully to this corner of electronics. This blog, with links to lots of review data, does a good job of explaining the joys of analog semis.
Mediatek launched its latest applications processor, the Dimensity 9400, and it looks really good. Mediatek has done an incredible job of moving up the technology ladder. Their products were once years behind Qualcomm, and now they are largely on par, if not occasionally ahead.
Intel’s GPUs are not selling particularly well. The current line of Gaudi semis appears to be at end of life, with Falcon Shores set to replace it. When exactly is Falcon Shores coming?
Networking and Wireless
Nvidia builds its ecosystem. Last week, Nvidia announced a partnership with Accenture to jointly develop enterprise AI software solutions, and then announced a deal with Foxconn who will build a server assembly line in Mexico to assemble AI servers. Nvidia is speed-running a process that it took Intel 15 years to build.
Cellular Tower owner Crown Castle is reportedly in the process of selling its fiber network. There is a lot of speculation as to whether this will happen and who the buyer will be. Crown has come under immense pressure this year from activist shareholders who do not like the amount of money Crown is spending on fiber and small cells. The fiber business has just always been brutally competitive, and small cells require immense investment, neither of which fit well with Crown Castle’s cash cow, yield-generating tower assets.
Germany is still deeply ambivalent about banning Huawei from its telecom networks. Mobile operator T-Mobile now says they are delaying their plans to deploy Open RAN systems for operational issues. All of which is to say this just drives home the point that Open RAN is not really an alternative to Huawei.
A new flavor of Wi-Fi specifically designed to operate over long ranges. Loosely speaking this is an amalgamation of LoRa and Wi-Fi, but people forget that Wi-Fi can already operate at fairly long range. This new “WiLo” approach could be useful for IoT deployments, but the devil will be in the details of how (if) this gets embedded into the broader Wi-Fi standard and how it is implemented in silicon.
On a related note, Cisco appears to be walking away from the LoRa market entirely. This could be read strictly as a Cisco cost-cutting story, but we imagine that if the business were doing better, Cisco would have cut elsewhere. LoRa is (was?) a really interesting standard, that could have been very useful, but it got lost along the way. The companies leading the standard did not seem to really understand how to build an ecosystem. And the price of cellular IoT, especially in Europe, is so cheap, that many are asking what future LoRa has.
Software and the Cloud
The Nobel Prize for Medicine went to AI researchers, in a heavily AI-themed award season this year. AI is useful for many elements of biotech research, but not for others, and understanding what the prize winners did can help scope out the solution space for AI in medicine. On a related note, Europe’s CERN is lending its compute capacity to cancer treatment research.
How to scale a 10,000 H100 Cluster for model training. Notice how many of the issues here touch on basic device configuration and networking issues, as opposed to compute.
Science and Climate
The costs of wind and solar energy in Australia are now less than 50% of the cost of coal and natural gas. Not available in all markets.
Diversions
The world of electrostatic energy.
Image by Google Gemini.
Thank you for reading D2D Newsletter. This post is public so feel free to share it.