Saving Private Intel
We look at how and how much it would take to save Intel without government subsidies. We also explore the complexities of getting GPUs servers into the data cetner.
In which we reflect on the ways to save Intel and do some rough math on how much that would cost. We also explore the growing pains of AI servers, take a look at Appleās iPhone event and recount some exciting new AI software projects.
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
Ben Thompson wrote a smart piece on Intelās future. We agreed with almost all of it, except for one key assumption. Thompson thinks there is no commercial justification for Intel Foundry. We respectfully disagree. Even setting aside National Security concerns, we think the industry badly needs an alternative to TSMC, a need which is only growing. Getting anyone to act on that is another matter.
Stemming from that exchange about Intel, we did some very rough math to determine just how much money Intel needs to survive, which we define as getting its 14A manufacturing process to volume production some time in late 2026. Adding up all their revenue and subsidies, then subtracting out their opex and capex needs, we think Intel needs $28 billion. Obviously a big number, but a bridgeable one.
We continued our exploration of the complexities of building servers. We viewed this first through the lens of AMDās acquisition of ZT Systems. One often overlooked fact is that silicon is actually a small portion of the cost of a server, 20% in CPUs, a bit higher in GPU systems. Someone has to design, procure and install all those other pieces. Nvidia has been doing that for years, and now that AMD is trying to compete with Nvidia in AI, they needed that capability as well hence the ZT acquisition.
Speaking of GPU systems, it turns out that actually getting them installed and running is challenging many seasoned data center experts. GPUs come with a whole new host of bugs, configuration errors, networking headaches and tools that sort of help fix those. And then thereās liquid cooling, which is causing something like 50% of install problems.
Apple unveiled their latest iPhones in what we felt was a fairly subdued event. They have some interesting AI features, nothing exciting but a few that look useful or fun, but also with a very un-Apple-like vague set of release dates. The event says a lot about the current state of AI, which is that we are firmly in the figuring out what it is good for phase.
OpenAI launched its newest O1 model, which is capable of āmulti-levelā reasoning. If you squint a bit, this points to some really interesting abilities for AI systems, somewhere in the future.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
CNBC reports on the ways in which AI is distorting the venture capital market. This report helped us put words to something we have been noticing for several months. VCs seem only interested in AI-related companies, but there is not enough of a market for this to support all that, at least in semis.
As we argue above, government support is not the solution to Intelās problems. That being said, the US government could do everyone some good by encouraging other big US tech companies to show some support to Intel.
A good explanation of how to calculate chip yield from fab wafers. Especially relevant in light of Intelās recent status update regarding its 20A and 18A manufacturing processes. 18Aās yield is not all the way there yet, but is fairly far down the learning curve.
Next Platform on the coming competition for AI inference. GPUs are not the guaranteed winner here, with all that means for Nvidia. By the same token, start-ups probably do not have much space to enter this market.
A study of tensor cores and how they implement CUDA. The links go deep.
Networking and Wireless
The head of ETSI says 6G will be an āevolutionā, a set of small improvements to 5G, using a variety of techniques including advanced modulation, beamforming, MIMO and a few more tricks. This is almost exactly how we would describe 5G. His point is that the carriers have no appetite for a massive capex spend after lackluster gains from 5G. The views of his organization are fairly influential, but the vendors will have a vote in the next few years as the standard takes shape, and they probably have a different view on carrier spending. See also, this interview with Rakutenās CMO about the āfailureā of 5G.
NTTās iOWN consortium built an all-optical link between Japan and Taiwan. This once seemed like science fiction, but there are now several important all-optical networks in production.
Amazon is now selling an Industrial IoT device, the Monitron. The device claims to detect abnormal machinery behavior. One telling detail, at this point it only connects to the Monitron smartphone app and not to AWS, so maybe not a strategic initiative.
Data center networking is shaping up to be a major area of competitive differentiation, especially among the hyperscalers. Meta now has a āspine-lessā network. All of these are looking for ways to remove networking layers in order to reduce latency, which is the biggest constraint in AI models.
A history of the commercialization of cellular Wi-Fi hotspots. Not surprisingly, the cellular operators stalled availability of these for years until they realized they were actually in the business of selling bits and not minutes.
A whole Substack series on Radio Frequency physics and chips.
Once more, cell phones are shown to not cause cancer.
Software and the Cloud
Two important AI projects launched recently. The first is Google Illuminate, which takes academic papers as an input and then returns a full podcast with two synthetic voices explaining the paper in a highly accessible way. The second is Googleās Game NGen which takes video game footage as an input and then returns new frames for that game. These are both in early days - Illuminate is in private beta, and NGen is just released as code. But looking at these it is possible to see a future for generative video. Of course, itās Google so maybe neither of these go anywhere.
The US government has started warning of the vulnerabilities of the Border Gateway Protocol (BGP). BGP powers pretty much all Internet networking, so this is a big one. We are glad to see governments taking this more seriously, it only took a decade of industry warnings and mishaps to get them here.
Gartner (via the Register) claims that Generative AI is its ābrute force phaseā, and once that ends demand for GPUs will fall. We think this is almost right. AI is very much about GPU-intensive training right now. The market for inference semis will be much more heterogeneous. However, the AI software market is developing so quickly, the end of GPU demand may go on much longer than people anticipate.
More customer complaints about VMWare post acquisition by Broadcom. Civo, a VMWare competitor, claims āmore than half of VMWare customers are looking to moveā. Add to that, news that AT&T is suing Broadcom for VMWare price hikes. Take this with a grain of salt. All this noise is a feature, not a bug, of Broadcomās playbook - cut costs, reduce product bloat, focus on major customers and raise prices. We think most companies will find it too painful to rip out VMWare, and the ones that do are probably not the ones Broadcom cares about.
Science and Climate
Over 90% of cars sold in Norway last month were electric vehicles. Good news for the environment, less so for incumbent European automakers.
Diversions
Researches tuned into chemical signals of mushrooms and harnessed those to steer a vehicle. We are deeply interested in plant/fungi signaling systems, but also see the appeal of this article to anyone who has every played Mario Kart.
Turning disposable vapes into a power bank. Some day soon, someone is going to wire this to a bank of raspberry pies and run an LLM, and probably play Quake on it via an AppleWatch screen.
Image by Microsoft Co-Pilot
Thank you for reading D2D Newsletter. This post is public so feel free to share it.