[UPDATE] D2D: The Practicalities of Data Center Strategies
In this issue, we look at the industry that builds servers as well as attempt to gauge what the hyperscalers are looking to achieve with their internal silicon.
Highlights from our blog including a look at Intel’s messaging problems, how that makes them vulnerable to a potential takeover, the changing face of how servers are actually built and a look at the silicon business needs behind the hyuperscalers’ silicon strategies. And more. Apologies for the update, we encountered some version control malfunctions yesterday. Thank you for your patience.
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
Intel faces a big problem in how it communicates for the next year or two. Their big story is the recovery of their manufacturing prowess, which seems to be both on track and fairly promising. But we will not really see the full results of that for some time, and in the interim there is not much to get excited about. We would caution that the company needs to be extra careful. In that interim period the stock is likely to be under considerable pressure, while at the same time many of its potential foundry customers will know a lot about how good those manufacturing processes are actually faring. If we wanted to be wildly speculative, we could see one of those customers buying the company once they feel it is sufficiently de-risked. There is really only one company that could even contemplate such a transaction - Broadcom. On paper, this actually looks fairly compelling, the reality is much more complex.
In the semiconductor industry, we tend to lump the hyperscalers into a single, monolithic bucket. But their ambitions and the things they need from silicon are widely divergent. We think it is important to keep their business models in mind to understand their silicon strategy as just one part of that.
We talk a lot about servers, but who builds those servers and how does their business work? We walk through how server owners design and build out data centers, with a changing constellation of ODMs, OEMs and silicon providers.
We take a look at the growing interest in “the edge”, that vast sea of devices where we suspect much of future AI workloads will actually take place. Qualcomm has some important advantages here that are often under-appreciated.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
Talk of RISC V has subsided considerably this year in the US. While it is still very important for a lot of functions, one of the obstacles to it gaining wider adoption in larger chips is the state of its tooling. A periodic reminder that open source anything has advantages but also some drawbacks.
On a related note, two years ago Google committed to porting Android to RISC V. In a recent move they deprecated what little of that they had already built. We do not think Google has really abandoned the effort, and the company is undergoing a lot of change, but this is a periodic reminder that Google is a fickle partner, and also that open source has some drawbacks.
In fairness, RISC V is still very important for a whole host of reasons. Not least because it allows for creation of things like $0.10 microcontrollers capable of running large language AI models.
Another emergent AI device, the Rabbit R1, turns out to essentially run on Android. This is not entirely surprising, Android is a common sense solution to a problem for device makers. On the other hand, it does open the question as to why anyone needs a standalone “AI” device when they already have a smartphone.
Remember when the Apple Vision Pro launched and we said at some point in the next year there will be a rush of articles saying it is has failed? Those articles are starting to emerge. It’s a $3,500 device with all kinds of advanced technology - it was never going to sell that many units and there would always be kinks to iron out in production.
Much is made of China’s strong position in battery components like Lithium. Every time someone discovers a lode of Lithium outside of China the news generates major headlines. But China’s real advantage rests in its dominant position in lithium processing, where it controls something like 95% of the market. So we think it is worth paying attention when academics discover a way to extract lithium from mine waste trailings.
Nvidia acquired Run:ai, an Israeli company that makes software for orchestrating workloads across banks of GPUs. This could just be a sensible bolt-on acquisition to help boost a few features of Nvidia’s products. Or it could be part of a broader strategy to run its own cloud.
Networking and Wireless
A long look at IMEI numbers, the true sacred name of each of our phones.
Software and the Cloud
Something important seems to be happening in the land of commercial open source software. Last month, database make Redis changed the terms of its licensing agreements, essentially looking to charge for many things it once gave away for free. To put it mildly, this was not well received. Last year, HashiCorp, which developed the critical multi-cloud support framework Terraform, made a similar change and saw its results crater. HashiCorp was acquired last week by IBM at a price well below its 2022 IPO valuation. Making money from open source is hard.
Timothy Morgan at Next Platform does some smart analysis on the business model behind AI Factories. These turn out to be pretty good businesses, at least as long as demand for GPUs is high.
The UK has put in place a new set of rules around security for IoT devices. In general, we have not been fans of UK tech regulators, but a lot of this seems highly sensible and necessary. IoT security is a mess.
We do not usually report on Valley gossip, but there is definitely something strange going on at Google right now. In addition to extensive layoffs (right ahead of their I/O developer event) someone seems to have an axe to grind with the man running Google search. There is a lot of commentary swirling around the team now, and it is odd to see such public scrutiny over executives other than the CEO.
We talk a lot about the difference between porting software to a new chip and optimizing for that chip. The former is fairly easy, the latter is much more labor-intensive but can make all the difference. Arm is stepping up its efforts to explain to developers how to do those optimizations.
In this week’s edition of sensational AI headlines. These researches highlight some interesting applications for LLMs in the field of financial markets. On the other hand, this practitioner finds that LLMs struggle with statistical modeling and probabilities. That seems a bit contradictory, but don’t worry because apparently “AI Surpassses humans in almost all performance”.
CERN’s data center stores 1 exabyte of data, that’s a 1 with 18 zeroes, or a gigabyte of gibabytes.
Diversions
Science Fiction author Neal Stephenson has a pretty good track record of predicting the future. Also, he is a good writer. He has a new book coming out later this year.
Sure the world looks like a mess, but researchers have created an mRNA treatment for glioblastoma brain cancer.
Image by Microsoft Co-Pilot
Thank you for reading D2D Newsletter. This post is public so feel free to share it.