D2D: Adrift at Sea, with No Land in Sight
In which we bemoan the current sorry state of venture fund raising, examine Apple's data center chip ambitions and the dreams of AI.
Highlights from our blog including a look at the logjam in venture LP fundraising, Apple’s reported server CPU designs and the way in which so much of the AI complex is dependent on a small number of companies coming up with something consumers actually want from AI. We also examine the ways in which Nvidia has mastered the server supply chain, one mobile operators’ first step into using AWS for its core network and more.
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
Venture funds are lost at sea, with no land in sight. Limited Partners (LPs) - the pension funds, sovereign wealth funds and family office who invest in VCs - are over-allocated to venture on paper, but everyone knows their portfolios are not marked to true value. No one is in a hurry to unravel that problem, and while it lasts and the IPO window is closed, fundraising is going to remain difficult.
There are now multiple reports that Apple is designing its own chip for the data center, or possibly using its own M-Series. There is sound logic behind this - Apple controls its own software stack and has a lot of data - putting them in a good position to build and run its own, novel, high-utility foundational AI model, with inference conducted entirely on device and thus highly private.
The whole AI industry is dependent on the advances of a small number of companies. The PC complex really needs Microsoft to come up with something magical at its upcoming launch event. OpenAI and Google both unveiled interesting products last week, but still seem to have some way to go in productizing them, and then there is Apple with their June WWDC event. We are not saying Ai is short of innovation, it’s just that we are not comfortable with how concentrated everyone’s hopes are.
If you like this content, you should check out our podcast The Circuit
Semis, Hardware and Deep Tech
The Next Platform crunches the numbers on the hyperscalers’ share of data center semis, ~50% today possibly going to 70% in a few years.
Amazon purchased a small stake in chip design firm AL Chip. So small that the deal flew under most people’s radars, except for Bloomberg’s Tim Culpan who posted the news on his LinkedIn feed. The stake is so small, about 0.2%, that it could just be a minor treasury management exercise, but that seems unlikely. Instead, we would guess that Amazon is planning to do a lot more work with ALChip and wants the stake to mark its position in line. Not really clear what this means, but worth keeping an eye on it.
DigiTimes has a very mixed track record, but they know a lot about at the intricacies of the electronics supply chain, so much of which flows through Taiwan. They recently published a piece looking at the way that Nvidia has mastered the server supply chain. We have been doing a lot of deep work lately on that chain and this piece resonated. So much of the conversation about AI and the cloud focuses on the headline chips - CPUs, GPUs, etc. - but those only make up 20% of the cost of a server, and the way in which the other 80% gets built is important but overlooked. Nvidia has much deeper experience with this part of the supply from years of building their own graphics cards. So it is worth paying attention to how they bring that expertise to the data center.
Bloomberg voices what he have been thinking for a while “Detroit [car makers] Makes the Same Mistake on EVs It Did With Japan”. Painful and surprising to watch history repeat itself this way.
Running AI models on FPGAs. We do not think many will build this kind of system, but it is important to remember that not all AI will run on GPUs.
Liquid cooling has become a hot topic. The new Nvidia systems use so much energy that air cooling is no longer sufficient. Expect more developments in this space.
A researcher in Singapore recently benchmarked a group of CPUs, they actually found that Alibaba’s YiTian 710 was the most performant. Benchmarks need to be taken with a grain of salt, but worth reading as this is the first time we have seen actual performance data for this part, and it definitely looks competitive.
Someone leaked a large, very private internal analysis from Dell of the new batch of laptop CPUs. These show Qualcomm’s chips have a big lead in power consumption and price. We reviewed the entire deck, and there is a lot of detail to back up the headline claims. Qualcomm has a shot at this market, but performance alone will not determine the outcome.
Networking and Wireless
Wireless operator O2 announced that they would begin running their core network for their German network on AWS. We believe this is the first time a major, incumbent operator has made this transition, and it is fairly significant. This looks a bit like a pilot program in that Germany is not a major market for the operator. Nonetheless, it will serve as a major test case. We have written a lot on this topic over the years - moving to the public cloud is highly tempting for the operators but also risks them getting trapped there. The industry will be watching this one closely.
We forgot to write about this from MWC, but AGC has developed a radio unit embedded in panes of glass. We are not sure about the economics of this, but it solves a major problem with millimeter wave small cells. Cities do not want any more street level network gear, but they probably won’t complain about resonant windows.
Radio astronomers found a new way to focus radio scopes to ‘filter out’ disruptions from Earth’s magnetic field and get much sharper images of the universe. Not exactly beam-forming, but close enough for us.
A look at the networking protocols used to communicate with satellites and some of the challenges inherent with connecting with a device 100 miles away. Or this start-up which is communicating with satellites via amped-up commodity Bluetooth gear.
Software and the Cloud
A detailed look at how much one software company spends on its own hardware versus what the equivalent would be if they were running on public clouds, saving 90% off their infrastructure costs. The author claims the company would not have been able to stay in business if they had run on the cloud. We could quibble with the math a bit, but there is a growing awareness of the drawbacks of the cloud.
Less publicized than Google’s recent I/O event, their DeepMind team seems to have scored a major breakthrough in using AI to decode human proteomics - the ways in which protein molecules fold. Very important implications for drug discovery.
Last issue we linked to one story saying AGI was imminent, and another saying it would never happen. Now we have an academic study that essentially asks “Can’t we all just get a long?”.
AWS has a new CEO, this is an informed take on what he needs to do differently, and it is not “more AI”.
Diversions
A map from the USDA showing changing patterns in home gardening. The great graphics and cute animations obscure some fairly sobering climate data.
Image by Microsoft Co-Pilot
Thank you for reading D2D Newsletter. This post is public so feel free to share it.