A lot of this is what I consider being from the Department of Obviousness Department, which I normally try to avoid, but sometimes, the timing is such that I appear to have powers above and beyond those of an ordinary ex-nerd thinking about the future. There was a time, once, when I was arrogant enough to reply – when someone asked me to predict the future – “I make the future, I don’t predict it.” Which was literally true but also JD Vance-level stupid.

Item #1: [arstechnica]

DeepSeek, the Chinese startup developing large language models that are competitive with those from US companies like OpenAI and Anthropic, is planning to enter the silicon business, according to Reuters.

Citing three people familiar with the matter, Reuters writes that DeepSeek has been working on a move into silicon for about a year. It has been meeting with potential partners in the hardware and silicon space and has been hiring engineers for the project.

The focus is on data center chips for inference, not training, and the goal is likely to reduce reliance on both Huawei and Nvidia.

Nvidia is the chipmaker for most AI companies in North America and Europe, but a United States export ban has prevented the company from achieving a similar presence in China. Huawei controls about half of the data center chip market there, and DeepSeek isn’t the only one trying to enter; Chinese tech giants like Alibaba and Baidu have been making moves, too.

Well, yeah. I constantly gnash my teeth in frustration at the way the US sits back, like a dumb and happy techbro, confident in its control of the market. Oh, it’s probably right enough, but every time it pushes the wrong part of the balloon, the other parts squish out uncontrollably. This is not a new move for the US; I first encountered ITAR (international traffic in arms regulation) back when I was working on crypto projects in the early 1990s. The US desired to control the global availability of high quality encryption, by regulating the export of chips that could run parallelized to attack (or build) Feistel-net cyphers like DES. Nobody at NSA or State Department or NIST believed that any of the export regime would slow people who were willing to break the rules but it sure was a pain for law-abiding researchers. I am developing a new principle – which is that government export controls always hit any target except the one they were designed for. (insert F-35 assembly in Turkey stories) In this case, I am mostly rooting for the Chinese, who are now more capitalist than the USA. I hope that they build a chip that can both process large AI trees concurrently, and mine for cryptocurrency. Those are not the same problem but if you’re making custom silicon, why not add a few more circuits; it’s waffer theen <- pronounced as formerly funny John Cleese.

The reason everyone is using Nvidia is twofold: they’re the most multi-core programmable chips on the consumer market, and ‘cuda. ‘Cuda is a programming language for talking to all the Nvidia CPUs on a chip, serializing inputs and outputs, downloading programs etc. ‘Cuda also handles bus-mastering memory and has powerful synchronization locking and queueing, so that a game can asynchronously calculate object trajectories without interrupting the CPU. Other than that, there’s no rocket science that makes AI love Nvidia, and the AI researchers would be perfectly happy with something that ran better and cost less. You can see where this is going: the Chinese develop a chip, start producing it with one of the government-influenced businesses, and the US freaks out and bans the chip from all US machines and networks, citing vague but terrifying backdoors. Because, apparently, Nvidia hasn’t thought of that …? Joking aside, the Chinese today are more pure capitalists than the US is. Witness a case in point: US networking companies created an alleged backdoor in Huawei networking gear and got it banned on critical and corporate networks. Do you think that Cisco Systems’ wholly-owned subsidiaries in congress were thinking of the competitive infrastructure landscape when they passed those directives? Not for a second! We have the best congress money can buy.

Anyhow, the big power-gobblers on a GPU are the screen RAM, the RAM cache for object models, and the multiple cores to drive it all. I forget how many cores my GPU is running, but it’s a lot and, uh, it gets things done. If someone produces a board with a bunch of RAM and many cores and a usable interface that is not tied up in non-disclosures, AI researchers will link their code – -define NONCUDA -lnoncuda and not only will the result be faster, cheaper, and it will gobble less electricity.

There has been considerable discussion about whether or not we are in a bubble. Of course we are! We are always in a bubble of something-or-other since US Steel, Ford Motors and Pullman rail cars stock got overvalued based on the cost of labor. The entire history of the US has been a sequence of bubbles: the Oklahoma land rush bubble, the railway bubble, the steam power bubble, the civil war weapons bubble, the cotton bubble, the slave bubble. But those are all examples of Great American Rip-Offs, where people paid ridiculous prices for growth stock, based on the obvious growth of an industry. The problem, like with the railways, is that AI data centers are something you can get the shareholders to invest in; remember it’s just the company diluting its stock to raise $1bn and build a new data center – but what that also does is raises the value of early insiders’ shares to catastrophic heights. When the “Lamborghinis and cocaine” guys start to glom onto the businesses, they’ll pretty quickly lead them astray. I know those guys. Those are the guys that’ll tell Facebook to buy some stupid company for a ton of money in order to move the stock (which they happen to hold) and the whole thing gets written down in the loss column 3 years later. Anyhow, yes, we are in a bubble. Companies like OpenAI are stuck because they have convinced the market that they need more money for GPUs and data centers – therefore they do a stock offering to the public for more GPUs and data centers, and raise $2bn. Then, the insiders liquidate a few of their shares, buy Lamborghinis and coke, and what the heck, may as well build a data center while we are at it! What happens if the cost balance implodes? Do you remember 2008? A lot of people got out over their skis, and face-planted. When people talk about a “bubble” – ignore them, they’re just rich guys wondering about the fate of other rich guys. The workers will get fucked either way, because capitalism needs a cheap work-force that is desperate, to threaten the expensive work-force that is comfortable. [No, I am not a Marxist but Marx’ understanding of the flaws in capitalism was profound.]

[Krea, the view from within a bubble, looking into a data center.]

Your Nvidia stock is good as gold for a long time because there are a lot of gamers and Intel hasn’t managed to cough up a good integrated chip with CPU and graphics together. They have been consistently out-marketed. Some day, they may implode but I doubt it, because that would take down Oracle, Microsoft, Cisco, and the rest of the US tech market. So, if the Chinese come up with something great, they’ll get buried in legislation and evaluations, etc. Where there might be problems is if a company that is ostensibly allied, (e.g.: Qualcomm) comes out with an AI chip and then the US will have to put on its stripper shoes and dance really hard. If you look at the history of the US stock market, the public gets a fleecing pretty much every 14 years. It’s not a hard-defined cycle but I think it has something to do with investors, and forgetting, and the memory effect of cocaine. Or something? [Actually, cocaine boosts my memory to uncomfortable levels, so much that I discontinued it long ago]

Item #2: [nvidia]

In another step forward, Equinix opened in January a dedicated facility to pursue advances in energy efficiency. One part of that work focuses on liquid cooling.

Born in the mainframe era, liquid cooling is maturing in the age of AI. It’s now widely used inside the world’s fastest supercomputers in a modern form called direct-chip cooling.

Liquid cooling is the next step in accelerated computing for NVIDIA GPUs that already deliver up to 20x better energy efficiency on AI inference and high performance computing jobs than CPUs.

When they say “born in the mainframe era” they mean “this is how real computers have always worked” – the CRAYs, IBM mainframes, CDC Cybers, etc. Another way of looking at it is that “room temperature” computing is a relatively new phenomenon which did not really exist! If you put an 80386-powered desktop in someone’s cubicle it’s not unbearably hot, or phenominally loud, but it’s a 400-watt toaster oven worth of heat and it’s only going into the office-space air conditioning. What we are looking at is incredibly lazy computing: the AI guys have taken basic computing and just scaled it insanely with software, without thinking of where they could improve it.

[It’s not a hot tub, it’s the output of a head-exchanger powered by a volcano. Iceland rules.]

In terms of cooling, closed loop is the obvious answer. There are more subtle answers, too, but most of them build on the closed loop concept of capturing “waste” energy. Also, if you think about it, a closed loop cooling system brings all of its members toward a sort of average temperature. A closed loop does not need municipal water to cool it – you run a big heat-exchanger and bury it in a lake, or underground, or offshore, or build an avocado farm on top of it. The icelanders have figured this out: they have a thing they call “the blue lagoon” which is the heat output of a geothermal plant that they charge people to swim around in the winter: win, win, win. Don’t tell me we’re not as smart as the icelanders, damn it. OK, most of us aren’t but some of us are.

In separate tests, both Equinix and NVIDIA found a data center using liquid cooling could run the same workloads as an air-cooled facility while using about 30 percent less energy. NVIDIA estimates the liquid-cooled data center could hit 1.15 PUE, far below 1.6 for its air-cooled cousin.

Liquid-cooled data centers can pack twice as much computing into the same space, too. That’s because the A100 GPUs use just one PCIe slot; air-cooled A100 GPUs fill two.

In raw Marcus-terms, liquid cooled data centers appear to be twice as fucking smart as air-cooled data centers. Because we’ve known about liquid cooling since the 70s. Even advanced PC case-modders know about liquid cooling. [many of them won’t run antifreeze in them because it’s conductive and will cause a big no-no if it leaks. So, instead they run water and their CPU chiller blocks become coral reefs full of obscure forms of algae.] I say, run vodka in the damn things, and confusion to the French!

Item #3: optimizing code

There are some indications that AI engines aren’t very well-coded. I don’t have references here to cite, but I saw one researcher who had actually done some cache effectiveness analysis on the AI’s keyword->vector cache and determined that it was ineffective. For example (making up things here) a typical query didn’t hit the same words often enough that caching the mappings did anything but eat a few gb of memory over time. When I saw about that, I was horrified. It appeared that some Computer Science PhDs had possibly been working on the code – and it’s stereotypically amateurish. A professional programmer (at least, any one that ever worked for me) would know “you do not pre-optimize your code.” The fact that someone just decided, apparently without measuring cache hits/misses versus updates, just put a big cache into production code – horrifying. Back when I managed development, you didn’t just add a feature, you documented an analysis of what it would do, and the benefits of adding it. When I see features being added uncomprehendingly, I assume I’m looking at Bullshit Code(tm) that can be reduced by 20-30% and improved by 100-500%. That’s just my industry experience, I could be wrong. But my experience also says, “I doubt it.” If code flaws like that are slipping into production, I can pretty much predict the software team is a “scrum” or some such bullshit, more focused on their Lamborghini than writing tight code.

That brings me to another philosophical point: the singularity. The premise of the singularity is that AIs will code better AIs and eventually they will tight-loop improvements until humans no longer comprehend what is going on. Well, not if the current generation of latte-swilling motherfuckers are teaching the AIs how to code. There is an interesting problem there, some of my friends and I have noodled about, which would be to start a business rating the code quality of available code, before AI ingests it. “Everything Paul Vixie writes has buffer overruns” etc. There will be no singularity if AIs are depending on the code of the last generation of latte-swilling motherfuckers who brought us PHP and JSON. It just won’t happen. Maybe we’ll get a reverse singularity, where AIs start acting as dumb as coder-bros and we have to nuke the whole mess from orbit, to be sure.

So, all in all, I think the bubble will burst for financial reasons of un-sustainability, and the technology bubble will continue forward. I’ve noticed that fads in computing often liberate previous incarnations to be art, instead of commerce. A lot of artists are complaining (fair enough, but their dying complaints are pro forma) that AI is collapsing the commercial value of illustration. Like illustration collapsed the commercial value of painting, and photography collapsed the commercial value of illustration. Each of these collapses is a blessing and a curse – mediocre artists who were making their money doing commercial illustration, are now liberated to pursue their skills as pure art – because nobody is paying for the old stuff. “But what if they can’t make it as pure artists?” you ask. Well, if they couldn’t make it as pure artists, they were kidding themselves about their commercial skills, too. Those folks will be unhappy but they were going to be unhappy no matter what. I’m sorry.

There are so many tropes tangled up in the anti-data-center memeology that it’s hard for me to sort out. The opponents come off, to me, like irrational anti-vaxxers: they are throwing everything at the wall and hoping something sticks. “Data centers are sucking all our electricity!” OK but what if the code was more efficient by a factor of 10, and the GPUs were replaced with cards that ran less power and did the same work? “Data centers are still bad because they are consuming all our water!” Sure I hear you but what if they were closed-loop and used the extra heat for growing yummy marijuana plants? “AI are using stolen data!” Well, you need to understand a bit more about machine learning, and study the differences between human learning and machine learning and ask yourself if reading a textbook on math is stealing from a mathematician or not. Fortunately for us all, runaway capitalism is taking the reins and what will happen is what will happen. I hope the AI turn out to be non-profitable, and are liberated as art. I am beginning to see some signs of that, but also signs of great transformation, so I am unsure where it all will go. The good news is that the technology will remain, because innovation is irreversible. I am often reminded of the AI in Brin’s The Postman – a genius-level gentle AI that was produced just as human civilization collapsed, which died because of lack of super-cooling for its chips.

Ironically, it probably could have rescued humanity if the power hadn’t failed.

As usual there is various stuff going on in my life that I have stopped talking about. My crafting work interests mostly me, and explaining it outside of myself is time-consuming and uninteresting. I don’t want to just post pictures of blades without their stories, but I don’t want to write their stories, either.

My mom died a month or so ago, and my last few weeks have been me and dad doing a memory tour (for him) retracing my childhood memories. I am pretty sure he enjoyed refreshing them. For me, since Paris was experiencing a supremely unlikely heat-wave (predicted by science) it was especially difficult. Holding a handful of what used to be my mother, and throwing it away, was a surreal experience. Fuck dementia. And, please, dementia, stay away from me.

I’ve been insanely busy with new stuff, some of which you might enjoy some of which you might not. I think that is what has held me back from blogging. I keep trying to think what is interesting, but I fundamentally realize that nothing is, any more. I have some fun AI art stuff and other things – painful beautiful things – I don’t know if I want to summon up the energy to write about. We’ll see. Stay tuned or fuck off, as you see fit.

@Pierce: I see your thoughtful comments about F-35s and the ongoing disaster there. I’ve been tempted to write about that, (concurrency and the radar is one, but also just plain flight-worthiness is not there) but it’s so painful to contemplate my tax money going for those bullshit aircraft which are only good at making short-range bombing runs against defenseless civilians. It makes me unusually sick.