What The Fudge for September 20, 2026: Design Fiction, Autonomy, and Kinetic Infrastructure

Welcome to What the Fudge for September 30th, 2026. This is the Design Fiction Edition. I've changed up the format a little bit this week, and I've looked at the workflow. So this time around, I'm going to go a little deeper into a couple of things, starting with macro-level views of what I've seen in the news and sort of what's come out in the design fiction I had been reading recently. One of the many blogs I subscribe to is called The Adjacency, which is part of Near Future Laboratory. If you're not familiar with that, I've linked to that into the notes. But there's a great copy of material that goes into this. It's kind of— it's provocative because it's fiction and it's design-centered, but it asks those kinds of future-facing questions. How would you— kind of negotiate or deal with some of the new technology that's coming into being. So in this case, it's looking at what would happen if the Consumer Price Index was pegged against AI inference or if we assume that free computation of tokens and the flow around it is like part of Maslow's hierarchy of needs. And so when you kind of go down that path of imagining what that might be like, you begin to think, wow, the future actually might be a little closer than it appears. It's a fascinating article. I've linked to it. What's interesting about fiction is, truth can be stranger than fiction. What's also interesting about some of the coverage this week is how on the money some of the fiction lines up with what we're seeing in different parts of the market. If you just look beyond the macroeconomic, just in terms of software engineering, we're now 4 years beyond what would have been the literally cut in half of demand for so-called software engineering talent. Whether or not you believe, as I do, that that's mostly related to post-zero interest rates, where you had a lot of overhiring, so now we're seeing more automation or supposedly claims of developer productivity going through the roof with all the AI and the orchestration. I look at this as maybe some of this is both true at the same time. You can think of it that way. If you just look at just the Bay Area, there is probably half of the software engineering demand there would have been, say, 4 years ago. But if we then pull back a little bit further and look at what's going on in the infrastructure world, I see some kind of familiar patterns. So the first is we now have an example of geopolitical events and the fragility of— one of the tenets of the internet was it would be resilient to war and the patterns of war. Routing around the problem, as they say. So what's interesting is AWS is now saying that they've lost data related to some of their facilities that were attacked in drone strikes. Those facilities, in some cases, are mostly or somewhat offline. And so if you think about this, we now have this kind of world of where is the next disruption going to come from? And so there's even plays that, just to highlight AWS one more time, they're going after different backup generators, how they're going to source those backup generators for other facilities. that I don't think they're intending to have attacked by drones, but it asks the question of what amount of a data center would need to be fortified from, say, drone attack, which might be one of the new patterns of kinetic warfare in our future. So if you then look back again into the past of patterns that I think are familiar, I was involved with converged and hyperconverged infrastructure from roughly 2010 with Acadia, which became the VCE or the Virtual Computing Environment Coalition, which was VMware, Cisco, EMC, and Intel. kind of putting together the cloud on a can solution. And so now we're seeing similar companies. Crusoe is one of those examples of, they kind of reinvented themselves a few times, but now we're seeing that same kind of factory-built or prefab data center design coming in. And the difference is now it's about things that are related to an AI workload. So inference on demand or inference cloud on a can, if you will. So then if you pull a little bit further back, a little further into the, supply chain or the value chain, you've also got manufacturers of chips, which I had predicted in 2021, we'd see a lot lower-cost chips. But that didn't clearly happen that way. It went the other direction. The demand was so high that now the price is really just through the roof. Now, we're seeing not just wanting to build more chips, but building more chips on a sovereign basis. There's a couple of examples here of You've got US manufacturing of chips that might be taking place where global providers are building into a United States manufacturing model. That'll be interesting to play out because you can expect that if the US is gonna do it, there's gonna be other nations that'll attempt to do that, whether that's Japan or China or other. But it'll be fascinating to see where there is an appetite to truly do chip fabrication, if that is indeed how we're funding— Yeah. or incentivizing new manufacturing to come into on a sovereign basis within a given country's borders. The next piece of this, of course, is what runs on all of this stuff, which is the models. And so this week, I think, was the week where it was a question— not a question of if, but when all of the major front-end labs would tell a story about how their model was so sophisticated and so complex that it somehow jumped out of the confines of the sandbox, whatever a poorly defined sandbox might be. And so this time around, it was Google. And so the question then becomes not just, is it a me-too marketing efficacy parity claim? Is it over-anthropomorphizing these agents that are not people, they're just software? But the question is, does the Frontier Lab know what it's doing? And so now there's this concept of, well, what if we bring in a trusted third party to do evaluation? And keep the Frontier Lab from quote unquote self-managing, self-governing, whatever it might mean, bringing that trusted third party to do that. And so the first example appears to be, it will be most likely Accenture, and that will be Accenture working with Anthropic. And so Accenture, the global system integrator, would be the, the kind of like the new evaluator of Anthropic so that Anthropic can, I'm assuming, outsource that to Accenture and then focus on the Anthropic things that Anthropic wants to do. So there's 2 ways to kind of splice or slice this. One is that now the GSIs might have a new lease on life, because I think AI was supposedly going to take down some of the GSIs. But now in order to support the thing that takes down the GSI, perhaps the GSI becomes the evaluator to ensure how quickly, I guess, the Frontier Lab and the AI could take out the GSI. So it is kind of interesting. But the last part of that where you really get into some interesting discussions is most AI to this point that's been consumer-facing or friendly has been around, you're chatting with it, you're working with it, it's producing things that are human interpretable. There's a new form of company that formed called Typesafe AI, and they got some news coverage this week. I think funding has a deal— has a lot to do with it. But effectively, it's designed around more numerical responses, probabilistic estimates, things that really kind of Separate what would be the hallucination of the LLM, or so-called the, you know, the higher the temperature, the more likely to be creative or hallucinate. So this is more of getting you back to the mathematical readout of what the response is. And a lot of that might have been under the covers for some of the earlier models. But as we got more consumer-facing, there was less appetite for so-called, you know, bits to fiddle with or knobs to turn or some of the kind of additional information that would be really lost on the end user. But as we get more sophisticated in our use cases, these highly numerical calculated or calculation-ready responses that give you probabilities and estimations are probably far more likely when we get into this concept of what we would consider to be a reasonable or acceptable response. So kind of interesting to see that playing out. And the way all this sort of ties back together again with what I was describing before is what happens when design fiction You know, the truth becomes stranger than the friction, the fiction, no pun intended. This is around the notion of server hardware. So I'm old enough to remember when Apple made servers. And I don't mean like the Studio Box or the, you know, the Mini that sits on your desk as a puck of silver goodness. I mean, actual ready-to-be rack-mounted 19-inch, you know, rack mount, not to be confused with the telecom guys. But the 19-inch rackmount Xserve, it was a really kind of slick-looking design, kind of like a polished stainless steel kind of a front. It was iconic looking, and it was really well-designed hardware too, but it didn't really last. And so, you know, the Dells, the HP, the HPEs of the world, etc., sort of took that rack mount business away, you know, the Supermicros of the world. But Apple did make X-Servers. And so it turns out that what's old might become new again. Apple might actually put their own servers together using M8 Ultra chips. Those would become localized AI servers. So similar to how people are doing with Mac Minis and Studios now, where they load them up with RAM and then put local models on them and then do private AI. This would be a building block for those AI factories. So as I said before, you've got these prefabricated, pre-engineered data center in a can models for inference, or designs for inference models to be run as workloads that are being produced by these other companies that are in the physical big build approach that's not as big As a large, massive data center, which has a longer timeline, these kinds of prefab, you know, pour a concrete plate, small, bring in power, maybe it's a gas line, you know, connect here, connect here, boom, up and running. You have private AI infrastructure, kind of, again, that kind of ISO 20, 40-foot container model. But here's Apple figuring out ways to take that same approach of localized AI. and put little AI factories in the hands of potentially another enterprise market that Apple really hasn't tapped into in quite some time. The Xserve is quite a bit old. I remember that back in the 2000s. So long story short there is, well, what could you fit inside of something like that? That's where companies like Prism ML come in, where they're still compressing even more of these AI models into smartphone kind of sized packages that would allow for very modest hardware to do the kinds of things that will be increasingly required for unlocking robotics, physical AI. And so there's a couple of interesting stories that I've linked to. One is the latest from PrismML. The next is the Semi-Analysis newsletter, if you haven't read that before. And it really goes into some detail around what's happening as these models become smaller. And because they're smaller, There's an ability to put that AI as close as possible to the physical events around us in the form of robotics. So it's fascinating. I know I went way over what was previously my 5-minute mark. In fact, I've gotten it to about 12 minutes this time. But again, I wanted to think about the format a little bit differently, start with a theme, and then look at how the week's coverage has supported not just the theme, but where it might be going in the future. And so I'm crossing over the 12-minute mark, but I wanted to thank you again for your time. For your listening. And I thank you for being a part of What the Fudge for September 20th. 25 of these things. And so I took 2 weeks off, but hope to be back in the same format. And I'll try to keep it maybe under 15 minutes. How about that? Okay, take care.