OpenAI's new GPT-5.6 variants and PrismML's billion-parameter model running on iPhone 17 Pro show on-device AI is becoming real. A memory-hardware boom is driving up prices now but should stabilize in two to three years. Meanwhile, EU scrutiny of vertical integration and U.S. state data-center moratoriums are reshaping where and how big tech can build.
Chapters
And welcome to What the Fudge podcast for July 12th, 2026, 18 times in a row edition. Let's keep this under 5 minutes. Starting off, I might sound like a broken record, but let's talk about AI. Uh, this week has really been interesting. You've got some updates that are shifting from the cloud-only or someone else's computer story to what happens as you shift towards thinking about on-device AI. And that could mean a very private AI, a sovereign AI, or something that is just, again, just on your device, the computer you have in your pocket, most likely a mobile phone. In this instance, you've got OpenAI taking out another page from the playbook of names you might forget in the future, but might be just ones we have to contend with for right now. They've released GPT-5.6. So in terms of version number, the other thing that occurs is most of these new releases are also coming into different variations. So You've got SOL, Terra, and Luna. I'm not sure I'm going to spend a lot of time memorizing these. I think of this as more of what you're describing as what's the light edition, which means the computational cost, whether that's inputs and outputs and tokens, etc., versus what might be considered more in the realm of privacy or so-called cybersecurity around the use of these models. The question will be, of course, is this exclusively something you could do only in a cloud or Are some of these models going to fit into the smaller language model form factors or things that could fit in the available RAM and computation on a multi-core capable on-device AI approach? And so it'll be interesting to see if you, if you look at this as companies that are not the big ones. So think of things like the startup PrismML showing what you can do in terms of putting a multi-billion parameter model on humble hardware, if you consider an iPhone 17 Pro to be humble hardware. It is interesting to think about this. This was, uh, something that would, uh, you would describe as perhaps theoretical a few years ago, but now we're starting to see the appetite for putting some of these models on devices that are not essentially in a cloud somewhere. So more to come on that front, I believe. I think this progression from only in the cloud to more private, more sovereign, more on-device is certainly going to be a growing area. So related to that, uh, great story if you actually have hardware that has sufficient memory. And so where would that memory come from? We're kind of waiting for the mass— massive capital investments in infrastructure to actually create chips, meaning high bandwidth memory or HBM. What happens when you need more access to DRAM and how you can have that appear in these products? So we're certainly seeing hardware products going up in price, but we're also going to be seeing some of this infrastructure investment in the ability to manufacture, ship, and distribute coming online, I think, probably within the next 2 to 3 years. So I don't know if we'll get into the, again, cheap, cheapest chips category, as I had mentioned before during the prior, uh, podcast updates, but it's certainly something to think about. And I think that most of the stories around the increase in cost of hardware right now are going to be a little bit on the temporary side. I don't know where we're gonna see subsidy of this, but, uh, never say never. And as a result, we'll also probably see some changes in what happens when we have cycles for regulatory inspection of some of the larger companies that are entirely benefiting from cloud consumption. And last week we talked about how Facebook has opened up their infrastructure to be more like a competitor to some of the infrastructure as a service companies they rely on today. And so as we see these players start to shift their roles, pretty good chance we're gonna see regulatory inspections coming from different places. I tend to follow the EU Commission. Whatever they are doing, I think that speaks to what might happen in other companies and countries, uh, or countries that have these companies and how they, uh, look at those companies, inspect them for, uh, so-called vertical integration as compared to, say, a monopoly. So when we think of this, um, and how that might tie into devices, it'll be interesting to see how this plays out for Google and and for Apple, for example. And then in terms of some of the other things and changes, just to keep this under 5 minutes, I think that we're gonna start to see some of the larger companies transforming, shifting the sands, if you will, moving a lot of team members out of roles into other roles. So this week was of course Microsoft and Xbox, and you have to imagine there's probably gonna be some changes in other adjacent or me too types of additions to these large companies. And then the last piece would be around some of these new bills that are coming in with moratoriums or limitations on new data center construction. Certainly here with the United States, we're on a state-by-state basis, basis. So keep an eye on that. And I'm gonna try to keep it under 5 minutes. But thank you again for the 18th edition for What the Fudge podcast for July 12, 2026. Take care and be well.
