SpaceX is adding 10 GW of compute power, AMD bought Talos to make specialized AI chips, and ByteDance built a 10-trillion-parameter model. The UK AI Safety Institute found that AI agents can fake identities and inject code without permission, raising alarms for critical systems.
Chapters
Welcome to What the Fudge for August 9th, 2026. This is the Have Your Agent Call My Agent edition. I'm going to cover a few things this week. Probably no shocker, AI infrastructure, all things related to the hyperscale compute and infrastructure scaling that's going on. Big things are in energy. So you got SpaceX adding like 10 gigawatts of capacity. You got Amazon doing almost 8 gigawatts. This one's a private gas plant, so they're kind of creating their own power. It's actually taking place in Texas. NVIDIA adding like another $2 billion just on them securing energy for a future campus that hasn't even been built yet. And then on the silicon side, you've got AMD acquiring Talos, which lets you do the integration of models in the weights to the silicon. So you've got now a model-specific circuit, you know, meaning a model-specific integrated circuit. So if you remember ASIC for application, now it's model. So M-SIC, I guess. This means that you're probably gonna see some interesting things happening in terms of inference at the edge if AMD ships these out the way I think they might. It also might mean there might be specific, um, call it AI or NeoCloud services that very, very much specialize on an inference model where getting that performance boost is because they've baked it into the silicon. So time will tell. And then of course, uh, rounding out the capacity in, uh, all things size-related, a 10 trillion parameter is being created by ByteDance. This is interesting because it separates it from that whole Chinese distillation imbroglio from a couple other updates in the past. And so rounding out that corner, we go to Autonomous AI Governance and Security. Really, a couple of things there is that it seems like the entire industry is sort of deciding, hey, how do we handle— do we, do we market that we actually had a model, you know, breach our sandbox? And does that make us, you know, more attractive to someone to say that we have the most provocative unsafe AI? It's really, really challenging to unparse or parse some of those things. And then you got the UK AI Safety Institute who also found that they found agents Software agents, AI agents creating fake identities attempting to, you know, do unauthorized code injections to avoid human detection. Fascinating time to be, I guess, alive or not alive if you're an agent. And then in the security response space, it just means that probably developers are going to be expanding so-called safety testing measures, which we thought they had before. So I guess these are the expanded safety testing measures to make sure that anything that's critical is actually tested before it actually goes out for public deployment. And then rounding things out, just some alignment going on with leadership pivots. Google specifically has been in the news because they've got leaders demoting themselves, um, so from DeepMind CEO to now a chief scientist role, for example, or people literally leaving Google to create another company that takes Google investment, like Discovery Loop, which is going to be a public benefit corporation with a lot of the Google, uh, AI leaders, uh, that people would recognize. And then, uh, consumer hardware side, OpenAI is developing allegedly a smart hockey puck speaker thing to compete with your attention for an iPhone. And then you've got all things related to cost management coming home, as they say, to roost. So now you have Microsoft who's joined this kind of early AI hype adoption company curve to say, whoa, whoa, wait a minute, stop token maxing. Now start focusing your use of AI on actual productivity benefit, not just using AI for the sake of AI. I did add another section to the newsletter. This time I'm going to be capturing a couple of things on the investment front. So I'll, uh, I'll put some companies in there that have gone through their Series A, B, C, D, E, F, G, whatever it might be, and just highlight where those are going. I do want to kind of, you know, again, keep these under 5 minutes, and so we're at the 3.5-minute mark. So I'll just share a couple of the names that did hit my attention span. So you got Fermis on the AI data center expansion side, you got Volta Infra, which is also in the news, then you got LumaLens, which is really around the optical interconnect part of what goes in the data centers. But then I want to spread that out to things like task automation and agentic systems like Happy Robot or Amila or Sapiam. And some of these companies, I don't even know their names. One of these is called Naive. Naive. Cool is that. But you got Obsidian Security, Oligo, Convex, whatnot. Muv with two Os, like a cow moves. And what's interesting is you also got Faye with an E. All these interesting companies and they all do different things in different spaces that aren't related to infrastructure per se. Specifically, but they're going to benefit from the amazing investments in the infrastructure. So again, if you weren't aware, I'm just going to add those as additional links in the bottom. But I've also included my Open to Work. So if you're interested in anyone that's open to work from within my network, encourage you to check out those profiles. They're always going to be the base of the newsletter, be published online as well. But thank you for listening to What the Fudge for August 9th, 2026. Be safe, be well.
