Broadcom raised $60 billion for custom AI chips while Anthropic hired Google's TPU lead to build its own silicon. OpenAI cut GPT-5.6 token prices 20%, and Stripe's $7.5 billion acquisition of OpenRouter will streamline routing across providers. Open-weight models now match closed models in half the training time, shifting the competitive landscape.
Chapters
Welcome to What the Fudge for Sunday, August 23rd, 2026. This is the Moats and Power Moves edition. This week we're going to take a look at some of the things related to debt, equity commitments, or all things related to the AI infrastructure around us. Some of the shifting economics around the models that are running on top of that infrastructure and how it's getting routed, as well as what's going to be taking place in the realms of AI development. So first off, if you look at what's happening with capital markets, uh, Boy, oh boy. Lots of deployment, lots of expansion. So you've got Broadcom going to secure apparently $60 billion in debt financing. That's just to support custom AI chip deployments. You've already got Nvidia in the wings making commitments around that as well. So if you look at what's going on with Anthropic on the actual services side of AI, they have decided they're going to secure their custom silicon future. So this is a potentially non-Nvidia view of the world. And they're going to be hiring Google's former lead of all things TPU. That individual will be joining. That probably means silicon, custom silicon in this case, for Anthropic. So that follows sort of what you've also seen with AWS, with, you know, Trainium Inferentia, kind of the extension of their Graviton lines, as well as what's going on with some other acquisitions, which we covered in prior updates. Then you look at what's going on with where the data centers for these chips would go. And you're starting to see some pushback on the municipal and state level where a officials are starting to examine what's going on with approving these data centers, you know, just beyond the dust permits, but just actually approval of a data center and how that's going to impact the tax structure, as well as what's going on with the power and the water that would go into it, not just the land usage. So pulling back from that and looking at the economics of the actual services being delivered, meaning training, inference, et cetera. If you look at what's going on there, you see some other interesting notes. First, open weight models are apparently now having the time or taking half the time it takes to reach the same performance parity levels or benchmarks for some of the closed models that are out there. So implication being that we're going to have more kinds of chips for more kinds of services and more options from open weight models that might be appropriate when You're just looking for something that's approaching what closed model performance would be. Then you've got this notion of lowering your prices. So we had covered some of the lower token prices from other providers. Now OpenAI has done that. They're gonna be doing that for GPT-5.6, SOL. I'll eventually probably memorize that until they get to 5.7 and give it a different name. But they've lowered by 20%, I think, for a few months. And so that means that roughly stated, you're gonna get— 20% off. $4 per million tokens for input and around $20 per million tokens output. And so if you are only using OpenAI, great. But if you're using multiple sources, where would you do your routing? Well, you're probably gonna use OpenRouter. Well, guess what? Stripe has definitely, we had talked about it previously, but they've definitely agreed to acquire OpenRouter. That's gonna be a $7.5 billion acquisition. And then just kind of rounding out, trying to keep this under 5 minutes, some of the elements on the developer front. Yeah. So on the developer front, what's fascinating about this is we talked about this infrastructure outlay. We've talked about where the dollars are coming from, where you're gonna be allowed to do it. We're talking about how the services are changing. And so on that front, if you remember Slack, which was acquired by Salesforce years ago, Slack has decided they're going to create an agentic AI coding channel. So as you enter your Slack experience, there's gonna be a dedicated channel that is just there for collaborating with you when you have AI particular coding needs. Now, you're also probably familiar with prior discussions around Cursor. Cursor has launched their own integrated agent-assisted workflows, so you can pull in things from GitHub. And this might be interesting if Cursor decides they're going to peel off potentially people that are using GitHub or Codeberg or other, and you're looking for another alternative for your Git needs that already has the agent-assisted workflows built in. So, you know, kind of stay tuned there. The second thing is that we talked previously about DRAM, NAND, flash, generally speaking, costs going through the roof, and that's just not going to change. There's actually the first indication that more or a wider swath of manufacturers are starting to pass along costs to consumers. And then the last thing I wanted to cover today was this, again, concept around that open model approach. Well, Google DeepMind, which is the Gemma open model, that's now exceeded a billion downloads. And it also means that developers have published just over 100,000 variants, and that's just in the past 2 years. So it's a blazing pace that we're at, and I'm trying to keep it under 5 minutes. So thank you for listening to What the Fudge. Thank you. For today, that would be August 23rd, 2026. Be safe and be well. Take care.
