00:24:11 · Jordi Hays, John Coogan
[00:24:11] S0: No. I think we I think we removed. [00:24:12] S1: I think it's too too too for the show. Oh, here we go. Tyler has the truth nuke. Gemini plutonium is coming in any minute now. No. The truth nuke is good. Really slow still. The longest soundboard effect we have. But this this all feels like, okay. A year ago, people were talking about, like, the naming schemes at the labs don't make any sense. OpenAI was probably the worst offender with, like, four o, and then, like, you'd go to four five, which was bigger a model, but it was slower. And then there was, like, Omni versus Fastmo, and it was, like, too much. And everyone was like, I don't get it. And then and then it was, like, five four, five five, five six, And then, eventually, OpenAI said, okay. We're we're going to use a human legible, not just purely numbers. We're gonna go Sarah, Luna, Soul, Astra. And that, I feel like, makes a lot of... A lot more sense to people. It's just more intuitive. And so I like that OpenAI has adopted that format. I like that Google has adopted that format. And I hope that the version numbers kind of melt away. And you're just in the world of, like, okay, like, big model for this task, small model for that task. And and it's a little bit it's a little bit easier to just intuitively get a grasp of whatever model is capable Anyways. [00:25:34] S0: But... Amazing benchmarks. Good luck. Excited to see the reactions. Yeah. And... [00:25:38] S1: And the ramp data. [00:25:39] S0: Yeah. In other news. Yep. OpenAI.
00:25:39 · Jordi Hays, John Coogan
[00:25:39] S0: Yeah. In other news. Yep. OpenAI. [00:25:43] S1: First, let me tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless real time experiences and new value with Cisco. Take me through it. [00:25:50] S0: OpenAI plans to buy Google's chips as part of its AI build out. The new... Like, first we had... So now we have TPUs. Mhmm. Last week, it was it was OpenAI models on GCP. Yeah. Feels like a start of a new era? [00:26:10] S1: Yeah. I mean, sort of a new era. I don't know. I mean, yeah. I guess OpenAI was founded because they were like, Google's bad. Right? [00:26:19] S0: Yeah. That's what [00:26:19] S1: I'm saying. Yeah. That's a good point. Now, it's like, they're not that bad. But this is the same thing when I get there off again, Elon. It's like, you know, very very diametrically opposed, you know, politically. [00:26:30] S0: But yeah. And... [00:26:31] S1: They got compute when you compute. Yeah. So Yeah. Business above all. Business solves everything. Let's give it up for business. [00:26:39] S0: Let's give it up for business. [00:26:41] S1: Business solves everything. [00:26:42] S0: Dylan Patel says semi analysis clients had this already. Woah. Screenshot. [00:26:48] S1: Little victory lap.
00:26:48 · John Coogan, Jordi Hays, Ben Liu
[00:26:48] S1: Little victory lap. [00:26:49] S0: AI loves largest lease ever more massive deals in the pipeline. AMD taking share, question mark. OpenAI's quarter is one eight gigawatt multi phase commitment in Ohio with NVIDIA backstopping 4.25 gigawatts. It's the largest single leasing commitment we have ever seen. Cool. In addition, we believe that OpenAI is discussing with NeoClouds for both GPU a as a service and TPU as a service. We think there's a one gigawatt TPU commitment, and OpenAI is looking for sites. Yep. Anthropic leased 1.7 gigawatts this quarter, with notably three sixty megawatts in Queensland, Australia. Let's give it up for Compute Down Under. Signed on the September 16 with operations starting in the first quarter of twenty eight. The largest overseas leased by an AI lab this quarter. The 1.7 gigawatts includes one gigawatt backstop by Google, but also a large site where we aren't sure yet who will provide the credit support. [00:27:42] S6: Do you [00:27:42] S1: think AI researchers are gonna be okay with doing inference in Australia? That's kind of controversial, right, because they're known for throwing another shrimp on the Barbie. That's rough. You're gonna have to do a truth. It's like, no more throwing shrimp on the Barbie. Just fosters. Just enjoy your fosters. [00:27:59] S0: Enjoy your fosters.
00:27:59 · Jordi Hays, John Coogan
[00:27:59] S0: Enjoy your fosters. [00:28:00] S1: And and stick to inference and fosters. No more shrimp. Anyway, let me tell you about CrowdStrike. Your business is AI. Their business is secure. CrowdStrike secures AI and stops breaches. Yeah. Oh, sorry. Just to close out that. It's it's just like every lab is using every part of the animal. Like, it's the, you know, they're fully... [00:28:19] S0: They're going whole hog.
00:28:19 · Jordi Hays, John Coogan, S2
[00:28:19] S0: They're going whole hog. [00:28:20] S1: They're going whole hog. They're using everything everything they possibly can. Every lab is using every chip and every data center and every hyperscaler and every Neo Cloud. They're using it all for all different things, and different chips will be better at different models and different part... Pieces of the inference stack, and, like, everything will be used. Very good for the... Like, if you've been in semis and and, you know, Neo Clouds and all this stuff, rising tide will, like, lift all boats as everything becomes... Like, there was a moment where it was, like, oh, if you made the wrong architectural decision, like, you would get smoked. That was the theory. And now it's, like, oh, actually, these models, at least a piece of the models can be, like, used in niche areas for different things. And so even if even if one chip is better at something, faster, but that... What's less smarter smarter but slower, like like, there will be an optimal way to to get a model running on that. And with AI agents, it's pretty easy to port to different things. I I think we saw that from the AMD coverage we did with Lisa Su where Anthropic vibe coded their AMD implementation, like, pretty quickly. So good for anyone who's betting on oligopoly in silicon, rougher if you were like, this will not be a car like market, but still very, very early. Still a lot of dependence on NVIDIA, obviously. So nothing too scary there, but just interesting that everything is getting used by everyone. [00:29:44] S2: So Yeah.
00:29:44 · S2, John Coogan, Jordi Hays
[00:29:44] S2: So Yeah. [00:29:45] S1: That's kinda my update. Anyway, open source open source... [00:29:50] S0: Yeah. So Dax over at OpenCode posted yesterday. He says, to put what I'm saying about demand in context, OpenRouter is the most successful model model router. You can almost serve a 100% of their deep sea traffic yesterday on a single GB 300 rack. This is worth 6 to 7,000,000. Even if there's 10x this demand, secretly, that's still less than a 100,000,000 of compute. [00:30:14] S1: This is crazy. This is crazy because I don't believe... I mean, I don't know. Maybe people are inferencing DeepSeek from Neo Clouds directly and not going through Open Router, but I would be very skeptical that there are large enterprises out there that are saying, oh, yeah. We have a million dollars a day of inference. Yes. A lot of [00:30:36] S0: people are saying, well, this doesn't... Yeah. This doesn't factor in people that have [00:30:40] S1: Their own CPUs on prem. [00:30:41] S0: Yep. But the the elephant in the room is the the base tens, the Yep. The models, the fireworks, the together AI. [00:30:50] S1: So it's going through that and it's not going on [00:30:51] S0: at Lambda's, you know, like a very very long list of Okay. Of other clouds that can [00:30:56] S1: Okay. [00:30:57] S0: That can serve these models. So I would expect that that is where that is where the that's where the billions of dollars of of inference spend is happening.