FULL INTERVIEW: Doug O'Laughlin Thinks Microsoft is OUT of the AI Race

Welcome to the stream. Doug, how are you doing?

0:00

I'm doing wonderful. How you guys doing?

0:04

Great to see you.

0:06

Can we pull up the video that John painstakingly made this morning that completely flopped?

0:07

I'm very excited to have you on the show. It feels like that scene in Sonic the Hedgehog three which I saw where Sonic and Shadow team up and join forces to talk about capex and agentic coding. What's new in your world? Is Claude code still the top of mind or are you still churning through the capex numbers from earnings?

0:14

Are you somewhat of an agent for Claude now like you work for Claude?

0:34

I do actually. I think I mostly just move my information back and forth. You know, I have pretty much like, I think of it as like my manager, you know, like it tells me what to do and then I go bring the information and I bring to my coworkers. I bring it back all day. I'm just on cloud code.

0:39

How many, how many prompts are you running right now? Do you have any threads going?

0:58

Okay, so. I have seven. I have seven threads.

1:04

Seven that are running right now or.

1:11

No, they're not running.

1:13

Waiting for your input.

1:14

I'm waiting for my input.

1:15

We'll let you get back to it.

1:17

Why don't you just have an eighth that just.

1:18

Let's talk about. Yeah, orchestration. Have you played with Gastown? Are you thinking about abstracting yourself to a higher level?

1:21

Okay, so Gastown is pretty intense. I don't think Gaston is going to work out. I think it's going to be agent storms.

1:29

Okay. Okay, explain the difference between Gastown and Agent Swarms.

1:35

Okay, so Gastown is probably the most forward looking thing I've read a lot, talked about how you created this self healing tool process to essentially to pass all these beads across and have all these workers and ways to self repair the agent workflow process. And I read it and I was like dude, this is brilliant and also fucking crazy. It's like, it feels like the ravings of a madman. And then I proceeded. Well, I was also in my badmin era like before, before the New Year's when you had 2x times usage. I pretty much was like literally railing Claude code constantly. I think I had, I think I had four 14 hour days.

1:41

Wow.

2:22

Yeah, it was beautiful.

2:23

It was beautiful. Okay, so talk to us about what you're actually building because we're talking about SaaS Apocalypse. It feels like there's a debate over build rebuild all your tools from scratch to save whatever your SaaS fees are.

2:24

Yeah.

2:38

And even yesterday was notable. OpenAI comes out with Frontier, which is. You look at like, you gotta look at this graphic, which feels like it was made for a Fortune 500 CEO to kind of. Or management team to kind of understand it. And it's like, here's more sass to replace your other SaaS. Right. It's like you've got the system of record down here, you have a bunch of agents in between, and then you've got different applications that you're using. And meanwhile, Anthropic is just like, we're making a really smart digital guy that can do whatever you want.

2:38

Yeah. So I think the two, there's two really interesting ways. I think OpenAI is like the Fortune 500, selling it from the top, if it makes sense. And then Anthropic is like, here's your cloud code agent, sell 20,000 of them. Did you see the, the Accenture Partnership? I think that's really interesting. So like, if you're, if you go back, there's. They're doing 30,000 people at Accenture.

3:11

What do you mean?

3:35

To like 30. 30,000 people at Accenture are going to learn how to Claude code.

3:36

Oh, and then that's.

3:41

And then be deployed into different companies.

3:43

Who knows? Who knows?

3:46

What else would they do?

3:48

Maybe they just replace Accenture. I mean, that's what they're going to be doing, I think.

3:49

Okay, wait, so, yeah, what do you mean? If Accenture folks are using Claude code, wouldn't they be using it on consulting projects internally to companies?

3:54

I think they're going to be using it internally. So they're going to be using internally and then they're going to be doing all these consulting things. Because if you think about it, one of the issues is like, you know, when you had SaaS, one of the biggest issues of changing from one CRM to another was effectively being like, hey, everyone, you have to quit your jobs for like 10 months to figure this out. The implementation, you'd have tons and tons of consultants do that. And I think that that's what the Accenture Partnership is. So essentially, people are going to be implementing Claude code and there's 30,000 people at Accenture who's going to do it. And then on the other side you have Frontier, which is like the 4 Fortune 500 way of being like, here's your plan, come to us and we'll build this whole thing, blah, blah, blah.

4:05

So, yeah, but is re implementing your CRM really the lowest hanging fruit for America's greatest companies? It can't possibly. There must be New ideas, new problems to solve, new tools to build. Why are we just going to shuffle the chips around the board instead of doing something productive?

4:48

Okay, so I think the system of record refresh is going to be really awesome because, you know, like, the big. I mean, it honestly does feel kind of boomer if you think about it. It's like the biggest data now. Everyone can have the big data now, but I think the. The automation that you've always dreamed of is actually going to happen and the system of record is just going to essentially have hooks out to all these other things are going to build on top of it, which is mostly like, you know, the frontier thing. Yeah. And essentially like all the information work is just going to be like, all on the agent and everything else is going to be like place where it lives and is stored for Fonzies. So instead of me having someone let me use my personal stack at Semi Alice, we use HubSpot, for example. So, hey, the sales this quarter, we need this like, quota or who did what or what products are selling better or like, you know, what's.

5:07

How many more podcasts should Dylan do this month? If we want to hit our goal, that's totally different.

5:56

We don't actually have like our, like, you know, the big. Yeah, sure. How many more podcasts? Shit like that. Right. Um, I could just vibe code it. I would just be like, hey, can you run this analysis for me? And in a perfect futuristic world, it'll go into the CRM, pull all the information of the. All of our inbounds.

6:02

Yeah.

6:18

Make be like, hey, the day after Dylan goes on a podcast, there's like 25 people who, who come in. The conversion rate is X. You could price it at this. Um, Dylan quit your, you know, stop working and effectively just like hit the podcast.

6:18

There we go.

6:33

You know, like. Yeah, yeah, exactly. So. So you could do this with anything though. Like, it's just information, man. It's gonna be pretty sick. But I think all the SaaS companies are going to essentially just become hooks for all the crap they've built on top of it.

6:33

Yeah.

6:47

Did you see Jensen yesterday was kind of defending some companies like SAP and ServiceNow and saying, hey, if I was a really smart humanoid out doing work in the world and I needed a screwdriver, would I just invent a new screen screwdriver or would I just take one off the shelf? And so that was like his defense. Tyler here took the other side of it and just said, there's gonna be a lot of situations where especially In a software only environment, it's easier to just build a very specific workflow that you need that you would have gotten from a SaaS provider versus you don't need to actually rebuild the entire platform.

6:48

Yeah, I think we're gonna be building a lot of screwdrivers. The thing that's important is, okay, you're not going to rent a truck, right? Like you're not going to build your own truck, but your own screwdriver, 100%. You're doing this big ginormous job. You need a hammer, you'll be like, okay, pull it out of my belt. Okay. But you're not going to be like, I'm about to move 700 tons of here to here. I need to rent a truck. You're not going to build the truck. And so that's what I think the system records are going to look like. They're going to look like places where actual data that cannot be like, cannot be vibed effectively. Like what's your inventory cannot have any fucking hallucinations, right, like your erp. But all of that will just be hooks for everything else because like all the information is just like pulling, retrieving, making the correlation, running the charts like, okay, but even that all the time.

7:29

But then how do you square the fact that a system of record is way less sticky if you have agents that can work around the clock to switch you over to a different system of record? Like, that still ends up putting massive pricing pressure.

8:19

So to be clear, I don't think it's good for everyone. I think my favorite analogy of this is there actually is a very old school type of software that's existed for a long time. All the shit on mainframes, it's all out there. Yeah. And funny enough, mainframes still grew 6% a year or whatever. Someone has the real number from 2002 to 2020.

8:34

That's crazy.

8:57

So they're going to grow, but it's just going to be a very different vision of the world that I don't think people are ready for. And the adjustment period is the big problem because all the stocks are priced like they're not going to be mainframes. And also, just for context, mainframes, there's like, hey, there's one of each company now. There's not like 10, just one each.

8:58

Yeah. So I think with all these, whether you're a system of record or you're some vertical software, you're going to need to show insane revenue growth in a truly AI native product. Otherwise investors, I think are going to continue to not be able to create a super compelling narrative why you should own it during this period of uncertainty.

9:21

Yeah, I mean pretty much what happens and we're going to go like investor brain. When anything goes X growth, the multiple goes massively down.

9:44

Yeah, yeah, that makes sense.

9:52

Eight times earnings.

9:53

Can you talk a little bit more about how what you're actually coding what you're building, like what the software is? Because from the demos that I've seen that you've posted, it feels much more like you have an agent that can do knowledge retrieval, data transformation, build dashboards, charts and like knowledge work as opposed to truly replacing software tools at this point. But have you built anything that's like long lived and runs like daily or is like something you keep revisiting because it's now a piece of software that does the job.

9:54

So the cloud code commits is now software that lives and runs every single day. Okay, that's like a scraper. Right? And then like that, that like lives in a database and that that will run forever. There's like a lot of other tracking price data tool stuff. Like a lot of the scraping that we're like that is not like publicly available. Like we do like a lot of that. Like we had a data team just do that and now effectively we can really accelerate that. So everyone can do that.

10:31

Sure, sure, sure.

10:56

There are other little things that I think are heuristics like little skills of I have blind spots that I consistently make over and over and I'm like, hey, I know this blind spot's an issue, blah, blah. You should consider this in this case. I don't think it's like the Galaxy Brain software and we're very far from there. Because if you actually play with these tools, a lot of context. Rod is real.

10:57

Yeah. So but how fast is it getting better? Because it feels like we're seeing the.

11:20

Meter graph scary fast, scary fast, scary, scary scary. So I started vibe coding with Claude 4 and it just wasn't or sorry Opus 4 on quad code and it just could not one shot websites in the way that 4.5 and 4.6 can. And if it just marginally improves from here, it feels like why would I pay for any kind of UI UX if it's just going to be generated at a good enough quality.

11:26

Yeah. How did you process the new models this week? Four, six, five, three. What's the review?

11:52

If you can't immediately notice the difference between 4.5 and 4.6, start polishing your resume. You are Cooked?

12:02

Yeah. You just got automated by an agent. I think 4.6 was a little disappointing, if we're honest with you. I think it might have been Sonic 5.

12:09

Oh, that's what people are saying. That's the conspiracy theory.

12:24

Right.

12:27

But what does that mean?

12:27

The original Sonic 5 leaks were that it's as good as Opus 4.5 but with 1 million context window and specifically trained for agent swarms.

12:30

Sure. So yeah. But practically does that just mean like same quality but faster, cheaper, at least.

12:39

For Anthropic and then they. And they make more money?

12:45

Oh yeah, yeah, yeah, yeah. Better margins. How are the margins looking for the labs right now? There's, there was a bunch of like FUD around it, but it seems like from all the leaks it's been like 50, 60, 70% pretty good.

12:48

Yeah. If you X all the free users it's always really good. Right. Honestly, Anthropic has no free users or like on a relative basis. So they, their margins are ironically on a like to like basis. Kind of not as good as you think.

12:58

Yeah. Can you break down a little bit more of the thesis of the Claude code is an inflection point article, what the key takeaway, who you're speaking to, what update you wanted to share and then I want to go into some of the pushback and your response to that?

13:13

Yeah, sure. So first I think the thing that makes me really excited is the first time since chain of thought I feel like we have a new scaling that feels very, very different and hardcore and I can actually see my entire life day to day change. I think I can expect some version of a cloud code harness to be effectively all my information work from now till the future. I am a daily user. I was not a daily user before and I expect to continue to be one.

13:33

And that's kind of what happened with the reasoning models. It went from like you could ask it stuff but it might hallucinate to like the answers are good, like you can pretty much rely and there's going to be citations and like it's going to be 99.999% usable for things. So you just have a question, you get that it may be not great at certain things, but in general, like it delivered on the initial chat experience that I think a lot of people were looking for and then they became dozers. Yeah.

14:05

How much do you think Anthropic cares more about winning in consumer than they've let on to date?

14:33

No, I don't think so. Everyone, everyone who works there is exactly like, like what you think it is. They're exactly who they say they are.

14:42

Yeah. They're software singularity pilled.

14:52

Yeah. And then I think cowork is what they're really excited about.

14:56

Sure. Makes sense. Yeah.

14:59

So they're not even thinking about a scenario where a bunch of people are using Claude in a work setting and say, hey, this is pretty great. And yeah, ChatGPT has ads. I'm happy to pay 20 bucks a month. I'll use it personally because I just think there's like an iPhone. I think the game to get to 3 billion users is like over. When you just look at the traction of like Gemini and ChatGPT and the fact that norm normal people aren't caring that much about the nuance, maybe that don't have that much to automate in their life, but there's like an iPhone size market. Like the iPhone wasn't the first smartphone to launch and it's possible. Like when I see this, like when I see this like the super bowl ad, the sort of like trust nuke I was calling it, right. Just like, hey, like it's really funny, they're like, you know, rage baiting open AI, but at the same time they're just destroying like trust around ads and LLMs potentially like permanently. Right. Because people, even when they start seeing ads that are more like display ads, they'll start thinking, well, like was the result influenced too? You know, it just like it hurts the trust. And so I think they're, I think my, my theory is that any product that like really catches on in the workplace could very well trickle over into life and anthropic could someday have a pretty big. They could have like a Netflix sized subscriber base for people that just want an ad free AI experience.

15:02

Yeah, that sounds completely right to me.

16:27

But you're saying your point is like, it's just secondary to them. They're like, it's a nice to have but like we don't. That's not our intention, I think.

16:30

Okay, so singularity pill. But I also, I think you have to pay for the singularity and I think it's going to be enterprise that does it.

16:39

Mm, Makes sense.

16:44

Yeah.

16:45

Yeah. I mean the other take on like you could wind up being like the Apple and like the premium, you know, privacy focused or you could wind up being like the DuckDuckGo which was like, yeah, you know, yeah, it was a counter to Google, but it never got to any meaningful scale.

16:46

Yeah, but I still think, I still think OpenAI is the apple. Like Apple was Synonymous with smartphones. When it really took off, like, what is the other smartphone? Nokia. Maybe you can argue this is like a BlackBerry. Yeah. I can't name it. Yeah, well like BlackBerry, right. It was known for work.

17:00

Yeah.

17:21

And then obviously like it swapped over. So I still think OpenAI is like the cognitive referent. And honestly, 5.3 cooks.

17:22

Yeah. Faster or just better or both?

17:28

Faster and better.

17:32

Faster and better. Okay. Talk about 5.2 token efficiency. Rune was pushing back on the article saying you're making the assertion that 5.2 token efficiency ruins long horizon planning and yet 5.2 tops the meter chart for long horizon planning half baked. What's the explanation there?

17:33

Didn't someone completely mock that argument? But he's kind of, he's like, sorry, I gotta find the guy. But it's like, I don't know what task is being done here. Like is art. Are they the same hardness?

17:55

Yes.

18:10

Did you just spam it to infinity and like you finished like a sufficiently long task of completion versus, versus, like, okay, let's just say we have two kids taking the SAT and one does a better job and finishes first and one does almost as good of a job and took seven times as long. And you're like, wow, that one's a smart kid.

18:11

Yeah.

18:30

No, dude, that doesn't make any sense.

18:32

Yeah, yeah, yeah, yeah, that makes sense. Yeah, yeah. An elegant solution delivered faster is uniformly better.

18:34

Yeah, 100%.

18:40

That makes sense.

18:41

Yeah. And so if you spam more tokens and you win and you're like, oh look, I mobbed them. And it's like, dude, what if you just use less tokens?

18:42

I think the benchmark is supposed to be for they have a reference class of projects that are supposed to take X amount of time. They would take a human developer six hours. And then they have all the models compete. And if you can compete the six hour task, then you get put at the six hour mark. It's not. Did you run for six hours? So it could be like implement a CRM product or you know, write a very complicated, you know, database or something. It would take, you know, a talented software developer six hours, two hours, one hour. And they have different tasks. And then you're trying to climb that hurdle.

18:50

Oh yeah. And then it's okay, it climbs higher and higher.

19:27

I think, I think that's loosely. It's because. Because obviously you could just say, okay, just reasoning, count to 1 billion and just go as slow as. And it works for days. And that's not impressive.

19:30

The. Yeah, I mean it's okay, so yeah, you're right. The, the different, the scaling thing. But like, okay, so one, the other thing I was doing, like now that we have vibe coding available to everyone, you can just have it do the same task and do like abc. Like I. I've been doing like a lot of internal benchmarking. Like everyone could benchmark guys like, dude, Codex 5.2 took so long and just never build for me.

19:42

Oh, interesting.

20:03

And it's like all the Codex hype during like it just never worked for me, man. Like never. It never one shot projects like Opus 4.5 did. And I'm just like, this feels like complete FUD. But that being said, Codex 5.3 cooks. I think I take everything back about 5.2 for 5.3.

20:04

Nice. Total reversal. Classic AI narrative. Just day by day, complete switching of the narrative. Talk about the NPM downloads because you said that you're now scraping them every day trying to understand how many commits on GitHub are related to CLAUDE code. And the pushback from Rune was that this counts NPM downloads as authoritative when Claude code numbers are hugely inflated. Because GitHub Actions does automatic CLAUDE code download every time. Continuous integration CI runs versus Codec's compute cloud. So maybe it's not apples to apples. What's more, nuance on the fast takeoff of CLAUDE code? Because honestly, when you said 20% of commits by the end of the year, I was like, that feels extremely low. I would expect like 70% and I would expect Codex to be at 30% and no more human commits because it's working.

20:22

So I wanted to make sure we have like a high standard, like a high 95% plus.

21:12

Sure.

21:19

I don't think like, sure. If it continues to grow on a week. On week basis. Like, yeah, it's like, you know, 100% by June or something like that.

21:19

Well, there's also the fact that like you could be writing code and still just like almost be using cloud code as like your linter or like your, your interface to GitHub. And if there's an abstraction layer there that people adopt, you're going to see the commits go through the roof, even if there's still a human in the loop meaningfully.

21:26

Look, look, look. Yeah, I'm not going to pretend like the cloud code commits thing is like the cleanest way ever. There's a lot of way. There's a lot of ways to fuck the data. For example, people who use who you could just say don't do this and won't do it. Number two, like private on a ratio is like five times bigger.

21:42

Sure.

21:59

That matters way more. And then like, I also think that, like, the way you consume it, like this doesn't count for cursor. People have been clearly using AI for like a long while and it doesn't show up. This is just the example. I can like, say like, hey, chart goes up really quickly.

21:59

Yeah, yeah, yeah, yeah, yeah. It's so cool.

22:16

It's not perfect. It's a data set that I create in a relatively short amount of time. Yeah, like. Well, seems pretty cool.

22:17

Yeah, yeah, no, it is.

22:25

Should we talk about Amazon?

22:27

Yeah, yeah. Let's move over to hyperscalers reaction.

22:28

Was the number two low?

22:31

Yeah. Bearish. Yeah. They're not taking it seriously. Yeah.

22:33

Future billing was crazy. Yeah, that was crazy to me.

22:37

Okay. Why?

22:39

Yeah, that was really shocking. You know, we do a lot of data center tracking and we do a lot of accelerator tracking and we were too up.

22:40

Okay. Are they, are they. Are they trying to play some sort of hype game where they're throwing out the biggest number and they're not actually even going to be able to buy enough equipment to spend it? Even if they are signaling the market. We're going to be hearing like, well, we wanted to buy this many Nvidia chips but we couldn't get them, or we have a delay at this data center because of regulation. And so they're just trying to project strength because they're sort of behind on the AI narrative a little bit. They don't have the big position that Microsoft does in OpenAI. They don't have, have, you know, a deep mind level team. And so they're saying we're going to go biggest on the dollar front, but then maybe they don't deliver on it. Or do you think at the end of the year we'll be like, yeah, they spent 200 billion?

22:49

I think at the end of the year they'd be like, yeah, they spent 200 billion.

23:33

Let's go.

23:36

They are the single biggest provider of power in the entire world. I think like the incremental and the AWS like supply chain can ramp a lot quicker than anyone else. And every, every example that we track in the data center, like the data center team, they are on time and can scale to like levels that are crazy.

23:38

Yeah.

23:58

Like Rainier's ramp is just like out of this world fast compared to everyone else. Every other gigawatt project is essentially delayed and they're going to be like ish on time.

23:59

Wow. So isn't that extremely just like Good for Amazon. Like they're properly positioned, they're properly transitioning. Like they. Yeah, like, who knows what happens to the rest?

24:07

Part of it.

24:17

But eventually this thing is just like part of the.

24:18

We were talking earlier, like, Jassy didn't exactly paint this, like incredibly exciting vision and share like, hey, we. You guys are actually underestimating demand still. Even if you're bullish on AI, you're underestimating demand. And we're in a position to actually try to get a more accurate read here. And that's why we're investing.

24:20

Yeah. And it's funny because they could have said one thing that would have made the call better and they'd be like, yeah, we see high 20s and like the stock would have ripped. But they're like, we continue to project to see this level of growth.

24:41

What percentage, what percentage of the 200 billion do you think will actually flow to Nvidia? Because Nvidia is rallying today. That's why we're wearing white suits. But it didn't rally immediately, in after hours.

24:55

I think a meaningful amount. I definitely cannot disclose what semi analysis thinks, but I think they're going to run out of trainium. And the answer is, what's the biggest amount of supply chain that's locked up? It's Nvidia.

25:08

Yeah, that makes a ton of sense.

25:25

Can you get Macron a free semi analysis plan? Because he came out this week with his big new initiative, 30 million euros for AI research. France is going to be the home of research.

25:27

How do you think all the hyperscalers will respond?

25:41

You know what's crazy is people have been trying to do a lot of work in France for a long time because they have this giant nuclear. Nuclear power plant. Yeah. It's kind of stranded and no one uses it. And like, everyone wants to be like, dude, I can get a gigawatt here. And then they like try to start building and they're just like, yeah, this is never going to happen. I'm just going to go back to the United States. Even though United States is like all fucked up, it's like, I can start there. And they're like, no, no, no, no, we'll start in like five years. And I can think of two specific projects that essentially did the same thing where it's like, oh my God, all this France data center power. And then like they started and like, never mind.

25:45

Yeah.

26:21

Why do you think GROK is climbing the charts right now? Any insight? It's like number three after get and after chatgpt in the overall App Store.

26:23

The iOS App Store, dude.

26:35

Actually, one. This tells you how locked in I've been with Claude code. I had no idea you're so locked.

26:37

In with Claude researching the AI race that you.

26:44

It's just the App Store is based on acceleration, but the GROK hype cycle of let's push all the Twitter users or the X users there, that sort of already happened. I don't know how this is happening because there isn't much hype about.

26:48

And it's happening. Yeah. It's happening off X. Yeah.

27:04

And a lot of people were like, yeah, like, you can talk to Ani and Valentine, but, like, is that really popular?

27:06

Might be on a singularity.

27:11

The real singular. Lonely people.

27:16

Yeah.

27:18

Oh, my God. Maybe I did see a video of, like, the Stormlight Archive thing and that. I feel like, like hit a broader audience in terms of video generation. And I think video generation, like, that always kind of wins. We actually did an analysis.

27:19

Yeah. Which is why you need to be bullish on the Disney OpenAI deal.

27:35

I think so.

27:40

Because, yeah, you have to be. We've seen the nano banana bump with Gemini, and this feels like it could be on an entirely different level.

27:40

Yeah.

27:49

Well, you know, my favorite thing is Gemini wasn't what actually, like, made it rip. It was.

27:49

Yeah.

27:57

Like, the ratio, like, really improved in terms of OpenAI to Gemini, like, way before, and then, like, Gemini, like, slightly helped, but I would say it's like 90% is. Banana. Banana.

27:58

No, no, no.

28:08

Doug in a bikini.

28:09

Yeah. You can just share an image and it's immediately apparent what is going on as a unique capability that you can't get anywhere else. They've cornered the market specifically on, like, the image editing. Not just the diffusion, but the, like, being able to take a photo, change the background, and have it actually look like your face or have the text look great. Like, it was. It was a unique, unique product. Really beyond a model.

28:11

Yeah. Oh, this is. By the way, this is my final steaming hot take in COD code, please. The reason why you should actually pay attention so much is because this is the first time, like, image models essentially always gain share. Video models always gain share, like Studio Ghibli moment. And then obviously chatgpt. This is the first, like, new moment. It's a new modality being the agent. And it's like, actually kicking off.

28:32

Yeah. I mean, how important do you think the cowork, like a desktop app, mobile functionality is to that? Because, like, the. You can have. You can have truly magic.

28:58

Chad has some insight. A lot of people using Grok Video to compete for a $1 million content.

29:08

Oh, that's right.

29:13

Free money. Free money.

29:15

Okay. Yeah. So it's free cash and then free money. The top two apps in the top three give you cash. That is a good fact check. Thank you, Chat. But yeah, my question about can you have a Studio Ghibli like moment if you have to open up a terminal just because there are so many normies that just will never open the terminal, no matter how magical the AI God is behind the terminal, it's just too much to go type one line of command.

29:16

That's why Cowork and Codex are going to probably be what actually happens. I think it's really fun to play around in the whatever 1% adopter, and I'm really enjoying it, but I just don't think. Yeah, it's going to be coworker Codex. And Codex is actually pretty good. Codex is, I think, a slightly more polished experience than Coworker.

29:45

Yeah. Last question for me, take me on the journey of what's going on with Microsoft. What you predicted, how that's changed, how their strategy has changed. Give me the proper way to understand Microsoft these. These days.

30:06

Yeah, Microsoft's not in the race, bro. Why? I know, but they're getting owned.

30:21

They have all of the ip. I don't. I agree with you. I don't understand why it's not like, oh, 5.3 launches. Microsoft's announcing it the same day and it's actually integrated and people are using it like on day one. It takes time.

30:28

It's a skill issue.

30:40

Yeah.

30:41

It's clearly something's going on. And honestly, the thing that makes me most bearish that is the fact that Satya is like, I'm not the CEO.

30:42

I'm.

30:48

I'm the. I'm the product manager of Copilot because I'm so boned if I don't get this figured out. Yeah, like, you can argue it is the mo. It is now existential. He's decided like, hey, my CEO job is getting this one thing right, otherwise we're screwed. And that is kind of worrying.

30:49

It does feel like they could potentially.

31:07

It also the pullback at the beginning of last year, kind of the quick pause is now looking silly in the context of Amazon coming in now and saying, yeah, everybody's on board now. Yeah, we'll see.

31:09

Yeah, yeah, we'll see. I think they have the most to lose.

31:25

What about GPU utilization? Brad Gerstner was hosting CNBC today, which was very cool, talking about how you know, in the dot com build out, the dark fiber was something only, like 7% of fiber that was being laid was actually being used. It was like, obviously, even at the time that. And yet now we're seeing GPU utilization rates, you know, maxed out.

31:29

Yeah.

31:54

Yeah. I think that's a pretty good counterpoint to anyone who's like, whatever. Like, at this point, H100 pricing has massively firmed up. B200 pricing definitely has super firmed up. And like, hey, there is. There's clearly demand. I mean, you know, whatever they're doing on the other side of it, that's like. That's the customer's issues. But, like, I mean, I still think, like, honestly, man, the codex thing or, sorry, my brain's all messed up. Claude code has been the most magical moment in technology for me in, like, my entire time. I think it just feels awesome, man.

31:55

Since the Game Boy.

32:36

Dude, this is better than Game Boys for me. I'm an information addict, though.

32:39

Yeah, makes sense.

32:43

I am.

32:44

Well, we appreciate you taking the time to come chat with us and writing about it and everything that you do. If you're listening, go hop on semianalysis.

32:46

Sign up for the 10 million a year plan. You get Doug's phone number. You can text him.

32:54

You actually can, actually.

33:02

Yeah, I know, I know.

33:03

You should just do it, though.

33:05

I mean, and it's underpriced. You're giving it away. You're taking away your time from all your different agents. So you got to price it.

33:06

Yeah.

33:13

Right. That's right. My manager will hate that.

33:13

Have a good day. Have a good weekend. We'll talk to you soon.

33:18