Ep. 004 - AI 数据中心对消费者电力成本的影响（数据中心、能源）| Jeremie Eliahou Ontiveros、Jordan Nanos、Doug O'Laughlin

All right, everybody. Welcome back to episode number four of "Semi-Analysis Weekly." I'm Jordan Nanos, here with Doug O’Laughlin again, and this week we've got Jeremie Eliahou Ontiveros. We've got part of his attention. The rest of his attention is back at the SemiAnalysis data center model, where Jeremie is Claude-coding 4 separate windows at a time right now. Can we get you to look at the camera for a second and introduce yourself?

Bro, let me vibe-code. It's more valuable. I have no time for this.

Yeah, keep avoiding me.

Yeah, sure.

Just before we get started, for all the vibe coders in the audience, if you're listening to this, we are hosting a hackathon the day before GTC kicks off, Sunday, March 20-something, I think. 14th, 15th? I don't know, whatever it is. The Sunday before GTC, we're going to be on site giving out grants for access to compute, as well as Claude Code tokens, and lots of cool prizes from lots of great sponsors. So if you're following Semi-Analysis and you're attending GTC, definitely sign up to be part of the hackathon. It's going to be fun. Are you guys excited for GTC?

You know what's funny? I'm not going, bro, so not me.

Doug, why aren't you going?

Because everyone's going to GTC, so I'm going to OFC instead.

Ah, okay. Cool, cool.

We have to have some presence at the other place, you know? We have a meetup happy hour, and they're like, “Oh, you're just going to send the dregs?” No. Someone's got to show up. I guess I'm the dregs, is what I'm saying.

No, but I think there's a lot of—I'll be honest with you, you guys are going to have it totally under control, I think. We're rolling pretty deep to GTC.

Yeah, we've got a big crew. The guys who've been working on Inference-X are going to be presenting a nice session. I'm real excited for the LPU heart GPU session presented by the former founder CEO of Roth, Jonathan Roth.

The LPU session is going to be, I think, probably the sold-out session, with people standing and whatnot. That's definitely the big thing.

People are getting really jumpy about SRAM because everyone is super long HBM right now. They're all scared that SRAM is going to take over HBM, and you have to sell the HBM. That's the dumbest narrative I've heard recently. People are so scared because they're so long HBM. So, LPU is going to be interesting.

Yeah.

Sweet.

For the show this week, we have a plan to talk about 3 big topics. The 1st is the article we put out on the free-tier newsletter. Jeremie is here to talk about that primarily. The title of that article is “Are AI Data Centers Increasing Consumers’ Electricity Prices?” The guy who leads our data center model, energy model, and industrials model—that whole team—is probably the best person to pick his brain on that.

Then we're going to talk about the Department of War versus Anthropic, with lots of news about that. Finally, we're going to talk about our internal usage of Claude Code going up. We're all working on dashboards these days. We used to sell these models, which are just spreadsheets and numbers, and now we need real-time visualization in a dashboard. That's the future.

We are building dashboards for business insight at SemiAnalysis. Functional dashboards for business insight.

So, what comes after dashboard? What's the future?

AGI, bro.

The next step.

The dashboard from SemiAnalysis was the Inference-X dashboard, and I think we already know what that is. The guys are piloting a feature internally right now to have a chatbot on top of the dashboard, so you don't even need to read the dashboard anymore. You can ask the agent to explain to you what the conclusion of the dashboard is.

I think the next step is to have some earbuds or glasses, and you just ask them, “Hey, what's going on?” You're walking down the street, and you're like, “Scrape this data set with 500 megabytes of data for me and do it ASAP, bro.”

How are you vibe-coding? Are you still typing, Jeremie, or have you graduated to audio speech-to-text?

Yeah, I'm still typing. I'm kind of a boomer, man. It took me a month to join the Claude Code mania.

I actually want to do speech-to-text, but I feel like the office is a disruptive place for it. I don't know if you guys are aware, but Jeremie has been to the office. Jordan's never been to the office. We're actually rolling so deep today; it is completely packed. David has to sit in the other room, so I can't imagine we're all just yelling at each other while we're vibe-coding. That's probably the biggest impediment.

If I could whisper-code, literally, or something, that would be my best.

That is the New York City office, too. Which office is bigger now, New York or Singapore?

In terms of headcount, I think Singapore still.

Okay.

Because there are 2 visitors today. Eric and Bowen are here.

Yeah, good point. Okay, let's talk about the article. I'm going to share my screen, and we can kick things off.

This article, I think, came as a bit of a response to a lot of people in the news talking about—well, really Trump, who was talking about legislation related to AI data centers and whether they're going to increase people's electricity bills. Obviously, this is the latest article here, and it's a nice one. 7,000 words. How many tokens, Jeremie? It's a big one.

Actually, that was my 1st Claude Code project. You can see some of the trials that are a little bit not in our usual style. That was the 1st thing I ever did with Claude Code. A lot of these analyses were made up, but, to be clear, that doesn't mean they're made up.

I spent a lot of time trying to understand the math and the data it processes. This is a great textbook example of how you can use Claude Code, because a lot of this data—electricity prices for consumers and commercial users, by state or by node or whatever—is available for free, actually. You can prompt it to get all of that, and then you can fact-check it yourself to make sure it's correct and make sure the math, capacity, and such are accurate as well.

Yeah, that makes sense. Can you talk through the thesis of the article? The high level is that there is an impact on the grid from data centers and from a bunch of capacity coming online, but it depends on how the grid is managed. It also depends on which data center or which timeline we're talking about in terms of what impact it can have on consumers' prices.

1. AI Data Centers Raise Power Bills

So, basically, the premise everyone is saying is: energy is the constraint, power prices are going to go up, and it's all the fault of AI data centers, right? That kind of makes sense, but there are big caveats to that.

People always over-index on the short-term data points they see, and there is a pretty flashy one, which is everything in the PJM area, including Virginia, New Jersey, Ohio, and some of these big states. Many people in these states faced a 15–20% increase in their bills versus 2 years ago. Obviously, when you have such an increase and it's in the era of AI, people are saying this is because of data centers. To some extent, it is.

2. PJM’s Broken Market Design

But the point we make in the article is that it's actually much more driven by market design. A lot of that is, to some extent, a function of errors. To some extent, I would argue, it's a design that is not properly adapted to AI. If anything, this is more a function of the last 20 years: the U.S. had 0 load growth. This is the 1st time in 20 years we've seen load growth, and it's pretty clear that this market was not designed for an environment where load is growing, if that makes sense.

Yeah, it makes total sense. When I was reading through this, I think the chart that stuck out to me the most was this one, which covered the PJM model they use to forecast what that load growth is going to be and how they've been making revisions. Because of the way they run an auction to let—

Yeah.

—energy providers set the price they're going to run power at 2 years ahead of time, if they miss on this forecast, then all of a sudden it costs everybody a whole bunch of money. We think, fundamentally, that the accuracy of their forecast is bad because they're not tracking the data centers properly, at least not to the quality with which we can when we use our data and our data center model.

And look, the other big problem, and this is where we contrast with ERCOT, is that this is an auction that is typically run once a year. Demand is based on their forecast. Supply is based on the way they analyze supply. All of that is subject to judgment calls.

And so then you basically go back to the debate, communism versus capitalism, right? To some extent, because in ERCOT you basically let market participants evaluate these items on a real-time basis. In PJM, you're talking about an auction, again, once a year. They can make mistakes.

And so basically, having the ability to react in real time enables the markets to react to the signals, right? Whereas, if you wanted to—the point is, ERCOT can see it coming ahead of time and can place generation ahead of time. Whereas in PJM, because the pricing signals are not set properly, everything is just late, and you just create a mechanism that is constrained and basically becomes bottlenecked.

Yeah. And maybe just to define for the audience, ERCOT is the utility in Texas where there's plenty of data centers coming online right now, and then PJM is the one—PJM stands for Pennsylvania, New Jersey, Maryland, or something—but it covers all of Data Center Alley through Virginia and Ohio, which is like US East 1 and stuff for all the major cloud providers.

So it's kind of the 2 most important grids when it comes to data centers coming online, and they have that completely different approach to how they set prices. One's real-time and one's a yearly auction, but it happens 2 years in advance. I don't even understand how that price gets set for PJM, to be honest.

Yeah. I mean, basically, via their demand forecast, they analyze what their needs are, and then it becomes a question of what is the marginal generator that can meet that need and what prices do these generators need to be able to meet the capacity requirement, right?

So if you have plenty of generation that can serve the specific requirement, then everything is fine. Prices don't go up like that. But if you have demand that surges and supply is tight, then that creates an imbalance, right? It's a typical market.

But the key here, and especially on the supply side, it's really interesting to think about how a regulator evaluates supply versus whether you can let the market price it in real time. And the point of the regulator evaluating supply is really one that strikes me because we put out that chart that shows how supply in PJM evolved over time.

And you see that in the last 4 years, it went down by something like 40 gigawatts, or minus 20%, if I remember correctly, off the top of my head. So it's a pretty massive supply cut in just a few years. And to some extent, it's legitimate because, for real, we had coal power plants that stopped producing power.

But if you look at the chart after that, we showed the bridge, and actually a substantial amount of that is driven by methodology changes. But then again, this goes back to their supply forecast: they evaluate supply. And to be clear, this is not to say they made the wrong decision. There's a bunch of unusual events that happened that led to this happening the way it happened.

So if you show maybe the next one, I don't know if you can change charts.

Yeah.

Yeah, this one. So the thermal accreditation reform, yes, that happened after a horrible winter storm that basically revealed that their power plants were not able to run to the same extent that they thought they could.

But again, this is typically the kind of thing that, when you have a market, when you have people evaluating their own risk and putting their money at risk, they're typically able to take better decisions. I guess that's the history of the United States of America over the last 100 years or so.

Yeah. So I guess 2 takeaways here. One is that, to be clear about what you said about the excess supply to match that demand over what they forecasted, it roughly means a power plant that sits idle for like 96% of the year and then needs to be paid a whole bunch of money to turn on for those times when you need it.

And therefore the price is really, really high if you haven't turned it on all year, for what seem like obvious reasons. The other thing is that, just to point out what you were just saying, CFP accreditation reform—meaning how they measure stuff—is responsible for 41%, or 14 gigawatts, of decline in their total.

To be clear, it's legitimate. There's a real reason. PJM used to be—has always been—an area that summer peaks.

But basically, the peak load of different areas—in some areas it's summer, in some areas it's winter, depending on how people heat or cool their homes and the weather, basically. And so PJM is more of a summer-peaking region, but basically they had a pretty bad winter in 2022. There was already a pretty bad one in 2021.

And what it made them realize—it's not actually a demand peak—is it peaked higher than expected. That's one thing. The other thing is they realized many of their power plants were actually stopping simultaneously. The reliability of power plants is materially impacted when weather conditions are bad, especially when it's cold out there.

And basically they didn't account for the risk that maybe 10 power plants could be not working at the same time. So to some extent it's legitimate, because the way this works is always to avoid the worst-case scenario. So you're forced to take these decisions to make sure you stress-test to the max of their ability.

And they realize, okay, this summer is really bad. Many power plants couldn't work as expected, so we have to reset our supply. So to that extent, it's legitimate. That had never happened.

Whereas the way they were measuring supply before was looking at historical—actually, at the power plant level, they were measuring historical risk. They were not measuring the clustering impact of when a summer is bad, like many power plants stop working at the same time, as opposed to just 1 power plant being offline for a moment for whatever reason, right?

But again, if you let that happen in real time, people can react to this faster and avoid this unfortunate situation.

It's probably—I think it's probably obviously a good thing to move from 170 total gigawatts offered to 156, because you never actually had 170.

Yeah. That's right. Yeah.

Right? So I don't think anybody's saying this revision is a bad thing to do. It's just so funny that at a time where I seem to see so much about people wanting to bring more power generation online, the numbers are moving in the exact opposite direction, to the tune of 14 gigawatts.

New gas capacity on this same chart in PJM is plus 3, and new renewables is plus 1.5 over the same time.

I have a question. Do you think this means that coal is going to be delayed? Essentially, are they going to delay retirements? I mean, people have been talking quite a bit about this, right? Like delaying coal. Do you think retirements are going to start to happen on the margin? This is kind of what this is

Yeah.

I don't know, high level telling me.

Yeah. So there are 2 things. One is, at the federal level, stopping coal power plants—stopping their retirement plan. That has happened. It's Fair Coal Order 206, if I remember correctly, or whatever.

A bunch of coal power plants were supposed to retire, and basically they're not going to retire anymore. And then there's another aspect, which is, from an economics point of view, some of these power plants are, whatever, 40 years old, 50 years old. They're obviously not the most efficient, so running them is not economical anymore.

But if prices are high enough, it becomes economical. And actually, the core issue with energy right now is that you have to balance the consumer expectations of prices staying low with the fact that, to incentivize new generation to come online, you need prices to actually be high, right?

So that's really one of the big challenges. One of the reasons—there are many others, but one of the reasons—why it's tough for generation on the grid to come online. Which is why we're moving to behind the meter, right? It's like you have way more control to build whatever you want for your needs, as opposed to waiting for the grid to perhaps decide to finally let more folks come in or change, reform something, right?

3. Data Centers Pay Their Share

Yeah. And do you think that's—so maybe the political aspect of this is that Trump was out there saying he's trying to have all the data center builders or the frontier labs sign on the dotted line that they are not going to increase consumers' electricity prices.

The grid operators are probably just looking for somebody else to blame when they have to increase people's prices. But the vast majority of the data centers that are coming online are using the new stuff, are using behind-the-grid or behind-the-meter generation, right? So that's—

Actually, not yet.

It’s going to be much bigger in 2027 or late 2026. Still, the vast majority of data centers coming online right now are grid-connected.

Okay.

And the majority are under, to some extent, old regulations. Increasingly, what you’re seeing is different states or utilities negotiating directly with their big hyperscale customers to build a custom tariff. It’s basically, “This is what you are going to pay,” which is different from the rate everyone else is paying, to make sure that their impact on the grid doesn’t flow through to everyone else.

So, if you take a step back, in which context could the AI boom impact retail bills? The core thing, especially on the transmission side, is that all of the transmission in the U.S. is regulated. Regulated means that it has a required return on equity, which means that if you build a gigantic asset and it’s not properly monetized, or the utilization rate is not what was expected, those costs actually end up being passed on to the consumer because the actual builder must turn a profit regardless. It takes no risk.

And so the massive risk here is that if we overbuild and we have regulations that force an unprofitable asset to be passed on to customers, then obviously that’s the worst-case scenario outcome.

Yeah.

And then the question is, how do you make sure this doesn’t happen? There are a few ways to do this. One is customized tariffs, as we discussed earlier, which is the hyperscaler telling the utility, “All of the upgrades that you’re doing on the grid, I’m going to pay for this,” or, “I’m going to pay for the majority of this. If you have to build a new power plant, I’m going to pay for it.”

And “pay for it” doesn’t necessarily mean paying for the CapEx. It just means that perhaps you can sign a 15-year or 20-year agreement with the utility. An interesting example was Oracle in Michigan, where one of the Stargate data centers for OpenAI is located. Oracle signed a $4 billion commitment with the utility. It’s a one-gigawatt data center, one-gigawatt IT, and 1.4-gigawatt gross, peak utility, which is pretty big in this landscape.

Imagine all of the gigawatts that we have. If all of them have that kind of commitment to the utility, probably revenue, CapEx, and all of it is going to go up quite a lot. But anyway, the point being, $2 billion—

What’s the timeframe for that use—$4 billion over 10 years, 15?

So it’s 4 years, of which—

Wow.

Sorry. There are 2 things. There’s an investment period, so that commitment has a 15-year or 20-year power-price commitment. But there’s also an investment commitment to purchase batteries and put them on the grid to basically increase the overall amount of capacity on the grid.

It’s actually a pretty interesting agreement. Oracle says, “I’m going to pay for batteries to improve grid reliability and whatnot. And in the end, we’re also going to sign a 20-year agreement to pay you a given amount of money, with a minimum demand charge.” This is actually very interesting, because it means that if the load somehow doesn’t materialize, they’re going to have to pay a charge regardless, right?

Yeah.

And that’s one of the new big things on the grid: making sure they pay their fair share to avoid a scenario where you build new generation and those costs are passed on to customers if the hyperscaler doesn’t consume the load it was planning to consume.

Wow. Okay. And I—

There’s a bunch of interesting stuff being negotiated, but there’s no easy solution. In these cases, with a big commitment like that, sometimes it’s just easier to build your own power plant.

Yeah, it makes sense. I wonder how long in advance that deal got negotiated for the site Oracle is going to do. It’s not like they can reach an agreement like that on a 6-month timeline, given the speed at which some people want to put up data centers these days.

That specific site, I think, began development about 2 years ago. There’s a third party involved as well. It’s a data center developer correlated digital. If I remember correctly, they started negotiations on that site about 2 years ago.

But Oracle signed the deal just a few months ago, I think in October. That data center is going to be operational in 2027, so that timeline itself is pretty short. The timeline from when you sign the deal to when the load is going to be operational is just kind of half.

Yeah. It’s somewhere between 1 and 3 years of negotiating with the utility for a 4-year, start-to-finish data center construction project timeline.

The hope is actually to get it done by 2028, the full data center. So there are different timelines here. For the battery investment, they said 4 years.

Mm-hmm.

But the hope is the data center will be operational by something like mid-2028.

Right. But when did they break ground or get the first permit to say, “We’re going to do this data center”? That was about 2 years ago, right?

Yeah, about 2 years ago. Yeah.

So 5 years total to get the whole thing online. I mean, a lot of people just talk about the queue with these utilities, trying to actually get a grid interconnection done and how long that takes. But some of these big data center projects are taking 4 years from start to finish.

It just seems like you get in line, and then you’re willing to pay for some batteries, and you can make some of this stuff happen.

Yes, but there’s a financing mismatch, which is that you have to take speculative risk. You don’t know yet whether the demand is going to materialize, right? To some extent, that’s what creates the mismatch. If you’re in the solar business or battery business or whatever, you have to make a forward bet, to some extent.

Yeah, makes sense.

Yeah.

What do you think? Is that demand going to materialize? Is Stargate going to power WarGPT because WarClaude is out?

I think they’re going to change the end user. OpenAI goes under and Anthropic takes it all, because all of the world is going to use $100,000 of Claude Code every day, just like us at SemiAnalysis. Fake numbers, because it’s not true. I’m not spending that.

You’re so Claude Code-brained now that you think OpenAI is going to go under.

Yeah, right.

4. Anthropic’s Revenue Surge

Our daily spend is kind of implying it’ll be… The thing that’s really crazy is just the rate of this, right? Let’s just talk about the Anthropic ARR crossover, which I’m happy about, but, dude, Jordan, I’m just going to say it: there’s a reason why I didn’t make this bet in 2026. We had a bet originally, but Anthropic is at the revenue run rate that OpenAI was at the end of the year, and it’s March, so it’s like a 2- or 3-month lead, right? It’s just—

That’s 2 months.

So close. Yeah, 2 months.

If you go back in time—not to victory-lap on something I didn’t win—but I was so impressed by Claude Code in September or October that I started telling Doug they were going to win versus OpenAI.

No, dude, that—

And Doug wants to make a bet.

No, no, that bet—no. That bet was way before, bro. It was way before the offsite. It was in the middle of the summer, man. Come on. I like betting, but—

Yeah, because the—right, because the bet would turn into a beer at the offsite. Yeah.

Which was in October.

Yeah. So it was way before September.

Okay.

I think it was way before.

Yeah, yeah.

I swear to God. Anyway—

Okay. Well, it’s here. It’s March, and it’s here. The exit ARR of February, we think, is close to Anthropic crossing over OpenAI—the whole company, right?

Yeah.

Claude Code is a big part of it, but—

Yeah, the whole number.

It’s not necessarily—Claude Code is still not the majority of Anthropic revenue currently. It’s growing at a very fast rate, but—

So let’s actually talk about that. I think it is. I think the thing is, it’s the—yeah, Jeremie could do it because he’s been model-grinding on this one, but it’s just how you attribute revenue, right? I think the way that they’re—

I think it’s—

Yeah. You can look at our usage dashboard. Claude Code attribution when using Claude Max in Claude Code is what they define as Claude Code. API usage is API usage anywhere, and they do not attribute it if it’s in Claude Code API mode or not.

Mm.

As fast-mode addicts, I think you can tell how we feel.

Yeah. Actually, especially the 1-million context. I think fast mode is on subscription, but 1-million context is just on API. But I guess the point is SemiAnalysis spending is like 90% to 95% going to API. It's not going to what they report as Claude Code. So when they said last month, “$2.5 billion is our annual run rate on Claude Code,” that's actually way too low.

That's super interesting.

I think they have to make that be like, “No, no, no, no, we—” I think it's more in their interest for generalized enterprise demand without having a ginormous vector be like, “Oh, all the enterprises are using it and God knows what,” instead of being like, “It's one hyperviral thing that people are yoloing $8 billion on.” But who cares? I mean, it clearly—the answer is revenue, and Anthropic is winning right now. Or, I mean, until OpenAI comes back. But it'll be so over, so back, or whatever this week or next month—who knows?

So what do you think is causing the extra spike? There was the Super Bowl thing; now there's the conflict with the Department of War. Do you think that, in a backhanded way, if the Department of War does not follow through with this supply-chain-risk designation for Anthropic, and they can keep doing business with everybody in the U.S. 6 months from now, and possibly even repair the relationship and get back into the U.S., that this tweet storm at Dario actually got people to—

Net promoter score positive. I mean—

Yeah.

Did you see the Katy Perry tweet? That is the net promoter score positive. Effectively, I'll make this—

I don't even know if it's—so maybe it's both, but to me it was just—

Wait, wait. Like—

Eyeballs. More people know what Claude and Anthropic mean now—

Yep.

—because they're following politics, not technology.

Yes. I agree with that.

Wait, do you guys actually think that it had any impact on their revenue?

Yes. Dude, you know what's the number-one downloaded app right now? It's above OpenAI, and that happened on the Department of War.

Probably not in France.

People don't care. They don't care about France, bro.

Back—

What do you mean? You don't like Claude?

Backwater country.

Claude is from Paris.

Claude is French. I forgot. I forgot Claude is French, yeah.

He's French.

There was a meme—I can't remember who made it. I think it was Near [?]. During the summer, Claude is lazy because he's on vacation.

C'est bon. Yeah. It is the French way. Currently Claude is protesting, and a little bit later he's going to go on a strike.

And then a strike afterward, yeah.

That is the way. That's how we do stuff, man.

No, but it's open to take on OpenAI, and that was, I think, on the Department of War. As long as their vendors don't essentially spit them out of compute, dude, I think the whole moment is gonna—

No, but let's talk revenue, because downloads could just be random people trying the app. But revenue is probably Claude Code, right? So that's most likely unrelated to that specific issue.

Well, it's Claude Code, but Cursor just claimed a huge acceleration as well, and they're using plenty of API tokens. I assume many of the other vibe-coding startups and apps, like Windsurf, Vercel, and Replit, are all using API tokens. And then they launched Claude for Finance, Claude for Legal, and Claude for Security Software, which tanked a bunch of SaaS stocks one day after the other. Somebody is buying something other than the first-party Claude Code Max subscriptions. It's not strictly a single vector consuming the tokens.

No, but my point is I think this acceleration—literally, they've added $10 billion of revenue in 2 months.

Yeah.

ARR.

They doubled ARR in 2 months, actually. I would think that's entirely driven by true business adoption, if that makes sense, as opposed to just temporary hype of people subscribing to and downloading the app in support of Claude, Anthropic, and everything.

No, I completely agree with that point.

There's definitely use cases. Like—

But, but hey—

Doug, you know this better than anyone else—enterprises, certain companies are removing Salesforce and building their own.

Well, yeah, no, I think it's a big deal, right? That's the real demand driver. But I agree, that's not what's moving revenue; it's just a cherry on top.

Yeah.

They do the API crossover—that's what really is driving all of it. Then, also, the cherry on top is they just hit number 1 on downloads, too. That's like, “We weren't even trying, and we got you that way.” That's a pretty hardcore mog, just because everyone's like, “Whoa, that's not even what they're known for.” It's like—

Yeah.

It's just a casual, “Hey, this is, just by the way, I also passed you on downloads this week.” I think that's why it's a big deal. It's, for lack of a better word, a hype moment. It's a—what is it?—a zeitgeist moment that I think really matters, and that's the difference.

Well, there's a complete narrative shift because we used to say, OpenAI, ChatGPT has a consumer market, and when you're the consumer household name, it's easier to penetrate the business world, right? Now we're actually seeing the opposite, which is no one knew Anthropic. They come from the business world, and now they're actually beating OpenAI on consumer.

5. Models Become Everything

Well, okay, what seems most clear to me is that all of these people who have been speculating about how the models are commodities, and therefore the value is going to be at the application layer where you can build these great experiences powered by whatever model is best—the best fit—are just completely wrong if you consider the fact that Anthropic has access to models we don't have. They have full control over the coding interface, and they can build a whole bunch of applications using that better than anybody else can, effectively for free.

So the idea that Anthropic is going to come for Salesforce, Harvey for legal, CrowdStrike in the SaaS security business, and on and on down the line, because they've already done this to GitHub Copilot on the coding side, seems right to me. It seems like the model companies will become the everything companies, as opposed to SaaS companies deciding to build models later or picking the best Chinese open-source one to power their product.

I think that universe could exist and might end up existing in the end state. But in the current state, when every 3 months the model gets better, there's no point, right? If there's an upgrade every year and you're building your thing on some upgrade every year, that's fine. Maybe you can plan it out, whatever. But if there's an upgrade every 3 months, there's no way. Your product cadence is going to be inherently slower than the model cadence.

I mean, dude, this is the history of Moore's Law, right? CPUs won because they got 50% better every single year, so it never made any sense for any specialization ever to happen. You might as well just bet on the underlying, better thing that's generally good at everything getting better next year, right?

But when Moore's Law ended and CPUs really topped out, that's when it became valuable to specialize because, hey, there's no more lift of performance this way. You have to go down a specialist route. So until the generalized models don't get better every 3 to 6 months, there's no point in doing any specialization. I think that's what we're going to see until scaling laws top out.

Yeah. I think in general, I agree with you, but the specific cases—like, if you're going to do specialization, the fact that they chose to do coding, which is upstream of so much other stuff, just seems like an obviously good choice. In other words, coding is the generalist thing to focus on, and you shouldn't specialize too much in that.

I have a question. Definitely, clearly, there was a strategy there. But do you think it wasn't because they're like the—I think coding is special in the case that the whole point of it is you make an autonomous agent who's better able to code a better model, and that's part of this. It's inherently a part of this. Now, if they thought marketing was the way they were going to scale, for lack of a better way, it would be a scaling marketing agent.

But it isn't, right? The only thing that matters is coding for making your core products better. I think that's the reason why they've always been so focused on coding.

Yeah. Well, the thing that makes your core product better is data, and the first obvious place to get that data was in the chat interface. People vote up and down on what's a good response and what's a bad response, or you see how much time they're using the chat. You can review the logs. Then that shifted when RL came out, because you need a lot more signal from 1 sample or 1 really long trace in RL than you do, and therefore chat isn't quite good enough. The quality of data in coding is much higher, and so the flywheel is in coding.

But it's not even coding, because you guys are using Claude Code for—not exactly. I mean, it's code, but it's more manipulation of your computer to build artifacts, like a screenshot of a plot that we just had on screen when we were talking about the article, right? You're not using it to replace software you would've written in the past. You're writing software to do something that you would've done in Excel in the past.

Coding is just general. I mean, it's just a higher-fidelity way to work with transistors, right? So the thing is, I'm going to software-big-brain this, right? Software historically is a layer on top of code for humans to interact with the machine, right? But since it's good enough that it can actually sit underneath software, pretty much humans are interacting with the agent that's now lower-level interacting with the computer, right?

Yeah.

You totally skip that portion because it's not valuable. What's really cool for us is we're analysts, right? We're analyzing crap. But we always had to use software that was made for us to do it. For the first time ever, I don't have to learn some new piece of software. I could use the computer to make information.

I don't have to learn statistics and stats and R to be able to do crazy analysis. I don't have to learn all this crap to scrape all this stuff. I can now work with every single type of software I've ever wanted to, with all this domain expertise that I never had to learn, and do whatever the hell analysis I ever wanted to do. That's the valuable part. Coding is a portion of that, but I think the reason why it's coding is because coding is close to the computer, right?

I really, truly, fundamentally believe LLMs, because of how compute-intensive they are, are kind of an extension of the computing paradigm.

Yeah. This is why Jeremie was joking about just walking around with your agent in your earphones, and then you can tell it to show you the dashboard and make an update or whatever you want.

And agents are kind of the next-level extension of compute. So, I mean, it's just all-in. Instead of us interacting with software built on top of compute, humans are just getting closer to the computer.

Why do we need to do software, dude? It'll just tell us. It'll be reading CSVs into my ear, bro. That's all I want to hear.

Maybe that's the bull case for a DGX Pod. You guys are all buying Mac minis to run your Claude Code.

I've actually come full circle, bro. VPS for life, bro. I have a DigitalOcean droplet. It's better.

All right. So is it over for Mac minis, or what's going on?

No, it's not.

I can remote-SSH into my actual laptop.

Yeah.

I can run a container on my laptop if I want some security.

I have remote SSH to my Mac mini, and then, obviously, the remote SSH to a DigitalOcean droplet. So sick, dude. I run cron jobs all day. More cron jobs. More cron jobs. Update the database more. It's so beautiful. Sorry.

Cool. No, go on.

Do you use Tailscale?

Huh?

We were talking about Tailscale, actually, because you were talking about how you used it for one of the projects you were doing, and I was like, “Oh, cool. I use Tailscale.” It's pretty nifty, dude. I like it quite a bit.

It's simple stuff, but it's good. The free-tier get-you-in option is just logging into your laptop from your phone pretty simply with a private network. But companies like CoreWeave use it for authenticating people to their clusters because it's much better than just public IPs and SSH and trying to whitelist IPs for people who move around the world.

So, yeah, Tailscale's awesome. There are lots of software companies that are very much accelerating because of this, and then there's other software that's just totally getting left in the dust. Tailscale's one of those. Obsidian's another one that you love, right, Doug?

I do. But you know what's funny? I think I'm graduating from Obsidian. Why fuck around with Markdown when I can just do it in the terminal? I don't care, actually. I've come to realize that in the beginning I was like, “Well, I want to make sure I could read all the Markdown and all this shit to audit it,” kind of looking to do some fine-tuning.

The only time I care about Markdown in Obsidian is essentially pulling up a document to edit it for a final time, and then copying and pasting it into WordPress or wherever the hell I go to. But I literally don't care about it being in Obsidian. I hope and trust that the model is going to be good enough at finding information within a thing that I think it should be abstracted away. That's my belief.

But the Obsidian CLI project is pretty interesting. I was very Obsidian-first in the beginning. Now I've kind of moved away from it. Maybe I'll go back. I don't know. But I think you could just have a vector database—QMD crap that Toby does—and search, and it'll be faster.

Yeah, absolutely. QMD—shout-out to Toby. Let's go.

Yeah.

These are all Canadians, buddy.

Oh, yeah.

Tailscale, Obsidian, and QMD from the ex-Shopify guy.

Wait, are they really?

Yeah, yeah, yeah.

Are they—wait, every single one of those is Canadian?

Yeah.

Wow, we should acquire Canada.

Yeah. The royal we, there, bud.

Sorry.

The royal we, yeah.

You and your army.

Hey, it wasn't even my idea, dude.

Yeah. Yeah.

And hey, and that also applies to anyone listening to this podcast. If you're a Canadian tech worker, we've had a lot of good luck. Jordan over here happens to work at SemiAnalysis. We moderately like the guy, so I think we're more than happy to hire more people from Canada. So yeah, we're always looking. So anyways, that's a complete aside. Always hiring.

Yeah, shout out to Rippling, another company that's going to accelerate as the AI wave takes over everything.

Where were we?

I think we're probably good to wrap up here, guys. I don't know.

Yeah, I think so.

Anything else from you guys?

I think so too.

Can we talk about this week?

Do you want to talk about Qwen? We haven't even posted about this, so I don't know what you want to talk about, dude.

We can talk about it. We can talk about it next week too.

Yeah, we can talk about it next week.

6. China’s AI Demand Emerges

Leadership changes—things are still shaking out. You got a hot take about Qwen?

I don't have a hot take. It's actually a pretty cold take after some tweets. I was like, “Damn, my take's pretty cold.” Essentially, they kicked him out because they didn't acquire enough users on Chinese New Year.

You definitely did write about that. That was a nice tokenomics update about all the Chinese New Year competition that was going on.

Yeah. And you know what's funny? My takeaway was Qwen wasn't bad, but I think they probably spent like $1 billion, and they're like, “Not good enough.” So now the KPIs are all user acquisition. I mean, Doubao—or ByteDance—won, for sure. And I think ByteDance is the only Eastern model that everyone at SemiAnalysis wants to use through CDance. Like, “I want CDance so badly, dude.”

Okay. Yeah, that was, I think, the thesis of that article, or at least the last little section of it that I wrote a bit about, which was that we've seen this with solar panels, drones, and BYD cars, where, on the margins, people are like, “Oh, that looks really cool, but my Tesla's still cool, so I'm not going to go out of my way to try to import a BYD.”

Everybody who I know who's seen the CDance videos is trying to figure out how to VPN their way into China and add something to that queue, which Eric posted this morning. How long?

It's like a 3-month queue or something like that. It's something incredible, dude. Effectively, if you want to sign up and you're trying to pay for stuff, you're like, “Yeah, dude, you might be able to generate a video on CDance in a few months.” It's kind of crazy.

There are 150,000 requests in the queue or something.

Yeah. No, it's 14,000 out of 30,000, and that's, I think, in the paid queue.

Oh my God.

Oh, wow. Dude, that's like old-school internet. You have to wait in line for your job to be done, and this is on the paid version.

Yeah, people who have—

Wow.

…already paid ByteDance are being told, “You can't even pay more to skip the line. You just have to wait in the queue.”

Pretty sick, dude. Also, my galaxy brain—I don't know how much inside baseball we want to do, but some of this might be the demand. That might be part of the demand that we're not recognizing well, that we're seeing in GPUs, right? The fact is that Chinese models—yeah, maybe the retention rates weren't so great for some of them, but ByteDance clearly got some retention, and clearly the acceleration of actual users in China is going up, right?

Yeah.

With CDance, with Qwen, with all this stuff. And China has way more people than the United States—hot take. If they're starting to actually adopt AI, all of a sudden that's a giant new demand vector—not completely untapped, but a huge new demand vector that I think is maybe underappreciated and is starting to leak to the rest of the world, because China domestically can't actually support compute. So this just ends up in Singapore or Europe or everywhere else around the world.

So the question here is: What percentage of Anthropic revenue and Claude Code comes from China?

Well, allegedly, it's all—

Any?

It's all—Okay, so you have to know that some amount of it is, because they're like, “Oh, malicious actors have been…” Because you know the fine-tuning thing, right?

Oh, with 150,000 prompts. Relationship with malicious actors, bro.

Yeah.

You know, all of this stuff about the distillation attacks—if you add up all of that stuff, it's less than one of our days' spend on the API. I really don't think that's a ton of—

Yeah, but then—

They're—

If I remember correctly, I think we had some data maybe last year on Cursor, and something like 15% of users were from China. They asked a bunch of people locally, and they were telling me, “Yeah, these tools are not blocked.” Unlike Google and some of that stuff, you can actually use Western AI agents because—

Hmm.

…the local ones suck. So I actually don't know exactly about Claude Code—if it's blocked or not. I have no idea.

So I think this one is well-documented, but apparently all the Chinese labs VPN into Japan and Korea. Then they move their sleep schedules forward so they can replicate an average Korean vibe coder. At one point last year, they were like, “Oh, Korea and Japan are 15% of Anthropic's revenue.” I was like, “Yeah, that's Chinese demand.” There's no way.

There's obviously usage. It's very clear who the Eastern models prefer to distill from—Claude, right?

Yeah.

The joke is—I think one of the best jokes I've seen so far—is that GLM—is it 4.6 or 5.0 or what? It's 4.6.

Yeah, 4.7.

The GLM-4.7's distillation of Claude is better than Claude can do itself.

It's sonic.

Yeah, it's sonic. That's so funny. That's hilarious. Just imagine if they had Claude, dude. And you know, the other meme too is that I think—

Okay, I think GLM does a lot of pretty good engineering, man. I don't know.

No, I'm being—Sorry. Let me put a little respect on their name. I feel like the bleeding edge is essentially razor-thin. I think it's very clear that if Eastern companies had access to Western compute, we would be mocked. There would be a new pretrain. If everything was equal in terms of compute, I think they would be ahead.

I can't wait for DeepSeek-V4, guys. Can't wait.

When is it coming out? When is this Enneagram? When is this memory-aware galaxy brain? I don't know. We'll see.

All right. I think that's a good way to end it. Jeremie, see you in San Jose in a week and a half.

Cheers, man.

Wow.

See you in San Jose.

I'm jealous, dude.

Talk to you next week.

Talk to you next week. I want some cigs, dude. Where are the SemiAnalysis cigs, bro? Actually, I want you to know that everyone gave me cigarette birthday gifts. I'm thinking of you, Jeremie. I'm thinking of you, bro.