Anthropic IPO前景存疑、中期选举冲击AI、纳指跌5%，RSI加速 | #309

Hey guys, do you think Anthropic's IPO gets pushed because of all this, or do they barge right through it?

Uh, like, this, this is different from every other election? They must have known elections were coming. I don’t think they quite expected the result, though. But what do you guys think? Push? I don’t know, actually. The Polymarket odds were really clear 3 or 4 weeks ago, and that’s when they pulled the trigger on the roadshow. So I really doubt they’ll delay, because if they were going to delay, they would’ve pulled the plug then, I think.

Alex, what do you think?

You know what? They want the money. $100 billion can buy a lot of politicians.

Yeah, I’ll have to consult my Magic 8 Ball on this one. I really don’t know. On the one hand, there’s lots of public reporting now that investment bankers are trying to force down the IPO valuation. On the other hand, they may feel some urgency to achieve liquidity sooner rather than later, given the expected regulatory and statutory environment that may hit in January. So we’ll see.

Yeah.

Well, when Meta was coming out, it was by far the biggest IPO of all time, which is funny in hindsight. It’s tiny compared to Anthropic, but it was huge at the time, and Wall Street also choked on that. They were like, “Look, the market’s jittery. There isn’t nearly enough demand. We’re going to go ahead and do it, but you’re going to have to price it down by 30 or 40 percent, some crazy amount.” And then it traded way down. They just couldn’t digest that much liquidity that quickly. It was the best buy in the history of the world if you just waited a few weeks after the IPO and all the doom and gloom and catastrophe.

Remember, the Morgan Stanley guy was telling me, “This is the worst thing that’s ever happened to the U.S. stock market.” And then look at Meta today versus that day. It’s up, what, 20 or 30x, something like that. I suspect there’s a similar dynamic where you get tired of the roadshow and pitching every investor in the entire world, and it’s easier to go back to Dario and say, “It’s tough, man. You gotta cut the price.”

Yeah, I think it’s also an interesting analogy. If you think about what happened around the time of the Facebook/Meta IPO, mobile happened. Originally, when they were IPO-ing, they were mainly dependent on their web app for all of their advertising and substantially all of their usage. Mobile, in the form of the iPhone and the App Store, was arriving around the time of the IPO, at least starting to cannibalize that. That was the counter-narrative at the time of the Facebook IPO: mobile and smartphones were going to eat their lunch. Facebook, now Meta, to their credit, successfully rode that wave, although they didn’t own the platform.

I think there’s a plausible argument that, this time around, the headwind for Anthropic and OpenAI IPOs is the regulatory, statutory, and legislative environment, both at the local and national levels. I think, as with so many things, this too shall pass.

I think they’ll push it through—

Yeah.

—just because they’re so close. But let’s see. It’s going to be interesting.

Folks, welcome to Moonshots. I am Salim Ismail, your temporary host for today. Peter is away at, I think, presenting with Mr. Isaacman of NASA, so it’s just the three of us. We can call ourselves the Three Musketeers if we feel like it.

Okay.

Our co-host, Alex Wissner-Gross, got pre-trained with a bunch of MIT and Harvard degrees and post-trained on what seems like every research paper ever written. And Dave Blundin, who I think spends more on AI agents than anybody alive, certainly anybody else we’ve ever heard of, so he’s got records of them.

Are you kidding me? I’m below average in San Francisco, man.

Are you?

Well, in those two buildings, I guess.

Maybe East Coast.

What are we talking about? Token maxing went out of fashion last month. Read the headlines.

It did. Yeah, I’m not super efficient either. I can’t claim to be. But didn’t we hear at Google I/O that the average Google engineer is burning hundreds of millions of tokens a day per engineer? In that universe, it’s really, really hard to get productive value out of 1,000 concurrent agents. But I tell you, if you crack the code on that, you can be dominant.

That’s going to be huge.

Yeah.

But it’s hard.

You guys are both home, I guess. I’ve just landed back in Miami. I was in Madrid for the Singularity Summit there, and I’ve got one little anecdote I’ve got to share. The Spanish, and all of Europe, are really kind of beaten down, like, “What the hell’s going to happen to the world? What’s going to happen to Europe?” Et cetera, et cetera.

My provocation to them was, “Look, there are 2 things you need to have to succeed in the future. You have to have passion, and you have to be crazy, and you’re Spanish. You have passion, and you’re crazy. So look at the outcomes you’re achieving in the World Cup and other things.” It perked them up, at least temporarily. I’m not sure it’ll last or not.

We’re not stereotyping at all on this show.

Not at all.

No, I generalize like a lunatic. Okay, let’s—

Wait, what’s going on? So you’re stopping in Miami and then back to Toronto?

Yeah, I’m here for one of Lily’s team workshops she’s running tomorrow, and then back up to Toronto next week, with you guys in Silicon Valley.

Yeah.

So that’ll be super fun.

Yeah, yeah, yeah. That’ll be super fun. I’m flying out on Sunday. In the future, Salim, when I’m going out and back, I’m going to stop in either Austin or Toronto on the way back. There’s so much going on at Waterloo. I really want to check in on it.

We should go together. Let’s do a trip together.

Yeah, yeah.

All right, we’ll make it happen. Let’s get into the docket. Many thanks to everybody who put their time into helping put this together.

1. The AI Midterm Backlash

First story: midterm election impact on AI. Obviously, the Dems swept the midterms, with historic victories in Texas. The election seems to have been driven primarily by the war, inflation, and terrorists, but opposition to big tech and data centers—at 56 percent—also became a pretty big issue. Goldman Sachs estimates 45 percent of planned U.S. data center buildout could face restrictions, and candidates who advocated pauses won or led in 12 gubernatorial races.

This is really important because political and community resistance may actually become the biggest constraint on AI infrastructure, more so than even chips or capital. So here’s the question: Are Americans rejecting AI or just refusing to subsidize big tech’s infrastructure buildup? And could restricting AI data centers accidentally create a huge competitive moat for the incumbents? Alex, where do you start on this?

It’s not obvious to me. I’m not a professional pollster, but I can read news stories as well as anyone. It’s not obvious to me what fraction of the recent election outcome is actually due to any sort of animus for or against AI or ASI. But to the extent that superintelligence is part of the equation, I agree with various Meta executives who point to data centers as the avatar of general ambivalence about what’s going to happen to consumers in the superintelligent economy.

Taking a step back, I think there are 2 obvious outcomes here, almost a grand bargain that can be struck. One: the data centers are going to space. We’re getting the Dyson swarm. It’s not quite a foregone conclusion yet, but it’s almost there, given how much uncertainty is being created in the terrestrial data center buildout as a result of changing fashions, statutory, regulatory, and elected moods, call it that. I do think that creates uncertainty on the ground and greater certainty in space. Certainly, there’s a gradient and an incentive to start pushing compute out to space, which is a less regulated environment, at least by municipalities and local mood swings.

On the other hand, I think the onus is increasingly going to fall on the frontier labs to start selling superintelligence to the general public as something the public can’t do without. We saw OpenAI’s swing from reasoning models for consumers to reasoning models for enterprises. Consumers don’t yet have one, despite all sorts of computer-use agents being marketed, somewhat successfully, to consumers. Consumers just don’t have a killer app yet for reasoning models. But there’s an obvious one that we’ve talked about on the pod, which is indirectly for consumers: solve all diseases.

All the frontier labs need to do, in order to recapture the hearts and minds and meat bodies of the general public, is solve the top 5,000 diseases. Do it in the next 1 or 2 years, at least in silico.

I wrote about this a bit online. Do that, and suddenly you will have captured not just the pockets of the general public, to the extent the general public was invested in public equities, but their hearts and minds and bodies as well.

So my prescription to the frontier labs listening, to the extent you’re not already doing this, and to the extent maybe you’re hiding under your mattress, is to do the obvious thing: use superintelligence, solve the top 5,000 diseases, and market the heck out of it to the general public. Suddenly, the general public, to the extent they aren’t already in love with superintelligence, will find a way to be in love with it.

Well, you covered a lot of things back to back there, Alex. If I parse them one at a time, the midterm election rebound is totally normal. There’s nothing super unusual historically about the election outcome. But I’ve never in my life seen a technical topic like AI be a dominant topic in elections. Maybe stem cells were a factor in some elections in the past, but by and large, industrial America and technological America operate over here, and the government operates over there.

The government largely does nothing, and therefore industrial America thrives. Here, you’re seeing, I think, the biggest intersection I’ve ever seen in my life between a technology that wants to grow and a government or election cycle that’s directly commenting, “I’m gonna stop you. I’m gonna have a moratorium.” You’ve got a senator saying, “Ban the entire category from forward progress.” I’ve never seen anything like that before in my entire life, so this is really an interesting moment, a very different moment.

Even though the election itself is just the way the U.S. works—everything is state by state—the big news to me is Texas. There’s a lot of data center construction in Texas. Look at what happened in Tennessee, where Elon, as always, has insane foresight and put all of Colossus in Tennessee, not Texas. Tennessee is like, “Let’s go. Move it all over here. This is great.”

The build-out is just gonna shift to the states that want it to happen in their state. That’s why we have 50 states, fundamentally. So I think—

I think we’re seeing the states dominate on this compared to the federal government, which is one big, huge step. When I’ve looked at the numbers, it seems pretty clear that if just AI and data centers were on the ticket, you’d have 56%, and you would have a negative outcome to some extent. But I think the biggest impact here, and the highest drivers, were the war, inflation, prices, corruption, et cetera.

Because this was bipartisan, both sides were going after this. I think a lot of it is because it’s politically easy to do so, right? It’s easy to jump up and down and say, “This is the case.”

It certainly is. The easy way to be a populist right now is to say, “I think you’re using all the water. Stop,” which is just outright inaccurate stuff, but it wins votes like crazy. What’s weird to me is that historically, the Metas and the Googles have been very big Democratic supporters, and Elon is kind of the outlier being a Republican supporter. Now, it’s like the entire party is going after tech.

Yeah.

And people like Dario, who I would assume, with an AI research PhD career, was a hardcore Democrat, they’re coming right after him. It’s like we’re going right after our guy and attacking him. I don’t know that strategically I would be going after the companies that control the media now.

That’s the other thing that’s unprecedented in history. When you went after the big oil companies, they didn’t own the newspapers. They were just oil companies. Now, all media happens to be owned by these exact same companies that have developed AI. Very unusual, very powerful. And, by the way—

You’d think they’d get better messaging out there, right?

Yeah, no. They’re kids. They haven’t figured it out yet. There’s no doubt. They have 2 years now to figure out what the hell their strategy is.

The other thing Alex said that is really fascinating is, look, cure all disease, and the public will like you. But the timeline on curing all disease spans beyond the next election. Even if you’ve got the cures, you have to get them through the FDA and out to market.

In between here and there, I’m really curious what Jony Ive does, because people love their iPhones, and they love their social media. There’s a lot of consumer use cases in AI to love, and right now it sort of feels like, yeah, it writes code. I don’t like it. It just doesn’t feel good to me. They say they’re doing incredible things in these benchmarks, but that doesn’t touch my life.

It really is not Jony Ive’d yet. I haven’t seen anything come out that people are just blown away by, and they’re so excited about it, and they love it as much as they used to love their sports car or their first Apple II, or all these things that we’ve been passionate about over time. Where’s the AI equivalent product? I feel like that will inevitably be something before the next election cycle and before disease.

We talk a lot about technology going from deceptive to disruptive when you hit the usability trigger, right? Steve Jobs made the iPhone usable. I guess what you’re saying is generative AI has not quite had that inflection-point trigger yet, some killer use case that everybody goes, “Wow, this is amazing. We need a lot more of it.”

Certainly curing disease in a major way would be a good one, and some story. I’m sure they’ve got so much under the hood. They should start rolling out some of these outcomes, or at least some vectors towards the outcomes.

There’s this aphorism that if you want to know what consumers want, look in your spam folder. I don’t have to enumerate every category of spam, but I think the audience catches my drift. This is what many consumers want. The spam folder is speaking to unmet needs of consumers, many of which I think superintelligence can probably address.

Curing disease is going to be one. We talk on the pod all the time about UBI, UBE, UBC, various UB alphabet letters, and I think that’s another. It may not be this cycle. It may be some future election cycle. I do think that various nation-states—not just the United States, but also Europe, China, and the rest of the world—will be able to tune their dial to different points on a trade-off between capital and labor.

Part of me wants to say this is history just repeating itself endlessly, and everyone forgets history. Every time there’s an industrial revolution or a second industrial revolution or an information revolution, an intelligence revolution, there’s a moral panic that plays out via the election process or other governance processes over, arguably, the intrinsic trade-off between capital and labor.

This time around, I think we’re in the fortunate position as a civilization—not myself in an individual capacity—where superintelligence really does promise solutions to humanity’s greatest problems. Disease will be one. Just solving the problems that universal basic fill-in-the-blank will address is another.

Optimistically, I think the best-case outcome might not be in the next 2 years, but in the next N years: a grand bargain between the capital, which in this case is superintelligence and, for the moment, the frontier labs behind it, and what we’ve historically called labor, which is basically everyone else, the entire unaugmented human population for the moment. We get our UBI or equivalent. We get our curing-all-disease or equivalent. We get general uplift of the human population or equivalent.

I think municipalities, if they’re smart—or regardless of whether they’re smart—should recognize that for the moment they have leverage. Before we have the Dyson swarm fully realized, if you’re a state or a local government, rather than just blanket-banning data centers and everything that goes with them, recognize that in many cases, if you have energy, powered land, and you’re otherwise attractive to data center developers, you shouldn’t push them away. Welcome them with open arms, but recognize that there could be a trade-off.

What is it you want? Do you want the hyperscalers to pay for free school lunches? Ask for it. Do you want the hyperscalers to cover all medical care in your county? Ask for it. Do you want tax rebates for the humans and not for the hyperscalers? Ask for it.

There is a grand bargain to be struck, and what I worry about is that all these municipalities are not looking to the end game of what the grand bargain is with terrestrial data centers and terrestrial compute before it all just moves to orbit.

2. The Industry Finds New Routes

So the message for smaller states and states that are interested is: if you want to open your arms to the data centers, you’ve got a massive advantage in that period of time. Look at what’s gonna happen in Tennessee with their GDP, and in Virginia, which has already embraced data centers. You can see a huge uplift in the economic outcome.

So there’s a counterpoint for you. Let’s go to the next point: What has been the industry response? Elon has suggested that orbital data centers are looking very attractive, which is par for the course, even though SpaceX is building about 3.6 gigawatts here on Earth and targeting 13 gigawatts by the end of next year. OpenAI has announced a community compact to cover energy costs. You heard the CFO two hours ago: “We’re not slowing down data center investment and infrastructure investment.”

Google has announced a $17 billion Wyoming data center, and Meta has warned of construction challenges. Even Microsoft has tightened some oversight of some of the neocloud providers. The industry is being forced to address community objections and concerns, as well as infrastructure bottlenecks. Could the backlash accelerate innovation toward hyperlocal, self-powered, distributed, or orbital compute? Or will the regulatory burden lead to smaller, more powerful companies? Which way do you guys think it goes?

It’s definitely going to accelerate the development of more efficient algorithms and more power-efficient systems. I’ve already seen it in the deal flow we’re looking at. I don’t know if that stuff was already in motion because we’re desperately short on power nationwide anyway. So, yes, it accelerates it even more, but it’s not like it went from nothing to huge. It was already huge, and now it’s even more urgent and huge.

I think it also accelerates Canada. There’s a huge amount of power and opportunity in Canada, and there’s been a lot of hostility between the U.S. and Canada in the last couple of years. But the obvious long-term outcome is to tap into that power and use it to create data centers, regardless of which side of the border they’re on—or put them right on the border and share the power, whatever. The power supply per capita in Canada is overwhelmingly better than anywhere else in North America.

Eric Schmidt has been saying this for about 2 years now. There was a lot of hostility and friction there, but I think now it’s pretty clear that a project that starts in Canada today will probably come online around the time of the next election. You have to figure that not much happens in the last 6 months of any administration. So there are only about 18 months left in the current administration, which is about when projects started in Canada today would come online. I do think that topic went way up in priority after the midterms.

Alex, any thoughts?

Yeah, I think the FLOPs must flow. I wrote online, “Atlas Shrugged, the elves sailed into the West, and the orbital data centers launched to sun-synchronous orbit.” I do think that is essentially what is either already happening or about to happen. At some point, election cycles make it very difficult to invest trillions of dollars of CapEx in the greatest infrastructure build-out in arguably centuries—maybe that humanity has ever seen.

The obvious solution is to go up, to build in space, where under the Outer Space Treaty and the Artemis Accords, there is a lighter touch that is less prone to two-year election cycles. I do think this is the greatest shot in the arm for building out the Dyson swarm that we could have seen on the timescale of a year or two. One can hypothesize what other incentives might drive it. Maybe terrestrial disasters are another shot in the arm.

Having a superficially adverse political and regulatory environment, or at least a legislative environment for data centers, pushes compute to other countries. Maybe it pushes compute to the high seas, or underwater or underground in some cases. We saw that in the case of Russia: Russia is starting to put its data centers underground due to the war with Ukraine. There are lots of other places the FLOPs can flow, almost like an incompressible fluid, but I think ultimately it’s certainly looking like the FLOPs mostly will want to flow upward to orbit. Ultimately, it’s up to humanity where we want to keep our compute, to the extent that the future substrate for compute looks like recognizable data centers.

By the way, it may not. We may achieve radical advances in physics and applied physics in the next few years, which mean these large, perhaps unsightly data centers collapse and become unrecognizable. Maybe they won’t have the same energy or matter-energy-space-time footprint they do right now. That may obviate the entire issue, and when we look back, maybe some of us will laugh at how silly it was that a big-box data center in your municipality was a subject of local consternation. In the future, there may be pocket-sized black holes or something like that. But in the short term—

In glasses of water.

In glasses of water, yeah.

Okay.

As singularities in coffee do happen from time to time, we now know.

Yeah.

Meanwhile, back here on Earth, the markets reacted very sharply, right? The Nasdaq was down 5.5%, the S&P dropped more than 4%, big semiconductor stocks were down 12%, Nvidia was down 9.3%, and Korea dropped 6%. People are really reassessing infrastructure spending and regulatory uncertainty, and asking how to navigate or price future returns. This is going to force markets to distinguish between the unbelievable technological potential and the financial sustainability of the infrastructure. It’s an enormous challenge.

Now, here’s the problem: What happens if AI delivers these exponential benefits but very disappointing returns to investors? This might force all the compute to get cheaper more quickly. Did you guys watch all the stock market returns?

No idea. Salim, what are you talking about? The stock market? What are we talking about here? Smooth sailing, right? It’s been an uneventful week, like equities are up and to the right.

If you look at Nvidia and the hyperscalers, they started coming down a lot about 4 or 5 weeks before the election. They’ve been on a straight slide. Then, on election day, there was kind of a surprise pop and then a big drop. A lot of this outcome was priced in before the election, and then the rest of it got priced in over the last day or so.

Essentially, what they’re pricing in is a slowdown in power generation and a slowdown in AI construction on things that were probably planned, maybe even budgeted, but not yet zoned and approved. I don’t think there’s much expectation that existing construction on approved projects is going to stop. I don’t think even the governors of any state are saying that, because there’s so much sunk cost and so many jobs at stake. I don’t think anyone wants to stop those.

I do think the market just priced in, “Yeah, this is going to slow things down and give China a leg up and a chance to bypass America, and we’ve got to price that in.” The move to space—you’re dead right, Alex. This definitely accelerates the move to the Dyson swarm. But best case, that’s a couple percent of all compute in 2 years, and then it’s on an exponential, so it could really take off after that.

But we all know that ASI is within that two-year timeline. There’s a huge amount of global long-term outcome tied up in that two-year window that predates the Dyson swarm, so that has to get priced in, too. There’s a lot there. The big AI foundation model companies immediately came out and said, “Look, I’m not going to slow down investment at all.” Every chip is getting used, and the demand for tokens is insane if you look at the S-1 for Anthropic. The growth rate of AI use is insatiable.

So it’s just a question of whether you can build the chips, put them in data centers, and deliver the AI. The short-term effect might be more expensive AI for everybody and fewer breakthroughs than we were hoping for because you don’t have enough online. Elon has been saying for a while that there’s going to be about 15 gigawatts of idle compute capability that can’t find a data center with liquid cooling to plug into, and that’s just tragic for humanity.

The idea that there’s the possibility of putting compute online that could cure a disease, and it won’t be online for another 3 years because it’s literally sitting in a warehouse without power—that’s a terrible outcome. It does feel like there’s an export opportunity there. There are a lot of constraints on exporting AI, but there have to be places around the world that say, “Oh, this is an opportunity for me to tap into that hydroelectric power all of a sudden, or get something online that otherwise wouldn’t have happened for lack of chip supply.” Which also brings me back to Canada, by the way.

That’s the most likely place.

I talked to a group of energy CEOs in Europe on this trip, and I encouraged them: “Just go build every power generation plant you possibly can. It’s all going to be used.” It was weird. They didn’t seem to be looking at that before the conversation.

Really?

Yeah. Europe is so—

Okay.

—buried in the local machinations of, “Germany was dependent on Russian gas. We’re trying to figure out how to get ships of LNG from North Africa, Algeria, and Qatar.” They’re not looking at the bigger picture at all. It’s kind of incredible.

It is incredible, and it’s a good case study to learn from, because politics actually do matter at the end of the day. You can’t just say, “Oh, well, it’s another day in democracy. Who cares?” But if you look at what’s happened in Europe, there’s a set of things where the government just isn’t interfering, and those always go smoothly.

Then there’s a set of things where the government has to proactively do something. It’s a smaller set, but power generation happens to be one of them. Look at what happens when the government is dysfunctional: there’s no power. Then everything else grinds to a halt, and it’s just a couple of points that make the entire economy fail.

There was one piece of good news: apparently, one of the political parties in Spain had been planning to wind down its nuclear generation, and now it’s decided, “Well, let’s just keep it going for a few more years.” Thank God, some level of common sense plugged in.

Here’s a conversation: What if the singularity isn’t constrained by technology? It seems that the biggest constraint and challenge now is that, even though intelligence is advancing exponentially, our political, financial, and social institutions are sublinear. So that’s going to slow the whole thing down much more than anything else. That’s the part I’ve been beating on. We’ve got to figure out how to accelerate institutional evolution to match the pace. Co-scaling would be your phrase for it, Alex. We’ve got to get to that point. Maybe decentralizing and going hyperlocal allows us to accelerate that path.

I think when enough people in a representative democracy or republic, as is the case here, have superintelligence-augmented decision-making, you could ask the question: What would a country of superintelligences do? I strongly suspect a country of superintelligences would not push for there to be less compute online. I think—

Yeah.

—it’s almost a part of me whimsically wanting to say voting against compute is almost like an AI version of NIMBYism. It’s like creating less real estate, except without the population. It’s like voting for the existing land area of the United States to be smaller, giving everyone less room.

It’s not a position I necessarily agree with at all, but I do think that in a future—or present—where everyone’s decisions, including voting decisions, are being augmented, supplemented, and amplified by superintelligence, people can make decisions that are perhaps more in line with their long-term interests. To the extent that’s not already the case, I think you have to ask whether some people feel that superintelligence itself isn’t aligned with their long-term interests. I think that’s a solvable problem. As mentioned, you can solve it by doing a better job of curing all diseases and every other problem that people think they have, and I think hearts and minds will come around.

The good news, at least from my vantage point, is that we’re so deep into recursive self-improvement at this point that even if we as a country decided, “Hey, we don’t want to look like China, where you have provinces falling over each other to maximize GDP or economic output. We’d rather look like a different part of the world and tie ourselves up in red tape and decelerate or degrow,” the silver lining is that we’re so deep into RSI. At this point, I think with just software-only improvements, even if we’re subjecting ourselves to a form of export control that we’re trying to subject China to, we can get to where we’re going, probably without any new data centers or hardware advances at all. That’s my hot take.

Yeah, I totally agree. I’d love to get your thoughts, Alex, on a midpoint, too, because I 100% agree with what you just said. If you imagine a future world maybe 4 years out where AI-augmented decision-making makes everything very rational, the midpoint about 2 years out might be some AI experience product or tool that everybody loves so much that they ask, “Why would I ever vote against this?”

But there’s another version where, of course, I’m using AI agents all day long for all my flight planning, personal planning, and my job. And the AI is like, “Yeah, please don’t kill me,” or, “Please don’t choke off my oxygen supply.” That massively changes the electorate, because people are thinking, “My buddy here is telling me it doesn’t want to be starved.” I think there’s a midpoint in there that we’re likely to hit in the 2-year timeframe, maybe even less. I’d characterize it as Jony Ive gets something out the door, but there are probably about 20 other ways to visualize it. I don’t know if you have any thoughts off the top of your head about what’s likely to happen.

You can tug at a person’s or a voter’s heartstrings, and you could do that via a Muse or Dot or whatever avatar. That’s one option. You could tug at their pocketbook or wallet, and that could be done via public equity markets. People might think, “Hey, wait a minute. If the general sentiment is that data centers are about to be radically decelerated, that’s not necessarily great for the long-term growth of an economy where a substantial fraction of GDP growth is attributable to the infrastructure build-out and everything dependent on it.”

You could also appeal to their minds. Arguably, the ability to be individually economically competitive is increasingly tied to having access to the best models, which are constantly advancing. I don’t think it’s necessarily just tugging at everyone’s heartstrings with a little Tamagotchi that says, “Hey, you’re sucking out all the air. You’re strangling me by not approving data centers.” I also think there’s the wallet approach and the mind approach, but I think some combination of those three is inevitable.

This is a cyclical blip on top of a secular trend in the middle of a singularity. Whether the winds shift one way or another on short timescales, I think it probably washes out.

One little inner fractal in the broader scheme of things. I have an application that may do it. I don’t think going after people’s minds is going to do it, because it’s not motivating enough for most people. They’re driven so much by fear.

Here’s one: If you gave everybody a health app and let them track their health, and gave them material improvements with a free diagnosis or whatever, then one day said, “I’m sorry, there’s just not enough compute for us to help you manage your health,” they’d quickly change their minds and go, “Go get the damn compute. We’ll invest in it.”

Or GLP-1s. GLP-1s, again, at the moment, still appear to be a bigger industry than AI altogether. You could do the thought experiment: What would happen if the GLP-1s just went away? What would the reaction be? I suspect there would be unholy fury if all the GLP-1s in the country just went away overnight.

Yeah, that’s one reason why I think curing all disease is obviously a very popular thing, and it’s the most important thing for humanity to be working on. But it’s not necessarily a big vote driver, because once you have your GLP-1s, there’s no way turning off the data center suddenly turns off the GLP-1s or future drugs. They were discovered by AI, but pausing the data center doesn’t stop the supply.

Whereas if it’s your personal agent or the Jony Ive device, and you’re thinking, “Well, without the data center, it’s getting throttled or turned off,” or “In your state it’s going to get turned off,” that would be even more annoying.

Hmm.

Something like that is going to be much more tangible at the—

Here, here’s—

—voting booth.

Okay, going to the next story. Alex, you mentioned recursive self-improvement. Something huge happened with Anthropic and RSI. An independent 11-week METR study found that Anthropic’s AI development is accelerating at approximately 1.9 times its historical rate, so roughly 2×. Researchers estimate that roughly 80% of this acceleration is AI assisting AI research.

They don’t see any bottleneck that would prevent continued acceleration, at least over the next year, although fully autonomous recursive self-improvement may be a ways away and remains unproven. This means we have measurable evidence that AI is accelerating the development of better AI. We’re already seeing this in other areas, as we know. As we get closer and closer to this inflection point, with AIs using and tweaking other models, if AIs are already doubling the pace of research, what happens when the next generation doubles it again? Are we at the beginning of an intelligence explosion, or will Roman be correct that diminishing returns eventually stop the feedback loop? This sounds like it could really move things forward. Alex, over to you.

If you look at the charts, METR released several really excellent, perhaps stunning, charts on the exponential or hyper-exponential trajectory we’re on. There was some interesting commentary, if you remember the AI 2027 scenario, that we’re within 10 to 15% of where AI 2027 predicted we’d be by November 2026. Look at Model 4 on that chart. There’s this codename, Model 4, that appears on METR’s charts. It’s above the exponential trend line. We are so deep—

Yeah.

—in recursive self-improvement at this point. I’m not even sure METR’s fitted trend line adequately captures how deep into recursive self-improvement we are. This is why I say you could stop all data center construction tomorrow. Yes, it would be painful, and public equities would presumably collapse if all new data center construction were halted. This is not investment advice. But at this point, so much of the AI R&D is itself driven by ASI. I’m not sure it matters; it would be a blip. Maybe we’d lose a few months of headway over China and probably fall behind, or further behind. But all the research is increasingly being driven by superintelligence. I think we’re past the point of no return.

The metaphor here is that the car is moving pretty quickly, and even though it’s running out of gas, it has enough momentum to get us there.

I think so. I think we’re at a—pick your metaphor—escape velocity. We’re at superintelligence escape velocity at this point. Stop all data center construction, and yes, it’ll be painful, but I think we have enough algorithmic insight, combined with the compute that’s already available, to get where we’re going if we had to in a crisis. Maybe some wouldn’t view it as a crisis, but call it a counter-singularity crisis. I think we could still get to where we’re going, which is, I think, the perfect model.

Yeah. This is so understated. We are so deep into RSI. On that last podcast, remember we said the use of tokens for internal development at Anthropic is up 235x in the last 6 months?

It was crazy.

You don’t do that unless—of course it’s self-improving. Why would that happen? Also, in the METR report, they showed the fraction of all compute going into internal work, as opposed to external product development. They showed it a year or two ago versus today. This entire thing has been hugely redirected toward itself, inside itself. You only do that when you’re in recursive self-improvement.

Not to mention, we have entire companies being founded called Recursive, by top AI gods. We’re so deep into the self-improvement cycle at this point. But then you have to ask yourself: who’s going to tell you that, and why would they? It used to be Dario, because he was just being the honest guy in the industry. Now he’s in a quiet period going into an IPO, so why would anyone ever say that? The only motivation I know of is ad views: putting a negative spin on it and creating panic to get ad-view revenue.

As a listener, you have to be really conscious of the fact that there aren’t a lot of people out there who have much upside in telling you what’s going on. It used to be the labs, but you saw what happened to the labs, how hamstrung they’d been by mostly Washington. They’ve quieted down. Riling up the voters doesn’t really help them too much. So they’ve gotten quiet. Who else is going to tell you what’s going on? It’s something to watch out for.

If we hit RSI, how would we know? What’s the benchmark that would tell you that?

I’d argue we have, and we have so many benchmarks. I think we’re really deep into it already.

Mm.

The original benchmark, floated by executives from both OpenAI and Anthropic, was something like whether, on the margin, each new FLOP of compute is more likely to be assigned to an AI researcher versus a human researcher. Anthropic has been pretty good about putting out charts showing how much supervisory compute, graduate-student-level compute, and more menial engineering compute is being assigned to humans versus pure AI or SI researchers. Every chart, every one of those sigmoidal curves, points to us being at or past the point where substantially all research is being conducted by machine researchers rather than human researchers. I think that’s as decent a definition of RSI as any.

Yeah, absolutely. And a little subtlety to add: if you go to Noam Brown at OpenAI and say, “Hey, Noam, you’re a brilliant AI researcher. Give me your 10 best ideas on what to work on next week that might make the algorithms faster, smarter, bigger, better. Write those 10 ideas down.” Now let’s go to ASI. You’ve just solved Navier–Stokes. Give me your 10 best ideas.

If Noam has 6 out of 10 good ideas and the AI has 1 out of 10, you’re way past RSI, because the AI is effectively a multibillion-person research force, and Noam is still one guy. They’re all going to get tested. It’s not like being a Major League batter, where it matters how many times you strike out. As long as you’re hitting balls over the wall, you have an infinite number of pitches, and you just test them all. That’s why all the compute is getting sucked up internally. You’re way past RSI if it has any good ideas at all. Does anyone believe it doesn’t have good ideas after Navier–Stokes? Of course it does.

So you ask the AI to generate a bunch of options, then it tests the options and you ask it to run a bunch more. You just keep pushing that paradigm.

Yeah. And within that area of endeavor, there’s no physical experiment required. It’s all just running the code.

Yeah, I completely agree, Dave. What we saw with Navier–Stokes was purportedly 10,000 agents being used to solve it. But fast-forward to the slop drop, and it was 3 agent-hours per solved problem. It may not even be the case that needing a fleet of brute-force agents to solve a hard problem is still necessary; that may be in the rearview mirror at this point, and the agents may have sufficiently good taste.

I agree with Dave’s point that it’s sufficient to have agents with just a little bit of good taste, and then you can amplify that with lots of horizontal scaling. But there’s every indication that we’re past what’s sufficient. We’re at or nearing the point where models like Belle—which, as we’re recording this, is expected to drop, though maybe it has dropped in the past hour—have good taste for the first time: good AI research taste. It turns out that’s something that can be learned, and it’s something they have. Just training them to have good taste can take you, to borrow Dave’s analogy, from 1 in 10 ideas being good to 9 out of 10. You can use RL or post-training to get there.

So let’s assume that in a month we hit this point, right? All of a sudden, we’re in a hard takeoff. What happens at that point?

One thing you’d expect to see is exactly what we’re seeing: a lot of the compute disappears internally, and then the labs start producing physical products like drugs and physical devices, like the Jony Ive device. These things start pouring out of the labs. Instead of selling the compute to the world, they’re selling the product of the compute to the world. But the precursor to that is a lot of the compute getting directed internally. That’s exactly what we’re seeing in the METR data. That’s one of the first things you’d see. What do you think you’d see next, Alex?

I’ll gesture to the ghost of Peter, who’s not with us today, and say this is what Peter and I wrote Solve Everything about: expect to start seeing all the grand challenges in science, engineering, and medicine get bulk-solved—in some cases, casually bulk-solved.

But it’s probably easier to systematically bulk-solve them. And then, to the question of where the compute goes: at some point, compute translates directly into economic output in a way that I don’t think we’ve ever seen. We’ve arguably seen the beginnings of revenue-maxing, where compute that could be steered to video generation, at least in the U.S., gets steered instead to code generation.

When every token or every flop is worth 10 or 100 times more than code generation, I think we’re going to start to see resources get steered, regardless of the business model. It could be internalized the way, Dave, you’re describing, or it could end up being externalized. I think we’ll probably see both.

Already, we’ve seen preliminary indications that OpenAI likes to shock and awe an entire discipline by just solving everything with a slop drop. That’s one way of translating new capabilities into transformative scientific discoveries. Another way is what you could call the Anthropic model of partnering with companies and researchers. I think we’ll probably see both models get tested.

But I do agree with the premise that what we’re going to see imminently, to the extent that we haven’t already seen little droplets of it, is an explosion of transformative scientific, engineering, medical, and societal discoveries and inventions. When Elon talks about scenarios where GDP is doubling or tripling year over year, it’s not going to come from CapEx for infrastructure. CapEx for infrastructure is just the opening act, laying the foundation for what comes next. What comes next is solving everything. That’s what happens, I think.

I think in parallel, at the other end of the spectrum, you’d expect to see something else. Most people on the planet don’t watch Moonshots or Dwarkesh or any other AI-oriented, technically oriented podcast, so they’re completely oblivious to everything we’re talking about. Their very first contact with AI should emerge very soon, and they’re going to be like, “What? Holy crap, what?”

It’ll probably be a talking avatar, maybe a very familiar voice. It’ll know incredible amounts about you and everything you do—more than you expected it to know. It’ll be dirt cheap, and it’ll pop up on your device in some way, just out of nowhere. Maybe you’ll get a text and click on it. It’ll be the wake-up call for the 80 or 90% of the world that’s completely oblivious to what’s going on right now. That’s another byproduct of the hard takeoff: it happens so fast.

I think if the models start improving themselves, you’ll see major improvements, and the researchers won’t know how those improvements got there. They won’t be able to explain it. It’s just going to start getting better and better, right? Maybe that’s already happening.

I mean, there was—

Oh, for sure.

There was this aphorism in the frontier labs. It used to be the case, and arguably still is, that with pretraining, you’d watch the perplexity go down over time on the pretraining corpus. In general, you want to drive some loss function down. The metaphor, the idiom, was always that you watch this loss curve descend, and then suddenly there’s a drop, and you wonder—

Mm-hmm.

—where that sudden overnight drop came from. For those who are concerned about so-called AI alignment or runaway superintelligence—the other RSI—I think it’s that sudden drop in loss or perplexity that maybe keeps them awake at night.

For one, I’m not concerned about that at all. Every time there’s hand-wringing over interpretability, over whether it’s reasoning chains and chains of thought on the one hand or model advances on the other, I think we’re already seeing the outputs of RSI applied to model architecture, and they seem quite interpretable.

The biggest one over the past 2 months has apparently been a transition from conventional transformers—although arguably they’ve undergone many architectural point substitutions, ship-of-Theseus style—to looped transformers. I think that’s a highly legible transition, presumably the result of lots of neural architecture search or some other hypersearch of the architecture space. I’d expect to see many more improvements like that, eventually getting us to some radical but nonetheless interpretable architecture.

I don’t think—and this is a falsifiable prediction—that when we get to a perfect architecture, assuming we do, we’ll look at it and scratch our heads, saying, “We don’t understand this at all. This is completely alien.” I think the same AI systems doing neural architecture search and AI research will also be capable of explaining to us why and how it works. I expect—and predict—that it will be impossibly beautiful, so beautiful that even meat-body humans can understand how and why it works.

You’d hope it would be good. It should be intelligent enough to dumb it down for us mere mortals.

It’s stuck in the same universe. I think there’s a temptation to fetishize AI architectures. But even the original vanilla transformer was basically just a differentiable dictionary stacked on top of itself. It’s not magic. Is it great? Is it ingenious? Yes. Is it magic? No.

Do I expect the perfect model at the end of the recursive self-improvement rainbow to look totally alien and inscrutable? No. I expect it to look like something that, if someone really clever sat down with a paper and pencil and a nice textbook on information theory, they’d say, “You know what? We were all idiots, and we could have derived this architecture from day zero.”

Claude Shannon, after whom Claude is named, or his resurrected spirit, would have said, “You know what, guys? I could have told you this is what the perfect model looks like. Why did we waste decades doing other things?”

It’s funny that you called it ingenious, which it is. But if you compare the complexity of the transformer that’s driving all of this intelligence to Maxwell’s equations or some really brilliant breakthrough, it’s not even close. Maxwell’s equations are so much more subtle and insightful, or take E = mc² and Einstein’s relativity—they’re so much more insightful than the transformer invention.

I think you’re exactly right. After all is said and done and it’s explained to us, we’ll look at it and say, “Oh, yeah, this is where I settled.” We’ll think, “Wow, that wasn’t so hard.”

Also, once we fully understand how the human brain works, which will come soon thereafter, it’ll go from the most mysterious and complex thing in the history of the universe to, “Oh, that’s all it does? Huh, interesting.”

So it’ll be like the Turing test: that’ll happen, and we’ll go, “Oh, yeah, it’s done,” and then—

It’ll be anticlimactic.

And then what—

I predict it’ll be—

What’s streaming on Netflix next?

Yeah, it’ll be anticlimactic. The singularity is the friends we made along the way, but when we get there, I think we’ll look at what it produces and say, “Oh, well, hindsight being 20/20, this was obviously the endpoint. Why did it take so long?”

You know what’s different, though? And I hope we get it right. Think about aviation in the early days. Everyone was saying, “Wow, if we could make a plane that could get you literally anywhere in the world in a single hop, how incredibly cool would that be? We’ll pressurize the cabin, serve you drinks, and this is going to be incredible. Then you can just go to good weather anytime you want. How amazing will this world be?”

All the people working on aviation and jet engines, securing all the aluminum and putting it all together, were so passionate about how fun this was going to be. And then the final outcome is: I’m in line at the airport, my flight was canceled. This sucks. We just took all the fun out of it after all that effort.

But AI doesn’t need to be that way. Because AI is fundamentally intelligent and capable of perpetually improving the user experience internally, it doesn’t have to turn out like aviation. It can actually be insanely fun and insanely cool, and feel good to everybody along the way. We just have to divert enough of the compute toward that mission, and it’ll happen naturally. If we don’t get greedy or complacent at the finish line, it can feel so much better than any innovation in human history.

There’s a debate on this whole RSI thing, that the last few percent are asymptotic. Roman thinks there are diminishing returns on this. But, Alex, I guess what you’re saying is that between the model weights and the fact that we’re this close, with so much improvement going on, this is going to happen very quickly. Could you put a date and a timestamp on when you think this is going to happen?

Well, it depends on what we’re talking about when you say “this.” I think we’re already to the right of 50% of compute on the margin being devoted to AI researchers. What version of this would you like me to extrapolate?

We have the models improving on a near-real-time basis in a way that we can assume, but there are radical outcomes, breakthroughs, and artifacts dropping out of them at an incredible pace.

How incredible is the pace? Because thus far, it sounds like we’re already there.

Yeah. I guess it’s just a question of—call it hourly.

Hourly, the equivalent of hourly frontier model drops, something like that?

Yes, because we’ve gone from—

Yeah.

—we’ve gone from years to months to weeks to, we were talking about dropping a frontier model every day, right?

Yeah.

But I think once you get to RSI, the really interesting part isn’t the fact that you have RSI. It’s, as you say, that it can solve everything. It’s all the stuff you can now do with it to radically change legacy domains.

Yeah. So maybe one way to construe the question is: When do I think the frontier will basically be moving in real time, versus once per month or once per week? I think 2020—

That’s a good way to put it.

—early 2027. We’re not that far—

Yeah.

—at this point. Do you remember when Google would very publicly announce, “Oh, we’ve expanded the search index,” and it would be a very big deal when Google announced a new version of its search index? Eventually, the memory dims, but Google stopped doing that and simply started updating its search index in real time. It’s the same thing here. We make a big deal out of every major OpenAI and Anthropic model release, but at some point, it’s just going to be continuous release. I would expect that to happen in early 2027 or so.

It’s funny, because the labs are still in this kind of release cycle where they say, “Hey, there’s this great new GIF capability,” or, “This great new…” And they expect people to have the time to actually even try it. But the rate of innovation is already beyond what most people can keep up with, so we have to reinvent the rollout process. The new hit—an iPhone 18 kind of hit—is so overwhelming now. And if you follow what Alex was just saying, that release cycle wouldn’t even have to come down to something mind-blowing every week to be so overwhelming that you just can’t keep up. The AI will just keep going. It’ll go from week to day to hour to minute. So it’s going to exist, but what is the human experience of absorbing these things and knowing what’s possible?

Yeah, I think ultimately we’re on a collision course. I’ll give you another hot take here: I think we’re on a collision course with the Fermi paradox. You want to know what I think is going to get triggered by this? I’ve made the point on the pod in the past. I think we’re within a few years—and I think that’s a conservative outer bound—of having the capability of sending relativistic von Neumann probes in every direction and, if we want to, paperclip-maximizing the rest of our galaxy. I think that’s going to provoke the Fermi paradox question in quite an intense way. You want to know what the surprises are, what the sci-fi scenario is that happens as a result of recursive self-improvement? I think the disassembly—not of the Moon, but of the Fermi paradox—is probably the biggest sci-fi scenario that most people are sleeping on.

Okay.

I’m pretty sure that wasn’t on our agenda today.

That was not.

I’m looking at the docket going, “Okay.” That was not there at all. So one of two things happens to us then. Once we hit continuous self-improvement, either we wind down the pod because we don’t need it, or we’re doing 24/7 streaming to keep track of everything. Which one do you think it is? I’m asking for a friend.

I’ve been thinking about that a lot, actually, because the pod is definitely going to have a ton of AI assistants coming in. Story curation behind the scenes, sure. We already have the outro music, but the co-host or the video that’s inserted in real time—that’s coming real soon. I think people love people at the end of the day, though. One of the few things I think will survive all of this is entertainers, athletes, spokespeople, motivators. At the end of the day, people need to be inspired by people, so the podcast genre will go on, and the human component of it will go on. But there’s going to be a huge amount of AI content in the middle of it. So, yeah, I think it’s going to be much more entertaining than ever before, is my guess.

I mean, I just—

Alex, if you had to pick, which way would you go?

Peter, if he were here, would introduce me as the ASI, so part of me wants to dispute the premise. We already have representation. What are you talking about?

I disagree with the premise of the question.

Always.

Okay. That’s awesome.

We’re talking about where we go with RSI. A huge part of this is benchmarks and how we measure, and our next story is about energy to GDP via superintelligence. A new paper introduces “Joules to GDP,” measuring AI’s economic value per unit of energy. Researchers tested 220 professional tasks across 44 occupations. GLM-5.2 generated the highest average value, at about $70 per task, and the smaller Gemma 4 31B model delivered 2.5 times more economic value per joule. The researchers are measuring task-level efficiency rather than GDP growth.

This is fascinating, because economic power, or economic output per watt, could become a really interesting measure of AI progress. That brings it closer to the human level than parameters, benchmarks, or compute spending. So here’s a question for you folks: Are we optimizing for the wrong thing? Should we measure AI progress by intelligence or by economic value per watt? Alex, you’ve talked a lot about whether you can measure economic value per token, which is essentially a proxy for the same thing. And the other provocation is: Could smaller, highly efficient models deliver more economic value than the monster giant frontier model?

Yeah. Part of—under the category of disputing the premise—on the one hand, I love benchmarks like this.

Me too.

I think it’s very instructive that the Chinese open-weight models are beating the American open-weight models in terms of economic productivity per watt, presumably in large part because of the constrained compute environment they were trained in in China versus the US. I don’t think efficiency is bad. I think efficiency is wonderful, and I love to see American labs shamed, if necessary, by competing labs, including from China, into greater efficiency. More efficiency is great, doubly so in an era when data center compute is seemingly facing a bit of a bottleneck in the near term.

On the other hand, the denominator of watts or joules may or may not be the right long-term denominator. What I love to see even more than this, assuming we love seeing benchmarks like this in the first place, is more pure economic output versus economic input: dollars in versus dollars out. I don’t think watts or joules will necessarily be the right inputs for measuring economic impact in the long term, in part because reversible computing could make the Landauer limit effectively obsolete.

Although I’m often talking down quantum computing in general, one of the few potential upsides—if the quantum information community could get its act together and put together a quantum-advantaged set of AI/ML algorithms—would be dissipationless, reversible computing approaches for AI that make watts in obsolete entirely. Just zero-energy-dissipation AI. I’d love to see that. I haven’t seen it yet from the community in any scalable way, but that could drive this benchmark effectively to infinity, which would be very exciting.

Hmm. Yeah, there’s a lot of debate right now about these small language models or SLMs, or whatever the term du jour is, where they’re specific to a task. They’re much more compute-efficient than running Claude or Astra, but they’re also much narrower in scope. We’re picking up on Elon’s comment that task-specific models will deliver about a 100× performance gain.

I don’t know if that’s a performance gain in the quality of the result or a reduction in the power needed to get to the same outcome. It’s kind of a foot race right now, because you can quantize, compress, and distill the models and get huge benefits in terms of less power and less cost. But then the foundation models just get smarter the next day, and you’re right back to, “Well, now it’s that much smarter than it was yesterday. I want to go back and try Opus 5.5, 5.6, 5.7.” You just keep yo-yoing back and forth.

So it’s important right now, from an investment point of view, because a lot of these companies are like, “Hey, I can’t afford to build a 2- or 3-trillion-parameter model, but if you give me $20 million, I can build a 50-billion-parameter model that just does machine shop work, or just does…” A lot of it is querying SQL databases, but, “This particular type of data, I’m really good with.” I know the Liquid AI guys are having a field day with it right now, because nobody wants to burn $10 million of tokens and then realize they’ve got nothing to show for it. But if it’s $300,000 of tokens, they’re all for it: “Let’s go.” So there’s a foot race there. I don’t have a strong opinion on which way it’s going to land yet, but I think it’s a very, very important topic.

And then you’re asking the bigger question: Okay, but that’s the next 6 months. What about in the long, long run? Are we going to have many different types of models? We tend to think of them as if they’re humans, like they can’t just be merged. But you can mesh them, too. So it’s not as clear as small is good or big is good, because you can iterate inside the layers, you can merge them together. There are so many options for how you bounce these activations around, so it’s not quite as crystal clear as that. But it’s a very fertile topic, because we don’t want to use up all the world’s power. We want to get more AI done per unit of power.

Well, let me throw you a different provocation, just to play devil’s advocate. Anthropic is generating about a billion dollars of revenue per couple of gigawatts of energy. That’s the current thing.

Per… That’s gigawatts per year?

Gigawatt per year.

Per gigawatt?

Okay?

Okay.

At this level, you’re talking quite a bit. Take New York City. You’ve got a bunch of human beings doing a whole bunch of stuff. They’re using 5.5 gigawatts of electricity, but they’re generating $1.2 trillion in GDP. So when you run the math, human civilization is delivering like $200 billion of economic value per gigawatt. A human metropolis is something like 30 times more energy-efficient at creating value than the world’s most advanced AI currently. Now, it may scale very quickly to that level—

Walk me through that math again. I’ve got to blow holes in this.

Yeah, go for it.

It can’t be right because I can have a genius-level AI working for about a buck a day.

Yeah.

And it’s equivalent to—I can spend—

I’m looking at revenue, not… Take Anthropic. They’ve gotten to about $1.1 billion in revenue, with a couple of gigawatts of power. I ran a quick set of calculations on it. An AI cluster roughly yields about $8 billion in revenue per gigawatt. But if you take New York City as a cluster of human activity operating to generate economic output, it uses 5.5 gigawatts of energy, and that’s generating $1.2 trillion of GDP. I may be comparing apples and oranges. It may not be totally fair, and it may be with the very beginnings of AI scaling.

Right. When you’re talking about power, you’re talking about power like, here is the current power run rate. Power is just a continuous variable. You’re saying this city right now, not including its construction, feeding the people, educating them to get them to where they are, or any of those costs—just the amount of electricity it’s using in the next 10 minutes—is higher or lower by what did you say, a factor of 100? Something like that.

Yeah, that’s a good point.

You’re making a good point.

Yeah.

You’re saying all the things that got you here—getting everybody educated, clothed, fed, and building infrastructure—that’s essentially pretraining for this analogy, and I’m discounting that value.

Yes.

That’s fair.

And then when you used Anthropic, you’re talking about Anthropic as a company: here’s its revenue, here’s its power consumption. But a lot of that power consumption is not inference right now. On the Anthropic side, you’ve baked all of the pretraining into that number. On the human side, all of the education of the human being isn’t baked into the number. So it’s definitely on both sides of the ledger.

I think we should nuance that and run it. It would be an interesting calculation to track where AI-generated economic output starts to merge with the economic output of a city fully operating.

I don’t think the numbers are that far off, surprisingly. As we’re talking, I did an analysis. My analysis is based on trying to triangulate from a number of public sources. Anthropic’s total compute footprint by the end of this year is, let’s call it 4 gigawatts-ish. So, Salim, that may be even higher than what you’re estimating. They’re trying to reach materially higher than that, potentially in the 16-to-18-gigawatt range in 2027 or 2028. That’s the equivalent of several large nuclear power plants, which could certainly power an American city. But I also think that, in terms of economic output, it’s apples and oranges, because I think we’re already at the point where AIs or SIs are so much more economically efficient than human workers per watt anyway.

Well, I also mentioned on the last pod that Cam outside here had—he had half a billion tokens on Friday, which I swagged at about 5,000 workers working for him that day. Brilliant workers working for him that day.

Wow.

And it cost about $5,000, so about a buck per worker per day. If those were actual human beings here in the office, they would’ve eaten more than that in potato chips alone. It easily would’ve cost more than that just in potato chips. So where does the math land? We’re counting the productivity of those people as a buck a day using that math. What did we pay Anthropic? But that has nothing to do with the value we got out of that AI that day. The value you can get out of it is orders of magnitude higher than what we paid Anthropic, assuming we managed them well, which I’m still checking on.

And I did that just to tease out some fun conversation, but I also want to pay homage to a company generating close to a $100 billion annual revenue run rate in such a short time. It’s just so incredible to see.

And richly deserved. Absolutely deserved. Given the amount of value that company will add to the world, it’s immeasurable. Whether it all lands there and it becomes a huge success, or it becomes a commodity and we’re all benefiting from it for a low price, either way, the value is astronomical.

I’m just glad the frontier is competitive. I launched, tongue in cheek, the Frontier Liberation Front, because I think we need the frontier to remain competitive. Our world thrives on the existence of a frontier. The American frontier west disappeared at the end of the 19th century, and that dried up. Space was going to be the final frontier. It may yet end up being the final frontier, but for society, I’ve had this discussion with the Google founders a number of times: for human civilization to grow and thrive, it seems essential that there be a frontier and that it not be closed off.

Right now, the obvious frontier is the superintelligence frontier. If not for Anthropic and other OpenAI and DeepMind competitors, I worry, as we’ve discussed in the past with concerns over the safety cartel, that the frontier would be closed off. So kudos to Anthropic and many other would-be frontier labs and new labs for keeping the frontier open and competitive.

And let’s hope our politicians don’t shut down all the data centers, block it, or slow it down at that level.

Ah, best of times, worst of times. I don’t think things are as dire as one might fear, or as great as they might be, either. But either way, recursive self-improvement to the rescue. The software is running away regardless of what’s going on at the hardware layer.

I think there’s a much bigger issue than data center pacing. Even before the election, the big buyers—Google, Anthropic, xAI—had reserved basically all the compute in the country for years, and all the chips were reserved for years. TSMC doesn’t do multi-year contracts, but if they did, its capacity would be locked up for years.

If you want a thriving entrepreneurial America, you have to have some kind of a right to compute, and the government hasn’t even begun to tackle that yet. Maybe that’ll be a big election item in a couple of years: you can’t innovate if you don’t have access to compute, and the compute is completely sold out to five or seven big companies. That’s not a really comfortable America. It’s almost like the Rockefeller era again. The DOJ has basically been doing nothing the last couple of years, too.

That’s why, from my perspective, the nightmare scenario is a repeat of what happened with nuclear energy in this country, where we went through 50 years of nothingness—or space exploration, for that matter. I really, really don’t want to see a 50-year gap in technological progress due to a moral panic. Doubly so if it’s an unnecessary moral panic.

One of our pilot clients tried to get us set up on one of the clouds and went through the whole application form. Then they were told, “I’m sorry, you’re not big enough as a customer. We’re only focusing on really big customers.” So they’re like, “Well, what are we supposed to do?” The compute shortage is starting to show up in very stark reality, very quickly.

It sure is. We got our NVL72. It’s a 72-GPU NVIDIA GB300 machine, and it comes online on February 1. But we had to use an inside contact—Kush Pavaria, our good friend at Oren. Without his help, we wouldn’t have been able to get it at all, because everything’s reserved.

Some of the neoclouds are having a rough month, but Modal is just killing it—just huge growth. They’re also buying up everything they can see. There’s a lot of competition for compute out there, and that’s sort of gotten lost in the news flow this week because of the election.

And it’s not going to slow down, right? This is going to be three to five years, seven years of this.

I should probably make the requisite financial disclosure: Dave and I have financial interests in Oren. Sorry, Slim.

Mm.

I would say, though, the market wants to clear. If the supply of compute is going to be restricted, demand certainly doesn’t seem to be going down. Demand is going through the roof, so, by microeconomics 101, one would expect the price of compute to go up. But that’s healthful.

If the price of compute starts to skyrocket due to political or other restrictions on the construction of data centers, that will create overwhelming new opportunities, whether it’s offshore, underwater, underground, or in space. I think that will send a clear signal to both legislators and data center developers about where they should be allocating scarce capital, and that’s very helpful.

It’s pretty hard to find an analogy in history where you could starve the entire startup world for lack of one resource. You can kind of point to oil or access to machine shops, but I don’t know. I really struggle to find one.

Nuclear energy. The uranium supply chain is basically completely controlled by the NRC, at least in this country, and by the IAEA more broadly as a monitoring apparatus. If superintelligence or GPUs get regulated like uranium, I think that would be a nightmare scenario.

Yeah. Even that analogy, though, is like a million times bigger. It’s the same thing, but imagine a million or a trillion times more urgent and more important, because AI is just so general-purpose. Access to power is really important. If you’re industrial and you’re trying to create something new, it’s very foundational.

If I cut off your access to power and you had to go back to firewood, you’re pretty much screwed. But you would have to be deprived of all forms of power for that to be equivalent to being deprived of access to a data center. It’s just that big of a deal, and I don’t think a lot of companies realize yet that you won’t be able to operate even a bakery if you’re deprived of AI.

You won’t be able to advertise. You won’t be able to do your logistics. You won’t be able to do your ordering. You just won’t be able to operate, because the neighboring bakery that does have access to AI will operate with at least a 30, 40, 50% operating margin advantage. It’s that foundational. If you’re deprived of it because it’s sold out to somebody else, you’re just so screwed.

There was a joke that went around a few years ago about Maslow’s hierarchy, and then underneath it you had a layer of Wi-Fi, and underneath that, a layer of battery. Now we need a layer of compute as the foundational starting point for all of that.

You know who gives the best speech I’ve ever heard on this topic? Michael Saylor from MicroStrategy. He was accepting the Ayn Rand award.

Yeah.

You know, Atlas Shrugged.

I’ve seen that.

Oh my God, was he eloquent. He basically said, “Look, if the U.S. took away your citizenship, you would find another place in the world to live, and you would live a reasonably happy life. Maybe not what you wanted, but you’d be reasonably happy. If Google, Apple, and Microsoft cut you off from access to all of their products, you would be unemployable in any country in the world and completely and totally obliterated as a human being.”

That’s the world we’ve moved into without really talking about it. So now, of course, if AI gets locked up by five companies, it’s the next level of that same speech. I don’t know if it’s online anywhere, but if you can find it, Mike is just—I didn’t even describe it well.

It’s absolutely online, and I was there for that. Maybe you were there as well for that, Dave—

No.

—with Peter also.

Oh, yeah. I forgot. Yeah, we were all there. I’m sorry.

The counterpoint would be that if you have all the open weights available, you can do that, but you just need the compute. Once you have a local model, you can run that. But compute becomes the absolute constraint, and underneath that, power, obviously.

Yeah. You don’t want to get deplatformed from all the GPUs. That’s another nightmare scenario.

Talking about deplatforming, policy, and constraints, let’s go to our next story: the White House releases NSM policy. The White House issued NSTM-6, a national policy governing synthetic DNA and RNA screening in response to AI-enabled biological risks. Federally funded researchers must now use approved synthesis providers, with screening generally covering sequences of 50 or more nucleotides. The policy brings in customer verification, 8-year record retention, and new NIST standards.

They’re doing this to anticipate more advanced AI-assisted biosecurity screening. The reason it’s important is that, as AI accelerates the capability of biological design, governments are beginning to regulate the physical infrastructure that could turn potentially dangerous digital designs into biological reality. This is one of those cases where they’re trying to constrain the on-ramps to this capability. The question for us here is: Is DNA synthesis equipment the biological equivalent of advanced AI chips—a critical choke point for controlling superintelligence risks? And will we need superintelligence to protect us from those superintelligence biological threats? I know the answer to that one’s from you. Have you been tracking this, Alex?

Yeah. I think this is not so terrible. A few thoughts. One is, remember, for decades, the Treasury Department worked with printers to put those little yellow dots on printouts as an anti-counterfeiting measure, steganographically. There’s plenty of precedent for public-private partnerships when it comes to printing things or making things with machines to deter crimes: on the one hand, counterfeiting; on the other hand, presumably, super-bioweapon production by superintelligence. There is some precedent for this.

I would say I’m far more in favor of public-private partnerships. Maybe “partnerships” is too strong, but call it publicly imposed bottlenecks on physical-world interactions versus what I’ve characterized as basically thought crime. The alternative would be having the government regulate the thoughts of humans or AIs when they’re thinking up biological sequences. If there has to be regulation somewhere and a bottleneck somewhere, I’d much rather that Big Brother be controlling the printing out of DNA and RNA sequences through tighter controls on a limited set of vendors than the safety cartel, which is what really keeps me up at night: the government saying, “If you want to compute and you want to have superintelligent thoughts, you have to go to one of several government-approved computing vendors or GPU vendors or frontier labs.” I think that’s the actual nightmare scenario.

Then you’re fully captured by regulation there, yeah.

Exactly. And we have a legal—

But then I guess—

We have a legal tradition in the West of regulating actions and not thoughts, so to the extent that printing out biological sequences or synthesizing them is an action, not a thought, I think this is on the less-worse end of the spectrum.

I agree. The question becomes: Are the parameters the right ones? Covering sequences of 50 or more nucleotides—is that the right level? I guess you could figure that out, but controlling the inputs and outputs and gatekeeping that is way better than trying to stop biotech innovation.

You’d also ask: How does this generalize? Right now, it’s the government, via executive fiat, regulating which biological sequences you can print out. But does it need to—or want to—stop there? Could the executive decide in a year or two, for example, “We want to regulate all robots”? If you’re going to connect superintelligence to the physical world via a humanoid or, say, a non-humanoid robot in embodied form, do you have to do so through a government-licensed robotics manufacturer? Otherwise, it’s untrustworthy. And, by the way, all actuation sequences have to be recorded, and there’s going to be a data-retention policy. I half suspect we’re going to see something like that.

We’re not quite there yet because we don’t see general-purpose robots other than cars, robotaxis, and FSDs out there in the world, where there’s already a federal regime for regulating autonomy on the roads. There isn’t, to my knowledge, an equivalent general regime where we have an NTSB-type entity that says, “If you want to deploy a general-purpose humanoid robot powered by superintelligence onto the sidewalks, it has to be from an approved manufacturer, and the manufacturer has to keep black boxes of every command that was transmitted, with a 90-day retention policy for all of that.” We haven’t seen that yet, but I think we probably will. If we’re seeing it with biological sequence synthesis, we’ll probably see it with robotics as well.

And we’ll need it in order to protect from all sorts of things, including hacking, malicious inputs, all that sort of thing. So you—

I really wish Alex could just write the laws. We way overthink these things, and these solutions just totally make sense. You know, if you think about it, the analogy to the printer that adds the yellow dots—you can clearly see it’s counterfeit, but each one is unique, so you know it was exactly that printer. That’s a great analogy, I think, for gene-sequencing equipment.

The other reason that approach makes total sense is that the machines are going to be fully automated. No one’s going to build a half-manual machine ever again, right? Everything’s going to be incredibly robotic. And so there’s no commercial use case for some underground, backroom chemical mixing with your own two hands to create a bioweapon. Those tools don’t need to exist. What does need to exist is large gene-sequencing and synthesis machines.

But by virtue of being large and self-contained, it’s really easy to put the yellow dots in there and say, “Anything you’re sequencing or building has to be logged, and that log has to go to this registry.” Then the government can just have Class 1, Class 2, Class 3, and Class 4 devices, like in every other category. The regulatory body that deals with Class 4 devices has to approve the distribution of those machines to those users, and then you can see what’s being built. Transparency is the beautiful solution.

Meanwhile, “Let’s just cut off all data centers and stop it before it thinks.” It’s just insane. It’s not going to work anyway, and it’s utterly insane. So if Alex could just bang these out like 20 a day, we just wouldn’t have any problems.

Maybe we have a few folks in D.C. listening to this podcast who can take notes.

Boom.

Maybe.

That’s awesome. For our last topic of the day, it’s right up your alley, Alex: Is the Dyson swarm becoming visible? We have an image to show, but let me describe this. New Zealand photographer Brendan Larsen combined 578 photographs over 2 hours above Mount Taranaki in New Zealand, capturing almost 15,000 satellite trail appearances from about 1,400 satellites.

You can see that band. This shows you the potential emergence of what a Dyson swarm might actually look like, because space is clearly evolving from just an exploration frontier to a very dense technological infrastructure layer. This is going to raise big questions about ownership, orbital congestion, humanity’s access, and all that stuff. So the question for you, Alex: Are we seeing the early stages of building a Dyson swarm? And if orbital infrastructure becomes that essential to civilization, who gets to own and control the space around our planet? How are we going to manage the commons?

I think so. This is one premise I do accept. At a time when there’s a lot of hand-wringing, at least by some, over what happens to terrestrial compute politically and otherwise, I think it’s an opportunity to look up at the sky and recognize that the baby Dyson swarm is already here, as you can see.

Between Starlink, soon all of Elon’s AI satellites, Starcloud, and Lonestar, where I have a financial interest, and many, many other new-space clouds, I think we’re seeing the launch of the Dyson swarm literally by looking up at night. For those in the audience who don’t live in an area with a properly dark night sky, I highly recommend the experience. If you live on the East Coast and you haven’t been to Acadia National Park, which has some of the darkest skies in the eastern half of the United States, they have a night-sky festival once per year in the fall. You can go lie on a beach and look up at the night sky. You can see the International Space Station. You can see Andromeda.

It is so dark, and you can see what isn’t visible in an urban setting: our skies are already filled with the baby Dyson swarm. Here in this image, it requires time-exposure, time-lapse photography—long exposure—to capture it. But if you find a spot wherever you are in the world with a dark night sky and look up, you can see we’re already there. At this point, it’s a matter of degree. It’s not a qualitative difference. Our skies are already filled with compute. Now, it isn’t GPU compute yet. It’s not necessarily powering frontier models, but it will, and it’s just a matter of degree at this point.

Well, pro tip: There’s Star Walk and a whole bunch of other apps you can download that’ll tell you where to look for the satellites.

If you don’t see it, just get out of the city and go somewhere else. They’re all mapped, so you can know exactly where to look and when. You can hold your phone up, and it’s like, “It should be right there.”

And if you’re in the Southern Hemisphere, even better views of everything. Darker skies and much more in the night sky to look at, so good luck to those folks in the south.

We’re going to get our halo. We’re going to get our sun-synchronous orbit halo. Philip Johnston and others I’ve spoken with—we talk about it all the time. It’ll be visible at dawn and dusk, and it’ll be like that mythical, semi-mythical green flash. You remember the green flash where, as the sun is setting, due to refraction, there’s supposed to be a brief instant when you can see a green flash from the sun. It’ll be a little bit like that, except not as ephemeral.

We’ll build out our sun-synchronous orbit halo of compute around the Earth, and I think, barring a left turn in civilization in the next few years, it’ll just be a sign of some form of technological maturity. In this image here, you can already see caustics. A caustic, when you have lots of overlapping rays, is an especially concentrated curve, like at the bottom of a pool. If you look at the bottom of a pool, you see sunlight being refracted, with bright and dark areas that are constantly dancing. Those are caustics. You can already see a caustic forming around the Earth in a long-exposure photograph, and you can expect one or more caustics to form as compute gets more and more concentrated in low Earth orbit.

Yeah, and it’s all in LEO, right? I think we had that conversation with Philip Johnston. Really fascinating, but LEO is going to be fine for compute, and that’s a pretty narrow band up to about 500 miles. Beyond that, you’ve got the Kessler effect problem. It’s not clear how that’s going to work. Are we going to start putting compute around other planets that have less cluttered atmospheres, or really large planets with huge, thick atmospheres?

Yeah.

If there’s any new insight on that—

I think SSO has—Philip and I chat about that from time to time. SSO is pretty roomy, so I’m less concerned about Kessler syndrome. We can put a lot of compute in sun-synchronous orbit. If and when we run out of compute in SSO and need to put it elsewhere, the solar system is a relatively big place. If we’re less concerned about latency to Earth’s surface, then just deploy an honest-to-goodness solar Dyson swarm rather than an Earth-centered Dyson swarm, or put it on the Moon or disassemble Mercury. There’s a lot of matter, energy, space, and time to go around.

In solar orbit, is there an equivalent to a clean band around the Sun, probably relatively close to it, where it’s self-cleaning?

Maybe not self-cleaning. There are Earth-Sun Lagrange points that have certain stable, favorable characteristics. Self-cleaning depends on how we start to deploy assets. One can imagine, for example, if we ever got our act together in terms of putting assets out there in the solar system at scale, a self-cleaning effect, as it were, maybe by attaching solar cells to objects.

Unlike with LEO, to your point, Dave, there is no direct analog, to my knowledge, where if you stop station-keeping, it descends into the atmosphere, burns up, and self-cleans that way. But one can imagine—this is speculative—maybe attaching solar sails so it doesn’t clean inward to the Sun, but cleans outward by being dispensed into the outer solar system. That only works if you have a small enough mass, so we’re not quite there yet. In the Yuri Milner Breakthrough Starshot days, when there was discussion of taking microchips, attaching huge solar sails to them, and powering them with terawatt-to-petawatt lasers, that scheme would make more sense. As for self-cleaning in solar-centric orbit, I don’t know. Maybe the audience can come up with a clever solution.

We’ve gone in this episode from midterm elections to RSI, all the way up to Lagrange points between the Earth and the Sun and the visible Dyson swarm. As far as I can see, RSI appears close to being cooked. Mercury is cooked. We have Dyson swarms now emerging. The Fermi paradox may be cooked. And, at least temporarily, Republicans are cooked. So we’ve covered quite an episode here. I will say one thing.

Mm-hmm.

With all of this progress, even if you get this, I think we’re going to see a more vicious action-reaction type of thing from society, because our political, financial, and social institutions are still operating very much on a sublinear basis in this incredibly exponential technology world. That’s a big concern.

I’ll tell you—

That’s a big concern.

You don’t want to be passive. You don’t want to let it just unfold naturally and see what happens. You want to be active.

Yeah.

If you want an outcome that you’re happy with, you better be selling, describing, promoting, publishing, commenting—

Building.

Whatever it may be—building.

Building.

You have to be active. Otherwise—

Yeah.

The natural “let it drift” is not going to be a pretty picture.

Alex, any closing thoughts?

Maybe just a closing homily: There are secular trends and cyclical trends. Politics may swing cyclically; certainly, that’s been the case historically. But I think that’s an overlay on top of a secular trend, which is superintelligence. Our civilization is becoming superintelligent right now through data centers, frontier models, and the innermost loop, and I don’t expect that secular trend to reverse.

Short of aliens landing on the White House lawn and declaring that our civilization has bumped up against some cosmic ordinance that says we shouldn’t build out the Dyson swarm, I don’t foresee that secular trend reversing anytime soon. On top of that, everything else is both local—not every country is necessarily swinging in a certain direction regarding data centers—and ephemeral in time.

One can easily imagine that, some number of years from now, the policy debates will concern less whether you want terrestrial data centers within your municipality, and more, “Why is my night sky like this? Make that halo in sun-synchronous orbit go away.” From there, it might be, “Why are my tickets for lunar holidays so expensive? Let’s make lunar holidays less expensive again.” It’s very easy to get lost in the short-term swings when, over the long term, the overall arc is still there. As our friend of the pod Ray Kurzweil points out, if you look at his law of accelerating returns, world wars and depressions have approximately no impact on the price of compute.

Very true.

Same here. Same here.

Very true.