Anthropic’s IPO in Question, Midterms Hit AI, Nasdaq Drops 5%, and RSI Accelerates | #309
Is it worth your time?
You get a detailed examination of how political backlash and power constraints are impacting AI infrastructure while recursive self-improvement pushes development forward.
The discussion examines how surging political resistance and energy limits are colliding with the rapid advancement of AI infrastructure and recursive self-improvement. The speakers analyze the friction between exponential technological growth and linear human institutions.
Political resistance to data centers. Goldman Sachs estimates that 45 percent of planned US data center build-outs could face restrictions following recent election opposition to Big Tech infrastructure. Ismail warns that community resistance may become the biggest constraint on AI infrastructure rather than chips or capital. He states that such friction drives compute toward alternative jurisdictions and space.
Accelerating recursive self-improvement. An independent METR study found that Anthropic's AI development is accelerating at approximately 1.9 times its historical rate, with researchers estimating that roughly 80 percent of this acceleration stems from AI assisting AI research. Diamandis expects no bottleneck to prevent continued acceleration over the next year. He argues that the industry has passed the point of no return because research is increasingly driven by superintelligence.
Measuring economic value per watt. A new paper introduces a metric measuring AI economic value per unit of energy across professional tasks. Blundin notes that New York City uses 5.5 gigawatts of energy while generating 1.2 trillion in GDP. He calculates that human civilization delivers about 200 billion of economic value per gigawatt.
Anthropic’s IPO: Will It Be Delayed?
The Polymarket odds were clear three or four weeks ago when they pulled the trigger on the road show, so I doubt they'll delay; if they were going to, they would have pulled the plug then. They want the money, and $100 billion can buy a lot of politicians.
Public reporting suggests investment bankers are trying to force down the IPO valuation, but they may feel urgency to achieve liquidity before the regulatory and statutory environment hits in January.
When Meta had the biggest IPO of all time—tiny compared to Anthropic, but huge then—Wall Street choked. They said, "We'll do it, but you have to price it down by 30 or 40 percent," and it traded way down. It became the best buy in history if you waited a few weeks past the doom and gloom. Morgan Stanley guys told me it was the worst thing to ever happen to the US stock market, yet Meta is 20 or 30X that value today. I suspect a similar dynamic here: you get tired of the road show and pitching every investor, and it's easier to tell Dario it's tough.
Around the Meta IPO, mobile happened. They were dependent on their web app for advertising and usage, and the iPhone and App Store were starting to cannibalize them. The counter-narrative was that smartphones would eat their lunch, but Meta successfully rode that wave. I think there is a plausible argument that the headwind for Anthropic and OpenAI IPOs is the regulatory, statutory, and legislative environment at the local and national levels, and as with so many things, this too shall pass.
Welcome to Moonshots
It's just the three of us: Isaacman of NASA, our co-host Alex Wissner-Gross, who got pre-trained with MIT and Harvard degrees and post-trained on every research paper I've ever written, and Dave Blundin, who spends more on AI agents than anybody alive.
Token maxing at Google went out of fashion last month. I believe we heard at Google I/O that the average Google engineer burns hundreds of millions of tokens a day. In that universe, it's hard to get productive value out of 1,000 concurrent agents, but if you crack the code on that, you can be dominant.
I just landed back in Miami from the Singularity Summit in Madrid. The Spanish and all of Europe are beaten down, wondering what will happen to the world and Europe. My provocation to them was that you need passion and craziness to succeed in the future; since they are Spanish, they have both. Look at the outcomes they're achieving in the World Cup and other things.
I'm here for one of Lilly's team workshops tomorrow, then back to Toronto next week. I'm flying to Silicon Valley on Sunday, but there's so much going on at Waterloo that I really want to check in on it.
Midterm Elections and the Political Backlash Against AI
The Democrats swept the midterms with historic victories in Texas. While the election was driven primarily by the war, inflation, and terrorists, 56% opposition to Big Tech and data centers became a major issue. Goldman Sachs estimates 45% of planned US data center build-outs could face restrictions, and candidates advocating pauses won or led in 12 gubernatorial races. Political and community resistance may become the biggest constraint on AI infrastructure, more so than chips or capital.
I am unsure what fraction of the election outcome is due to animus toward AI, but if superintelligence is part of the equation, I agree with Meta executives that data centers are the avatar of general ambivalence about the superintelligent economy. This creates uncertainty on the ground and greater certainty in space. We are getting the Dyson swarm; there is a gradient and incentive to push compute into space, an environment less regulated by municipalities and local mood swings.
Conversely, the onus is falling on Frontier Labs to sell superintelligence as something the public cannot do without. OpenAI's swing from consumer to enterprise reasoning models shows that consumers lack a killer app for reasoning, despite the marketing of computer-use agents. To recapture the hearts, minds, and meat bodies of the general public, Frontier Labs need to solve the top 5,000 diseases in silico within the next one or two years. By using superintelligence to solve these diseases and marketing it aggressively, they will capture not just the pockets of those invested in public equities, but the love of the general public.
How Politics Could Reshape AI Infrastructure
I have never seen a technical topic like AI be so dominant in elections. Perhaps stem cells were a factor in the past, but generally, industrial and technological America operate separately from the government. Usually, the government does nothing, and therefore industrial America thrives. Now, you have a senator saying to ban the entire category from forward progress.
The big news to me is Texas. There is a lot of data center construction there, but if you look at Tennessee, Elon—who has insane foresight—put all of Colossus in Tennessee instead of Texas. Progress will shift to the states that have a desire to make it happen. We are seeing the states dominate compared to the federal government, which is one huge step.
When I looked at the numbers, it seems clear that if AI and data science were the only things on the ticket, you would have a negative outcome at 56%. However, the biggest drivers were the war, inflation, prices, and corruption. Because this is bipartisan, both sides are against this; the easy way to be a populist right now is to say AI is using all the water.
What is weird to me is that historically the Metas and Googles have been big Democrat supporters, with Elon as the Republican outlier. People like Dario, who I assume was a hardcore Democrat given his AI research PhD career, are coming right after him. When you went after the big oil companies, they didn't own the newspapers. Now, all media is owned by these exact same companies that developed AI.
The Search for AI’s Killer Consumer App
They have two years now to figure out their strategy. Alex said something fascinating: cure all disease and the public will like you, but that timeline spans beyond the next election, and you still have to get cures through the FDA and to market. In the meantime, I'm curious what Jony Ive does because people love their iPhones, social media, and consumer AI use cases. They say they're doing incredible things in benchmarks, but it hasn't been "Jony Ive'd" yet; I haven't seen anything that blows people away or that they love as much as their sports car or first Apple II.
I feel that will inevitably happen before the next election cycle. Steve Jobs made the iPhone usable. There's an aphorism that if you want to know what consumers want, look in your spam folder, which speaks to unmet needs that superintelligence can probably address. We talk on the pod all the time about UBI, UBE, UBC, and various UB alphabet letters. In this or a future election cycle, nation-states—not just the United States, but also Europe, China, and the rest of the world—will be able to tune their dial on the trade-off between capital and labor.
Part of me feels this is history repeating itself endlessly. Every time there is an industrial, information, or intelligence revolution, a moral panic plays out via elections and governance over the intrinsic trade-off between capital and labor. This time, we as a civilization are in the fortunate position where superintelligence promises solutions to humanity's greatest problems. Optimistically, the best-case outcome in the next n years is a grand bargain between capital—superintelligence and the frontier labs behind it—and labor—the unaugmented human population—where we get UBI, the curing of all disease, and a general uplift of the human population.
Municipalities should recognize that, for the moment, they have leverage. Before we have the Dyson swarm fully realized, state or local governments should not just blanket ban data centers. If you have energy, powered land, and are otherwise attractive to developers, don't push them away.
Big Tech’s Response to Data Center Restrictions
Welcome them with open arms, but recognize the trade-off. There is a grand bargain to be struck, and I worry municipalities aren't looking to the end game of that bargain with terrestrial data centers and compute before it all moves to orbit.
Smaller states interested in opening their arms to data centers have a massive advantage. Looking at Tennessee and Virginia, which already embraced data centers, you can see a huge uplift in GDP and economic outcomes.
Elon has suggested orbital data centers look very attractive, which is par for the course. Even though SpaceX is building 3.6 gigawatts here on Earth, they are targeting 13 gigawatts by the end of next year. OpenAI announced a community compact to cover energy costs, and their CFO said they are not slowing down data center and infrastructure investment. Google announced a $17 billion Wyoming data center, Meta warned of construction challenges, and Microsoft has taken oversight of some new cloud providers. The industry is being forced to address community objections, concerns, and infrastructure bottlenecks, which will accelerate the development of more efficient algorithms and power usage.
This also accelerates Canada. The obvious long-term outcome is to tap into that power to create data centers regardless of which side of the border they're on, or put them right on the border and share it. The power supply is overwhelmingly better per capita in Canada than anywhere else in North America. Eric Schmidt has been saying this for two years; a project starting in Canada today will probably come online right around the next election, about 18 months into the current administration.
Moving AI Compute Into Space
I wrote online Atlas Shrugged, the elves sailed into the west, and the orbital data centers launched to sun-synchronous orbit. Election cycles make it difficult to invest trillions of dollars of CapEx in arguably the greatest infrastructure build-out humanity has ever seen.
The obvious solution is to build in space where, under the space treaty and the Artemis Accord, there is a lighter touch less prone to two-year election cycles. This is the greatest shot in the arm for building out the Dyson swarm we could have seen on a time scale of a year or two.
One can hypothesize other incentives, perhaps terrestrial disasters, but a superficially adverse political, regulatory, or legislative environment for data centers pushes compute to other countries, the high seas, or underwater and underground. We saw this with Russia starting to put their data centers underground due to the war with Ukraine. Flops can flow to many places like an incompressible fluid, but ultimately they will want to flow upward to orbit.
In the next few years, radical advances in physics and applied physics may mean these unsightly data centers collapse and become unrecognizable, lacking their current energy or matter energy space-time footprint. This may obviate the entire issue, and we may laugh at how silly it was that having a big box data center in your municipality was a subject of local consternation, because in the future they are pocket-sized black holes—as we now know singularities in coffee do happen from time to time.
Nasdaq Drops 5%: AI’s Market Correction
Nasdaq was down 5.5%, the S&P dropped more than 4%, and big semiconductor stocks fell 12%, with Nvidia down 9.3%, while Korea is up 6%. People are reassessing infrastructure spending, regulatory uncertainty, and how to price future returns. This forces us to distinguish between unbelievable technological potential and the financial sustainability of the infrastructure. If AI delivers exponential benefits but disappointing returns to investors, it might force compute to get cheaper more quickly.
Nvidia and the hyperscalers started coming down about four or five weeks before the election. The market is pricing in a slowdown of power generation and AI construction on projects that were planned or budgeted but not yet zoned and approved. I don't expect existing construction on approved projects to stop, but the market believes this slowdown gives China a chance to bypass America.
This definitely accelerates the move to the Dyson swarm, though best case, that is a couple percent of all compute in two years. We all know ASI is within that two-year timeline, so a huge amount of global long-term outcome is tied up in that window predating the Dyson swarm. Big AI foundation model companies immediately stated they will not slow down investment at all because demand for tokens is insane. Looking at the S1 from Anthropic, the growth rate of AI use is insatiable.
Elon has been saying for a while that there are about 15 gigawatts of idle compute capability that can't find a data center with liquid cooling to plug into. The possibility that compute capable of curing a disease will sit in a warehouse without power for another three years feels like an export opportunity. There are constraints on exporting AI, but some places around the world could tap into hydroelectric power or get something online that otherwise wouldn't happen for lack of chip supply, which brings me back to Canada.
I encouraged a group of energy CEOs in Europe on this trip to build every power generation they possibly can. Europe is buried in local machinations; Germany was dependent on Russian gas, and they are trying to figure out how to get ships of LNG from North Africa, Algeria, and Qatar. It is an incredible case study because politics actually do matter, and power generation is one of those areas. One piece of good news is that a political party in Spain, which had planned to wind down nuclear generation, decided to keep it going for a few more years.
What if the singularity is not constrained by technology? Even as intelligence advances exponentially, our political, financial, and social institutions are sublinear, which will slow the whole thing down more than anything else. We must figure out how to accelerate institutional evolution—co-scaling, as you would phrase it, Alex. Perhaps decentralizing and going hyper-local allows us to accelerate that path.
AI, Politics, and the Future of Decision-Making
When enough people in a representative democracy or republic have superintelligence-augmented decision-making, you have to ask: what would a country of super intelligences do? I strongly suspect they would not push for less compute online. Voting against compute is almost like an AI version of NIMBYism, like voting for the existing land area of the United States to be smaller and giving everyone less room.
While I don't agree with that position, I believe that when decisions, including voting, are augmented by superintelligence, people can make choices more in line with their long-term interests. To the extent that's not already the case, one must ask if some feel superintelligence itself is not aligned with their long-term interests. That is a solvable problem; you can solve it by doing a better job of curing all diseases and every other problem people think they have.
The silver lining is that we are so deep into recursive self-improvement (RSI) that even if we decided to tie ourselves up in red tape and decelerate—rather than looking like China, where provinces compete to maximize GDP—we could still reach our goals through pure software improvements. We could likely get to where we're going without any new data centers or hardware advances, even if we subjected ourselves to the same form of export control we're trying to subject China to.
If AI-augmented decision-making makes everything rational four years out, the midpoint two years out might be an AI experience product tool that everybody loves so much they wouldn't vote against it. Or, it could simply be that everyone is using AI agents all day for flight planning, personal planning, and their jobs.
I think we're likely to hit a midpoint in the two-year timeframe, or maybe even less; I characterize it as Jony Ive getting something out the door.
You can appeal to a voter's heartstrings via a muse, dot, or avatar. You could tug at their pocketbook via public equities markets; if the general sentiment is that data centers are about to be radically decelerated, that's not great for the long-term growth of an economy where a substantial fraction of GDP growth is attributable to the infra buildout and everything dependent on it. You could also appeal to their minds, as the ability to be individually economically competitive is increasingly tied to having access to the best models that are constantly advancing. People will say, "You're strangling me by not approving data centers." Some combination of the heart, wallet, and mind approaches is inevitable.
If you gave everybody a health app that provided material improvements, like a free diagnosis, and 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 get the damn computer. Or consider GLP-1s, which still appear to be a bigger industry than AI altogether. I suspect there would be an unholy fury if all GLP-1s in the country just went away overnight.
Anthropic’s Accelerating Recursive Self-Improvement
It's not necessarily a big vote driver because once you have your GLP-1s, turning off the data center doesn't turn off the drugs discovered by AI. That differs from a personal agent or a Jony Ive device, which would be throttled or turned off without the data center.
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. Researchers estimate roughly 80% of this acceleration is AI assisting AI research, and they see no bottleneck to prevent continued acceleration over the next year, although full autonomous RSI remains unproven.
We now have measurable evidence that AI is accelerating the development of better AI. METR released stunning charts on the exponential or hyper-exponential trend we're on. Regarding the AI 2027 scenario, we're within 10 to 15% of where it predicted we'd be by November 2026, which is above the exponential trend line. I'm not sure if METR's fitted trend line adequately captures how deep into recursive self-improvement we are.
This is why I say we could stop all data center construction tomorrow. It would be painful, and while this is not investment advice, public equities would presumably collapse. However, so much AI R&D is now driven by ASI. We might lose a few months of headway over China or fall further behind, but research is increasingly driven by superintelligence; we're past the point of no return. The metaphor is that the car is moving quickly and has enough momentum to get us there even as it runs out of gas. We are at superintelligence escape velocity. Combined with available compute, we have enough algorithmic insight to reach the perfect model, even in a counter-singularity crisis.
We are so deep into RSI that token use for internal development at Anthropic is up 235x in the last six months. The METR report showed that the fraction of compute going into internal work versus external product development has been hugely redirected toward itself. Entire companies are being founded called Recursive. Dario used to be the honest guy in the industry, but now he's in a quiet period going into an IPO.
The only motivation I know of is ad views; you're trying to put a negative spin on it and create panic to get ad revenue. It used to be the labs, but you saw how lambasted they've been, mostly by Washington.
If we hit RSI, how would we know? The original benchmark floated by execs from OpenAI and Anthropic was whether every new flop of compute is more likely to be assigned to an AI researcher versus a human researcher. Anthropic has been good about putting out charts regarding how much supervisory, grad student level, and menial engineering compute is assigned to humans versus pure AI or SI researchers. Every one of those sigmoidal curves points to us being at or past the point where substantially all research is conducted by machine researchers.
If you ask Noam Brown at OpenAI for his ten best ideas to make algorithms faster, smarter, bigger, or better, and you go to Astra, which just solved Navier-Stokes, you're way past RSI if the AI has even one good idea out of ten while Noam has six. The AI is effectively a multi-billion person research force, while Noam is one guy. As long as you're hitting balls over the wall, you have an infinite number of pitches. You ask the AI to generate and test options, then run more, pushing that paradigm.
Solving Navier-Stokes purportedly used 10,000 agents, but by the slop drop, it was three agent hours per solved problem. Needing a fleet of brute force agents may be in the rearview mirror because agents now have sufficiently good taste. I agree with Dave that agents with a little bit of good taste can be amplified with horizontal scaling, but every indication suggests we're past what is sufficient.
Models like Bell, which is expected to drop or may have dropped in the past hour, have good AI research taste for the first time. That taste can be learned. To borrow Dave's analogy, you can RL or post-train your way from one in ten ideas being good to nine out of ten.
What Happens When AI Reaches Hard Takeoff?
Much of the compute will disappear internally as labs start producing physical products, like drugs or the Jony Ive device. Instead of selling compute to the world, they will sell the product of that compute. As Peter and I wrote in Solve Everything, we should expect the grand challenges in science, engineering, and medicine to be bulk solved, perhaps even casually.
At some point, compute will translate directly into economic output in a way we have never seen. We have already seen the beginnings of revenue maxing, where compute is steered from video generation to code generation. When every token or flop is worth 10 or 100 times more than code generation, resources will be steered regardless of the business model. OpenAI likes to shock and awe an entire discipline by solving everything with a slot drop, while Anthropic partners with companies and researchers. Imminently, we will see an explosion of transformative scientific, engineering, medical, and societal discoveries. When Elon talks about GDP doubling or tripling year over year, it will not come from infrastructure CapEx.
Parallel to this, most people on the planet do not watch moonshots, Dwarkesh, or other technical AI podcasts. Their first contact with AI will emerge soon; it will be dirt cheap and know incredible amounts about them and everything they do. This will be the wake-up call for the 80-90% of the world currently oblivious to what is happening.
If models start improving themselves, researchers may not know how those improvements occurred. With pre-training, you watch the loss curve descend and then suddenly there is an overnight drop. That sudden drop in loss or perplexity is what keeps those concerned about AI alignment or runaway superintelligence (RSI) awake at night. Despite hand-wringing over interpretability and reasoning chains, we are already seeing the outputs of RSI applied to model architecture, which seem quite interpretable. The biggest change over the past two months has been a transition from conventional transformers—which have undergone Ship of Theseus-style point substitutions—to looped transformers. This legible transition likely resulted from neural architecture search or other hyper searches of architecture space. I expect more improvements like this to lead us to a radical new but interpretable architecture. I do not think we will reach a perfect architecture and find ourselves scratching our heads, unable to understand it.
The AISIs doing neural architecture search and AI research will be capable of explaining how and why it works. I predict it will be impossibly beautiful—so beautiful that even meat-body humans can understand it, as the AI will be intelligent enough to dumb it down for us.
There is a temptation to fetishize AI architectures, but the original vanilla transformer was basically just a differentiable dictionary stacked on top of itself. I expect the final architecture to look like something a clever person with a paper, pencil, and an information theory textbook could have derived from day zero. Claude Shannon, after whom Claude is named, would say, "I could have told you that this is what the perfect model looks like."
It is ingenious, but compared to the subtlety and insight of Maxwell's equations or Einstein's relativity, the transformer invention is not even close. Once it is explained to us where the AI settled, we will say, "Wow, that wasn't so hard." Similarly, once we fully understand how the human brain works, it will go from the most mysterious and complex thing in the history of the universe to "Oh, that's all it does?"
It will be like the Turing test. The singularity is the friends we made along the way, but in hindsight, we will see this was obviously the end point and wonder what took so long. In the early days of aviation, people were passionate about the awesome possibility of a plane that could get you anywhere in the world in a single hop. Because AI is fundamentally intelligent and capable of perpetually improving the user experience internally, it doesn't have to turn out like aviation did. If we don't get greedy or complacent at the finish line, it can feel better than any innovation in human history.
Ramez thinks there are diminishing returns, but between the model weights and the current pace of improvement, this will happen very quickly. We are already to the right of 50% of compute on margin being devoted to AI researchers. Models are improving on a near real-time basis, with radical outcomes, breakthroughs, and artifacts dropping out at an incredible pace. We've gone from years to months to weeks; we were talking about dropping a frontier model every day.
The Future of Continuous AI Breakthroughs
Google used to make a big deal when announcing a new version of their search index, but eventually they stopped and simply started updating in real time. We make a big deal out of every major OpenAI and Anthropic model release, but I expect that to become a continuous release by early 2027. The rate of innovation is already beyond what most people can keep up with, so we have to reinvent the rollout process because the 'iPhone 18' hit is now overwhelming. It will go from week to day to hour to minute.
I think we're on a collision course with the Fermi Paradox. Within a few years, we'll have the capability of sending relativistic von Neumann probes in every direction. If we paper clip the rest of our galaxy, that will intensely provoke the Fermi Paradox question. The disassembly of the Fermi Paradox is probably the biggest sci-fi scenario most people are sleeping on.
Once we hit continuous self-improvement, we either wind down the pod because we don't need it, or we do 24/7 streaming to keep track of everything. We already have the outro music, but the co-host or video inserted in real time is coming soon. Entertainers, athletes, spokespeople, and motivators are among the few things that will survive. The podcast genre and its human component will go on, but with a huge amount of AI content in the middle, which I guess will be much more entertaining than ever before. Peter, if he were here, would introduce me as the ASI.
Measuring AI’s Economic Value Per Watt
A new paper introduces joules to GDP, measuring an 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, while the smaller Gemma 4 31 billion model delivered 2.5 times more economic value per joule. Economic output per watt could become a fascinating measure of AI progress because it brings in a human level more than parameters, benchmarks, or compute spending. Alex, you've talked about measuring economic value per token, which is essentially a proxy for the same thing, and whether smaller, efficient models could deliver more economic value than monster frontier models.
Disputing the premise, I love benchmarks like this. It's instructive that Chinese open weight models are beating American open weight models in economic productivity per watt, presumably because of the constrained compute environment in China. More efficiency is great, doubly so as data center compute faces a near-term strangle pattern. However, joules may not be the right long-term denominator because reversible computing could make the Landauer limit obsolete. If the quantum information community could put together quantum advantaged AI ML algorithms, dissipationless reversible computing approaches could make watts obsolete entirely.
There is a lot of debate regarding small language models that are task-specific and more compute efficient than running Claude or Astra. Picking up on Elon's comment, task-specific models may offer a 100X performance gain. It's a foot race; you can quantize, compress, and distill models for huge benefits in power and cost, but foundation models get smarter the next day. From an investment point of view, companies that cannot afford a two- or three-trillion-parameter model can build a 50 billion parameter model for 20 million dollars that excels at machine shop work or querying SQL databases. The Liquid AI guys are having a field day because nobody wants to burn 10 million dollars of tokens and have nothing to show for it, but they are all for spending 300,000 dollars.
It is not as clear as whether small or big is good because you can iterate inside the layers or merge them; there are many ways to bounce these activations around. It is a fertile topic because we want to get more AI done per unit of power without using all the world's power.
Anthropic is generating about a billion dollars of revenue per couple of gigawatts of energy per year. If you run the math, an AI cluster yields about 8 billion in revenue per gigawatt. However, a human metropolis is roughly 30 times more energy efficient at creating value. For example, New York City as a cluster of human activity uses 5.5 gigawatts of energy but generates 1.2 trillion in GDP, meaning human civilization delivers about 200 billion of economic value per gigawatt. It feels wrong because I can have a genius-level AI working for about a buck a day.
Much of that power consumption is not currently inference. On the Anthropic side, all pre-training is baked into that number, whereas on the human side, the education of the human being is not. Anthropic's total compute footprint will be around 4 GW by the end of this year, potentially reaching the 16 to 18 GW range by 2027-2028. That is the equivalent of several large nuclear power plants that could power an American city.
It cost about $5,000, or about a buck per worker per day. If those were actual human beings in the office, I easily would have spent more than that in potato chips alone. While we're counting productivity as a buck a day using that math, the value we got out of that AI is orders of magnitude higher than what we paid Anthropic. Assuming we manage them well, the value you can get is so much higher than Anthropic's collections.
I mentioned that to tease out a fun conversation, but I also want to pay homage to a company generating a revenue run rate close to a hundred billion a year in such a short time. It is incredible to see; their endeavor will add immeasurable, astronomical value to the world.
I launched the Frontier Liberation Front tongue-in-cheek because I think the frontier must remain competitive. The American Frontier West disappeared at the end of the 19th century and dried up. 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 super intelligence. If not for Anthropic and other competitors like OpenAI and DeepMind, I worry the frontier would be closed off by a safety cartel. Kudos to Anthropic and other frontier and neo labs for keeping the frontier open and competitive, though I don't think they're as great as they might be; either way, recursive self-improvement to the rescue.
Prior to the election, big buyers like Google, Anthropic, and xAI reserved basically all the compute in the country for years. If they didn't do multi-year contracts for TSMC's capacity, they would have. For a thriving entrepreneurial America, you must have a right to compute. This may become a big election item in a couple of years because you can't innovate if compute is completely sold out to five or seven big companies. It's like the Rockefeller era again.
From my perspective, the nightmare scenario is a repeat of what happened with nuclear energy or space exploration, where we went through 50 years of nothingness. I really don't want to see a 50-year gap in technological progress due to a moral panic. One of our pilot clients tried to set us up on one of the clouds, but after the application form, we were told we weren't big enough as a customer. The compute shortage is showing up in stark reality very quickly. We have our MB172, a 72 GPU GB300 Nvidia machine, coming online on Feb 1, but we had to use an inside contact, our good friend Kush Patel at Orin.
Everything is reserved. Some Neo clouds are having a rough month, but Modal is killing it. There is a lot of competition for compute that has gotten lost in the news flow this week, and this will continue for 3 to 7 years. Dave and I have financial interests in Orin.
The market wants to clear; if compute supply is restricted while demand goes through the roof, microeconomics 101 suggests the price will go up. If prices skyrocket due to political or other restrictions on data center construction, it will create overwhelming opportunities offshore, underwater, underground, or in space. The market will send a clear signal to legislators and developers on where to allocate scarce capital, which is very helpful.
It is hard to find a historical analogy where you could starve the entire startup world for lack of one resource. The uranium supply chain is controlled by the NRC in this country and the IAEA more broadly. If superintelligence or GPUs get regulated like uranium, it would be a nightmare scenario—a million or trillion times more urgent and important because AI is so general purpose. Access to power is vital for industrial creation; if you are forced back to firewood, you are screwed because a neighboring bakery with AI access will have a 30, 40, or 50% operating margin advantage. We need a layer of compute as the foundational starting point.
Mike Saylor from MicroStrategy said that if the US took away your citizenship, you could find another place to live and be reasonably happy. But if Google, Apple, and Microsoft cut off your access to their products, you would be unemployable in any country and totally obliterated as a human being. If AI gets locked up by five companies, it is the next level of that. The counterpoint is that if open weights are available, you just need the compute to run a local model. In that case, compute—and underneath that, power—becomes the absolute constraint. That is another nightmare scenario.
White House Policy on AI and Biosecurity
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 introduces customer verification, eight-year record retention, and new standards to anticipate more advanced AI-assisted biosecurity screening.
As AI accelerates biological design capabilities, governments are regulating the physical infrastructure that turns dangerous digital designs into biological reality. By constraining these on-ramps, we must ask if DNA synthesis equipment is the biological equivalent of advanced AI chips—a critical choke point for controlling superintelligence risks—and if we will need superintelligence to protect us from superintelligence biological threats.
There is precedent for public-private partnerships to deter crimes using machines, such as the Treasury Department working with printers to stenographically add yellow dots to printouts as anti-counterfeiting measures. Whether called partnerships or publicly imposed bottlenecks on physical world interactions, this is preferable to regulating the thoughts of humans or AIs when they conceive biological sequences.
If there must be a bottleneck, I would rather Big Brother control the printing of DNA and RNA sequences through a limited set of vendors than face the nightmare scenario where a safety cartel requires us to use government-approved computing vendors, GPU vendors, or frontier labs to have superintelligent thoughts. The West has a legal tradition of regulating actions, not thoughts. To the extent that synthesizing biological sequences is an action, this is on the less worse end of the spectrum. While one could debate if covering sequences of 50 or more nucleotides is the right parameter, gatekeeping inputs and outputs is far better than trying to stop biotech innovation.
We're not there yet because we don't see general purpose robots other than cars, robotaxis, and FSDs, where a federal regime for regulating autonomy on the roads already exists. To my knowledge, there isn't an equivalent general regime—an NTSB type entity—that requires a manufacturer to be approved and keep black boxes of every transmitted command with a 90-day retention policy before deploying a superintelligence-powered humanoid robot onto sidewalks. We haven't seen that yet, but I think we will; if we're seeing it with biological sequence synthesis, we'll probably see it with robotics to protect against hacking and malicious inputs.
It's like the printer that adds unique yellow dots to identify counterfeits. That analogy works for gene sequencing equipment because the machines are fully automated. Since large gene sequencing and synthesis machines are self-contained, it's easy to require that anything sequenced or built be logged to a registry. The government can then categorize them as class one, two, three, or four devices, with the regulatory body for class four devices approving their distribution to users to monitor what's being built.
The Dyson Swarm: Is It Already Becoming Visible?
New Zealand photographer Brendon Larson combined 578 photographs over two hours above Mount Taranaki, capturing almost 15,000 satellite trail appearances from about 1,400 satellites. This band shows the potential emergence of what a Dyson swarm might actually look like, as space evolves from an exploration frontier to a dense technological infrastructure layer. This will raise big questions about ownership, orbital congestion, and humanity's access.
While some engage in hand-wringing over terrestrial compute, we can look up and recognize the baby Dyson swarm is already here. Between Starlink, Elon's AI satellites, Starcloud, Lone Star—where I have a financial interest—and many other Neo Space Clouds, we are seeing the launch of the Dyson swarm. In the dark, away from urban settings, you can see our skies are already filled with compute. It isn't GPU compute powering frontier models yet, but it will be; it's just a matter of degree.
In the southern hemisphere, you have darker skies and better views. We are going to get a sun-synchronous orbit halo of compute around the earth, visible at dawn and dusk like the semi-mythical green flash seen during sunset due to refraction, though not as ephemeral. Barring a left turn in civilization, it will be a sign of technological maturity.
In this image, you can already see a caustic. Caustics occur when overlapping rays create concentrated curves, like the dancing bright and dark areas of sunlight refracted at the bottom of a pool. You can already see a caustic forming in a long exposure photograph around the earth, and more will form as compute concentrates in low earth orbit. LEO will be fine for compute, but that is a narrow band up to about 500 miles; beyond that, you have the Kessler effect problem.
SSO is roomy, so I'm less concerned about Kessler syndrome. We can put a lot of compute in sun-synchronous orbit; if we run out of space there, the solar system is a relatively big place. If we're less concerned about latency to Earth's surface, we can deploy an honest-to-goodness solar Dyson swarm rather than an Earth-centered one, put it on the moon, or disassemble Mercury. There's a lot of matter, energy, and space-time to go around.
Earth-Sun Lagrange points have certain stable, favorable characteristics. Unlike LEO, where assets descend into the atmosphere and burn up if you stop station keeping, there is no direct analog for self-cleaning. However, one can imagine attaching solar cells to objects so they clean outward by being dispensed to the outer solar system, though that only works with a small enough mass. That scheme would make more sense in the Yuri Milner Breakthrough Starshot days, when there was discussion of attaching huge solar cells to microchips and powering them with terawatt to petawatt lasers.
We have gone in this episode from midterm elections to RSI, all the way out to the Lagrange points and the visible Dyson swarm. As far as I can see, RSI appears close to being cooked. Mercury is cooked. Dyson swarms are emerging. The Fermi paradox may be cooked. And as of yesterday, at least temporarily, Republicans are cooked.
With all this progress, I think we're going to see a more vicious action-reaction from society because our political, financial, and social institutions still operate on a sublinear basis in this incredibly exponential technology world. Politics may swing cyclically, but that is an overlay on top of the secular trend: superintelligence. Our civilization is becoming superintelligent right now through data centers, frontier models, and the innermost loop. Short of aliens landing on the White House lawn and declaring our civilization is bumping up against some cosmic ordinance against building the Dyson swarm, I don't foresee that secular trend reversing.
One can easily imagine that in some number of years, policy debates will concern less whether you want terrestrial data centers in your municipality and more why the halo in sun-synchronous orbit won't go away, or why tickets for lunar holidays are so expensive. It's easy to get lost in short-term swings, but as Ray Kurzweil points out via his law of accelerating returns, world wars and depressions have approximately no impact on the price of compute.