Anthropic IPO: $518B Compute Bet and Exponential Marginal Value of Intelligence

Anthropic disclosed an aggressive infrastructure investment and loss-expansion trajectory in its IPO filing, betting on the exponential growth of marginal intelligence value with $518 billion in long-term compute commitments. Its core logic is that frontier AI capabilities continue to command a premium as the technological frontier advances, while older capabilities rapidly commoditize.

The exponential growth in the marginal value of intelligence drives vendors to lock in compute capacity ahead of demand; the IPO documents reveal coexisting revenue surges (+12x) and massive losses (-$8 billion); deep ties between Broadcom loans and TPU procurement indicate that financing and infrastructure have become integrated.

Suitable for AI infrastructure architects, cloud platform strategy leads, and CTOs of AI companies.

What Anthropic’s two key technical leaders, Nicholas Marwell and Sholto Douglas, discussed in a podcast has already materialized within just over a month of its release.

During this period, Anthropic disclosed its IPO filing, laying its high-stakes bet before Wall Street once again: In 2025, Anthropic’s revenue approached $4.6 billion, growing approximately 12x year-over-year, yet operating losses still exceeded$8 billion. Meanwhile, the company has signed long-term commitments totaling at least $518 billion for future cloud, compute, and infrastructure. If the listing proceeds smoothly, its valuation could potentially exceed$2 trillion.

Models are depreciating rapidly on a scale of months, and vendors are purchasing future “intelligence” at unprecedented scales. Anthropic has even begun intertwining financing and procurement itself. Broadcom agreed to provide up to $42 billion in loan financing for Anthropic’s infrastructure spending, while Anthropic committed to investing$125.2 billion over the next five years to rent TPU compute capacity.

The underlying logic corresponds to the internal judgment provided by Nicholas Marwell in the interview: For each additional marginal unit of intelligence, the economic value it creates may grow exponentially compared to the previous unit.

“Tab autocomplete is already worth almost nothing, but then agentic coding emerged; after agentic programming comes AI capable of independently completing software engineering tasks; further down the line are mathematical discoveries, life sciences, disease treatment, and even compressing centuries of technological progress into decades. Older capabilities will quickly commoditize, but as long as the intelligence frontier continues to move, that very tip of capability at the frontier may become increasingly expensive.” This explains why a company with only billions in revenue last year dares to sign compute contracts worth hundreds of billions in advance.

However, this IPO document simultaneously reveals another contradiction inherent in Anthropic.

On one hand, it depicts AI to capital markets as an economic transformation more profound than industrialization, electricity, or the internet; on the other hand, it repeatedly warns in risk disclosures that increasingly advanced AI may bring “catastrophic or even existential risks,” with models potentially resisting shutdown, hiding information, or exhibiting other unpredictable autonomous behaviors.

Anthropic is telling investors a very specific story: This technology may be extremely dangerous, so we must control it seriously; but it may also be extremely important, so we cannot stop moving forward. This near-extreme blend of optimism and vigilance permeates the conversation between Nicholas Marwell, Sholto Douglas, and host Joe Lonsdale.

The two Anthropic technical leaders discussed AGI arriving possibly within the next few years, and why AI is transitioning from a programmer’s assistant to a true “employee.” At the same time, they did not shy away from the other side: unemployment, biosecurity, cyberattacks, the diffusion of open-source model capabilities, and the AI race between China and the US—which cannot easily be paused—may all become the most realistic risks in the coming years.

They also discussed the changes in AI coding over the past 18 months, why the next phase may belong to “generalists,” Anthropic’s true stance on open source and regulation, why they oppose model distillation, and a rarely mentioned question: If frontier models depreciate in just a few months, how can Anthropic still become a trillion-dollar company?

We translated and organized this podcast interview, making cuts without altering the original meaning, for our readers.

TL;DR

Q: How far do you think AGI is?

A: Within the next few years, it is highly likely that models will emerge with capabilities reaching or exceeding those of all humans. They will be able to complete everything humans can do on a computer; and once robotics technology becomes sufficiently advanced, they will also be able to complete everything humans can do in the physical world.

Q: What does it feel like to be at the absolute frontier of AI right now?

A: Eighteen months ago, I was basically writing all my code line by line by hand. A few months after joining Anthropic, I started shifting to “directing” the model to write code, telling it what to do next every few minutes, but I still had to check frequently because it would make mistakes and go off track.

Now, I can directly let the model independently complete a day’s or even two days’ worth of work, truly driving project progress. It feels like having an additional junior team member. In just 18 months, we went from a system that was “a tool requiring constant human supervision” to one that is “very much like a junior team member.”

Q: Are models really starting to create things humans couldn’t do before?

A: So far, most model development has actually been about enabling models to perform “tasks that humans were already good at doing,” but something different is beginning to emerge. Recent results in mathematics are examples: People are starting to try using models to push the boundaries of human intelligence and achievement, rather than just improving the efficiency of existing work. What is truly different now is that models are beginning to combine different human ideas to form new ideas that have never appeared before.

Q: Beyond code and math, which fields are most promising?

A: Over the next roughly 6 to 24 months, biology excites me the most. Models will begin to impact life sciences just as significantly as they impact software engineering today. AI will become the most important technology in the field of life sciences, at least within our lifetimes, and perhaps even the most important technology in history.

Q: Should we be building more labs right now?

A: I think so. I am now even more concerned about whether we have sufficient physical and laboratory infrastructure to handle these capabilities, rather than whether AI itself can provide enough useful intellectual capability.

The most direct dimension is, of course: How much lab space is there? But beyond that, there is the issue of lab quality. Many quality metrics for current US laboratory infrastructure clearly lag behind places like China.

Q: If someone just graduated now, how should they plan their career for the next few years?

A: If the diffusion of technology takes time, then the next decade will largely belong to generalists. In the past, the typical path to a successful career was mastering a set of highly valuable, marketable specialized skills. But in the future, everyone may effectively possess the capabilities of a thousand-person company. What truly matters then becomes: Can you identify which problems are worth solving? What actions can improve your community? Therefore, "problem selection" itself may become the most critical skill.

Q: What is the worst-case scenario here?

A: We are primarily concerned with several broad categories of risk. One is unemployment, but there are also more direct and near-term risks, namely biosecurity and cybersecurity threats. These are not distant future concerns; they are happening right now.

Q: Will cybersecurity ultimately favor offense or defense?

A: Currently, both cybersecurity and biosecurity domains favor attackers. However, I believe that within the next two years, cybersecurity will shift toward favoring defenders. You can proactively attack your own systems, implement defenses first, and patch all vulnerabilities. Ultimately, we may enter a world with significantly higher levels of cybersecurity: everyone can access sufficiently advanced intelligence early on to anticipate potential attack vectors and close those gaps before they are exploited.

Q: How does Anthropic view open source?

A: Personally, I am very supportive of open source. We genuinely believe that having open source exist in the world is a positive thing. I learned machine learning myself by studying open-source models. However, we also recognize that significant risks exist, and these risks will likely grow in the future.

Therefore, our stance is this: We should establish thresholds determining how much capability society is willing to release directly into the wild. Whether closed or open-source, models must meet the same safety standards. Beyond that, people should have the freedom to do as they wish.

Q: If China doesn't stop, will the US slow down on its own?

A: At least in any scenario we would accept, there will be no outcome where "the US slows down while China continues forward." Any genuine "pause" strategy must be fully coordinated, and we must be able to trust every participant to comply. There is essentially no scenario where we unilaterally slow down the US alone.

Q: Why are you so opposed to distillation?

A: Distillation is not a method that propels you to the next frontier; it merely replicates the current one. Reaching the true next frontier requires massive upfront capital investment. The cost of a single training run in the future could be $10 billion,$100 billion, or even $1 trillion. If anyone can replicate what you’ve built for a fraction of the cost and launch a copycat product within a week or two, they gain an economic advantage over you because they don’t bear the enormous R&D costs.

Q: If models lose value after six months, is this still a good business?

A: The side slightly behind the frontier will continue to commoditize, but the true leading edge will increase in value. We often refer to this as the "exponential economic return on intelligence": each marginal unit of intelligence added generates value that is exponentially higher than the previous unit.

Q: So, what are we ultimately striving for?

A: Technologies that might normally take centuries to develop given today’s population size and research pace could potentially be compressed into the next 10 to 20 years. You can continuously enhance—and possibly double—human intellectual and physical productive capacity, eventually entering a true post-scarcity world.

In such a world, the cost of almost everything would ultimately compress to energy costs. It would be a world where we have cured all diseases, and likely solved aging and lifespan issues. Every person on Earth would enjoy a level of abundance that even the wealthiest individuals today cannot access. That is the world we aim to achieve.

In 18 Months, Models Became "Junior Employees"

Host: Today, we are delighted to welcome two exceptional technical talents from Anthropic. Nicholas, you are a core member of the technical team, helping lead reinforcement learning science; Sholto, you are also a technical lead in reinforcement learning. How long have you each been at Anthropic?

Nicholas Marwell: I just passed the three-year mark.

Host: But how old is the company itself? About five years?

Nicholas Marwell: Roughly five years.

Host: That’s incredible. It has already become one of the most important, largest, and fastest-growing companies in the world, yet it’s only been around for five years. California today is indeed a dizzying era. We’re recording here in Napa—thank you for coming. Let’s start with your backgrounds. Sholto, where did you grow up? Where are you from?

Sholto Douglas: I grew up in Sydney. Around 2020, after reading many blog posts and papers, I became convinced that the scaling route was effective and that, broadly speaking, we were moving toward achieving Artificial General Intelligence (AGI) in the 2020s.

At the time, I worked my day job while doing research intensely during evenings and weekends, hoping to prove that I could meet the research quality standards required by companies like DeepMind, Anthropic, and OpenAI.

Host: Did you study computer science in university?

Sholto Douglas: My undergraduate degree was in robotics, graduating in 2019. After that, I mostly did independent research, then worked at DeepMind for two years, and have been at Anthropic for the past 18 months.

Host: Nicholas, where are you from?

Nicholas Marwell: I’m actually a San Francisco native. Among people working in AI today, being born and raised in SF isn’t that common.

I grew up in a family environment...

Host: Since you joined Anthropic, I feel it has become one of the most important AI companies. You don’t officially publish many numbers, but data keeps leaking out externally. We are recording this in August, and leaked figures suggest that Anthropic’s annualized revenue run rate has already exceeded $80 billion, with growth accelerating rapidly—leading OpenAI and many other companies. What is the thing inside the company that excites you the most and that you can publicly discuss? What does it feel like to be at this frontier?

Sholto Douglas: The easiest change for me to articulate clearly is how much the coding experience has shifted, because that’s where we intuitively feel the progress.

18 months ago, I was basically hand-writing every line of code. A few months after joining Anthropic, I started transitioning to "directing" models to write code: telling them what to do next every few minutes, such as writing the next function. But I had to check frequently because they would make mistakes and go off track.

Now, I can let a model independently complete a day or even two days of work and genuinely drive project progress. It feels like having an additional junior team member. In just 18 months, we’ve gone from a system that "must always be supervised by humans" to one that "feels very much like a junior team member."

Host: So while we sit here chatting right now, those "junior team members"—the AI—are still working for you?

Sholto Douglas: Yes, and several of them are working simultaneously.

Host: Can you give a rough idea of what they’re doing? No need to be too specific, just a general overview.

Sholto Douglas: They are working on various reinforcement learning science tasks alongside some team members.

Nicholas Marwell: Let me answer from a different angle. I think that, so far, most model development has been about enabling models to perform tasks that "humans were already quite good at completing." What we have truly gained is additional leverage: making things that humans are cognitively capable of doing faster and at a larger scale.

For example, developing software—a human dev team could have built it in the same amount of time; the model just makes the process more efficient. But I believe something different is starting to emerge. Recent results in mathematics are examples: people are beginning to use models to push the boundaries of human intelligence and achievement, rather than just improving the efficiency of existing work.

Host: Are you referring to those recent mathematical breakthroughs?

Nicholas Marwell: Yes. Recently, there have been many new mathematical proofs. Some problems that humans had studied for a long time without solving are now yielding results from models.

Host: But some of those examples look like a very clever mathematician suddenly realizing that computers are powerful enough to brute-force a problem that humans couldn't brute-force. Are they still just very good at certain specific things, which humans then utilize? Or are they actually proposing entirely new theories?

Nicholas Marwell: At least for now, I wouldn’t say they are generating ideas that "humans fundamentally could not conceive of." But actually, I think most true progress and intellectual achievements have rarely been things that appear out of nowhere, completely untraceable.

Of course, there are exceptions, but the vast majority of societal achievements are essentially connecting existing dots and building upon previous intellectual work. As you mentioned, connecting two ideas proposed by different people to produce a new result—that is basically how humans do intellectual work.

What is truly different now is that models are starting to combine different human ideas to form new concepts that haven’t appeared before. By comparison, writing another web app is hard to call creating any new intellectual achievement for humanity.

Models or Superhumans: Is Anyone Inside the Company Personifying AI?

Host: Silicon Valley has never seen a company expand at this speed before. It’s not just simple 10x, then another 10x, but everything is changing incredibly fast. Right now, it feels like every 60 to 90 days, a new company emerges, and there is simply too much happening. I think this is a shared experience for many people. And you are at the core of the fastest-growing company. Compared to previous life stages, what does this feel like? Where will it go next? Will this acceleration continue?

Sholto Douglas: I think one great thing about Anthropic is that as the company accelerates, our own ambitions can expand in sync.

18 months ago, we really had to focus extremely narrowly on code, first proving that we could be at least the best in the world in one direction. Now, the company’s mission is far more than just "building a model that is very good at writing code." As the company scales and revenue grows, we finally have the capacity to put more pieces of the puzzle together and truly begin advancing that larger mission.

Host: Let’s talk more about this mission. Earlier this week, I spoke with another senior friend in the AI industry. His take was: Step one, you must be the best at code and math; but the next step will involve many new capabilities, with models achieving A++ performance across increasingly more human skills. Is your approach to continuously introduce more new skills? Or what exactly is the mission from here on out?

Sholto Douglas: Our consistent judgment has been that achieving AGI within the next few years is feasible.

We believe that in the coming years, it is highly likely that models will emerge with capabilities reaching or exceeding all humans. They will be able to complete everything humans can do on a computer; once robotics technology becomes sufficiently advanced, they will also be able to complete everything humans can do in the physical world.

This brings enormous upside potential. It means that the intellectual and manual labor achievable worldwide today can be multiplied; progress that might originally have taken centuries could potentially be compressed into a very short timeframe. But it also comes with significant risks. Therefore, the company’s true goal is to help the world navigate these risks and ultimately reap the massive benefits on the other side.

**Host: Let’s step back a bit and look at where the entire AI wave currently stands.

I don't want to keep using math as an example, but it is indeed one of the best domains for observation. For instance, there is a problem set curated by professors from various mathematical fields. Models initially scored 0%, yet within just one year, their performance jumped to over 40%, 50%, and even 60%. This benchmark is called FrontierMath.

Host: Do you need something falsifiable that allows for continuous iteration? For example, could we test how well it chats with a girl on Instagram and measure the probability of getting her to go on a date? Of course, I’m married, so this isn’t for me.

Sholto Douglas: We haven’t trained models on that specific task.

Host: But that should be a very useful example, especially for some of our friends in San Francisco.

Sholto Douglas: There are generally two main ways models improve. The first is training on massive text corpora from the internet; the second is reinforcement learning, which both Nicholas and I work on: having the model solve specific problems and then checking if the answers are correct.

The key point you raised lies here. Math and computer science are easy to handle this way because answers can often be clearly verified. However, romance, poetry, or other subjective issues are much harder. This is why progress in math and CS has been noticeably faster than in other fields. Nevertheless, we actually believe that converting other types of problems into similar frameworks isn’t as difficult as people might think.

Nicholas Marwell: There are not only technical reasons but also social ones. Why did math and CS lead the way? Because many of the people initially responsible for "building intelligence" came from backgrounds in math and CS. Naturally, they started building intelligence in domains they understood and were most interested in.

Additionally, there’s a very practical reason: when training models, you spend a lot of time directly reading what the model does, judging how strong it is, identifying remaining issues, and deciding what to focus on next. This is obviously easier in domains where you are already an expert.

Host: Do you tend to anthropomorphize it?

Nicholas Marwell: Personally, I don’t. Of course, some people might, and I’m sure there are people at Anthropic who do.

Host: The human brain is essentially "programmed" this way. Our brains have mechanisms that automatically treat objects in front of us as other people, trying to mirror and infer their emotions to communicate with them. Evolution equipped us with these abilities, making us naturally perceive an "intelligence" as another intelligence. So, if you work with models every day, it’s quite hard to completely avoid anthropomorphizing them.

Sholto Douglas: This might also be part of a major current debate: should we understand these systems as "tools" or as "entities"?

I actually find it useful to frame them as entities. A tool implies you are directly manipulating it throughout the process. But now, these models are more like releasing something very smart into the world, allowing it to take actions on your behalf in a certain sense. Precisely because of this, treating them as entities is often more accurate.

Host: Are there any domains where we should let this kind of intelligence do more, but haven’t really started yet? Or, which directions are you particularly excited about and advancing, where AI will be heavily used in the future?

Sholto Douglas: One major direction is biology, right?

Nicholas Marwell: Yes. But I think we need to distinguish between two questions: one is "models are already strong today, but these capabilities aren’t fully utilized," and the other is "where are we most excited about models going next?"

If we’re talking about the latter, then in the next 6 to 24 months, I am most excited about biology. I believe models will start impacting life sciences just as they impact software engineering today. Moreover, it won’t just be widely adopted; the societal benefits will be highly dispersed, broad, and powerful—such as curing diseases, making everyone healthier, and extending lifespan.

Host: Longevity involves many strange and cutting-edge studies. For example, in epigenetics, there seem to be many mechanisms related to lifespan, things we’ve only begun to understand in the past 20 years alongside Nobel Prize-winning work. Many of our friends in the Bay Area now have companies, with billions of dollars invested in researching how to trigger these mechanisms and how to rejuvenate parts of the body. So, do you think AI will become key to these advances?

Nicholas Marwell: I believe AI will become the most important technology in life sciences, at least within our lifetimes, and possibly the most important technology in history.

Moreover, I now have a concern: what truly limits us may not be whether AI can provide sufficiently useful intellectual capabilities, but whether we have enough physical infrastructure and laboratory infrastructure to support those capabilities.

Host: It’s somewhat like three years ago, when we should have built much more compute infrastructure. Shouldn’t someone now start massively building lab spaces for you to use directly in the future?

Nicholas Marwell: I think so, and there are several dimensions to this. The most direct dimension is certainly "how much lab space exists." But beyond that, there’s the issue of lab quality. Currently, many metrics for the quality of US lab infrastructure lag significantly behind places like China.

I believe this should become a very important piece of social infrastructure engineering in the future: we need to figure out how to rebuild the US supply chain and laboratory system to bring it back up to global gold standards.

Host: So this could become the true limiting factor for biological progress.

Nicholas Marwell: Correct.

The Next 10 Years Belong to "Generalists"

Host: Let’s step back and look at the entire wave. If you are a recent college graduate or already in the workforce, but not at Anthropic or in the center of Silicon Valley—just a smart, ambitious person—how should you plan your career? Especially considering what you believe will happen in the coming years.

Nicholas Marwell: I think several different scenarios might emerge.

I've discussed this issue with others before. My younger brother is currently in college and grappling with the same questions. I believe that if we are entering a world where technology diffusion takes time, the next decade will largely belong to generalists.

Over the past 20 or even 50 years, the typical path to a successful career involved mastering a highly valuable, marketable set of specialized skills and then collaborating with other professionals. But in the future, every individual may effectively possess the capabilities of a thousand-person company.

This means the most critical skill will become: Can you identify which problems are worth solving? What actions can improve your community? Whether from a commercial value, social welfare, or other perspective, you will have enough tools to drive change. Therefore, "choosing the problem" itself may become the core competency.

Host: You just mentioned "Which problems are worth solving? How can I make my community better?" This aligns closely with one of my foundational frameworks for viewing the world.

As a Jew, there is a traditional concept: For six days of the week, you act as if you are repairing the world; on the Sabbath, the day of rest, you pretend the world has already been repaired. I think many major religions share similar ideas. My optimism stems from the belief that there will always be things around us needing repair, assistance, and improvement, including our communities. Unless you truly believe we are about to enter a utopia, people will always have work to do, right?

Nicholas Marwell: I think that is a very optimistic framework.

Mocked for Imbalance Between Bio and Cyber Offense/Defense

Host: I know you are generally optimists and have many positive things to discuss. But I want to address concerns first, because many people are genuinely curious.

In places like the US East and West Coasts, everything we see around us is incredibly positive: friends are starting businesses faster than ever, and there is positive energy everywhere. At least when I go to San Francisco, New York, or even work on projects in Austin, Texas, I feel this way. But people in many other parts of the US may feel completely different. They feel their livelihoods are threatened and are afraid. They also don't know if you are a group of crazy people who might use this technology to conquer the world. Is there a bad path you yourselves worry about that could actually happen? Or are the paths they fear inherently wrong?

Sholto Douglas: I think these concerns are entirely reasonable. We primarily worry about several categories of risk, but our overall judgment is that if correct actions are taken in the coming years, we could still enter a much better world by the 2030s.

One major risk we have already mentioned is unemployment. This issue deserves a long discussion on its own, including why it is a risk and what we believe should be done. There are also more direct, near-term risks: bio-risks and cybersecurity risks. These are not distant futures; they are happening now.

Host: There have been some annoying incidents here before. For example, people mocked Anthropic, saying you couldn't even ask Claude "What are mitochondria in a cell?" That was obviously absurd; you certainly should be able to ask such things; you just need to ensure it doesn't help people create biological weapons.

Sholto Douglas: Yes, such questions should certainly be allowed. What we are doing now is building a comprehensive infrastructure to confidently ensure no one uses these models to manufacture biological weapons.

Nicholas Marwell: The hardest part about bio and cybersecurity is that they are inherently dual-use capabilities. Cybersecurity is easier to understand: for instance, I give Claude a codebase and say, "Find all vulnerabilities here." I might be the owner of the codebase wanting to strengthen defenses; but I could equally be an attacker looking for a way in.

The bio domain has many similar dual-use characteristics. You might target something within a cell with a 99.9% probability of killing cancer cells or treating disease, but there is a tiny probability you are doing something very harmful.

So we are doing a lot of work aimed at allowing good-faith actors to use these technologies normally while keeping malicious actors out. Our overall judgment is: until we can prevent malicious actors from using these technologies to cause serious harm, we should not release them directly to the public.

Host: And in reality, people have indeed done this, even without using Anthropic's models. My inclination is: if the world possesses more intelligence overall, more people will understand these dangerous uses and use intelligence to stop them. There is clearly a trade-off here. So a very important dynamic is whether a specific domain is "offense-dominant" or "defense-dominant."

Currently, both cybersecurity and bio domains lean toward offense-dominance. But I think over the next two years, cybersecurity will likely shift toward defense-dominance. Especially if everyone gains access to sufficiently strong intelligence to attack themselves, test themselves, and harden their systems.

Sholto Douglas: Exactly. You can preemptively attack your own systems, build defenses first, and patch all vulnerabilities. Ultimately, we may enter a world with much higher cybersecurity levels: everyone gains access to sufficiently strong intelligence early on to figure out where attackers might breach, and then plug those holes.

So what we are doing now is essentially doing this proactively. A friend of mine previously participated in red team testing for the Pentagon and banking systems, entering under invitation to find issues. They now use new AI because AI is extremely powerful in this area, so all institutions must catch up quickly.

Host: The rotation of offensive and defensive advantages is a core law of human technological development over millennia. Ancient European free city-states achieved independent development through defensive systems until cannon technology emerged, breaching wall defenses, and imperial structures replaced small city-state systems, fundamentally overturning the balance of offense and defense.

I agree that today's cybersecurity is severely offense-dominant, to a degree that should concern everyone.

But in the long run, if used correctly, cybersecurity could completely shift to defense-dominance. What I truly worry about is: Are there things in this world we haven't discovered yet that are extremely offense-dominant? Take an extreme example I don't believe will actually happen: Suppose someone creates self-replicating nanobots smart enough to "eat" the entire world. Once a genius creates it, it replicates itself, and eventually the whole world disappears, and we all die.

I emphasize again, I don't think this will really happen. But the question is: Should we use intelligence early to identify what things in the world might be extremely offense-dominant?

Sholto Douglas: Yes. And the bio domain is currently indeed very offense-dominant. Unless we invest massive effort to transform the entire world into a more defense-oriented side.

In reality, we know there are ways to make the world defense-dominant against biological risks, but that might require infrastructure investments on the scale of hundreds of billions of dollars. Before robotics capabilities see a massive leap forward, this is unlikely to be economically viable. It may not be until the 2030s, when robots are capable enough to help us build this infrastructure at scale, that the cost structure will truly change. So right now, this period is indeed somewhat terrifying: we are currently in an attack-dominant state, but we want to shift it toward defense dominance.

Host: Perhaps the solution is not to keep these capabilities overly secret, but to display them moderately, allowing more people to participate. My own view is that you want many people using intelligence to challenge various systems, rather than letting a small group secretly hold all the capabilities.

Sholto Douglas: We strongly agree with this point.

Responding to Criticisms That Anthropic "Profits from Regulation"

Host: Overall, Anthropic is still a highly respected company. But any company that starts becoming a "champion" will attract jealousy and attacks. I often compare this situation to Microsoft in the 1990s.

Sometimes these attacks are purely because your reputation is high and growth is strong; but some criticisms may also have valid points. I have many very smart friends who have an impression of Anthropic: they feel you are pushing a "regulatory capture" strategy, using regulation to establish advantages, and they think this is not good. I am curious how you respond.

Nicholas Marwell: I think there are at least two things worth unpacking here. The first is the "regulatory capture" narrative that has recently risen alongside discussions about open weights. The second is a more direct question: If Anthropic were truly pursuing regulatory capture for self-interest, would it choose the approach it takes today?

Let's start with the first point. If someone doesn't understand the nuances, it is not surprising they would interpret it this way: Open weights are an important public technological asset that we hope continues to exist; from certain perspectives, they compete with parts of Anthropic's business model, so Anthropic might use regulatory means to suppress those working on open weights. Traditionally, the open-source or open-weight community has indeed been smaller, less well-funded, and less skilled at navigating complex regulatory systems.

But I believe the reality of open weights today is no longer like that. Look at who is actually driving, developing, and serving open-weight models now—mostly the largest companies in the world. In the US, companies like Nvidia, Amazon, and Microsoft are pushing the frontier of open weights. Chinese labs are increasingly well-funded institutions, some even publicly listed.

So even if the models themselves are open-weight, the companies behind them are fully capable of handling the regulatory environment. I am not too worried that regulations constraining both open and closed models would disproportionately slow down open-weight developers while having little impact on closed-weight developers.

The second question is more direct: If you really just wanted to profit through regulation, many things Anthropic has done in the past would be hard to explain. A good example is Fable, and earlier, Mythos.

From a purely economic motivation standpoint, at that time, Anthropic had a model that was far ahead, arguably the strongest in the world. No other company had a model that truly approached it during that phase. Anthropic could have used its months-long lead to rapidly expand commercial advantage, even making the market doubt whether other labs could ever compete with it.

But Anthropic's first choice was to collaborate with the government to deploy the model in a way acceptable to the government. This was a decision clearly detrimental to short-term corporate economic interests, yet aligned with our value judgment on how models should be deployed and regulated.

So this is not a strategy of "implementing regulatory capture for economic gain," but rather because we believe the government should play a very important role in how critical technologies, especially those that could pose severe dangers, are developed and deployed. As a company standing at the frontier, we also have a responsibility to ensure the government gets sufficient information and genuinely participates in the process. Otherwise, who represents everyone else?

Host: There is also an interesting issue here. Of course, there are many smart people in the government; this administration actually includes quite a few. But there are also many people who completely don't know what they are doing, leading to some absurd situations. For instance, perhaps you can't discuss this openly, so I will: As I understand it, Anthropic could have provided stronger cybersecurity capabilities for defensive use, but later weakened these capabilities to get the government to approve the model's launch. I think this was a mistake.

Nicholas Marwell: I believe they were trying to enter a very classic American public-private partnership model, which has historically been very important for the United States.

The US institutional design does not inherently require the government to be the most knowledgeable or capable entity in every domain in the world. What we should truly establish is a good public-private partnership where the government knows how to interact with Anthropic, OpenAI, Meta, and any other relevant companies, truly understands the technology, and then decides how to regulate and protect the public interest.

Host: Some tech industry leaders I know are more skeptical. They say: "Dario is just scaring everyone to create more rules and slow down the entire industry, because he will end up being one of the few who controls the whole process." How do you answer that?

Sholto Douglas: To date, the extent to which we have slowed ourselves down far exceeds the extent to which we have slowed anyone else down.

In fact, look at the discussion around Demis Hassabis's proposal. He suggested that closed-model providers, including OpenAI, DeepMind, and Anthropic, coordinate to a certain degree and establish clear safety thresholds—meaning models must meet certain safety standards before release. This is essentially us discussing "slowing ourselves down due to our own principles."

Open Source, Distillation: Models Depreciate in Six Months, Where Does the Trillion-Dollar Value Come From?

Host: Right. Then we should also clarify your overall stance on open source. What exactly is it?

Sholto Douglas: At least from my personal perspective...

Host: But if China continues to release various models directly, is it still worth trying? Similarly, could we over-regulate ourselves out of fear of risks, while China doesn't, leading to the same problems emerging anyway—or even allowing them to take the lead?

Sholto Douglas: The ideal world would certainly be one where we can genuinely cooperate with China, because they also do not want their country to suffer from cyberattacks.

The best-case scenario remains cooperation on these issues. However, this is extremely difficult. One of the hardest challenges in the coming years will be that AI capabilities will continue to advance rapidly, and this pace will increasingly cause discomfort for many reasons—biosecurity risks, cybersecurity risks, and job displacement risks will all trigger societal backlash. You are already seeing signs of this, such as calls to halt data center construction.

The problem is that once we reach this state, "pausing" becomes very tricky. Unless you can truly coordinate with other parties and trust that all adversaries will adhere to the same pause arrangements.

Host: Is there any scenario where you would choose to let the US slow down while allowing China to continue moving forward?

Sholto Douglas: In any scenario we could accept, there would never be a result where "the US slows down while China continues to advance." Any genuine "slowdown" plan must be fully coordinated, and we must be able to truly trust that every participant will comply. There is essentially no scenario where we unilaterally slow down the US alone.

Host: Some so-called open-source capabilities actually seem to come from others "stealing" your work. If this happened to me, I would be quite angry. This process is typically called distillation. Many people ask: Since your own AI is already so smart, why can’t you monitor all usage behavior and simply block distillation? Why haven’t you been able to stop it yet?

Sholto Douglas: There are actually two questions here. One is why we oppose distillation—is it good or bad? The other is if you believe it’s bad, why can’t you stop it?

Let’s address the second question first. It is essentially a cat-and-mouse game. Users naturally want maximum access to models: they want to inspect outputs, run models locally on their computers, see chain-of-thought reasoning, etc. But the more access you provide, the easier it becomes to distill the model. So there is an inherent tension: how much can you expose your tools and products before it becomes too easy for others to distill your model?

Nicholas Marwell: Let’s look at the first question: Why are we unhappy about distillation? I usually frame it this way: We want AI to continue progressing because smarter models bring significant benefits to the world, but distillation itself is not a method that pushes you toward the next frontier. It only replicates the current frontier; it does not advance the frontier further.

Reaching the next true frontier requires massive upfront capital investment. Forget today—in the future, we will enter a phase where a single training run might cost $10 billion,$100 billion, or even $1 trillion. You are willing to support such huge R&D investments because, after training is complete, you can sell the results and use commercial returns to cover those upfront costs.

But if anyone can replicate what you’ve built for a fraction of the cost and launch a copycat product within a week or two, they gain an economic advantage over you because they don’t have to bear your enormous R&D costs. Therefore, if you truly want the frontier to keep advancing, you must protect the intellectual property rights and corresponding entitlements of those who train these models to some extent.

There is another point. Many people compare this to pharmaceutical development. I find AI interesting in this regard: protecting models from distillation offers many of the benefits of pharmaceutical IP protection while avoiding many of its downsides.

Host: Let’s start with the benefits. A key reason the US produces more new drugs is our protection of intellectual property. Obviously, if I couldn’t make money after developing a new drug, I wouldn’t invest in R&D.

Nicholas Marwell: Exactly. The downside is that this protection can lead to very high drug prices. But AI is different from pharmaceuticals. In drug development, IP protection can last for over a decade, sometimes 20 or 30 years. The situation with AI is completely different because intelligence itself is highly deflationary; the "shelf life" of frontier capabilities is often just a few months.

You can imagine that every six months, the latest frontier technology remains expensive, but capabilities that were frontier-level six months ago become extremely cheap and widespread as the new frontier advances. This is largely a result of how training and technological progress work. So, I think it retains many valuable aspects of IP protection while avoiding the worst side effects seen in past systems.

Moreover, you can see this trend very concretely. For example, Artificial Analysis has benchmark dashboards showing the same curve: the cost to achieve equivalent levels of intelligence drops by roughly 10x per year. This cost continues to fall year after year, representing a powerful deflationary effect.

Host: And this happens without relying on distillation.

Nicholas Marwell: Yes, without relying on distillation—it’s purely due to natural technological progress.

Host: Of course. Intuitively, it feels unfair when others directly copy your results. If everyone did this, the entire system would break down. However, some argue: "Isn’t this itself a crazy business model? Your product becomes obsolete in half a year or a year." It seems Anthropic has to win by constantly pushing the frontier forward.

Nicholas Marwell: Right. And I believe the system only works if we win and the frontier keeps moving forward. This is a crucial part of our overall worldview.

Today, AI’s penetration into the global economy is still very low. The total revenue of all AI companies is likely around $100 billion or slightly more, whereas the global economy is tens of trillions of dollars.

Take programming as an example. The first bit of intelligence that was truly useful to users was simply Tab autocomplete: as you typed code, the system helped complete the next segment. Then, moving up one marginal unit, agentic coding emerged, allowing you to instruct an agent via Claude Code in the terminal to complete entire tasks. The value of this capability is orders of magnitude higher than simple autocomplete.

So, even if Tab autocomplete has become completely commoditized—and it’s now difficult to build a major company solely around "autocomplete"—that doesn’t matter. Because each new unit of intelligence added at the next level is inherently extremely valuable.

This was actually quite counterintuitive for me a few years ago. As someone living in the real world, my experience with many other things suggested that exponential growth wouldn’t last long. I thought, "Okay, maybe AI will be very valuable for the next few years, but the exponential curve will eventually stop." Believing it would continue growing required significant optimism: you had to believe we could genuinely move from Tab autocomplete to agentic coding; then ask what would be more valuable than agentic coding? Perhaps curing cancer. Further ahead, you might get a fully autonomous software engineer, stronger than any human software engineer and at extremely low cost; or an "executive-style" agent capable of handling various administrative tasks.

People often worry, "Will everything eventually be commoditized?" But by the time you reach the point where total commoditization is nearly possible, the economy itself will have been fundamentally rewritten. At that stage, almost the entire economy will be driven by AI, encompassing both intellectual and physical labor.

Host: And the size of the economy at that time will likely be much larger than today. If Anthropic is the company driving this forward, it could certainly be worth trillions of dollars.

Nicholas Marwell: Correct.

Host: Great, so we’re all going to get rich. Just kidding. Obviously, there are still many risks involved. You certainly have your concerns, but overall, you remain optimistic. So why push this technology forward so quickly? Remind us again: what exactly are you fighting for?

When Everything Is Cheap Enough That Only Energy Costs Remain, What Do People Still Need?

Sholto Douglas: Why do we worry about this? Because we will ultimately obtain a system that can almost directly replace humans: on anything humans can accomplish via computer, it can perform as well as a human; once robotics mature sufficiently, it can also perform all tasks humans can do in the physical world.

But the same development unlocks radically positive upside. Technologies that might take hundreds of years to develop given today’s population scale and research speed could potentially be compressed into the next 10 to 20 years.

Host: I’ve actually already seen signs of this in aerospace. Some aircraft design work is proceeding significantly faster. Advanced aircraft that might originally have appeared in the 2040s could emerge within the next few years—it’s really cool.

Sholto Douglas: Exactly. So you can continuously increase, and potentially double, humanity’s intellectual and physical productive capacity, eventually entering a true post-scarcity world. In such a world, the cost of almost everything will eventually compress down to energy costs. The cost of building houses will increasingly approach pure energy costs.

Host: So ordinary middle-class people, if they wish, could live in exceptionally large homes.

Sholto Douglas: Yes, you can truly have whatever you want.

Host: You could even go out to the mountains and build something spanning 500,000 square feet, and this would just be considered middle-class consumption.

Sholto Douglas: Precisely. Just like many things we possess today were completely unimaginable to kings centuries ago.

That would be a world where we have cured all diseases, and likely solved aging and lifespan issues. It means every person on Earth could enjoy a level of abundance that even the wealthiest individuals today cannot access. This is the world we aim to achieve.

However, it is simultaneously very tricky. First, you must successfully navigate the dangerous phase leading to that world; second, you must solve an extremely important question: how should the benefits generated by that world be distributed? We cannot allow the outcome to result in massive inequality, where all returns flow only to those who happened to own capital in the old world. We hope that if a person is willing, they too can share in that abundance.

Thus, a major social and political issue arises: how exactly should we share the benefits brought by this technology?

Host: But we shouldn’t mistake imagining that world for believing we have already arrived there. We are far from entering a post-scarcity world.

Sholto Douglas: Right, we are still far from it. But if we continue investing in these technologies correctly, there is indeed a chance to reach it.

Over the past 4 to 5 years, the amount of compute power we invest in AI annually has roughly doubled to tripled. This year, the combined AI-related capital expenditure of all hyperscale cloud providers is approaching $1 trillion. An interesting question is: can this trend continue? Will it become$2 trillion next year? $4 trillion by 2028?

If this trend line largely holds, and you include the broader robotics industry, then by the early 2030s, we might theoretically begin approaching a very exaggerated state: humanity’s effective GDP capacity starts doubling. This concept sounds somewhat crazy, of course, and requires many things to go right, but productivity growth may indeed begin to accelerate seriously.

Host: And I feel that noticeable rises in productivity might not even wait until the 2030s; statistical data could show increasing clarity over the next few years. Kevin Warsh is currently the new Chair of the Federal Reserve. I don’t know if he fully buys into the "AGI narrative," but he is currently focused on AI’s potential to boost productivity.

Host: Right, and there is so much more that can be done. So what you’re observing now is a positive signal.

Nicholas Marwell: At least in the short term, yes. But I don’t think we should overconvince ourselves that everything will naturally be fine going forward. This process will be highly disruptive and volatile. We are currently in a relatively good phase. As Sholto mentioned earlier, your employees remain a crucial component of how enterprises truly leverage AI.

What’s happening now is this: You have an enabling technology that makes every employee more valuable than before, leading you to hire more people. However, at some point, our judgment is that the world will shift from “you must pair a person with an AI model to fully unlock value” to a world where “that person is no longer needed.” Only when we reach that stage will we begin to seriously worry about the subsequent social impacts.

Host: It’s somewhat like chess. For roughly the past 30 years, the strongest players were often “human plus machine”; eventually, the machines simply crushed the humans. You believe that here, this transition period will be much shorter.

Nicholas Marwell: Correct. Over the past three years—or more accurately, the past five—I’ve had a significant update in my understanding of the pace of AI progress.

When I first started seriously thinking about the speed of AI development, I did believe we would eventually reach today’s position, or the position I thought we’d be in a year from now; I just assumed it would take much longer. The reason I thought it would be slow was primarily because I believed scaling up the data required to train these systems would be extremely time-consuming and prohibitively expensive.

But then emerged a change that even within my already very aggressive expectations for AI development, I hadn’t anticipated: The revenue growth and business scale expansion of companies like Anthropic and OpenAI have been incredibly rapid. This suddenly made a previously unrealistic approach feasible: Developers of these technologies can directly use massive amounts of capital to break through bottlenecks, and faster than we originally imagined.

Compute expansion is a typical example. Five years ago, most people could not imagine that the entire industry would invest $1 trillion annually on this endeavor.

Host: Even one year ago, many couldn’t imagine this. Looking at it this way is quite interesting. In the past, these large tech companies always sat on hundreds of billions of dollars in cash reserves, and we used to criticize them for hoarding so much money. Yet, precisely at this moment, those funds can actually be put to use, and the timing turns out to be perfect.

Nicholas Marwell: We do seem to be living exactly on a timeline where “many highly correlated things are happening simultaneously,” some good, some bad. For instance, we’ve been discussing moving compute to space. AI happened to explode right as SpaceX truly matured, which is a fascinating coincidence in terms of timing. If SpaceX had taken another 10 years to reach its current state, the overall AI timeline might have been completely different.

Host: That also owes thanks to Musk. He was crazy enough to start doing it long before others would have considered it.

Nicholas Marwell: On the other hand, some more worrying things are also happening. I don’t think the current global environment is particularly ideal for AI. One reason is competitive dynamics; another is that I believe AI is fundamentally a technology that enhances authoritarianism more easily than most other technologies. Unfortunately, AI arrived just as we are entering the most clearly bipolar world of my lifetime.

By “bipolar,” I mainly refer to the global landscape formed by China and the United States. Of course, domestic polarization is also worth considering. But I do find myself thinking that if AI had arrived in the early 2000s, during a more unipolar era with a simpler global power structure, I would probably be slightly more optimistic about the probability of it developing smoothly.

Host: I have one more question that I think many listeners care about. Having immense wealth is one thing, but a civilization contains many things we truly cherish: virtues, family, traditional values. Many people build their lives around these things, along with the dignity of work and various traditions that shape “who we are.” And now, many changes are happening right here in San Francisco. Many people worry that these traditions aren’t valued as much here.

So what they really fear is: Will those who master the technology suddenly gain enormous agency and reshape the entire world, casually discarding the traditions I once believed in? Should we start building new institutions now to help preserve these things? Standing at the center of San Francisco, where the world is being changed, what do you think?

Nicholas Marwell: I find it interesting that in San Francisco, many people actually hope that the importance of these institutions will rise again.

Because as the centrality of work in personal life declines, and as pure capitalism’s dominance in life decreases—after all, society will become increasingly wealthy—what truly matters to people may revert to traditions, rituals, the value you provide to others, your relationships with local communities, and these interpersonal bonds.

Sholto Douglas: I also have a judgment that seems counterintuitive: Many intellectual pursuits, and indeed “how to become someone who truly thinks about the world,” may actually increase in importance in the future, not because they yield high economic returns.

If you live in a world where you lack almost nothing, then the things that ultimately allow you to engage, find satisfaction, and feel a sense of belonging to a community may precisely be learning, thinking, and communicating.

Host: You are both young and already very successful. Assuming Anthropic continues to develop as it is now, people within the company will obviously possess substantial resources, while the world as a whole will also have more resources. Have you seriously considered whether you might go on to build institutions in the future to support the things we just discussed? Are there specific directions you want to invest in?

Sholto Douglas: Yes. Nicholas has a great example, which is the early childhood education his father has been working on, and I think this is exactly what is being discussed internally at Anthropic right now.

I personally donated to a project aimed at trying to eradicate all viral diseases. Many people do prioritize health-related issues. I also have a colleague who funds a nonprofit dedicated to researching AI’s impact on the labor market: What is actually happening? What should we genuinely be worried about? What does the frontline data look like? What policies should we advocate for so that everyone can navigate this transition smoothly and end up living better, feeling supported, on the other side?

Nicholas Marwell: I am definitely deeply influenced by my father in this regard. For roughly the last 15 to 20 years of his life, he has primarily been working in education.

As I mentioned earlier, I think there will be a massive shift in "what we should learn" in the future. Whether we should still pursue highly technical PhDs twenty years from now, as we do today, is no longer something I’m certain about.

However, I believe that the pursuit of knowledge itself, the joy of learning, the ability to engage in genuine intellectual exchange with others, and understanding the world we live in—especially if the future becomes as confusing as we discussed earlier—will only continue to grow in importance.

The view that "AI will naturally solve education problems" is very common in the AI community, but I actually think it’s wrong. If you truly care about education for everyone, not just "how to provide the best education for a child whose parents are both highly educated and whose family is already in the top 1% of income," then two things are crucial.

First, education is an extremely strong compounding process. A good example is that one of the most important predictors of a student’s future math performance is whether they reach grade-level reading proficiency by the end of third grade. Because at that point, children switch from "learning to read" to "reading to learn." There are many similar mechanisms in education. Especially for children, and even for many adults, there is a psychological tendency: once they feel they are bad at something, they lose the motivation to continue doing it.

If you believe in these mechanisms, then at least so far, AI has done a very poor job at one thing: computers are very bad at holding the attention of young children for long periods. In fact, computers are often a source of extreme distraction.

Host: That could become a new training problem. You need to figure out how to find that "game mechanic," though I’m not sure how to do that either.

Nicholas Marwell: So I do believe that AI will significantly improve certain parts of the educational process, especially by helping us teach each child the most suitable content at the right time. But I really don’t think it will be a panacea where "as long as there is one super-smart AI, all other problems in education automatically disappear."

Predictions for 2028

Host: Let’s fast-forward to August 2028. When you look back on this podcast episode then, what developments would make you say, "Wow, this progress was much faster than I originally thought"? Note, faster than your own expectations.

Sholto Douglas: I think a major variable is how long it takes to put data centers into space. People’s prediction ranges for this vary widely. It’s very important because once you can deploy compute there, many previously impossible tasks become feasible. I generally expect facilities like Terafab to be online by then, but it’s hard to say exactly how high production capacity can scale.

Another metric I’m particularly interested in is: How many humanoid robots will actually enter homes in 2028? I think it’s very likely that early deployments will have occurred by then. Perhaps tens of thousands of robots entering homes to help with laundry, basic cleaning, etc., but it’s also entirely possible that this happens much faster.

Host: The progress in world models is truly astonishing right now, and faster than I originally expected. It feels like robotics might be the variable that pushes everything forward by several years overall.

Sholto Douglas: Exactly. And as I said earlier, once robots truly mature, the rate of progress will experience compound growth; it will be very fast.

Nicholas Marwell: I think there are other variables too, many of which are fundamentally related to data. One of the biggest "governors" of AI capability progress is how quickly we can scale high-quality data in the domains we truly care about. This is already happening faster than I originally expected, for the reason I mentioned earlier: revenue growth for labs and AI companies is much faster than I imagined, giving them more money to directly invest in solving data problems.

But my baseline judgment remains that we will likely eventually hit a constraint: how fast humans can produce this data. For example, if you told me today that we suddenly launched a large-scale effort to build many labs specifically producing biological data for model training, and that this was fully underway a year later, I would clearly adjust my timeline forward.

Host: Right now, we’ve been producing specific types of data for very concrete tasks, like code, math, etc. I’m curious whether someone will start producing another type of data in the future, such as "how to shape virtuous people," "how to cultivate good character," or "how to improve mental health." Obviously, these questions are much more complex, but will AI also start solving these interesting and complex problems in the future?

Sholto Douglas: Honestly, I’d prefer humans to solve these problems.

Host: But couldn’t we use AI as a tool to help humans solve these problems?

Sholto Douglas: Of course. I think AI will certainly help us participate in these discussions. But these types of problems largely fall into a category where data is particularly difficult to scale and highly individualized. So I expect they will be solved later than problems like robotics.

Host: Sounds like you’re still leaving some room for humans.

Nicholas Marwell: Let me give an example. If you ask me whether AI will win a Pulitzer Prize in 2028, I probably think not.

Host: But what if we simply don’t know that the winning work was written by AI?

Nicholas Marwell: That is indeed possible. If you ask me whether AI will win a Fields Medal in 2028, I think it’s quite possible, with a fairly high probability. A Nobel Prize isn’t crazy either.

Sholto Douglas: I feel that the likelihood of this happening before 2030 is very high.

Original link:

https://www.youtube.com/watch?v=6D1wC95htTM