Anthropic早期往事：22家VC中21家拒绝它｜AI的四大瓶颈｜Anj Midha

AI alignment, don't get me wrong, is hard, but not the hardest problem. Human alignment is really the problem right now.

If we don't secure frontier-model inference, or what I would call state-of-the-art inference, behind a coordinated Iron Dome, I don't think we have a sustainable shot at staying at the frontier over the next decade.

There's no saturation in superconductor discovery at all.

Anj, I am so looking forward to this. Dude, I have stalked the [__] out of you for the last 3 or 4 days. I spoke to Bing Gordon. I had a catch-up with Bing before this. It was very nice to speak to him. Thank you so much for joining me today, dude.

Thanks for having me. It's been too long. It only took us what—8 years, 9 years? I forget when it was.

I was 12 when we last did it.

12 in startup land is 25, right?

1. Are Scaling Laws Dead?

You know, I'm confused. Help me out. I had Demis on the show the other day from DeepMind. He was like, “No, I'm not sure if we're seeing scaling laws, but we are definitely seeing slightly diminishing returns in performance as we scale.” So, potentially, are we getting to a stage where increased compute is no longer leading to increased performance?

Oh, no, absolutely not. No, that's not true at all. In certain domains that are well explored, like coding, for example, yes, there's an increasing amount of compute required to get an incremental gain in some evaluation that's super-saturated. But if you said, “Anj, what about materials science?” I'm sitting here at Periodic Labs' office. This is my incubator—my latest incubation is called Periodic Labs. I spend 3 days a week here in Menlo Park.

We have a 30,000-square-foot facility where we have LLMs that predict new materials and new superconductors. We then have robots synthesize those new materials, and then we have physical machines, like X-ray diffraction machines, that validate whether those materials have the properties that were predicted by the LLMs. Then we pipe that verification data back into our training run, however many times we need.

2. The Four Bottlenecks Holding AI Back

I can tell you, throwing more compute at the problem is probably producing super-exponential gains right now per iteration. So, it depends on which domain you're talking about and which modality. There's no saturation in superconductor discovery, for example, at all. The bitter lesson is holding and is alive and well.

I totally get that. Can I ask you: When we look at the bottlenecks around performance and progression today, what are the bottlenecks that persist most significantly to you? Is it algorithms? Is it data, or is it compute? Can you help me understand which is lagging most?

There are 4 or 5. It's context feedback, which I'm happy to talk about. It's compute. There's capital, which you need to continuously deploy into the compute and context-feedback loops. And then there's culture. I think that culture might actually be the most important bottleneck of all time, but those are the 4 I would say.

Algorithmic innovation, I think, is basically a function of culture. If you have the right culture, you attract the best researchers, and the best research talent then wants to work on pushing the frontier. Algorithmic innovation just falls out of having a really good team that's very flexible about what kind of architecture they want to use.

If you have the right culture, the algorithmic innovation bottleneck solves itself. The researchers aren't focused or tied to one architecture versus another. They're not going, “I'm all in on LLMs or transformers versus diffusion models.” The best scientists and researchers just want to solve the problem, the mission.

If you have a very mission-driven culture where they're saying, “We want to move the frontier of coding or the frontier of materials science,” the algorithmic stuff takes care of itself. So, that's actually not the bottleneck anymore, in my view.

2–3 years ago, that was a huge bottleneck, when we were trying to figure out which algorithms scale and whether there are some limits to the transformer architecture versus diffusion models. What I've come to realize is, if you solve the culture problem, you can solve the research and algorithmic problem.

Then the bottleneck of context feedback—what is the data you need to keep doing frontier research over and over again?—is step number 1, because I think that is also where you have the most business and commercial advantage. I think there's lots of alpha and value to be gained in pretraining, mid-training, and so on.

But that last mile, where you deploy a model or an agent in some new domain, collect feedback on how it's performing in real time, and then feed that back—like I was saying, here we do physical verification of materials science at Periodic Labs—that's where you probably have the biggest bottlenecks on capabilities.

What you should be doing, if you're trying to advance the frontiers, is going, “Okay, these models suck,” for example. About a year ago, as an example, I realized there was a lot of talk about models being good at physics and chemistry—AI for science.

I was a scientist in the Applied Physics Department at Stanford, and we started benchmarking these models—Claude, Gemini, and so on. And, surprise, they sucked. They were so bad. I was like, “There's this disconnect between the marketing hype of AI for science and the reality.” At the time, at least, they were starting to get good at code, but they were terrible at scientific analysis.

The conclusion was pretty simple: They were just missing a lot of the physics and chemistry data you need to reason about the physical world. But to do that, we don't have enough of that data on the internet, because the internet is mostly pretraining data about things like blogs and coding.

If you need physics and science, that's a real bottleneck, because that data is locked up in national labs and academic labs. It's locked up in physical semiconductor manufacturing plants. How do you get that data in? The bottleneck I realized was really the critical part of getting these models to reason about the physics and science frontier, which is something I care about deeply.

The way we solved that at Periodic was to set up a physical lab with robots doing all that. You could apply that same recipe to whatever domain in which you want to see more and more progress. Then you ask, “Okay, how much compute infrastructure do you need to keep that RL loop, or the physical verification loop, scaling at bigger and bigger scale?”

Then you need the capital to fund all this. You need equity, debt, and a whole bunch of different structured finance vehicles to get land, power, and shell. So, that's the compute bottleneck.

3. Why AI for Science Sucked

Lastly, culture, because if you have all of those 3 things but you don't have the right team and the right mission-driven culture, the whole thing falls apart. And so, those in my mind are the 4 bottlenecks I wake up every day trying to figure out how we unblock for the best teams.

If we just go through them, when we look at that context feedback on the data side, will we then see a generation of vertically integrated foundation-model companies like Periodic for a ton of different things?

Yeah. When I went to grad school for machine learning, I went to Stanford for bioinformatics, which was machine learning applied to health care. The space was not as good at marketing as it is today. Superintelligence, love it. At the end of the day, what are we talking about? We're talking about very powerful models within some domain, and we're seeing, within distribution, very powerful capabilities that you can definitely call superhuman.

There's no way, for example, I as an individual scientist could analyze the reams and reams of data coming out of the lab here without AI models. There's just no chance. The fact that you can take all the data from training from a physical lab and just throw it at a bunch of AI models and ask them to analyze things is a superhuman capability. We didn't have that before.

Okay, fine. Let's call that superintelligence. Within coding, within materials science, within each of these domain distributions, we are seeing capabilities that are superhuman. We didn't have them before.

In fact, I would say we're even starting to see automation of those tasks, especially where there's coding involved, starting to be somewhat recursive, right? If you have a good coding model, then you can say, “Okay, let me automate data analysis. Let me automate data cleaning,” and so on. Some people would call that recursive self-improvement. Totally happening.

4. Sovereign Data & the Cloud Act

But it's not like I can just say to a coding model, “Please bootstrap a physical R&D lab for me in Menlo Park. Get all the permitting. Go raise—go find Andreessen to raise money from. Go set up the physical infrastructure and just bootstrap all this data.” That's just an entirely different kind of frontier, execution, and problem.

My question to you, then, is: how do I determine what is not going to get cloudified in that vertical model company build-out? Because you could look at Cursor and say, “Well, they’ve built their own vertical model end to end, and it’s been cloudified, if we’re being blunt.” Periodic Labs won’t be, because of the physical data that’s being produced in the labs. How do I know what will be cloudified versus what won’t in that model there?

Yeah, this is a good question. If we want to unlock frontier progress generally across a bunch of domains, then where are the bottlenecks, and where will the value accrue? Context is not necessarily the moat. I would not say that yet. I think venture capitalists are very quick to analyze moats, but I would say context feedback loops, where you have unique and differentiated access, are where progress will be most legible to you. If there are other teams who don’t have access to that context, it will also be where you have a superior business model.

Here’s an example I give in the class, right? Sovereign data. Are you familiar with the CLOUD Act?

Yeah.

Yeah, okay. So, the U.S. CLOUD Act says that, hey, if there are any cloud workloads running on infrastructure that is managed by an American company, then the U.S. government has to be able to access that data.

Now, if you happen to be running military defense mission-critical workloads in Europe on AI infrastructure that is managed by an American company, well, that context, which is supercritical, can’t be sent across the border. That’s an example of a unique and sensitive context that needs to be run locally.

And so, if you’re ASML or CMA CGM, which is doing logistics at scale, and some of this logistics is with mission-critical supplies, you can’t have your supply-chain data being processed by an AI bot that’s running on servers that are subject to the CLOUD Act.

So, what do you do? You look for local infrastructure partners. You start going, “Hey, who are the AI infrastructure providers in Europe that we trust?” Well, it turns out there aren’t that many who can actually handle mission-critical infrastructure at scale for AI.

So, you call up someone called Arthur Mensch, who is a French scientist from DeepMind turned entrepreneur and started a lab called Mistral, which is running massive workloads, and you say, “Arthur, would you actually build infrastructure that can be secured locally?”

And that’s why, suddenly, in July 2025, at VivaTech in Paris, you have President Macron and Jensen Huang standing onstage next to Arthur, a 33-year-old scientist, unveiling a gigawatt AI infrastructure facility in Paris. Why? Because the context, the mission-critical context of those workloads, is so important to be run locally that you can’t run them on Amazon Web Services, GCP, or Azure. And it’s the first time in 15 years that hyperscaler dominance is up for grabs for startups.

With the greatest of respect, is that the core investment thesis of Mistral for you?

For me, yeah. Independence at scale of every part of the AI infrastructure stack: land, power, shell in Europe—that’s sovereign, it’s local; compute infrastructure that’s local; and models that are trained locally. By the way, they’re fully open, so they can be deployed and customized globally wherever needed, but certainly in Europe. The full independent stack is the bet. Yeah.

5. The Investment Thesis Behind Mistral

Do Anthropic and OpenAI just accept that and roll over? I don’t understand, because government is a mega-portion of their efforts and workload today. Both of them, when I speak to them, are like, “Oh, we’re absolutely coming for Europe.” So how do they get around that?

Well, I can’t speak for OpenAI too much because I’m not involved there directly, but Anthropic, I will say, the mission and vision have always been very American-aligned, right? They’ve always said, “Hey, America is the crown jewel of the world in terms of innovation. This is where we’re located.” Again, Anthropic is located in Silicon Valley, and I think the company really, really wants to do what’s best for the American government and the American way of life, which is democracy and freedom.

It turns out the world’s largest enterprise customers are governments and Fortune 500 companies, and many of those that are overseas need these workloads to be running locally.

6. The Brutal Early Days of Anthropic

You said you’ve obviously been involved with Anthropic since the earliest days. I’m just fascinated. I think people kind of forget about their early days almost. People talk about, “Oh, SBF investing early and what a visionary he was.”

Right.

What was Anthropic and Dario like in the early days?

Well, I’ve known Dario forever. Dario was one of the lead authors on GPT-3. We’ve been friends for many years. Tom gave me a call and said, “Anj, for various reasons, we want to leave and start this new lab called Anthropic. We’re going to need a lot of capital. We’re going to need compute.”

I had already sold Ubiquity6 at that point, so I’d kind of gone through the founder journey. Dario, Tom, and I started doing these weekly sessions in early 2021 to try to figure out how to turn what was really a research hypothesis, right—which is to scale the scaling recipe—into a business hypothesis. I would say it took 12 to 24 months.

They did a lot of the hard work figuring out how to really operationalize the idea of this AI pair programmer, right? Where you take the context feedback loop of the local repository, the files, the directories of programming, and, in a very methodical way, make predictable progress on the capabilities of software engineering.

If anything, my biggest flaw as an investor and as a founder is being too early to things. That was my lesson with Ubiquity6. I was early to computer vision, which is now obviously blowing up the whole multimodal generative-modeling space. Since then, I think I have updated my strategy on how to get timing right.

But at the time, our recipe was pretty simple: raise some money, buy some compute, get a little bit of context data on programming, put out a basic version of the model, deploy it with teams that we trust who are doing coding, and then pipe that feedback loop back into the training run over and over again.

When you do that with inference, it gives you two things, right? It gives you revenue to buy more compute, and it gives you the context feedback to keep improving the capabilities curve. I was like, “Great, this makes total sense, guys. Let’s go raise money.”

I invested a bunch of my money—that was just life savings, which was not much, given I was a poor founder at the time, and most of my net worth was tied up in Discord stock. It pains me sometimes to look back at the emails of friends.

I introduced them to 22 friends up and down Sand Hill Road, and so there were some investors there. We got 21 nos. I was like, “What are you guys thinking?” And they said, “Well, this recipe sounds good in theory, but where’s the proof?”

I said, “Proof? These are the guys who invented GPT-3. How much more proof do you want?” And they said, “What’s GPT-3?” I said, “Oh my God, how do you go about educating somebody who doesn’t even understand the technology and the breakthroughs that are happening in the machine-learning community?”

I was lucky because I had that training from grad school. I started a computer vision company, so something that was super legible to me was just a completely different world for those investors.

For those investors, remember, we originally tried to go out and raise $500 million and then had to re-anchor to only raising a $100 million seed round, which at the time felt like a lot, but of course was tiny compared to how much OpenAI had raised, because by then I think OpenAI had already raised $1 billion.

The whole idea of compute multipliers—where we could, for every dollar of venture capital raised, produce a unit of intelligence for 6 times less—was not something the VCs understood. That’s why, over the next 24 months, the people who got it were people like some of the folks in the ML community who also had an overlap with the effective-altruist community, like SBF, but also Amazon.

This was very legible to Amazon because they were watching what was happening with Azure and OpenAI. They were like, “Well, this is super aligned. If you guys can actually create a bunch of state-of-the-art models that are hosted on Amazon, that’s super accretive to the AWS business.”

That’s why it resulted in deep compute and capital for an equity partnership with Amazon that was originally $4 billion. A lot of this is public now, but at the time it was a really tough journey.

I would give Dario, Tom, the other co-founders—Daniela, Jack, Sam McCandlish, and Jared Kaplan—credit. It was such a brutal time getting this company going. People don’t understand.

Is there anything you would have advised them differently, knowing all that you know now?

I’m not sure I would, because the world is a very different place today. At the time, it really did feel like there was no one they could trust.

Is it not impossible not to be hauled up in front of Congress if you reach a certain scale?

Oh.

Whether you’re Google, whether you’re Facebook, or whether you’re Anthropic fighting against the Pentagon.

And you get to a scale where it is impossible not to have that conflict?

Oh, absolutely. No, what are you talking about? Look, I started Amp as a public benefit corporation because I think it’s actually a very aligned model. Have you heard of REI? REI is a public benefit corporation. They make billions of dollars in revenue and profit. Have they ever been hauled up in front of Congress? No.

Like Ben & Jerry’s, a public benefit corporation. Have they been hauled up in front of Congress? No, because they self-moderated. At the time, they said, “Here’s our mission, but we have to build a business.” As long as you hold those 2 things in balance, those things are not in conflict long-term.

7. Public Benefit Corporations: Mission vs Profit in the Age of AI

If your goal in life is to push humanity forward in some stable, reliable way, then there are always tensions where you have your mission, and then you have your profit motive, and you’ve got to be able to moderate between those 2. I think public benefit governance allows you to do that. I think we need more public benefit charters in Silicon Valley and in technology, and I think we will get there. If you look at the arc of infrastructure businesses, for example, right?

I actually had a chat with a mutual friend of ours who asked not to be revealed.

They said, “For fuck’s sake, all these PBCs—public benefit corporations—will these startup founders not just fucking win their market first?” How are they feeling? Are they investors in Anthropic?

No. Okay. So, tell them to give me a call when they’d like to be investors in the world’s fastest-growing business of all time. Then they can lecture me about public benefit governance and market-share adoption.

Public benefit governance gives the leadership the ability to make decisions that sometimes are not legible to shareholders as best for them.

What decision can you foresee with Amp that is aligned to your mission but does not put the profit motive first?

There are many up and down the stack because we see ourselves as a full-stack scaling partner to the best frontier technology teams. We also see ourselves as a little bit of an institution whose job is to propose independent standards for AI and try to evangelize the adoption of those standards through profit-generating businesses. We have a venture capital business. We also have an infrastructure business.

A good example of this for now is that we’re actually giving away most of our compute at cost. Now, if you’re a shareholder, you’d go, “Wait, you have billions of dollars of compute infrastructure that you’re giving away at cost?” Yes, because we think that’s the right thing for humanity. We think that’s the right way, long-term, to have a healthy, independent ecosystem, which is what our mission is.

Our mission says Amp is a public benefit holding company. Our vision is to ensure there’s a healthy, independent frontier technology ecosystem. Our mission is to maximize the world’s frontier output.

To do that long-term, we think the teams that are truly doing innovation—truly pushing the frontier of science and engineering—need access to compute. Many of those teams today can’t afford to pay price-gouging, extraordinarily high prices for compute infrastructure. So, we’re happy to provide them access in a way that’s mission-aligned.

8. The AMP Grid: Building the Electricity Grid for Compute

Anj, how do you secure the compute supply? Maybe I should know this, but it’s the most starved resource today. How do you secure a resource that no one else can seemingly secure?

Well, step 1 is you get there first, before people realize how valuable it is. I’ve been beating the drum on this for 4 years now. When I got to a16z as a general partner, the first thing I did was sit down with Marc Andreessen and say, “We need more compute. We need compute access for these incubations I’m going to do.”

9. Coordinating Defense Against AI Distillation Attacks

They said, “No problem, Anj. Let’s set up a program. What do you need?” So, we used our balance sheet to start procuring compute through the Oxygen program. That gave me the ability to build pretty deep relationships with the industry and build trust with compute partners, with whom we now have lots and lots of relationships that we’re scaling in ways that would be very hard if I didn’t have that time and flexibility to understand what is required to really get that infrastructure right.

We’ve talked a little bit publicly about what we’re building, which is the Amp Grid. Essentially, what the electricity grid did for electricity, we’re trying to do for compute infrastructure. We see ourselves as an independent system operator of the grid.

We’re not a cloud provider. We don’t own our own data centers. We’re not a traditional venture capital firm, either. We see ourselves as an independent system operator, which means our job is to coordinate capacity across the ecosystem in a way that allows the best independent teams to provision for their base load, not their peak, so they don’t have to overprovision.

When they want to be able to spike up and down for training runs or inference needs, they feel secure that the capacity exists. We are roughly in industrial-revolution England in 1885 right now, where the steam engine has been discovered and you can use steam to produce all kinds of new products.

Many of these frontier labs are factories running their own generators in their backyards at half capacity, and I’m going, “This makes no sense. Let’s all pull our generators so that a shoe factory can spike up during the day, a steel factory can spike up during the night, and then you maximize utilization and, ultimately, output.”

10. Co-Founding Companies Like Kleiner Used to

When you think about allocating it, are you not using compute and the cost of compute as a loss leader for your venture fund business, which then comes in and says, “Okay, you name any of your incredible businesses that you own, whether it’s your Anthropic, your Mistral, or your Black Forest Labs, and say, ‘Okay, you’ll get the compute at cost, but for that, we need $300 million invested.’” Is that your way of winning?

That’s not at all how we make the deal. That’s not the deal.

The deal is, I incubate new companies like Periodic Labs one at a time. That’s all I can do; I can only do this one at a time because I like to team up with scientists or engineers who are at the forefront of their field. It takes a lot of work to create these new companies from scratch.

In many ways, I have the privilege to realize that we are entering a back-to-the-future era of venture capital. If you think about the birth of modern industries—semiconductors, gene editing, the biotech industry, or self-driving cars—Silicon Valley in the early days of the founding of what I call these frontier industries was very different.

The way you start the most iconic companies is very different from how companies were funded for the last 10 years in the ZIRP era. Intel, for example, was a very close partnership between a couple of scientists and an investor called Arthur Rock, who was a founding investor and was at the office every day. Arthur literally wrote the stock incentive plan. He used to run all-hands meetings at the company every week.

If you look at Genentech, which was incubated in the basement of Kleiner, its co-founders were Herb Boyer, a professor at UCSF, and Bob Swanson, who was an associate at Kleiner. I got to apprentice in that mode of venture capital because when I got to Kleiner, I was 20. I was wrapping up grad school at Stanford Med School, but I was working nights and weekends at Kleiner on the investing team.

Brook Byers, who was at KPCB, had an office next to me, and he had some free time, so I’d go to him and be like, “Brook, teach me your ways.” He regaled me with all the stories of how Genentech was being founded. I was like, “Wait. So, you’re saying basically Bob co-founded Genentech here in the basement of Kleiner?” He’s like, “Yeah. That’s what it meant to be a partner.”

I said, “Well, that’s not what happens here anymore. We write a bunch of checks to SaaS companies, and then they go off and do stuff.” He was like, “Different times.”

Are they mutually exclusive? What I mean by that is, can you have a venture ecosystem where you have a bunch of people writing a bunch of checks, as we have done for the last 10 years, and a next generation—or, to your point, a back-to-the-future era of venture capital—where you co-found a business side by side?

Can they run side by side? Or are we actually entering an era where value accrual is in the co-founding and incubation side?

I think it’s very hard for them to coexist inside of 1 person. It’s very hard to coexist sometimes inside of even 1 firm because there’s a reason I’m sitting here at Periodic Labs. I work here 3 days a week.

Every day from 8:00 a.m. to 8:30 a.m. for the last year, Liam Donohue and I have had a stand-up every morning where we go through the priorities of the company. Then we make them, prioritize them, and go execute.

The compute team at Amp is sitting upstairs procuring compute for the Periodic guys. My role models have always been the Arthur Rocks, the Bob Swansons, and the Mike Markkulas of personal computing. Effectively, the first CEO for the first year of Apple was Mike Markkula.

He was an angel investor. He was the one doing all the capex, supply chain, capital, and all of that stuff that allowed Steve and Woz to focus on the product and the engineering. That kind of deep partnership is what I get really excited about.

Can I get back to something you said before, which is that we're at the Industrial Revolution stage? I was like, okay, help me understand that. If we're at the Industrial Revolution stage, what does that mean for where we're going, and how should I be acting as an investor today?

You have to hold 2 things in conflict, which is going to seem paradoxical. This is the most important thing I learned from Marc and Ben: the future is not determined. Anyone who tells you that they can predict the future with certainty should be taken with a healthy dose of suspicion.

Instead, I try to approach things like a scientist and ask, what are the biggest bottlenecks? Let's come up with a hypothesis for how these bottlenecks will be solved, and let's run multiple experiments in parallel. Whichever one emerges, you just have to be very truth-seeking and be willing to say you're wrong.

I would say, as an investor, your job is to come up with a hypothesis for where the future is going, be willing to run multiple different experiments that are aligned with your mission in parallel, and be willing to be wrong. You have to be honest with your LPs that some of them may be wrong.

What do you say to a Brian Singerman of the world, who always said, “I'm not smart enough to predict the future, but my job is to pick founders that are able to do so”?

I think the safest way to predict the future is to invent it. Do the hard work, come up with your point of view on what happened next if we're in Industrial Revolution England, and what were the emerging properties of the businesses that became valuable institutions over the next 50 years after 1885.

Then figure out which part of that world, and which figure from history of that era, you look up to the most. Go read about their lives, the businesses they ran, and the tensions that emerged in the practice of their businesses later in life, because they made mistakes when they were young. Try to learn from their mistakes, and then go and execute.

What's the parallel trajectory from the 1885-onward time frame that you think will play out in the next era?

Obviously, in the world of infrastructure, I think we need something like the grid for computer infrastructure. That's what I've spent most of my days on: a coordinating mechanism that allowed the transition of coal and electricity from being resources that were being hoarded to being stable, reliable commodities that the best engineering teams and the best factories had access to. That's what I think about a lot.

Since you're so talented at media and storytelling, and your mission is to push the European continent forward, I think one of the things I would do if I were you is try to figure out how we educate the leading capital allocators and infrastructure allocators in Europe about the coming era. Whether that's through media or educational programs, get them to understand their role in unblocking the bottlenecks for the best scientists and engineers in Europe.

And largely a lack of pension fund reform in a lot of cases, to be quite honest.

You spend your time on pension fund reform.

How much more cash do we need in Europe for frontier AI to be what we think it can be? Is it 2 times? Is it 10 times?

That's a good question. I would try to go about it from a top-down approach and a bottoms-up sizing approach. For us at Amp, when I look at the grid we're building, which is sort of reasoning by analogy, we have started securing about 1.3 gigawatts of compute infrastructure. That's roughly $40 of cloud spend over the next 4 years, and that's financed with about 20% equity. That's about $10 of equity capital; the remainder is debt capital.

We have a bunch of partners that help us put together these equity and debt packages to secure compute infrastructure for our companies. In Europe, I would talk to Arthur Mensch and figure out how much he thinks is required for the independent ecosystem over there.

If you're doing your atomic unit of math in gigawatts, from a top-down perspective, I think Google is roughly at 12 to 15 gigawatts, as far as I'm aware, of infrastructure for internal and externally deployed needs. They have a huge land, power, and shell pipeline coming.

If Europe does not have access to Google-level infrastructure, then what are you guys even doing? That's roughly what the continent needs for full sovereignty: to have at least as much infrastructure locally as there is within the Alphabet Holdings pool over the next 4 years.

What's easier, the equity raise or the debt raise?

I would say the biggest challenge has been figuring out the right, aligned financial structure across both in a way that's legible to capital allocators at scale. It took me about a year to really get all the pieces right.

There are very large equity pools—the long-term, mission-aligned balance sheets in the world that want to help frontier scientists and researchers and university labs get access to the compute they want. But they don't have OpEx. They don't have cash to spend on the compute.

If you can find a way to align equity and debt balance sheets in a way that's de-risked, the fundraising is not a problem. It's actually a systems-design problem, which took me, again, a year. It probably took me 4 years to get right, but now that we've figured it out, it's not been a problem.

11. We're in a GPU Wastage Bubble, Not an AI Bubble

Do you think we are still underinvested in data centers today?

We are deeply underinvested in secure compute. Let me put it this way: we are not in an AI crisis, and we're not in an AI bubble, for sure. That's the question I keep getting asked.

We are definitely in a GPU-wastage bubble, where there are stranded buckets of compute—billions of dollars of compute sitting unutilized. If we could pool them together on a grid across the independent ecosystem—

Why are they unutilized, sorry?

For a couple of different reasons. One is that compute is not fungible. Unlike electricity, which had to go through a process of standardization—AC/DC, where megawatts are megawatts are megawatts—compute is not fungible today.

Forget fungibility of compute across different manufacturers, like NVIDIA and AMD. Within a manufacturer, NVIDIA chips, for example—H100s, GB200s, and GB300s—are all completely different chip types. If you have 1 cluster where you're doing a training run on H100s, and then you want to do continued post-training of that, or have it do a distributed training run of that training workload on GB200s, it doesn't work.

There are just stranded pools of compute because FLOPs are the atomic unit of computation. I wish FLOPs were fungible, but not all FLOPs are born equal today. If you provisioned a cluster 2 or 3 years ago with H100s, and now you want to run some of those workloads on the newer-generation models, you're memory-bound by H100 chips.

You can't unlock the benefits of the Blackwell chip without basically buying a new cluster. Suddenly, you have this H100 cluster that you don't want to train on anymore because it's old school. The chip doesn't have the right memory properties to train your frontier models.

It's very hard for any individual company to see all of this stuff, but when you're on 7 or 8 boards like I am, you've been doing this for 15 years, and you start to see patterns emerge, you think, “Wait a minute. Why is there all this unutilized compute sitting here and there? This is a loophole for everybody.”

12. Why Compute Isn't Fungible

Are our frontier models moving faster than the pace of chips? As you said, with H100s, you have newer and newer models, and then you're training them on older and older chips because that's what's free, so it's not moving in lockstep. Is that the problem that we're articulating?

No, no, no. The problem we're articulating is that compute is not fungible. There are no standards for fungibility, and there are no institutions enforcing enough standardization of compute. We are in the pre-standardization era of compute today, which was the pre-standardization era of electricity in 1885.

I hope we can self-regulate, self-standardize, and self-enforce standardization so that we can skip the boom-and-bust cycles that happened with electricity over the next 50 years. This happens with every infrastructure cycle in the pre-standardization era. It happened with electricity in 1885, it happened with steel, and it happened with railroads.

Every time you have this boom-and-bust cycle, wars are fought, companies backstab each other, and it's super painful.

It's annoying. And my view is that compute not being fungible is what's resulting in all this talk about AI—the AI bubble. But what people forget is that we don't have an AI capabilities bubble. The capabilities are extraordinary in every domain. We have an infrastructure wastage crisis right now.

It's because there are no open standards. There's no open protocol for how FLOPs from one data center can flow to somebody else who needs them across chip types and secure boundaries. It's resulting in a lot of pain for the ecosystem.

If we have compute standardization in the way that you said, will we remove the boom-and-bust cycle, or is that just one part of it?

I think that will go a long way in preventing this and instead just allowing this.

I'm sorry for asking. You're like, “Jesus Christ, Harry. I'm a professor at Stanford, and you waste my Omniverse,” which is a fat statement. A British accent goes a long way there. What is the biggest bottleneck or barrier to compute standardization that you want to achieve?

It all goes back to alignment, man. Misaligned incentives up and down the stack.

How are Silicon Valley and D.C. not aligned?

For one, I don't think we've standardized on whether AI should be regulated, treated, and procured as just good old-fashioned software or like a new kind of system. Look, I went to grad school for machine learning, and what you learn in Machine Learning 101 is that models are statistical. They're not deterministic, right?

There are some properties of a statistical system that are different from a spreadsheet. A spreadsheet is deterministic software, and a statistical model today is not. So, should the procurement of a spreadsheet be the same, from an IT perspective, as that of a statistical model? Open debate. That is the core debate. That's the problem.

AI alignment, don't get me wrong, is hard, but it's not the hardest problem. Human misalignment—human alignment—is really the problem we have in the world right now. We need technologists who understand the difference between deterministic software and statistical systems to propose a set of standards for how procurement for this should work.

Then we need standards people in D.C. We have this thing called NIST. We have various bodies in the government that should get together and say, “Thank you guys for proposing the standard. This is where it makes sense. This is where it doesn't.” This is called an RFC process, and we're going to standardize on this definition of procurement.

This is what happened with TCP/IP and the internet. It happened with AC/DC and electricity. We have not done that yet for the model era. Unfortunately, these are called open standards, and the standardization process is being confused with marketing.

13. How China Is Winning the AI Race

Now, President Trump is actually, I think, trying to do his best, from what I can tell, in at least giving America enough freedom to innovate that these standards can even be discovered in our labs here. First, you need somebody to actually pioneer and figure out what the standards should even look like. I think there's just a lot of noise.

Do you worry that the CCP is subsidizing a generation of Chinese models that are now being used by American companies, whereby they have frontier models to essentially set where model capabilities can be, and then have a real effort to make the open-source Chinese models as close to those benchmarks as possible, much, much cheaper?

The engineering execution right now, up and down the stack in China, is extraordinary. Here's what's happening: What they realized is that the AI scaling race is not a chip race. It's a full-stack systems co-design race.

If you can't compete head-to-head on chips for now, what do you do? You compete on systems design. You say, “Okay, we don't have leading-edge chips here yet, so let's try to compete on systems.” You co-design the chip that you have, which might be Huawei, with the compute infrastructure and the training run. Then you design that to have a bunch of performance improvements at every layer of the stack.

Then you do adversarial distillation at scale, where you take Western models and, from various different endpoints, distill the state of the art. Then you try to get as many performance gains as possible from that data, release that back out to the world as open models, see what people react to, get feedback, and do the next one, and the next one, and then catch up.

At the point when you catch up, you say, “Wait a minute, we're starting to be at the frontier. Why do we need to open-source anymore? This is good enough for our local domestic needs.”

It's beautiful. And that has actually, by the way, resulted in innovation. They're innovating at every step of the cycle. That's why Huawei chips are able to produce capability improvements today in China that rival some of the best chips here when integrated up and down the stack.

In a sense, it's the Google strategy, right? Google is integrated: land, power, silicon, TPUs, Borg, XLA, Google DeepMind, Gemini. The systems co-design there, up and down, results in efficiency that gives you huge performance gains at the end of the day. China has replicated that strategy, using open source as a sort of bootstrapping mechanism to catch up. It's extraordinary.

Does that concern you?

Are you kidding? Absolutely. That's why I think what we need is a Western grid where all frontier inference is served through an Iron Dome. If there are any adversarial distillation attacks against any one of our teams, we coordinate together.

Because I'm on 7 boards, I'm in group chats where I get texted by one founder saying, “Anj, is anyone else noticing today that there's a huge spike in distillation from this region?” Then I put them in a group chat, and we coordinate. It's very informal right now.

You said before that state-sponsored attacks on frontier AI labs are getting worse. What do we not know that we should know?

We should know that there are insider threats. We should know that there's distillation happening across the U.S. and Europe that's taking advantage of us all not being united. That distillation is taking advantage of our political systems. Our mission-critical infrastructure is quite vulnerable, especially data centers that are serving workloads being used by enterprises.

I think that, from a business standpoint, if we don't secure frontier model inference, or what I would call state-of-the-art inference, behind a coordinated Iron Dome, I don't think we have a sustainable shot at staying at the frontier over the next decade.

I'm sorry, what does that mean—an Iron Dome for inference—in terms of sustaining it?

It means that all inference, no matter which company is serving it, is served through a shared proxy that can tell each other when there's an attack happening on one part of the frontier. Think of it as an Iron Dome across the entire Western Front, right?

Just because you're here, you're in one company, you can't see that your model being served through this other company is being distilled. So, it's a deployment coordination protocol. It's basically my group chat that I've got with a bunch of different founders, but scaled, where people go, “We're seeing this attack today,” and others go, “We are, too. Let's coordinate on defensive response.”

I'm sorry for my lack of cohesion on the question, really. I feel guilty, and I don't blame you for leaving this interview thinking, “God, he's got worse over the 8 years, not better.”

But I was watching this interview—you were speaking of inference—with someone I think from Base10, and they were saying that the demand for inference has grown not linearly, but combinatorially. And that is how we would see it progress over the next 3 to 5 years. Do you agree with that?

If we keep scaling capabilities, that will definitely happen. The problem is that there are a couple of bottlenecks on scaling capabilities that are quite existential. The 4 core bottlenecks to capabilities progress we've talked about are context, compute, capital, and culture.

Capital allocation is a huge problem. We've got to educate people on why these capabilities are extraordinary. This is the biggest financial bonanza of all time, if you know where to allocate. There's a reason why I invested in Anthropic in the seed round, and now, as you pointed out, the returns of all the body of work I've done in the last 4 years are attracting LPs at the highest levels.

But we're just getting started. I think some of these projections you see are correct if we unblock the bottlenecks along the way.

In compute infrastructure, secure compute infrastructure that's fungible and standardized—that's the biggest bottleneck. I think if there's any reason why OpenAI, Anthropic, Gemini, and so on don't hit their revenue targets over the next few years, it's because they won't have access to enough compute.

I'll say there's a related bottleneck. When I was at Stanford many years ago, as a kid, I took this class that Peter taught, which I think was turning into this book called “Zero to One.” This is Peter Thiel. I was an editor for The Stanford Review, and he had this quote, right? “Competition is for losers.”

And now, having done this for 15 years, I've kind of updated my theory of business. I think he was not wrong, but he was insufficiently precise: I think perfect competition is for losers.

14. Perfect Competition Is for Losers

What does that mean, “perfect competition is for losers”?

It means that if you have 50 companies all doing LLM training or doing coding models, that's a losing proposition. It's like perfect competition is like restaurants. There's no defensibility. That's why restaurants go out of business all the time. It's very hard for them to differentiate.

On the other hand, monopolies are mafias. Once you have a monopoly at one part of the stack, they stop innovating, and instead they try to go up or down by using the balance sheet to acquire. They start hoarding resources. They start saying, “You give me this, and I will force you to basically subsume yourself to me,” and I'm seeing that kind of behavior up and down the stack. Mafias are not good for innovation.

I think what we need is optimal competition. The optimal competition setup is that you have 3 or 4 teams in every frontier that are making extraordinary progress. If you invest in them, you get extraordinary returns, but they're not so comfortable as to be a monopoly such that they can stop innovating. That's important, because when they stop innovating, as humanity, we're fucked.

I believe that, with optimal competition, we need to transition to the era of optimal competition in frontier technology. I think we need leaders, stewards, venture capitalists, politicians, and educators to remind the world that we have already lived through this era of boom and bust and so on.

These companies—what's going to happen, right? Like you said, Andrej Bystrov and Inference [?], all these companies. Inference is an extraordinary growth curve ahead, but it's not going to be an extraordinary growth curve if there are 50 Inference companies all competing with each other in a race to the bottom, which is kind of what's happening right now. It's not clear to me that we need 50 Inference companies. It's not clear to me that VCs are smart enough to realize that they're just lighting hundreds of millions of dollars on fire in a category where having 4 or 5 really good, trusted inference providers is net good.

But will VC subsidization of 50, 60, 70, whatever the number of companies is, not make it impossible for the good companies—the 4 or 5—to progress through that cycle?

It's a bit of a self-destructive mechanism, because if you have 50 different companies all competing for scarce compute resources, then the folks who are actually innovating can't get it, and so they can't do their next round of product innovation and so on. And that's the problem when you have that.

Is that where we are now, though?

That's where we are right now. The best inference teams are calling me up. Actually, all inference teams are calling me up and saying, “Anj, do you have compute for us?” Because their product is reselling compute. But it's been hoarded. It's been hoarded by the hyperscalers. It's been hoarded by people who are not innovating but are sitting on compute.

It's so obvious to me now that I've left a16z and I'm an independent ecosystem public-benefit corporation that the existential threat to innovation in this category is lack of compute. That's why I started procuring compute for the independent ecosystem a while ago. We're trying to find a way to get these teams enough compute that they need to keep innovating.

The 4 or 5 inference companies that win versus the others that don't?

Supply. Access to supply.

Is it that simple?

Yep. Compute supply. If you don't have compute, how do you do inference, man? What are you selling? You need a product to sell. If you're making a steam engine, you need coal.

One of your former partners, likely Martin Casado, tweeted last night that we're going to enter a time where only model creators access the most powerful models. That will power, obviously, the services in the application layer, or the apps that they provide. Do you believe that will be a world in which we exist, where model providers inherently safeguard the best models for their provisioning of apps, à la Claude, potentially, or not?

What Martin is suggesting is that, in competing cases, they will offer a worse model, which gives them an advantage. As an example, ElevenLabs, which serves a huge number of application-layer companies, will reserve their latest models so they can offer the best customer support, and then sell their older models to Sierra and Adecco, so they have a lower-quality model, retaining the best for themselves.

The embedded assumption there, right? What we have learned empirically over the history of technology is that if you have a general-purpose product like the iPhone that works for everybody, then the natural incentive is to amortize the cost of product development over the largest number of users. So, if you have a general model that's good for everybody, it will be available to everyone.

If you have specialized models that are good for some people, they will be priced through product segmentation. I think what this is telling us is that if there are many custom models, some of them will be accessible and others will not be. If anything, I think we should see the fact that frontier model labs are saying, “Hey, here's a new model we have. It only makes sense for some large enterprises to access this,” as vindication of the ecosystem truth that there's going to be an ecosystem of different models of different types.

There's no one large god model. If there was, I think there would be a market desire to have prime ministers, presidents, and students all use the same iPhone. Inherently, you can raise the most money and invest the most product-budget dollars in a general product and amortize the cost of that across everybody.

But if you have specialized models, I don't think they're going to be accessible to everybody, and they don't need to be. I think this open- and closed-access thing is somewhat overblown. Empirically, from a systems perspective, if you look at the history of technology, if you have general products, they're distributed to the masses. If you have custom products, they have enterprise segmentation. Some are accessible to the enterprise, and others are not.

How many of those foundation-model-layer companies that are yet to be built will be worth over $100 billion?

Oh, so many. I'm saying Periodic Labs is one. I'm sitting in one right here, right? But they're not foundation-model companies. I would call them frontier-systems companies.

This is the problem. Every time I kept calling and trying to educate people, 4 years ago they'd be like, “Anj, Anthropic is a foundation-model company, and Mistral is a foundation-model company.” I said, “No, guys, that's just one part of what they do. Maybe they're starting there because that's a core competence.”

But there's a reason why Anthropic also has something called Claude Code. There's also a reason why Mistral has something called Mistral Compute. There's a reason why Microsoft, who's a cloud, also has a Copilot business. These labels—the categories of foundation models—need to be viewed with more suspicion than they are.

What matters is the full systems co-design, the full-stack frontier research loop that you need to run with customers. Later, when that happens, when you say, “Oh my God, Anj, Anthropic is now—they were a model company, and now they're launching a product called Claude Code,” I'm like, “What do you mean? That was part of the plan all along. Of course, you need to have a pair-programmer interface for a model. Why would you assume otherwise? Oh, because you just weren't paying attention, and you had your neat market maps that your associates were giving you, and you thought that was truth.”

The commercial community has forgotten how to build businesses, and they've forgotten the difference between first principles and marketing. That's the problem. That's one of the other misalignment problems. The ground truth of these machine-learning-systems businesses is that they've always been frontier-systems businesses. They were never just foundation-model businesses.

Now, if you had to package that up and tell your LPs that because it was legible to them, then I can't blame you, I guess. But the LPs I work with, I'm very upfront with them. I say, “Look, these categories are going through huge reinventions. If you want to partner with me, what you get is a full-stack partner, and I will tell you the first principles of what's going on. These first-principles insights will change over time, but you've got to be comfortable with huge capex outlays in businesses that end up winning the entire category.”

That's what frontier technology is. I think foundation models have been deeply misunderstood. This is part of why I started the class 4 years ago. I just thought AI at scale was going through a bunch of reinvention. Then we reinvented the class to be Infrastructure at Scale last year, and this year it's Frontier Systems, because not enough people realize that, to keep the capabilities frontier moving, you need to think about these projects—these companies—as frontier-systems projects, not foundation-model projects.

Does that make sense?

It does, but when I hear about the capex required, I respectfully ask: do you have enough money? I mean, the $1.3 billion was—

No.

Yeah. How much money do you need?

Well, for the 1.3 gigawatt, which was kind of our proof of concept, that capital is not a problem. I think the question is if you want to scale beyond that.

Mhm.

Yeah, we need way more capital to be deployed across the Western front in the United States and U.S.-allied countries.

How much do you think you need?

As long as the capabilities frontier keeps moving and we want a healthy, independent ecosystem, we'll just keep raising more capital. There's no end to that. The day machine learning stops working as a systematic way to give humanity more capabilities, that's when I'll say, “We have enough, Harry.” But that's so far out, I don't even know how to reason about that.

I could talk to you all day, but before we do a quick fire, how will Vannevar be fundamentally different in 5 years' time than it is today?

Well, again, go back to history. I think there will be a few people like Arthur Rock, Bob Swanson, and Mike Markkula who turned their practice into institutions. There will be others who don't. If they don't evolve themselves for what entrepreneurs of this era need, then I think they should get out of the venture capital business, because we don't need more bankers.

One of the beautiful things—I have a friend, Vlad, who runs Robinhood, and he floated this venture-fund thing on Robinhood recently.

Yeah, yeah, Robinhood Ventures, I think it is.

Yeah. But when you have software that can play many of the coordinating roles of venture capital firms, why do you need somebody who's just a pure—to borrow a Marxism—wrapper on LPs?

Look, here's what I'm most concerned about with the capital ecosystem: not enough of the wealth-creation opportunity that's happening in frontier AI is being shared with the public. And that's not good for anybody. Because if you don't share this wealth-creation opportunity with the people who are supposed to be welcoming this technology into their lives, which is ultimately the public, what are they going to do? They're going to say, “I don't want these data centers.”

With the greatest of respect, a lot of the money in venture capital funds comes from endowments, pension funds, and teachers' funds. That wealth distribution should ultimately trickle down, if we believe in that.

But how many venture capital firms were in the seed round of Anthropic?

Ooh, none.

That's the answer for you. And that's happening again and again and again. There's a huge misallocation of public capital into venture managers who are not capturing enough value in frontier AI. Instead, they're investing in a bunch of stuff that's not going to exist. And the public's going to be mad.

Did you put $300 million into Anthropic in one go?

I've had the privilege to invest many hundreds of millions of dollars into Anthropic across several rounds, from the first to the most recent one. So, I consider that lucky. I intend to give most of it away to public-benefit causes and public-benefit education programs. And I think we're at the very beginning of Anthropic's journey on commercial progress.

15. Quick-Fire Round

Dude, I'm going to do a quick-fire round with you, because otherwise it's going to take all day. If you could advise an LP investing in venture funds on one thing, what would you advise them?

Educate yourself. Take the class. Do the readings. Don't skip the hard work. Too many LPs are outsourcing their hard work—the work they're supposed to be doing as capital allocators, which is understanding what's actually going on. Then decide which venture managers and allocators you think have a unique, defensible advantage on the bottlenecks. I would be investing in the bottlenecks.

Basically, too many GPs aren't doing the work. The amount of GPs who've never built anything with AI is astonishing.

I agree. Completely agreed.

You'll laugh at me: I've built with every different vibe-coding provider. I'm trying to turn my media company into an AI-first media company. It's pathetic compared to the shit that you do, but at least I'm trying. I'm seeing the bottlenecks of Supabase integrations and everything that comes with them. And you learn by building. I think if you're not doing that in the beginning, you shouldn't be investing, period.

I completely agree. There was a sovereign country that came to me at the end of last year and said, “Anj, we want to bring 26 of our ministers to your house and do a 1-year program where we educate them. It's a frontier AI program where we learn what's going on in AI from lectures and so on. And then we want to do a deployment project where each of our ministers actually builds AI agents.”

And I said, “You know what? That's like—if you didn't see it, take Stanford CS 153. That's a microcosm of this course I'm doing with this country, the sovereign fund that we partnered with.”

That's the way. You have to do the work: read the literature, understand what's going on in research, and then deploy yourself. Build tools.

The class project—the Stanford CS 153 class project—is the one-person frontier lab. I genuinely believe that what would have taken 50 people to do 4 years ago, now, with the right AI tools, you can do with 1 person. And as a leader, if you haven't played with these tools, deployed yourself, and built your own agent, I don't think you understand what's going on. I'm not letting the ministers who are taking this class with me graduate until they build and deploy agents. I've told them that they're not getting their graduate certification.

Have you told your wife that 26 ministers are coming to your house?

She let me co-host them at our house tonight.

Tonight, Anj?

She let me co-host them at our house in San Francisco a few weeks ago. And I'm very lucky to have Viv—I don't deserve Viv, I'll tell you that. But she's very mission-aligned, and we both believe that the best thing we could be doing with our time is educating at scale.

What makes Dario so good that other people don't see from the outside?

One

sheer scientific brilliance. Truly world-class technical ability in his domain. Two: an obsessive desire for truth-seeking—to keep reasoning and keep doing experiments until he's right. He's a physicist at heart. I think Dario is a physicist at the end of the day; he's not actually a computer scientist.

A world-class physicist tries to derive general laws of reality by looking at data and running empirical experiments. He's an empiricist, and he has an obsessive desire to be a good empiricist.

The third is mission-alignment culture. He says, “This is our focus. This is our mission. No drift. We won't take shortcuts. We are willing to make huge trade-offs to hit this mission.” And that attracts the best talent. Incredible talent.

In the face of criticism from people saying, “You're a mercenary. You're blah blah blah. You're just doing this for profit,” no, actually, it turns out there's a ruthless desire to stay focused on the mission. And that results in hard trade-offs and priorities. If you're not aligned with that mission, then you'll just think he's crazy or he's evil or whatever. It's crazy how many ad hominem attacks I've seen against him. But he's about that clarity of mission.

What have you changed your mind on in the last 12 months?

The biggest one is health. I've had some health experiences, both with family members and myself, that made me realize we just don't know how much time we have on Earth. And that makes you stop taking for granted how much time we have. So, I started taking time much more seriously.

But I would say—and this is something I say in every lecture I do at Stanford—we talk a lot about scaling laws and technical stuff, but I also give the kids an Anj's life-scaling-laws lesson. I'm very inspired by Richard Feynman. Feynman's lectures always combine technical education with a little bit of life coaching for them.

My number-one scaling law for the students was: take life seriously, but don't take it so seriously that you forget what makes it worth living, which is having fun with friends, working on interesting projects with people you love, and not taking relationships for granted. It's humans that make the world go around.

If you're so focused on your next fund or your next raise or whatever, you just take for granted the one thing we all don't know how much we have, which is time with each other. And so, I just started valuing my time more, and my relationships with people.

There are so many people—I mean, my parents. I left my parents behind in India to move to college at Stanford, and I've gone weeks of my life without calling or texting them. And now they're in their 60s, and I—

I would give you a hug if we were in person. Fuck, I'm so sorry, man. I—

Don't worry, it's okay.

Jesus. You know, the first money we ever made from the show, we made it because my mom has MS, and we couldn't afford treatment for her.

The only way that I could pay for it was by putting adverts in the show. That was how we did it, and they still pay for it.

Yeah, man. The trade-offs. The sacrifices.

Parents are amazing.

Parents are insane.

How do you escape the money treadmill? I didn’t have money when I grew up, and I was like, “I’ll be happy when I get, you know, X amount of money.” It bites when you’re escaping that money treadmill.

I was very lucky that I went to Singapore on a government scholarship. Lee Kuan Yew, who was the founding father of Singapore—I’m a big Lee Kuan Yewist—realized that they didn’t have many resources. They didn’t have money as a founding nation. They had nothing, basically, other than themselves and their location, their strategic location.

He realized, “We need to build a talent program. We need to run this country like a company.” We would recruit the best talent from across Asia. Because I think I was in the top 10 or something in a public exam when I was in the 10th grade in India, I was tapped to be a scholar in Singapore. I took it—I was a government scholar.

I didn’t actually have to take it; I was lucky enough that my parents could have paid for it. We had a family business in telecom. But it was very important to me to be independent from my parents, because in Indian culture and in a lot of cultures, if you don’t have financial independence, you’re always kind of beholden to somebody else. In the case of community cultures like India, there’s a lot of pressure to adhere to their values and so on.

And I think I did that because, subconsciously, I’m very lucky to have a sister who lives in London and who fought my battles for me. She was 7 years older, and I got to see that she was a rebel. She wanted to do all kinds of things, including going to fashion school, and they didn’t want her to, so she had to go to law school because they were paying for it. I think she actually fought some of my battles and made me realize that the more independent I was, the more freedom I had, and freedom matters to me a lot.

I’ve always found that I’m willing to make big trade-offs in money to retain my independence. I’ve always defined my financial goals through independence. That’s what has always mattered to me. Anytime I find my independence feeling threatened, I go, “You know what? I need a change.” That’s what matters to me.

I think you need to figure out what your mission is—what matters to you more than anything else that you’re willing to just turn down all kinds of money and job opportunities and so on—because that clarifies a lot where you spend your time, basically.

My mission—my mission is really to enrich the already very rich family offices of Europe.

Okay. Well, then you’ve got a ways to go on the treadmill, brother.

I was out of the loop. I didn’t read the memo.

You know, I also think that people are just not very funny anymore. We lack a little bit of humor in a lot of society. It’s so sad. I was with my partner the other day, and I said, “You speak like AI.” And he’s like, “I know, and you know what? I talk to my wife like I talk to Claude, and she fucking kills me.” I’m like, “Yeah, that’s not a good thing.”

Dude, final one. It’s a bit morbid, but what do you want to be remembered for?

You know, Viv asked me this 3 years ago at a party. I think it was at our anniversary or something. We were with 10 of our closest friends, including some of the co-founders of Anthropic. It was one of these classic San Francisco dinner parties, and she put me on the spot.

I think what she had asked was, “What do you want it to say on your tombstone?” And I blurted out, “He was right.” The room just went dead quiet, and they were like, “Yep.”

It’s because I have this obsessive desire and need to try to learn where the future is going and then tell everybody about it. Everybody thinks I’m some snake oil salesman or whatever. Now that’s changed, because it turns out I was so right. I made LPs so much money that they’re all now asking me, “Anj, can you...?”

I was invited back to teach this class at Stanford, right? People are finally, I guess, realizing, “Okay, Anj might know a thing or two about the future. Let’s go get his take.”

I went to this boarding school in India called Rishi Valley.

7 years of no tech.

No tech.

7 years.

No wonder you’re a happy and adjusted person.

It’s taken me a while to get here, but yeah.

Would you let your children have social media?

Yes. I think it’d be crazy not to let them have access to social media, but I think it has to be done in moderation. Most parents have a really hard time moderating it with their kids, and then it’s really hard to moderate.

With Rishi Valley, I had access to a computer once a week. You need to enforce something like that, where you don’t take it for granted. It’s within a structured sort of environment, and then you develop good habits, protocols, and practices to not be dependent on it.

I would plan my Wikipedia sessions. You got 1 hour a week in the computer room at Rishi Valley, so you’d really have to plan the highest use of that time. You’re not dependent on it, but you just use it as a high-leverage strategic asset. That’s how I think technology should be viewed.

You shouldn’t take it for granted. The problem is, people keep saying we’re going to have the singularity soon. I’m like, “Have you realized we’ve been at this for 10 years? Half of you have outsourced your brain and thinking to this device anyway?”

Dude, I so enjoyed doing this. Thank you so much for putting up with me. You’ve been utterly fantastic.

No problem, man. Thank you for the consistency with which you’ve kept this up. I mean, you’re an institution now, right? The hard thing is—I mean, I can’t believe this. You’ve been at this. We’ve gotten old together, right? When you were, like you said, you were a kid. I was much younger.

Pat Grady and I—Pat Grady was one of the first people I met in venture, and he was an associate, and I was a 17-year-old. I laughed with Pat. I said to him the other day, “Dude, you’ve gone from associate to, like, head of Sequoia, and I’ve gone from podcaster to podcaster.”