AI Fund 普通合伙人 Andrew Ng：LLM 是下一件地缘政治武器，AI 时代毛利率还重要吗？

Andrew, I've been an admirer for a long time, so I've been really looking forward to making this happen. Thank you so much for joining me today.

Yeah, thank you, Harry. I've watched a bunch of shows, and I really enjoyed your recent one with my friend Martin Casado as well. That was very memorable, so I'm actually thrilled to be here. I love Martin. He's a very, very special man.

I want to start with something that you've said before. You said AI is the new electricity, and when I think about electricity and where we are today, I want to understand the bottlenecks. Everyone seems to suggest that it really is about data, compute, and algorithms. Are those the 3 parameters around which we should think about bottlenecks? If so, which one do you think is the biggest bottleneck?

I would say the 2 biggest bottlenecks right now may be electricity and semiconductors. I think electricity is one of them. In the US, I am honestly worried that many data center operators are stuck in permitting. I know that local community support is important, and some people don't want a data center there, but once we build roads and railways as infrastructure for a certain generation, data centers are the critical infrastructure for building the digital economy.

The lack of electricity in America and in a number of Western countries is a problem. In contrast, I see China building power plants left and right, including nuclear. That will be an interesting dynamic. Semiconductors are another bottleneck. AI is so complicated that we also need more data and better algorithms. All of it is worth working on, but in the short term, there are constraints with electricity and semiconductors.

Can you talk to me about the constraints around semiconductors that you think are most pressing, that most people don't realize?

In my career working on AI, I have yet to meet a single AI person who ever felt like they had enough compute. Give us any amount of compute, and we'll use it all up and say we still don't have enough. This has been a constraint for the last 20 years or so.

What I'm seeing is that, with the rise of generative AI, there are very valuable workloads. For example, AI-assisted coding is fantastic. It's making us so much more productive. But if you use Claude Code enough, sometimes you get rate limited. I find that many companies really have excess demand, which is a very rare problem to have. So many people want more LLM inference and more tokens generated, but we just don't have the semiconductors, data centers, or electricity to meet the demand.

There's a lot we could do with AI token generation, and it's frustrating when, on the supply side, we can't supply enough to people who want it. On the demand side, you get very limited if you use too much.

How should I think about that insatiable need for more compute and the improvements that come from it, with the recognition that many people say GPT-5 was an example that scaling laws have been reached to a certain extent and that a focus on efficiency has been the transition? How should I balance those 2 seemingly differing opinions?

It is true that token generation is getting more efficient and cheaper. In fact, if you look at OpenAI's open-weight model, they actually released models that are very efficient to run. I think they did a good job with—was it like 120 billion parameters or something, with, I think, 5.7 billion active? So it's actually a very efficient model to run. Despite the cost of token generation falling, our demand for it is insatiable.

One interesting thing that's happened in AI is that, if we look at where the buckets of value are, one of the big buckets of value is AI-assisted coding. I think this hearkens back to an earlier era, a previous generation. I think Google came to dominate horizontal information discovery, like web search, but there was room for lots of verticals to be built out when the internet was being built. So we wound up with Travelocity and Expedia fighting it out for travel, a bunch of folks fighting it out in retail, and a bunch of others fighting it out in transportation, social media, and so on.

What we're seeing now is that ChatGPT has such a strong consumer brand. ChatGPT seems to be the dominant player in new-generation horizontal information discovery, although I think Gemini, with its channel advantage through control of Android and Chrome, is a serious player as well. But if that's where horizontal information turns out to be, then there's still plenty of room for lots of verticals to be built out.

One of the clear buckets of really valuable verticals is AI coding assistance, where Claude Code—I use that every day; I love it—and OpenAI Codex have a lot of momentum as well. It's clearly making developers so much more productive and efficient that the demand is just through the roof for people to use more and more of this. One thing I find exciting is that I often look at AI coding assistance as a harbinger for what might happen to other job functions as AI marketing tools, AI recruiting tools, and AI finance tools become more efficient. I often look at AI coding assistance as a foreshadowing of what may happen in other sectors as the tools get better for them, too.

I had Joelle Pineau from Cohere, formerly of Facebook, on the show recently, and she said that AI coding assistants are in the same place that image generation was in 2016 or 2017 in terms of maturity. Do you think that's a fair statement of the environment today, or do you not think so?

I don't know. I think it's further along. In 2016, image generation wasn't super valuable. I don't remember it being that valuable back then, but I think today AI coding assistance is really working well.

At AI Fund, my head of engineering recently said, "Hey, let's think about standardizing on tools." Basically, he said, "I need these tools, and you have to pry them out of my cold, dead hands." I think our developers feel really strongly about them. I myself don't ever want to have to code again without our AI coding assistance. I think the tools are really working well, but there's still a lot of headroom for how much better they can get.

I do want to go back to the core bottlenecks. We said they're about electricity, and we said they're about semiconductors. When we look at the build-out of data centers today, as you said, regulation has been a big part of preventing that in a lot of ways. Do you think Trump has done more to help or to hurt the progression of AI in the United States from an infrastructure perspective over the last few years?

Over the last few years, the US federal government has done some good things and some less helpful things. I feel like clearing out unnecessary regulations has been a very good move. Even last year, the bipartisan Schumer AI Insight Forum—I think there were a lot of people lobbying the US government to pass stifling regulations. There were a lot of hyped-up AI safety narratives saying AI could lead to human extinction, which is kind of a ridiculous statement, to try to get stifling, anticompetitive regulations passed, often to try to shut down open-source and open-weight models.

Fortunately, we beat back a lot of that. I think the bipartisan Schumer AI Insight Forum did a really good job digging into the truth and concluding that America should be investing in AI rather than passing unnecessary regulations to slow it down. I think Trump did a good job, and then his whole team—David Sacks, Christian [?], and so on—did a good job clearing out unnecessary regulations.

On the flip side, one of America's huge competitive advantages has been its ability to attract talent, including high-skill talent as well as, frankly, young talent that may not currently be high-skilled but could be high-skilled in the future. To the extent that America is not investing as much in attracting talent, I think that would be an unforced error.

Lastly, I think investments in science are important. Helping our institutions of higher education have the resources to train our graduate students and invest in scientific technology is really precious. Anything that damages that would also be very unfortunate.

If I gave you a regulatory magic wand, Andrew, what would you change that would have the most significant, needle-moving impact?

America is fortunate to have a lot of very smart people wanting to come here to do really challenging, really tough problems. Many of our Nobel laureates are immigrants. Einstein, for example, was an immigrant. I think continuing to cultivate America as a place to attract great talent to work together in a democratic nation that respects the rule of law would help us move ahead.

I think securing the semiconductor supply chain would be very valuable as well. I have a lot of friends in Taiwan. I love Taiwan. America's dependency on TSMC is concerning in case anything happens.

Frankly, there's one very funny thing that happened in society. There was recently a Pew report showing how much Americans think AI would be good for them and how enthusiastic they are about it versus how many are not enthusiastic. Even though a lot of AI technologies were invented in America, a lot of people don't trust or don't like AI.

The joy of what I do, Andrew, is that I get to speak to incredible people and then cross-reference what they say.

David Cahn from Sequoia said, “A really useful barometer for effectiveness is: can AI replace the bottom 5% of the capabilities of what a workforce does?” Joelle from Cohere said, “No, that’s crap.” The real question is: can it 10x people’s ability? Forget the bottom 5%. Can it 10x? How do you think about a barometer for the success of the workforce with AI, with those in mind?

In the case of software engineering, it is accelerating the writing of code. There are so many projects that used to take 6 engineers half a year to build that today I or one of my engineers can build in a weekend. I hope that we never have to go back to coding without our AI assistance again, because the acceleration, the productivity boost, is incredible.

For example, one weekend I thought, “I want flashcards for my daughter to practice multiplication,” and she wanted flashcards. So I thought I could either drive to the store and buy a bunch of flashcards for her, or I could just use AI to write code for me to generate and print out a bunch of flashcards. And so I did the latter. This is a very low economic value to AI-assisted coding, but I could get that done very quickly.

Do you think vibe coding is an enduring market? Do you think everyone will want to code and accessibility is important, or do you think it, bluntly, just allows builders to build better and more efficiently?

I think we need all of the above. I’ve had mixed feelings about the term “vibe coding,” but nitpicking terminology aside, I think everyone should learn to code. What I’m seeing is that, for a lot of job roles that aren’t just software engineering, people who can code can get more done than people who can’t code.

For example, my marketer wanted to run a user survey once, and she wanted something for people to give live feedback. She looked in the App Store, couldn’t find anything, so she said, “You know what? I’m going to spend 2 days to code it up.” It took 2 days, but my marketer then built a little mobile app where users could swipe left or right to give feedback on some marketing messages we wanted to use to test. Because of that, we were able to run user experiments, get feedback, and it helped her do her job better as a marketer.

In contrast, a marketer who couldn’t code a little app to let people swipe around and give feedback would not have been able to do this, would not have gotten feedback, and would not have been able to move forward. Today, my best recruiters not only screen résumés by hand; they are also writing prompts to get AI to help them screen résumés. It’s been interesting.

Which is amazing. But going to your point about people not being fearful—and they are fearful—you understand that that would lead to efficiency gains, which mean headcount reductions. I’m not into this kind of fearmongering, but if you can screen a lot more with AI, I don’t need my 3 other analysts. I think there’s a small subset of jobs that, frankly, are in trouble.

Actually, here’s one thing about hype: AI is amazing. There’s a lot of stuff it can’t do. This phantom AGI someday where AI can do everything a human can do—I think we’re very far away from that. I would say no, decades away, maybe even longer.

The trick is, if AI could do 30% of a recruiter’s job—who knows, maybe 50%, although that feels a little bit high—there’s another 50% to 70% of stuff that we still need the human to do. But it’s also clear that if you use AI and someone doesn’t, that’s actually a huge difference in what you can accomplish. You’re much better off using AI. But because AI can’t do everything, there’s still plenty of work that we need humans to do for a lot of job roles.

Do you not think we have a white-collar talent pipeline problem, though? Whether you’re a consultant or a legal associate in the junior ranks, a lot of what you can do is being replaced by AI, and they are actually cutting juniors. You’re seeing this across the board. The fear is we’re going to have this talent hole where, in 10 years’ time, there are no juniors to go up into senior roles because we’ve replaced them.

I don’t think it’s as dire as that. I think there is a big problem, but I don’t think it’s exactly that problem. So let me tell you what I’m seeing in software engineering.

The most productive engineers I know are usually not fresh college grads. They are people who have 10 or 20 years of experience or whatever, are really on top of AI, know the AI tools, and understand AI. Those experienced people who are on top of AI move faster than anything the world has seen, even 1 or 2 years ago.

One tier down are fresh college grads who are really on top of AI. I’ve hired quite a few people—fresh college grads who, for whatever reason, through the social network community, really learned the AI tools—and they move really fast, but they’re not as good as experienced people who know AI. One tier down from the fresh college grads are people with 10 years of coding experience who had a comfortable job and, for whatever reason, are still coding like it’s 2022, before ChatGPT. Those people I just don’t hire anymore.

There are people who had a comfortable job, kept coding the old way, and simply did not learn AI. I think those people may get into trouble at some point. There’s one other tier that’s in trouble: fresh college grads who don’t know AI.

One unfortunate thing is that university curricula are slow to change. I actually feel pretty bad that even today there are universities graduating CS undergrads who have not made a single call to a single API on the internet. Imagine graduating a CS undergrad who has never heard of cloud computing. It’s like, “What is a cloud? Oh, I just need to run things…” That’s weird. You just can’t be a CS major and not know how to do things on the cloud.

I’m getting to a point where I don’t think it’s right. I feel like we’ve got to not train CS majors without also making sure they know how to use AI to help them with coding and without also making sure they know the AI building blocks. University curricula are slow to change, and that’s a cohort of students entering the job market that’s really struggling. But the fresh college grads who know AI—we can’t find enough of them. So many businesses love to hire those fresh college grads.

I just want to touch on the 10x, 100x engineers that you said are just amazing. We’re seeing pay packets and compensation bands larger than they’ve ever been—$3.5 billion in certain cases for a single engineer. Are these justified pay packages, given the impact they are having on companies’ enterprise value, or are these bubble-like pay packages that we should be concerned about?

I don’t know. It is really hard to say. I know a number of people who have gotten really huge pay packets. I’m actually very happy for them. I think it’s great, the funding going into paying AI people really well.

I mean this nicely: do you think it’s like $100 million for an engineer? I worry that you’re just not going to be as productive. If I give you $100 million overnight, God, you might buy a nice house and go on holiday, and you lose a bit of efficiency.

I don’t know. I have a lot of Silicon Valley friends who, for whatever reason, have made a little bit of money. Many of them just keep working really, really hard, both before and after they wound up making a little bit of money. I find that a lot of the tech culture is that we do stuff because it’s fun, because it lets us hopefully help other people, and it’s a way to change the world. I find that wealth makes people become lazy much less than one might guess.

I’m intrigued to see how you think about this. You said all the different ways that it could impact many different verticals, and you said we overhype doomsday scenarios and everything in between. Andrej Karpathy recently said AGI will just blend into 2% GDP growth, which I thought sounded a little bit unexciting, to be honest, Andrew. I wanted some seismic shift in productivity increase. Do you think blending into 2% GDP growth is what you expect, or do you expect a much more significant 5% or 6%, like Masa at SoftBank expects?

I hope we can get much closer to 5%, 6%, or more in GDP growth. When looking toward the future, it turns out one of the most expensive things in today’s world is intelligence. This is why it’s so expensive, at least in the US, to hire a highly skilled doctor to advise us on a medical condition or hire a highly skilled tutor to patiently teach our kids, because that intelligence—training up that wise doctor, wise teacher, or wise adviser—is very expensive.

But with AI, we finally have a path to make intelligence cheap. In the future, if everyone can be assisted by an army of smart, well-informed staff on all of these topics under the sun that currently only the relatively wealthy in society can afford to hire people for, then individuals will be so much more empowered and able to get so much more done. That highly empowered individual’s life will be so different, and GDP growth will be massive.

I totally get that and agree, especially when it comes to the democratization of knowledge and the benefits that come from it.

You used the word “open” earlier when talking about the open-weights ecosystem we’ve seen. We’ve seen a reversion to a closed world in a lot of cases. How do you feel about that reversion to closed models, and how do you analyze the state of play today between open and closed?

It’s still very dynamic. For a lot of American companies, the leading frontier model is often kept closed, and then the one-tier-down model—not quite as good—is released as open. I think it’s much better than nothing. I’m actually grateful to all the teams that are releasing open-source, open-weight models.

The other dynamic is that China, especially, has been taking the lead—or getting up there—in terms of releasing tons of really good open-weight models. I would say this is not what I would have predicted a decade ago: that China’s AI industry would end up being more open than America’s AI industry.

Why do you think China wants an open AI world?

It turns out that openness is great for a country’s development. When a team releases open-source software, the circulation of knowledge is much faster through the nearby community. What I see is that when a team in China releases an open-weight model, yes, of course, Americans can take advantage of it. But the Chinese economy benefits even more from it because once something is open, it’s easier for teams to call each other and say, “Hey, buddy, how does this really work? I’m having trouble with this part of the model.” That circulation of knowledge is really valuable for innovation.

When the U.S. has more closed models, and when teams are trying to pay these $100 million salaries to extract talent, that circulation of knowledge becomes very slow, and it slows down the rate of American and European innovation.

With the commoditization of the model layer and the opening of it, it actually increases the premium on manufacturing and the ability to manufacture at scale, which China has a much greater ability to do than the U.S. Do you not think that actually leads a lot of their thinking around why they want to reduce the strength of U.S. models, in addition to increasing innovation and the circulation of knowledge, which open-weight models help with?

I think open-weight models are a tremendous source of geopolitical influence. For example, if someday some kid in a developing nation asks a question about a politically sensitive topic, or asks, “Hey, what are the national borders in this case?” or “What is the history of this event or that event?” the country of origin of the model they end up using will be delivering some answer. Whether that answer is skewed toward one nation’s values or another nation’s values is actually a tremendous source of influence and soft power.

Like it or not, open-weight models are a key part of the AI supply chain. China releasing free, low-cost models into that key part of the supply chain means it is really starting to build up a lead and a commanding user base. That, too, will be a source of influence.

This is why I think nations with a strong media and entertainment industry have disproportionate influence. South Korea has vastly disproportionate influence because of its leading entertainment industry. People listen to whatever K-pop or whatever, and that buys the nation a lot of influence. Hollywood was a tremendous source of soft power for America. It paints a certain vision of the American dream and talks about the values of freedom and democracy. I think this is another frontier of communications and soft power.

You have the most fascinating perspective, having obviously spent many years at Google and then, obviously, Baidu as well. Having been on both sides of the table in certain respects, we have this strange binary polarization of the AI race between China and the U.S. Do you agree with that positioning of China versus the U.S. in an AI race?

I think there’s a lot of room for cooperation and also some places that will be competitive. First, while people—sometimes even me—talk about the AI race, there’s no single finish line. It’s not one race. AI is a general-purpose technology, and you could be better or worse at coding, better or worse at answering questions, better or worse at helping with marketing and finance, and so on.

AI has many different capabilities, and there’s no one finish line. I think that in any one capability, we’re going to keep improving for a long time. I feel like AGI has been hyped up as if it were a finish line, but I don’t think it’s a finish line. It’s just continually improving capabilities for decades to come.

Having said that, nations with stronger AI capabilities are going to be more powerful. Citizens will be more prosperous, and economies will grow faster. To the extent that different nations’ incentives are not aligned, nations with more powerful AI capabilities will be able to do more.

If a country has a fantastic electricity grid and another country has power outages, one country can use the electricity grid to do more manufacturing, more industrial work, and just do a lot more that way.

Do you not think we still underestimate China’s ability? I think we definitely do in Europe, but I think in the U.S., respectfully, I see a lot of U.S. arrogance around its positioning. Then you go to China—and you’ve been to China and spent huge amounts of time there—and you realize the speed and the intensity with which they move. It’s a different level from both Europe and the U.S.

Yeah. To be fair, I think the U.S., Europe, and China all have problems as well. Having said that, I think the work ethic and the velocity, when China’s government makes an all-nation commitment or an industrial commitment, are actually a very powerful force.

There are state-level investments in semiconductors and in the education system, with K-12 kids being trained to use AI. Businesses also use AI, share knowledge, and sometimes build this stuff and sell it internationally. That whole-of-economy, whole-of-country effort is actually a very powerful force, along with control over rare-earth elements. I wouldn’t underestimate it.

Given that we shouldn’t underestimate it, do you think it’s right that we have export controls on chips? Obviously, NVIDIA has had a lot of export controls back and forth. Do you think that’s right or not?

I think the export controls on chips have largely backfired. The way the U.S. first put restrictions on Huawei, and then later on exports of NVIDIA, AMD, and other semiconductors, really incentivized China.

Before the export controls, semiconductor development in China was not—frankly, it wasn’t moving that fast. It was a nascent area, and there was some investment. But when America did that, China really accelerated its semiconductor development. America incentivized China to do this, and it is paying off for China.

A number of Chinese companies are building offerings where the individual chips are less powerful but perhaps much larger in number, trying to build offerings competitive with certainly the last generation of NVIDIA, and maybe increasingly the current generation. If I were to analyze this purely from the perspective of U.S. national self-interest, I think that caused China to accelerate its semiconductor industry in a way that may not be helpful to the U.S. long term.

I sit in Europe, and you obviously live in London. You told me you were born in London. My question to you is: It transparently feels like we’re very far behind, and people say we’ve already lost. How do you feel about Europe’s position in a very new world, and what can Europe do to regain some semblance of equality between the U.S. and China?

If I had one wish for the European regulators—I’ve spoken with quite a few European regulators, and I was hearing things like, “We want to be leaders in regulating AI, and that’s a competitive advantage.” With all due respect, that’s not a competitive advantage.

My one wish for Europe is to stop regulating so much and just focus on investing and building. It’s so early in the days of AI. It’s still early in the game, and Europe has plenty of smart people. Let people work hard. Don’t force them not to work hard. Let people who want to work hard work hard, and stop overregulating. Just go and invest and build stuff.

Where do we most need to be investing where we’re not investing enough?

There’s tons of capital going into data centers and infrastructure. We can debate whether there’s a bubble or not. We definitely need a lot of investment. Are we getting to the point where people are using such esoteric financial instruments to find cash for it that there’ll be a bubble? We could debate that. We definitely need a lot of investment, but when does it become overinvestment? That’s an interesting question.

The other place that I think we need to invest in a lot is not just the infrastructure, data center, and foundation-model layers, but the application layer. Because others have spent billions of dollars training these AI models, we can now access them for hundreds or thousands of dollars, or even tens of dollars. It’s wonderful to build tons of applications that just weren’t possible before.

From a VC investment perspective, what I’ve heard from multiple VCs is that, bizarrely, the cost of trying something out is so low that there are fewer ideas.

It’s not quite clear where to put massive amounts of capital to work at the application layer. In fact, if you look at a lot of the application-layer investments, sometimes it feels like firms are putting in $100 million so that they can pay OpenAI and Anthropic, so OpenAI and Anthropic can pay NVIDIA, which is where all the money is ending up.

Having said that, there are so many valuable bets to be placed at the application layer to build things. But the dilemma is that you could do it in a very capital-efficient way. If someone wants to say, “I want to put $10 billion to work,” yes, you can build $10 billion worth of data centers. We know how to spend that money, but how do you spend $10 billion building applications?

The problem is almost that it only costs me $1 million to try an idea. So how do I spend $10 billion? It’s kind of a problem and also not a problem, but I think we should invest in it.

Well, does it? Because when you look at AI margins, what are the margins for AI application-layer companies? They’re terrible. They make no money. They cost a lot of money to build because you have large engineering teams that build them. They don’t cost less; they cost more, not less.

I think it still varies. I’m seeing a lot of green shoots of software applications that were not that expensive to build, if your LLM token usage is not the majority of your expense. If you look at Replit or Lovable, 80% of their pass-through is to Anthropic.

Mm-hmm.

Yeah. The dynamic that I’m excited about is that, as LLM token costs continue to come down, we’ll see how the economics change. Right now, tokens are just expensive, but hopefully that will change, and the value created is really large.

I remember an earlier era, in the early days of food delivery. I saw this in both the US and China. There was a lot of VC-subsidized eating. It was great; we could eat food that was delivered, and it was basically VC-subsidized.

I think we’re seeing that right now with a lot of VC-subsidized AI computing. The laws of physics, or the laws of finance, say that at some point this can’t go on forever. But where it settles down will be, I think, that there will be some very valuable businesses that are not perpetually VC-subsidized.

Navigating this crazy VC-subsidy world to get to a good outcome takes a lot of skill. Having said that, I still want to say that a lot of smaller applications that are not yet doing hundreds of millions of dollars—maybe they’re doing millions or tens of millions of dollars in revenue—haven’t been that expensive to build and operate. I think we’ll continue to see them grow.

Speaking of the smaller niches, so to speak, that continue to grow, how do you think about the question of—you mentioned earlier, brilliantly, that articulation of horizontals and then the verticals beneath them, with Google and now OpenAI being the horizontals. How do you think about a world of large, monolithic models versus much smaller, much more efficient, much more specialized models? How do you think about that, and has your mindset changed around which will be more dominant?

I think it’s clear it’ll be all of the above. We’ll have large models, midsize models, and tiny models. The reason I’m confident about that is because the nature of intelligence is diverse.

Sometimes we do intellectually easy tasks. If someone asked me—yesterday, my daughter misspelled the word “butterfly,” so I needed to tell her how to spell “butterfly.” That’s an easy intellectual task. Sometimes I’m sitting down thinking for hours about a complex technical problem, and that’s really hard.

Intelligence has a range of things we want to do, and the set of things we want AI to do has a huge range. If you want AI to do basic grammar checking and spell-checking, you don’t need a trillion-parameter model. Just a tiny model, maybe running locally, can do that. But if you want it to do complex reasoning to write a piece of code, then yes, having a powerful model is going to do better.

I’m very confident we’ll end up with a huge range of models, small and large, to do the huge range of tasks—just like we have humans do a range of tasks of different difficulty. It’s the same with AI.

Does that mean that you disagree with Andrej Karpathy when he said that useful agents—importantly, useful agents—are a decade away?

I disagree with that. I think we’re seeing useful agentic workflows right now. The AI Fund team has built so many agentic workflows for so many tasks where we simply could not have done the tasks without agentic workflows.

Can you give me an example? I’m fascinated.

Over a year ago, after one of the Biden–Trump debates, we thought that tariff compliance might become an issue. Unfortunately, we turned out to be right.

Last year, I think it was around August, we started exploring the idea of building technology to help with tariff compliance. By the way, I don’t know if you’ve seen these tariff-compliance documents, but frankly, when I look at what it takes to file this paperwork, it makes me think, “Oh my God, what is this?”

You say, “Import a bicycle,” and then you look at the bicycle’s specifications: how much it costs, the size of the wheels, and all these rules and regulations for importing a bicycle. It makes me think, “Oh my God, are humans really doing this?”

So we built agentic workflows to carefully read tariff-compliance documents, carefully get the specifications for what someone wants to import, and try to match them and make suggestions. This is now one of our portfolio companies, Giga Dynamics. Because of the increased complexity in tariff compliance, it has been doing pretty well.

I find that we simply could not have done this without agentic workflows. With medical assistance, we have different AI Fund portfolio companies, including a medical AI assistant operating in India and an AI assistant, Katus, helping process legal documents. Many of these workflows we simply could not have done otherwise.

The large businesses are doing this, too—not just our startups. When we look at the hyperscalers and I talk to friends at large businesses, there are a bunch of internal workflows that we simply could not be doing without these AI agents.

When we think about the core of a business, it’s margins, and most of these businesses don’t have margins. Do you care about margins when investing today? With absolute respect—and this sounds disrespectful—do you take the utopian view that it will correct itself with time and efficiency gains?

At some point, the laws of physics—or the laws of finance, or something—mean that margins do matter. But one of the tricky things about AI is that we know the technology is going to change. We don’t build assuming the technology will be stagnant; we build assuming the technology will evolve.

One obvious example is that token prices have been rapidly falling. Depending on who you believe, they’re falling 80% a year, or whatever. Frankly, when we build prototypes, we routinely don’t worry about token costs, because the first and most important thing is, “Let’s build a product that users love.”

What we find is that, after we’ve built something and users start to use it, our API bill starts climbing. Then, every few weeks, you look at it and go, “Wow, this is getting really expensive. This is costing me the salary of 1 engineer. It costs me more than 2 engineers; it costs me more than a whole bunch of engineers.”

But fortunately, whenever that has happened so far, we’ve been able to use techniques to bend the cost curve back down even faster than the rate at which token prices are falling in the market.

I find that absolute margins are important, but when you have a view for where the technology is going, it lets you not build for the margins today, but for what you can forecast them to be in the future. I think that’s an important distinction.

But we don’t take a blind, utopian, “AGI, blah blah blah” view either. I think that’s also overly simplistic.

How do you think about defensibility in an AI world? A lot of people suggest that the time to copy is reduced significantly and that defensibility itself is being questioned in AI. Do you agree with that questioning of defensibility today, or not?

Moats are changing. I find that moats tend to be a function of the industry rather than a function of the technology. AI as a technology doesn’t really offer an answer to the moat question for most businesses. If you’re building AI for drones, legal, or whatever, the moat is more a function of that industry.

One thing that is changing with regard to moats is that software used to be a moat. If you had invested 10 years in building software, it was really hard to replicate. That moat is much weaker than before.

But other moats remain. Are you trying to use AI to accelerate the building of a two-sided marketplace, which can be very defensible? Are you building for consumers or enterprises? Are there brand and reputational effects that can help you build defensibility there?

I find that the software moat has changed, but other moats tend to be analyzed based on the industry.

Okay. So software moats have changed. Fantastic. We now have margins that matter, but we have a little bit more elasticity there. The software moat has changed in terms of the ability to stay relevant for large enterprises. What are the single biggest barriers preventing large enterprises from implementing AI aggressively and preventing themselves from becoming extinct?

I think the biggest barrier in most large enterprises is actually people and change management.

Not data.

It's not data. I think it's definitely not data. Not that data isn't important, but it's definitely not the bottleneck. I think data has been overhyped. The interesting thing about AI hype is that there's almost always a gem of truth in the hype; it's just been hyped up 10 times more than the reality.

Maybe let me give one example, then I'll come back to data. There's been this buzz about, “With AI, we'll have unicorns with 1 employee.” It's fine if you want to build a unicorn startup—a billion-dollar company—with 1 employee. It's a good thing to do, but frankly, if you're at a $1 billion valuation, you could afford to pay 2 employees or even 10. So why do you need to hype it all the way up to say, “Let's do this with just 1 employee”?

It is true that team sizes are shrinking and that people can get more done with smaller teams. That is true, but the hype is then saying, “Let's build a unicorn with 1 employee.” I find a lot of AI hype so hard to disentangle because there's a gem of truth in it; it's just been hyped up a lot more.

On data, data is important, but it turns out that data is very verticalized, and you don't need as much of it to get started as you think. For example, Landing AI does a lot of work with financial institutions and healthcare. A lot of financial institutions have plenty of transaction data. Take the PDF file, turn it into LLM-ready Markdown text, process that, and find value in it.

For example, we could take SEC filings and large, complex financial tables and very accurately turn those financial tables into Excel spreadsheets. Then you can have your analysts or your AI analyze that and draw conclusions. So often, with a bit of scrappiness, you can look at internal data and public data and get some things going.

It turns out that a lot of internet data is general-purpose data. Most of the world's data is actually private, and a lot of business data is valuable transaction data: sales data, product data, manufacturing data, logistics data, and all of that. With a scrappy team that knows how to use it, you can actually start to build something and get value out of it. Not to say more data wouldn't be even better, but you're not stuck from taking the first few steps because of a lack of data.

Andrew, I speak to many CEOs of businesses this size, and they say, “Harry, are you kidding me? You think we can get security and permissioning for our data and our enterprise? No. We don't have Slack. We don't have Notion. Everything is custom-built.” You're seeing the likes of JPMorgan and Goldman Sachs absolutely refuse any ChatGPT use and build internal systems. Is that the world that we inhabit for enterprise AI adoption?

I think we'll get there. I find that a lot of enterprises are adopting LLMs, RAG, and many others. Today, there are still businesses that are on-premises rather than on the cloud.

But we're making progress, and it'll take time. One thing about this AI hype—that we'll have AGI in 2 years or whatever—I think that's just ridiculous. For most reasonable definitions of AGI, that's just not going to happen. Just as we're now well into the cloud era but still have an awful lot of on-premises workloads, I think AI adoption will be wonderful and there will be tremendous GDP growth, but it's also going to take much longer than the hype says it will.

I actually think that a decade from now, we will still be working to identify valuable applications in enterprises and building them. Having said that, we will make a lot of progress over the next 1 or 2 years, but we're not going to be done even 10 years from now.

What else does everyone think they know about AI and its adoption and implementation that they get wrong?

Even earlier this year, we saw some senior business leaders advise people not to learn to code on the grounds that AI will automate it. We'll look back on that as some of the worst career advice ever given. As coding becomes easier with AI assisting us, a lot more people should learn to code, not fewer.

I'm already seeing it. I mentioned the marketplace example just now, with building an app for feedback and swiping. For a lot of job functions, people who know how to tell a computer exactly what they want it to do, so the computer can do it for them, will just be more powerful. For the foreseeable future, the language of precisely telling computers what you want them to do is coding.

That doesn't mean you should write code by hand. Writing code by hand is becoming obsolete. Really, don't do that. But get AI to write code for you, and people who can do that will be more effective, more powerful, and have more fun.

If we're that early, where in a decade's time we're still going to be looking for and identifying areas where AI can improve meaningfully, do we have enough money to fund both the energy and the compute requirements for that 10-year period? Sam Altman has said he needs $1 trillion. He needs the energy of Japan. If we're 10 years out before we have still not that much improvement, do we have the money to fund it?

I think we'll see plenty of improvement over the next 2 years, but I think we still won't be done getting even more improvements 10 years from now. One place where this is super promising is AI-assisted coding. We're seeing real productivity gains and real returns. It's really changing the way software is written, and it's been fantastic.

Frankly, for so many of my friends, coding is so much more fun with AI to help us out with it all. So we are seeing returns, just to be clear, but we still won't be done growing this 10 years from now.

But if you look at the TAM, the secret to success in AI investing is whether we'll see a transition from human labor budgets to software budgets. If we have that, then that's the holy grail. You and I will make a lot of money with our funds, and it's fantastic news because the TAMs have massively increased, or the spend has massively increased. If we're not actually going to lose any people, then we don't see that transition from human labor budgets to software. Do you think we won't see that transition?

To me, the question is: is AI mostly for cost savings, or is it for growth? I know that it's difficult to change workflows, and a lot of companies tend to think about cost savings. But here's the problem. There's actually one pattern I see. Let's say I have a workflow that has 5 steps, and let's say each step takes 20% of my effort. Maybe I'm doing underwriting approvals: do I approve this loan or not?

For simplicity, let's say there are 5 steps, each taking 20% of my effort. If you can automate 1 of those steps, that's a 20% cost saving, which is really nice. It could be great if you're a low-margin business, but it doesn't feel like a game changer.

What I find is that the more valuable uses of AI often require rethinking that workflow. The pattern I see is that instead of taking a 20% cost saving, which you could do—there's nothing wrong with that—the 2 patterns for getting to growth are either doing more or doing it faster.

In the case of underwriting and making loans, instead of saving 20% of my human labor, I can rework the workflow to turn around my decision-making time. Instead of someone needing to wait 2 weeks before a loan officer looks at it, we can just give you an initial answer in 10 minutes. That changes the product and lets you drive growth.

There's also the “more” pattern. There are a lot of businesses that could offer high-touch customer service only for expensive, high-end clients. But if you can now serve a much larger group of people—or, let's say, deliver high-touch financial advice to a small group of people—if you can now deliver that quality of service to a lot more people, then that again changes the product and lets you drive growth.

Instead of cost savings, if AI lets you do something way faster or lets you take a task and do it 1,000 times more, instead of serving a small number of people, you can serve a lot more people because it's now economic to do so. These are the 2 patterns I've seen to drive value increases, and I think that will be important for unlocking a lot of this GDP growth.

You said “economic to do so.” Do you think it's crucial that we see vertical ownership, in terms of seeing NVIDIA own models as well as the chip layer? We're seeing Facebook build out data centers more than anyone, and we're seeing everyone build out data centers. Is it important that we own every layer of the stack, or will we see individual participants own horizontal layers of the stack?

I think it does evolve over time. I'm going to make an analogy. In the early days of the computing industry, it was the vertical players that won because, if you wanted to connect the keyboard to your computer motherboard, which is the CPU, was it okay if your keyboard had plus or minus 5 volts and your CPU had some other voltage? Was that okay or not?

We didn't know where the API boundaries were. If your CPU had your memory laid out a certain way, your compute and your math accelerator needed to interoperate with each other.

Before we wound up having a clear conception of where to draw lines and where the API boundaries should be, the integrated players—IBM back in the day—could solve all the problems and build valuable, working products. But as the industry matured, we started to have standards. For example, now we have a USB standard. Before that, there were other standards.

Now you make a computer, someone else makes a keyboard, we plug them together, and it all works. When an industry is immature, it turns out that where to draw those boundaries to let different participants do their part and still interoperate is less clear. But as the industry matures and there are more standards—for example, if I want to publish a compressed LLM model on the internet, what's the file format for that?—then that makes it easier for individual players to do something and still have it fit into the broader ecosystem.

So do you think Zuck and Sam are right to be spending as much as they are on data centers, or should they be patient and wait for the maturation of the industry, where they can then be horizontal?

I think clearly OpenAI's investments have paid off to date. It is possible to overinvest at some point, but I don't know if that is the point. I think, also, that the financial instruments being used by many players to shift risk around have been really interesting. I find that overly complex use of financial instruments to shift risk sometimes increases the risk of there being a bubble at some point. That's something to watch out for.

Do you worry about the circular deals?

It's something to keep an eye on. I'm not alarmed by them, but I think things could be more frothy or less frothy. Things could be more of a bubble or less of a bubble, and these are signs of things feeling a little bit more bubble-ish.

When does a sign turn into a big concern for you with these?

I think you mentioned the Sequoia article on the $600 billion problem of AI. I am concerned about that, but it's interesting: my concern for different layers of the stack is different. What I'm seeing is that, for the application layer, there is very clear ROI. I think it's fantastic. Someone else trained these models, and you can build applications for $100,000 or $1 million and now generate ROI.

I think calibrating to the right level of infrastructure investment is tricky. But having said that, it is also, at the same time, very clear that we do need more electricity, more data centers, and more semiconductors. That, too, is very clear, so we should be investing a lot, and I'm glad we are. But what exactly is the right amount to invest? I think that's the tricky question. It should be a lot, though.

Do you get annoyed by the bubble discussion?

I don't get annoyed by the bubble discussion. I do get annoyed by the hype. I don't like it when regulators call me up and say, “Hey, we heard AI could lead to human extinction.” Thankfully, there's much less of that now than there was a couple of years ago. Instead, the conversation should be, “How can we upskill the workforce? Where can we invest?” Not, “How do we slow this thing down?”

I think the hype has really distorted public perception of AI. One downside to the hype, too, is that without public support of AI, things slow down. One of my friends works a lot with high school students, and he told me that he was talking to a high school student about maybe pursuing a career in AI. She said, “You know what? I heard AI could have something to do with human extinction. I don't want to have anything to do with that.”

This hype turned a high school girl away from working on AI at a time when it would be so promising for her to leap into AI. I think this really causes people to make weird decisions, both at the individual student level and at the community level. When a community shuts down the building of a data center, even though data centers could be good for the community and good for the world, I think that's also unfortunate.

I'd love to move to a quick-fire round where I say a short statement, but staying on that thread, the first question is: what's your biggest advice to educational institutions to make sure they equip students for a generation of AI?

Embrace it. Update curricula. Teach them as much AI as possible. Students are going to live in a world where they will be using AI and having it help them. We've got to teach students to do that. I think it'll be different for different fields, but one thing that is clear is: get all your students to learn to code.

What's one thing you've changed your mind about AI in the last 12 to 18 months?

I think my favorite tools keep changing. If you ask me every 3 months over the last year what my favorite coding tool is, my answer would have kept changing.

Do you think Anthropic will beat OpenAI in the coding wars?

Really hard to say. OpenAI has a very strong consumer brand, and that's very defensible. In contrast, developers are more likely to switch coding tools on a dime. I love Claude Code. I think it's fantastic, but I find myself using OpenAI Codex much more over the last month.

I think OpenAI Codex has actually gained real momentum. I'm also keeping an eye on Gemini CLI, which I think is getting better, maybe at a faster rate than people have given it credit for. In the coding-dev-tools and API-tools market, the moat is weaker than having a strong consumer brand. I think that's something that companies have to sort out.

Tell me, what was your biggest takeaway from BYD? It's such a different company from anything that we're used to in the West. What was your biggest takeaway?

I really appreciated the speed and intensity of BYD and also of the China ecosystem. I think it's really unfortunate that, in some parts of the United States, advising someone to work hard is viewed as politically incorrect or something.

In Europe, I'm chastised for it.

Oh, okay. All right, great. Hopefully the European viewers won't hate me, or hate us both, for that. Frankly, I wish people could work 4 hours a week and be wildly successful, but the practical reality is that when people work hard, they get more done.

I want to acknowledge that not everyone, at every point in their life, is in a position to work hard. The week after my kids were born, I didn't work that hard. I took time off and spent time with the kids for more than a week. We need to respect people in all walks of life, including people who, for whatever reason, are not in a position to work hard at that moment.

But if someone wants to work hard and, to quote Steve Jobs, “make a dent in the universe,” let's empower them and celebrate that. If someone, for whatever situation, can't work hard, let's also respect that and maybe celebrate that. But I think this is a moment in time when there's so much stuff we could build. People who work hard to learn a lot and build things will accomplish a lot.

Did you do 996?

The term 996 wasn't an explicit term that I used. These days, I just really love what I do. It really doesn't feel like work, but I work on a lot of my weekends. I'm sitting in a coffee shop coding away because it's the most fun thing I could do on a Saturday. I don't bother to keep track of my hours. It's probably a lot.

What's the hardest transition element moving from operator to investor?

One thing about AI Fund: yes, we call ourselves a fund, but frankly, the way we run the fund day to day, we act much more like operators than investors. AI Fund is a venture studio, and I believe our skill set is actually in building, not just in capital asset allocation or whatever.

We work really hard to screen ideas. We talk to customers; I'm sometimes on customer calls myself. Then we bring in founders to work alongside us. We're reviewing the product, giving feedback on the product, and arguing about pricing. My day-to-day life is much more like an operator's. Yes, eventually we have to do the financial diligence, I write a check, and we do follow-ons. We do all that, but there's a lot more to it.

I'm really sorry, Andrew. Then are you a fund or are you an incubator?

We call ourselves a venture studio or a venture builder. We don't usually call ourselves an incubator. Incubators usually bring in founders who already have an idea. We go earlier than that.

We often work with our investors and partners to come up with an idea, and only after we have an idea do we go and try to find the best founder to co-build and co-found the company with us. So we don't call ourselves an incubator.

How much ownership do you have, then, when you make those original investments and seed the company?

It depends. We end up with some common stock for the sweat equity of building the company, and our first check is usually around $1 million at a $4 million cap—so, kind of 20% ownership on a SAFE.

And so we're basically getting 20% to 25% ownership on entry, with some common stock.

Yes, plus some common stock for the sweat equity.

Totally get you. What do you think is the biggest—

But to me, the reason we do this is because, while there are VCs that do the competitive deal-flow thing and make a lot of money that way, I think my team's biggest contribution is not fighting over hot deals. It is finding ideas and creating companies that would not exist but for the fact that we and a founder got together to co-found them.

So I think we just create more value in the world by creating new companies, rather than only discovering hot companies to try to put money into.

What concerns you most today, Andrew? I love your optimism and your open-mindedness. What concerns you on the flip side?

The difficulty of bringing everyone along with us. In previous ways of economic disruption, like when our nations went from mainly agriculture to non-agriculture, someone who was a farmer could keep farming until they retired, but their kids had to learn a different trade, maybe move to the city or whatever.

The change is so fast this time around that we need people who are alive today to learn new skills, as opposed to needing their kids to learn new skills. That’s actually very challenging. Historically, I don’t think we’ve ever been good at that.

You do a lot of interviews, Andrew. You speak to many journalists. I’m not a journalist; I’ve never actually had a job. Do you find the quality of interviewers who ask you questions to be good?

I think media has an important role to play in curating and disseminating knowledge. I think the quality of the questions that reporters are asking has been very clearly trending up over time. But there is still the hype element that keeps distorting the information ecosystem.

Unfortunately, there are financial incentives, regulatory capture, legislative benefits, and other types of incentives that drive certain types of hype. That’s actually one pattern that I’ve seen. I won’t name any companies, but I find that companies with something to lose have become more moderate in their statements over time. I find that, as an established company, you just say more sensible things.

But there are some companies that I think are at greater existential risk, and I find some of those companies, which I don’t want to name, to be the worst sources of hype because they’ve got less to lose. They’re just saying a bunch of random stuff in many respects. It’s a lashing out in desperation.

When you look at Demis Hassabis, obviously a brilliant leader, or Sam Altman or Dario Amodei, all of them, I think, have moderated their positions significantly with the maturing of their companies.

No comment on individuals, but I think that when you have something to lose, you say more sensible things. When your company faces greater existential risk, sometimes people say weird things for fundraising.

I like to finish on a tone of optimism. What single thing are you most excited about when you look forward to the next decade? For me, for example, my mother has MS. I think we’ll have incredible medical discoveries in diseases where we haven’t really made much advancement in years. That excites me. What excites you?

I’m sorry to hear about your mother. What excites me is that I want to empower everyone to build AI. I think the distance between having an idea and building it is now much shorter, and we need not just software engineers to be creating.

In the future, I hope that a lot of people, instead of saying, “Is there an app for that?” will say, “I built an app for that.” Instead of just being software users, they’ll be software creators. When we get there, people all around the world will be much more empowered, get more done, and have more fun.

Andrew, this has been such a joy to do. Thank you so much for putting up with my prying and my pressing. You’ve been amazing, and I really appreciate the time.

I really enjoy your show, so it’s a privilege to be here, Harry.