DeepSeek、Stargate与AI的6000亿美元问题：红杉资本 David Cahn 对谈

You're listening to Gradient Dissent, a show about making machine learning work in the real world, and I'm your host Lukas Biewald. This is an interview with David Cahn, who is currently a partner in Sequoia, where he invests in AI. He was an early investor in Weights & Biases when he was at Coatue, where he also invested in Hugging Face, Runway, and many other well-known AI companies. David always has an interesting perspective on the industry. In the first part of this interview, we talk about his famous article about AI's $600 billion question, and also get into topics around his religion, how that affects his perspective on AI, and how he thinks about his career. After we did the interview a week ago, two big pieces of news hit around Stargate and DeepSeek's much cheaper AI model, so we wanted to go back to him and get his hot take on these new topics. We also cover those at the beginning of this interview. This is a super fun conversation, and I really hope you enjoy it.

1. Discussion on recent news: Deepseek and Stargate updates

We actually recorded the episode last week, but then there was so much news to talk about that we didn't want to release it without getting some perspective on the new things that have happened. Maybe starting with the DeepSeek launch, David, you should describe it first, but then I'd love to hear any quick takeaways that you have from that.

It is wild how much changes in AI in 1 week. I guess every day right now in AI seems to be pretty dramatic.

As a reminder, DeepSeek is this Chinese AI model that was launched late last year, and it was trained by a hedge fund in China. The news this week is that the market is really focused on these small models. DeepSeek is a smaller model that is a distilled model, which means that it's trained on some of the outputs from existing models.

2. Deepseek: The rise of smaller, cheaper AI models

The thing that everyone is focused on right now is just how cheap this model was to train, and the idea that it may be a signal that AI is now commoditizing, or that these foundation models are now commoditizing. That's something people had been talking about for the last 6 to 12 months, and this was just clear evidence, or a clear data point, that this is starting to happen.

Do you think that there's a major breakthrough here? What do you think is going on that made this possible?

I don't think it's incredibly surprising. I do think it's an update in people's mental models about how the AI space is working right now.

You had this moment in mid-2023 when GPT-4 came out, and in my blog post about the end of 2024 and predicting 2025, something that I talked about was this idea of everyone getting to parity with GPT-4. The big event of 2024 was Google, Meta, and other companies, including xAI, getting to parity with GPT-4 in terms of the quality of their models.

What we've all learned in the last week is that China has now caught up. I think this was a trend line that the industry and the category were on: everybody was catching up to GPT-4.

The bigger question in my mind—and even more important to me than the DeepSeek news—is Ilya Sutskever's talk that he gave at NeurIPS in Q4 of last year. I've been telling people, "Watch that 3 times if you want to understand what's going on right now."

Ilya Sutskever is the guy who basically invented modern deep learning. He invented the scaling-law race that we're now on. There are a few people, of course, like Dario Amodei, who trained GPT-3, who are a big part of that, but Ilya is probably the scientist who is most often credited with this and is most known for it.

3. The evolving role of scaling laws and their implications

Basically, what Ilya said was that pretraining is dead, or that scaling laws are dead. That sound bite has been going around for the last 6 months, but I think listening to the full talk and considering the ramifications of it really begs the question: We knew people were going to catch up on the existing models. That's always how technology works. You come out with a breakthrough, and people catch up. The question is, where is the next breakthrough going to come from?

I think DeepSeek may be overhyped, and what Ilya said may be underhyped or underfocused on.

As an entrepreneur in the space, I've had the experience for 15 years of people continuing to find ways to make this stuff cheaper. My investors kind of freak out and say, "Is this commoditizing?" Then it seems like people always end up using those cost savings to spend more in aggregate.

I'm actually a little surprised that the stock market is down so much today. What's your perspective?

It's funny. In my first "AI's $600 Billion Question" post, I basically said that when this stuff gets cheaper—and it's inevitable that it's going to get cheaper—this is really good for startups.

I may have a bit of the opposite view in terms of the ramifications for a company like Weights & Biases, or the ramifications if you're the founder of an AI startup. I do find it funny that the market seems to freak out at the wrong moments, and it's because these things are lagging. People are actually reacting, in my mind, to news that was pretty clear 6 months ago.

What is really going on is that as these models get cheaper, it's great for the application layer and great for builders. If you're somebody building in AI, and before you had to pay X dollars for all these GPUs and train these giant models on these GPUs, now inference is much cheaper. I think that's really good.

This is a theme that Satya Nadella, the CEO of Microsoft, has been saying for a while. People have not been listening, but Satya has been saying for a while that these models were going to get distilled and become a lot cheaper. In his view, that's really good for Microsoft because Microsoft can host these models.

I don't think this is quite as doomsday as people would have you believe.

Should we talk about Stargate a little bit? Do you have a perspective there? Maybe you should also describe what that is.

It's wild to have 2 major updates in 1 week, and in many ways, updates in opposite directions.

4. How cloud companies are funding massive AI investments

Stargate is this effort to spend $500 billion on data centers. DeepSeek is this idea that maybe we don't actually need to spend hundreds of billions of dollars on data centers. Part of why folks are focused on this, as they should be, is that these are 2 very different directions that AI could take.

The broader market has been pricing in or thinking about the Stargate direction over the last year or so: models are just going to keep getting bigger, we're going to have bigger and bigger data centers, and those bigger data centers get us better models. Those are scaling laws.

DeepSeek, of course, is a move in the opposite direction: these things are going to get smaller and cheaper, and we're going to get this renaissance of building at the application layer, where we're going to generate value for end users.

With Stargate, everyone saw the news. It's a very big financial commitment, and it's potentially great for the AI ecosystem. You have new data centers popping up. We've known that there's going to be this big race to build data centers, and that's continuing.

To me, the new information from the Stargate launch was actually the revealed preference in Microsoft's behavior. Microsoft has had a right of first refusal on new data center building for quite a while.

What do you mean by that?

If OpenAI wants to build a data center, Azure and Microsoft have the first shot at building that data center. If there's a data center for a new model, Satya is the one who's going to build that.

Satya came out after the Stargate announcement and said, "I'm good for the $80 billion that I've committed to AI data centers, but I'm not upping my commitment." Ben Thompson, in his newsletter, had a really good analysis of this.

5. The competitive dynamics driving AI spending

We're basically moving from equity-funded data centers—you can think about Microsoft, Google, and Amazon funding these data centers using the cash streams from their existing businesses—to credit-funded data centers, or leveraged data centers. Instead of funding them off the balance sheets of these big companies, they're going to have to get funded from the cash flows that they generate.

That's a big shift in the market, and a very logical one. It's a late-market, mature-market shift. Ben Thompson compares this to the dot-com era and how that was the last leg of the dot-com push, where you move from equity-funded to credit-funded investments.

What are the ramifications of that?

I don't know. I think we're all going to find out.

Potentially, they could be very positive. If you believe that these data centers are ultimately going to yield AGI, and I think a lot of people do believe that, then potentially this is an incredible move in the right direction.

If you think that the DeepSeek direction is actually the direction the world is going to go, and these models are going to get smaller, then maybe it counters that narrative. There is a tension between these 2 stories about how the world works. Frankly, none of us know where the world is going to go, and that's what makes it so interesting.

Does this stuff affect Sequoia in real time? Do you guys get together and say, "Do we need to update our worldview today?" Or does Sequoia take more of a longer view, where you wait and see?

When you're investing in private companies, you don't have to react so quickly necessarily.

I would say we tend not to be reactive in general, and we tend to take a very long-term view. Day-to-day movements don't really matter for us, and we don't think about them. They don't affect our decision-making.

In this case, these updates were very much in line with our broader thematic thesis and just how the world is moving. Like everybody else, we were surprised by the extent of these changes and the extent of the reaction to them, but I wouldn't say it has dramatically reshaped our perspective on AI.

One of the ways I've seen you come up outside of knowing you has been this first article that you wrote, which was called "AI's $200 Billion Question," and then I think it got updated to "AI's $600 Billion Question." I imagine a lot of folks listening to this have read the article, and of course we'll post it in the show notes.

Let's not assume that someone has read it. Maybe you could lay out the case. It feels like your article might still be a little bit confusing to someone without your finance background, so try to dumb it down for the AI audience here, if you don't mind.

Happy to. To start, the basic question I was trying to answer—and this was from my own curiosity at the time—was: Where's all the revenue?

6. Breaking down “AI’s $600 Billion Question”

We're spending hundreds of billions of dollars on AI. The hyperscalers are each building out these massive data centers, and I just didn't have a great intuition for the scale of this investment or the scale that the AI ecosystem needs to reach in order for these investments to make sense.

"AI's $200 Billion Question," which became "AI's $600 Million Question," was basically some form of napkin math that I had come to or iterated on over time in terms of thinking through how we go from data center investment to the amount of revenue that's required to pay back these investments in AI.

This number is the total spend on data centers in some period, right?

Let me walk through the basic math.

In 2024, when I published the $600 billion piece over the summer, we thought that NVIDIA was going to do about $150 billion of run-rate revenue in 2024. For every dollar that you spend on GPUs, you have to spend a dollar on the data center, energy, generators, power, and all of that stuff.

Assuming that the people who are buying these GPUs are using them to create AI systems, you can apply this 2-to-1 multiplier and say, "For $150 billion being spent on GPUs, we're going to spend $300 billion on data centers."

Then the question is: We have this GPU capacity and these data centers, but now some startup somewhere, some enterprise somewhere, or some company is going to use that data center capacity and deliver a service. You're a startup using OpenAI's API; it's calling a data center somewhere in Microsoft's data centers to run that model. You want to earn a margin.

I assumed, for the sake of argument, that you, Mr. Startup, want to earn a 50% gross margin on your product. That means that for every dollar you spend on AI, you need to generate $2 of revenue to get a 50% gross margin. You get this second multiplier.

You take the $150 billion, double it to $300 billion to get the amount of investment in data centers, and then double it again to $600 billion. You say, "Okay, $600 million is the amount of money that all the AI startups combined, enterprises, and all the companies using AI combined need to generate to pay back 1 year of investment."

That's maybe 1 final important point: $600 billion isn't a total number into infinity. If in 2024 we spend $150 billion, that's $600 billion of required revenue. If in 2025 we spend another $150 billion, now it's $1.2 trillion. It goes up and up and up. It's almost this debt that we've invested in and now have to pay back over time in AI revenue.

The latest revision of the article goes back to this summer, and obviously things have been changing a lot. How would the article change if you were going to write it today?

This may surprise people, but it wouldn't change very much. I'll break it down into the 2 pieces again: the cost piece and the revenue piece.

From the cost side of the equation, there's been a remarkable stabilization in the amount that the big hyperscalers—Microsoft, Google, and Amazon—are spending on their data centers. In my most recent piece, which I published in December, I laid out the math.

Microsoft is now spending roughly $20 billion per quarter on new data centers, and that started to stabilize in Q3 and Q4 of the last calendar year. Google is spending about $13 billion per quarter on data centers, and that's also started to stabilize.

7. Stabilization of AI infrastructure costs and the revenue gap

Meta and Amazon haven't fully stabilized, but I expect them to stabilize around the same levels. You'll see Amazon stabilize in the low $20 billions, similar to Microsoft, and Meta stabilize in the low teens, similar to Google.

What that means is that AI's $600 Billion Question is not going to become a trillion-dollar question. We've reached a state at which it's no longer growing exponentially.

The second piece is the revenue piece. When I first published the piece, I said, "We need to generate $600 billion of revenue. Are we doing that? Are we succeeding at that objective?"

I tried to count as generously as I could. I said, "There are some billions of dollars of revenue from OpenAI. You have a bunch of startups that are each generating tens or hundreds of millions of dollars of revenue. Let's assume the big tech companies are generating $5 billion of revenue," which I think was a very generous assumption.

In the summer of last year, I said there was a $500 billion hole. There was a huge gap between where we needed to be and where we were. If you fast-forward to today, that's also pretty similar.

OpenAI is bigger, and Anthropic is bigger, but OpenAI is still the lion's share of the revenue that's been generated in the AI ecosystem. The big tech companies have not fully unlocked AI revenue in their businesses. We saw this week that Google announced that you have to buy its AI product through Gmail, so it's trying to figure out how to distribute this AI product.

We're actually in a very similar place to where we were in July 2024.

Where, then, is this money coming from to buy all these data centers if the revenue is not being generated? There certainly isn't enough venture capital investment to fund hundreds of billions of dollars of spending, so doesn't it have to be inside enterprises?

This is a great question. All of these pieces that I'm writing are just me thinking out loud. I asked the same question. I didn't know the answer, so I went and tried to figure it out.

I published a piece about what I call the prisoner's dilemma of AI. I basically explained that the cloud business itself is about a $500 billion business. The cloud business is the golden goose for Amazon, Google, and Microsoft, and it's a very profitable business.

8. Where the money for AI infrastructure is coming from

We now live in a world—and I think this is remarkable even in the history of business—where 7 companies represent 33% of the S&P 500: the Magnificent 7.

What's basically happened is that you have these 7 oligopolists, or monopolists, whatever you want to call them, and they all have this golden goose: their cloud business.

If you're sitting in the board meeting of Microsoft, you're telling yourself, "We don't want to fall behind Google. If we do, we're going to lose our position in this amazingly lucrative oligopoly."

You have this sense that everybody has to spend in order to compete with one another. To answer your question very directly, the dollars that are going to fund these data centers are coming out of existing profits from what is a very lucrative business that's been built over the last decade.

I see, so it's the cloud companies that are just using the profits from their existing businesses to fund most of this. Do you have a sense of what they're trying to do with these huge GPU spends?

I think they're trying to win the AI race. There is genuine optimism in these companies. I do think they really believe that there's this huge potential for AI.

At the same time, I think they're afraid. They're afraid that if they don't spend, their competitors will spend and they'll fall behind. In Q2 of last year, shortly after I published the $600 Billion Question, Mark Zuckerberg went on Bloomberg and basically said, "The opportunity is immense, but people are also spending a lot of money. We might not exactly pay it back, but we have to, because if we don't, we're going to miss this incredible opportunity."

9. Cloud companies' strategies to dominate the AI race

Then Sundar Pichai, on Google's earnings call that summer in August, basically said something very similar: "We have to spend or we're going to fall behind."

I think it's a competitive race, and everybody hopes and believes that the revenue is going to catch up. For what it's worth, I also believe that the revenue is going to catch up. I'm very optimistic about AI.

I'm not calling it out because I think it's never going to happen. It's just a very, very big number, and it's good to have an intuition around where we have to get. I think it also helps us drive, if we're building or investing in startups, toward building things that are going to be massively impactful to people in order to make these investments work.

Do you have a sense of where those areas are?

I do. I'll give you 1 example that I'm very excited about right now. In the 2025 predictions piece that I wrote, I basically said, "I think AI search is the killer app."

I've become really obsessed with Perplexity as a user. I'll use that as an example, but then I'll explain why I think this spreads further than just 1 use case.

I use Perplexity 10 to 20 times a day, and I think it's an incredible research tool. Think about search before AI: It was a navigational tool. It found where the answers were and told you to go read the answer somewhere.

10. The emergence of AI search as the next killer application

Now, with AI search, you basically have informational search. It can go through the index of the web, and it can go through the information that it learned from pretraining, synthesize that information, and just give you the answer directly. That's an order-of-magnitude improvement in terms of the capability that I'm getting.

The thing that makes me feel as though I've seen the future a little bit with Perplexity is that I think Perplexity is actually mapping itself onto my theory of mind. I'll explain what that means.

People use the term "cognitive architecture" a lot to describe how AI agents work, and I've sort of inverted it. I have a cognitive architecture, and you have a cognitive architecture. Depending on what we do and our profession, the way that our brain works is changed by that. Our patterns of thought are changed.

If you spend a lot of time being an investor, you start to think a certain way. If you spend a lot of time being an engineer, you start to think a certain way. You spend a lot of time being a lawyer, and you start to think a certain way. The same is true for a doctor.

If you have friends who are in these professions, you get a sense for some of the patterns of thought they have to have in these careers, and the way that they solve problems.

One opportunity that I'm really excited about is the proliferation of AI search engines for different professions and different types of human beings that map themselves onto our patterns of thought, or theory of mind, if you will.

What makes Perplexity so powerful for me is not just that it's doing this search. It's not just that it's using the models pretrained by OpenAI and Anthropic. What makes it so powerful is that it understands how to deliver the information in a way that makes me very productive.

I think Perplexity has incredible product-market fit with people who are essentially researchers, and as an investor, part of your job is essentially to be a researcher.

To extend that to people in the audience and other people who are not researchers, I think this will be true in a lot of professions. Harvey is doing this in the legal field. If you're a lawyer, you have certain ways of processing information and solving problems. I think you're going to use Harvey, and they're going to train it on a bunch of proprietary data, so they're going to have better insights for your profession.

You're going to use that at work maybe 10 times a day or 20 times a day. OpenEvidence is a company that's doing this for doctors. You can search and go through all the medical literature, and it understands that doctors are trying to solve a patient problem. The way they want to see the answer to that question is not how I probably want to see it when I'm doing a search.

I think everybody is going to end up having this AI search pal or AI search assistant. I think it's going to make us as humans and knowledge workers much more productive. I've seen that uplift in my productivity, and I think we're going to start to see that proliferation.

I think that's going to drive incredible productivity and revenue uplift over time.

It's interesting because with Harvey, I feel like the important part I would have pointed to is that it has proprietary data sets that you want to look at. I also like Perplexity, and I use it for a lot of personal research, maybe a little less professionally than you.

It's an interesting point of view. It makes me think that maybe Perplexity is going to be even more successful than people say. Don't you think most jobs are—if you're going to define research this broadly—primarily research?

I don't know. I know that my job has a lot of research, and I agree with you that a lot of white-collar jobs are essentially research jobs.

I guess the way I think about it is that there are 4 dimensions on which you can differentiate these AI products. The first, and I think the one that's most underestimated, is intent extraction.

I think Perplexity does a really good job of understanding what my intent is when I make a query. If you specialize by domain, you're probably going to be better at intent extraction than if you don't. Because Perplexity understands, to some degree, who its users are, I think it's able to respond to them.

I've talked to Aravind about this, and I think they're doing a lot of work on it. This is hidden work. These AI products look simple, but there's a lot of work that goes into these intent-extraction models.

11. The proliferation of AI in professions: Customizing workflows

The second is proprietary data, which you mentioned. I think we're in the first inning of this. Now that we understand that we can get so much value out of this data, I think every company that's competing in these fields is going to enter a race to collect proprietary data.

Maybe some of it is going to be synthetic, maybe some of it is going to be human-generated, and maybe some of it already exists and we just need to find it. I think that's a second dimension.

The third dimension is formatting: How do you actually process that data? If you ask ChatGPT a question—as I've done many times—it gives you this long prose answer. As a consumer, maybe I'm asking it about philosophy or history, and that's exactly how I want to process the information. I do want a long prose answer.

If I'm asking a business question, such as, "Tell me the market size of the auto sector," I just want the answer, and I want it as fast as possible. The formatting is different across these 2 platforms.

The final thing is this cognitive-architecture point: To what extent are you mapping yourself onto who your user base is? I think that's beyond intent extraction, because you can actually design the product and the experience such that a certain user gets an incredible experience.

I want to take a step back in your career and understand your thinking when you started and how you're thinking today. You were really early with Weights & Biases and getting interested in AI and developer infrastructure. Can you talk about what you were thinking then and maybe what surprised you along the way?

I'll start with developer infrastructure, because that came first. I would say that was my first real, strong investment thesis.

I studied computer science in college, and I always thought of myself as a 7-out-of-10 software engineer. But I had this perspective that there were going to be a lot more software engineers like me, and fewer who looked like the Netflix engineer who builds everything themselves. They're brilliant, they understand how the compiler works, and they understand everything at a low level about the operating system.

I had this perspective that it was so economically valuable to be a software engineer that we were going to see a proliferation of those people. The original investment thesis was: What if you could make developers more productive?

12. David’s early perspectives on developer infrastructure and AI

That led me to Supabase, which is now used by more than 30% of Y Combinator companies. It led me to Replit, which now has tens of millions of users. It also led me down the open-source path, where you think about things like Confluent, Databricks, and Starburst, which build on these open-source projects.

If you're at one of these big tech companies and invested hundreds of thousands or tens of thousands of hours in building very robust infrastructure, now you can make it available to every developer. Not every company has to build it themselves.

That was my first investment thesis. I came to AI a bit later, and maybe this is interesting for you: I articulated the AI thesis after we invested in Weights & Biases.

I can pinpoint what I think of as the origin of the AI thesis. I was looking at Snowflake at the time, and we loved Snowflake. I had this intuition that Snowflake was on a path to become a $100 million company.

What does Snowflake do at the bottom, if you really break it down? Snowflake is a company that makes charts so that you can make good decisions for your business. That's really economically valuable if you're going to make the right decisions.

I had this moment, or this aha moment, where I basically said, "What if you could do more than that? What if it's more than pretty charts that are going to get made with data? What if you can make decisions? What if you can actually automate processes with this data?"

Naively, I said, "Maybe that's a trillion-dollar opportunity. Snowflake is a $100 million opportunity." That was the articulation of the thesis.

I think it was also influenced by the work that you and I were doing together. At that point, we were probably a year into working with Weights & Biases. I had been going to these board meetings, and you had shared anecdotes of companies that were using Weights & Biases.

I distinctly remember you telling me about John Deere. You said, "John Deere is using Weights & Biases, and now they're using AI to automate how they spread fertilizer." I thought, "Wow, that's such a phenomenal example of how you can use technology."

I remember talking about Qualcomm in the early days. You had these real companies using deep learning, which was really nascent at the time.

The combination of seeing it in action, having a sense of what was possible with Weights & Biases, and looking at the broader market opportunity led me down the AI rabbit hole. Of course, we've all been shocked and surprised by how far it's come in the year since then.

I want to get more specific about your investments. You've obviously had some big wins, and maybe some that weren't wins, which you didn't want to name, but we can perhaps guess who they are.

I was curious whether you have takeaways from investing in AI companies for probably longer than most people. Maybe the patterns are the same as investing in non-AI companies. I'm curious about that too. What's your perspective?

I'll give a few learnings. I'll start with Weights & Biases, because that's an easy one.

The big lesson from Weights & Biases for me was patience. I had known who you were because I used CrowdFlower, your last startup, in college for some machine-learning projects. I thought CrowdFlower was pretty cool.

At the time that we invested in Weights & Biases, the product was already really good. I think OpenAI was already using it at that time, and basically everybody legitimately doing deep learning was using Weights & Biases. But deep learning was this tiny niche corner within the machine-learning world. It wasn't the de facto way that people were doing things, so the market was still very young.

I remember that for the first year, the market was just maturing. Then, at some point in early 2020, things started to hit and take off.

13. Lessons learned from investing in AI startups

When I look back at that with hindsight, if you have a great team and a great product and you believe in the market, you just have to wait, and eventually the market comes. In our case, the market came with much more force and vigor than we would have anticipated at the time.

Even before ChatGPT, Weights & Biases was taking off, and deep learning was taking off. We were in the right place and just had to wait for it to work.

I'll give you a second one: Hugging Face. We invested in it later. At the time, most people knew that it was fascinating and that it had a huge community.

It's memorable for me because it's probably the most competitive investment I've ever made. You remember this, Lukas, because I think we talked about it at the time. They had 13 term sheets or something crazy. Everybody wanted to invest in this company, and I knew I had to do anything I could to win.

I remember flying down to Miami and camping out in a hotel for a week. I remember building an AI bot at the time. It was a chatbot built on the Transformers library, and it basically said, "Here are the 10 reasons that we should work together."

I had all 10 founders call Clem. It was memorable because sometimes it's clear which companies are working, and you just have to know, get in front of them, and win.

I'll give you some of the mistakes and the learnings. There have been a number of companies where there's a consistent thread in terms of the mistake I made and hopefully have learned from. That was underestimating how hard go-to-market is.

You meet a founder, and they're an amazing founder. The product is fantastic. Over the years, you experience viscerally, by being in the trenches, that it's hard to sell.

I think there are 2 takeaways from that. One is that, of course, you need a good sales team. You need a commercially oriented founder who desires to sell and wakes up every morning thinking, "I want to grow this business. I want to grow sales."

But I think there's a second piece: You need to put yourself in the shoes of the buyer and say, "I'm super busy. The buyer has a lot of things going on."

If your product is 20% better, forget about it. The buyer isn't going to wake up in the morning and allocate the time to go buy your product. If the market is really competitive, then you have to somehow convince the buyer that you're different.

I've experienced this as well, where you actually have a better product but just can't convince anybody that that's the case. It's not only about having the right product and the right team; you need to be able to unlock the market.

I'm so surprised to hear you say that. I feel like what I've observed throughout my life is that the product matters more and more.

We sort of—I actually think it was somebody at a Sequoia event who said, "All problems manifest as go-to-market problems, but it's almost never a go-to-market problem." That really struck me as, "Yeah, that's definitely consistent with my experience."

Do you actually prefer founding teams that have go-to-market-oriented people on them?

I think the person you're talking about is Doug Leone. Doug has this concept of a product life cycle and really digs deep with founders.

14. Balancing product development and go-to-market strategies

He has this great workshop where he'll go deep and say, "Okay, your problem is manifesting as go-to-market, but is your product good? Is your marketing good? Have you figured out who the audience is?" He'll go deep on that.

I think that's absolutely right. There are a number of problems that can manifest, but a lot of those problems are also go-to-market-related in the sense of: Have you narrowed in on the audience? Can they spend the money that they need to spend?

The fix may be, "I'm going to change my product to narrow in on the audience." But ultimately, I think that is fundamentally a go-to-market problem: You didn't build the product for a certain audience.

To answer your question about what I look for in founders, I like technical founders. I think all types of founders can succeed, but I do like technical founders. I like to invest in technical things, and you want to invest in people who are much smarter than you if they're building something really technical and challenging.

I don't think it's necessarily about having a go-to-market background as a founder. I think it is about having a passion and excitement for go-to-market, and understanding that as you scale as a founder, you're going to have to change and evolve.

Brian Chesky has this great quote that I distinctly remember from a Sequoia Base Camp event. He basically says, "The rate of growth of a company is constrained by the rate of growth of the founder."

You're on this growth curve, and something that I try to experience with my founders is that I'm also on this growth curve. We're all on this growth curve together. One of the things that's meaningful in my job and in working with founders is that you get to grow together over the journey of the company, as you and I have over the last 6 years.

I do think there is this evolution, and I've even seen it in Weights & Biases. You wake up, you're passionate about go-to-market, and you care a ton about whether our customers are happy. You're a technical founder, but I don't know—today, are you a technical founder and a go-to-market founder? Maybe you're all of the above.

Someone recently thought that I came up through go-to-market, which is the first time in my life, as opposed to through tech. I was a little alarmed by it, but also kind of flattered. I wasn't sure what to make of it. It had never happened to me before.

How would you describe your evolution over the last 6 years?

Good question. I would say that the biggest thing that's shaped me over the last—I'll say 8 years—is that I've seen a full startup cycle now. I've been doing this job for 8 years.

In the beginning, I was really focused on, "I need to look for great founders. I need to look for great companies, and my job is to get in business with them." Then things happen. I'm going to work with them on the board, but ultimately the company and the founder are much more important than I am, so things are going to take the trajectory that they're going to take.

The biggest thing that's changed in the last few years is that I've come to the idea that you have to be a true believer. You have to go native once you invest in a company.

15. David’s evolution as an investor and embracing belief in founders

Sure, at the point of investment, everybody believes. But you should ask yourself: At the point of investment, if in 5 years everybody gives up on this company and says it's not going to work, am I still going to believe? How deeply am I going to be passionate about this idea, this company, this founder, and this mission?

I think that's a really high bar, and the thing I've learned over the last 8 years is that you owe that to founders.

The venture ecosystem can sometimes be very frothy. You end up in situations where everybody is clamoring to get into a company, and you have to ask yourself, "What is this going to look like in 8 or 10 years? This is going to be a really long journey. What does that look like?"

I think that's deeply affected my investment decision-making, in the sense of holding myself to that standard and making sure that is what I'm doing.

That's an interesting answer. I'm kind of surprised by it. Do you worry that, with that perspective, you might continue to engage with something that's not working?

It seems to me that maybe a good way to do venture would be to bounce around quickly and go deep on whatever is winning right now, and shut down companies when they don't work. Surely the model of early-stage investing is that most of them are not successful.

Do you want to be digging in after 5 years with companies where things really aren't working, or encouraging the founders to move on to something else?

I think it's true for both companies that are working and companies that don't work. For companies that are working, you might not actually make the investment if you don't think this way, because in the short term they don't look that compelling. But in the long term, you think to yourself, "I really believe in this problem. I think we need to solve it, and I believe in this founder."

Even if there's a chance that it doesn't work, I'm still going to do it. Let me give an example of that affirmative investment decision, because I think that's actually almost more surprising.

I remember investing in Runway before Stable Diffusion. You meet Cristóbal, Anastasis, and Alejandro, the 3 founders, and they just deeply cared about this problem. They hadn't even come up with Stable Diffusion yet, so it wasn't even part of the story at that time.

I felt that these guys really resonated with the problem. Video was taking off within the enterprise, and I had been using Loom every day at that time. These founders, if they weren't doing this, would be artists. This is the thing that they live and die and breathe by.

I had this feeling: "I want to work with these founders." Again, they hadn't even invented the thing that ended up mattering for the company, but they cared about the problem, and I cared about the problem. I felt that in 5 years, I would still want to work with these guys.

I think that's a good example in the sense that you almost wouldn't make the investment decision if you weren't thinking that way.

A second example is even more radical: a company called Form Energy. I met the company at the Series B. The founder had run the energy business at Tesla, and he had this idea of storing solar and wind power in giant batteries—these billion-dollar batteries on the grid.

That's a radical idea. I think most VCs see that and think it's too hard. At the time, this was early in the company, so the science wasn't fully proven out. It is now.

I thought to myself, "In 5 years, would I want to work with this company?" Of course. This is probably 1 of the 10 most important problems facing human civilization right now: Can we get cheap energy storage? Of course, now with AI, that's even more important.

The founders were incredible. This founder had run the energy business at Tesla, and his co-founder, Yet-Ming Chiang, was the chief science officer. Everybody agreed that he was 1 of the top 3 battery people in the world.

Rather than deciding that it was too hard, I decided to dig in. I hired a Harvard professor, and we spent months diligencing the technology. At the end, the Harvard professor said, "Actually, can you pay me in the form of stock in this company? I really believe in what they're doing."

Fast-forward 5 years, and the company has built a billion-dollar factory in West Virginia. They're manufacturing these batteries, and it has the potential to really change the world.

Again, you end up with an affirmative investment decision because you're thinking 5 years from now, as opposed to thinking in the moment.

It also sounds from your stories as though it was a factor that you wanted to work with these people in 5 years regardless of what happened. It seems like maybe these applications were not as important as the humans here.

I'm not sure. I think it's the general domain plus the humans. Do I care? Am I going to care that this problem gets solved?

I subscribe to this concept from Buffett, which basically says that you have this punch card and you get 20 punches on the punch card for the rest of your life, and those are your investments. Do you really want to use one of those?

I think the problem matters in the sense that great people working on a problem I don't care about are still great, and I'm happy to be friends with them. But I don't want to dedicate the rest of my life to that.

16. Exciting opportunities in global supply chains and robotics

Of course, you hear me say this often, and you know me well: At the end of the day, people are the single most important thing. A fascinating problem with the wrong people also isn't interesting. Those are both necessary conditions.

What are other domains you're interested in where you haven't met the team you want to invest in?

One area I'm fascinated by is the future of the global supply chain. That's another big, big problem.

We're in a situation right now where the United States is trying to decouple from China. I meet a lot of companies that are making "Made in America" products, and you go 6 layers down in the supply chain, and even the Made in America version is using Chinese suppliers 6 layers down.

China is just so much cheaper for so many components. I call this the Shenzhen problem. You go to Shenzhen, and everybody is running these amazing automated factories. Not only is the labor cheaper, but the machinery is much more advanced, and the automation is much more advanced.

You also have these agglomeration economies where everybody is working together. One problem I'm fascinated by is that it's not enough to just assemble things in America. We're going to have to rebuild a lot of the supply chain.

To pinpoint 1 specific thing, we're going to have to build some version of Shenzhen in the West. Is that in America? Is it in Mexico? Is it in Canada? I don't know.

But I think that's an incredibly important and strategic problem, and we're just beginning to scratch the surface of it. Companies that are manufacturing things in these categories are just getting started.

Skilled manufacturing is fascinating. You look at America today, and we've gone through this phase of globalization for the last 70 years. We outsourced a lot of our manufacturing capability, and rebuilding that muscle is going to be really difficult and really important.

It does seem like there's going to be a new genre of this kind of thing with the advances in robotics that seem clearly to be coming.

I agree. The AI and robotics space—to your point on robotics—is very clear. There is going to be some application of this new AI technology in robotics.

How is that going to manifest best? I hope founders figure it out. I'm probably not going to figure it out, but I think the intersection of AI and robotics is going to be very fruitful.

It would shock me if, in the next 20 years, we don't have some version of these robots in our homes, factories, and elsewhere.

One thing I really admire about you, and have always admired about you, is how quiet and humble you are. I think this interview might not even capture how humbly you typically show up, and that's in contrast to most of your peers.

17. The intersection of AI and robotics in the next decade

I had a perception of venture capital that you really needed to be aggressive and visible in order to be successful and win these competitive deals. I'm curious because you bring more of a computer scientist, or sort of a nerd, personality to the industry you're in.

People might be interested to hear how you've leaned into your authentic personality while also having an incredibly successful career and winning deals like Hugging Face, which was quite competitive.

First of all, thank you. That's very nice.

There are so many ways to be successful in any career, and venture capital is of course one of them. I try to be authentic to who I am. I grew up with the people I'm investing in. They're the people I grew up with and was friends with in high school and as a kid.

I try to develop individual, 1-on-1 relationships. Maybe that's why I'm more boisterous on this call: We know each other, we're friends, and we've been friends for a long time.

One thing I've tried to do in my role is see each 1-on-1 relationship with a founder as the atomic unit of work to be done. The work to be done is to build a relationship with a founder, understand what they're doing, believe in what they're doing, and then be with them through thick and thin over time.

18. How David’s authentic, relationship-driven approach fuels success

I think if you do that, you can build really strong relationships with people. At least in my experience, you don't need to have a huge public persona and be highly visible. It's not necessary.

In the 1-on-1 context, you should be aggressive, and I think I am somewhat aggressive. I can say, "I really think that what you're doing is great," or, "I really think that you should think about this."

But I think the context in which that should happen is not the board meeting, in front of all these people, or as some public thing. It should be a 1-on-1 conversation, a walk around the block, or a conversation over coffee.

I deeply care about people. I think I deeply care about all the people I work with. I have this phrase that I use all the time, this mantra, if you will: Everybody knows everything.

I just think humans are so much more transparent than we think we are. I think you know what I'm thinking right now, and I think I know what you're thinking right now. We can connect at this individual level.

We don't need to do this. There doesn't have to be a divergence between me as an individual relating to you as an individual, and then me as an individual somehow relating to a group of people or relating in a larger group.

I guess you also are incredibly competitive. The story about winning the Hugging Face deal by actually using its technology is a pretty badass story of a particular kind of sales that we can all learn from.

But I appreciate how you can authentically be nontransactional in these relationships. I'm happy to see someone with that style be so successful.

I actually didn't know you thought this, but when I was looking at your website and reading your blog posts in preparation for this, you said that you think there's a trade-off between looking smart and being right. I thought that was a really interesting perspective.

I've thought a lot about that, especially as a founder when you're trying to present to VCs. There's this thing where you're trying to simplify what you're doing, but it's hard to simplify it. How much should I bend reality in order to make something digestible?

I've always felt that being able to handle that complexity and not need to simplify things can lead you to better decisions. I was curious how you came up with this concept, and whether you have any stories about how it's helped you or how you've engaged with it.

I got this from my dad. My dad is a serial entrepreneur and professor, and he's tried nobly throughout his life on many things.

I remember growing up and going to school. You have the teacher, and they know everything, and you're just the kid. My dad always had this view: "The teacher is just a person. They know a lot of things, but they're a human being. They know some things and not others."

You go to college, and you have the university president wearing their robes and degrees. They stand in front of the entire student body and make all these proclamations. Then we saw on October 7 that they're just people. That's not even to criticize them, but just to say that they're human, and humans are flawed and fallible.

19. The trade-off between “looking smart” and “being right”

I grew up with this idea that there's a distinction between form and substance. Form is not indicative of substance. Just because something looks 1 way doesn't mean it actually is that way.

You have to be penetrating and try to look through form. So much of what we do as a society is put up form in order to present ourselves a certain way or be a certain way, when the substance is actually much deeper.

I try to do that. One thing I love about Silicon Valley is that I think Silicon Valley embodies some of this. You think about the story of Steve Jobs showing up to Sequoia without shoes, barefoot, and somehow Sequoia invested. Don Valentine thought it was a substantive enough idea that he was able to look through the presentation.

I try to do this in my day-to-day investing. Coming back to looking smart versus being right, one way to look really smart is to invest in things that everybody wants to invest in. That generally works, and you look pretty smart.

There are plenty of investments I've looked at where I said to myself, "If I do this, everyone's going to think I'm brilliant for 2 years, but I don't really believe that in 7 years. I just can't. At a deep level, in my gut, I don't see how that works." So I wouldn't do it.

On the flip side, there are things where, for 1 reason or another, you're going to look dumb for a couple of years. People are going to think, "This person overpaid," or, "This idea will never work." There are a bunch of reasons why you might not look great for a couple of years, but substantively you believe that it's going to work.

I experience this tension between looking smart and being right very viscerally on a day-to-day basis, because the decisions I'm going to make are quite different depending on which of those I'm optimizing for.

One thing I've discovered—and I think it took me years to figure out and really hone—is that the people I want to work with are also people who are running that algorithm. This is something I really care about: I don't want to work with people who are presenting something that isn't the true substance of who they are and what they're doing.

The intimacy that you can build with someone when you're both presenting your true selves and being authentic to what you believe in allows you to build trust, and that trust is very durable.

When I look at founders, I also want to work with founders who are just going to tell it like it is. We're going to have direct conversations. Those direct conversations can also be serious conversations, because if you're not doing this form-versus-substance thing, I can say, "I really don't know about this. You should talk to somebody else about it. It's not my area of expertise."

In a conversation, you can also say tough things: "I think this person is not the right fit for your team." Because we're both presenting the real substance, we can have this intimate dialogue.

Maybe that could segue into another question. I was curious to get your opinion on whether your religion influences your thoughts on AI.

You're one of the few people at Sequoia who talks about religion more than most. I do feel like AI has almost religion-scale implications when you think about AGI and things like that.

Maybe that's a naive question, but I wonder if there's a connection there for you.

I'm deeply fascinated by religion, and we've had conversations about religion in the past. I find all religions interesting because I think religion is trying to answer the question: What does it mean to be human?

20. Closing thoughts on AI’s future and humanity’s role

That question has a lot of relevance in AI. I've read the major books of Christianity, Judaism, and Islam, the Mahabharata, the Tao, and the books of all the major religions in my own personal quest and curiosity to understand what it means to be human and what this experience is.

I think of religion as a package. Institutional religion is the wrapping paper. Once you tear away the wrapping paper, what's inside is similar across different cultures and contexts.

What's inside is this moment of confrontation: confrontation with consciousness, confrontation with God, confrontation with truth, and confrontation with things that are really essential and important.

Bringing it back to AI, you probably don't know or remember how much this influenced me, but you sent me Gödel, Escher, Bach one year for Christmas. That's a book about AI by an AI researcher, and he offers his own hypothesis—a nonreligious hypothesis—which says that this unique element of human consciousness is the idea of self-reflection, or recursiveness.

He runs that through mathematics in the work of Gödel, through art in the work of Escher, and through music in the work of Bach. He says that this is essentially the uniqueness of the human condition.

Why does that matter for the world today and the work we're doing in AI? There's an open question of whether AI can be self-reflective. There's an open question of whether we're going to get a very intelligent AI whose IQ is higher than mine.

I think we're going to get a very high-IQ AI, and maybe it will be better at making decisions than the rest of us. You have this science-fiction image of the computer that's the president of the United States. You can imagine a world where AI is far more intelligent than humans.

But will that AI have this self-recursive, self-reflective property? Will it have this mirroring in consciousness where it knows that it exists, understands that, and uses it as a lens through which to perceive the world?

Humans very much have that at the core of how we experience the world, not just rationally and intellectually but also emotionally. We can experience awe, and we can experience empathy.

Empathy is fundamentally a result of this self-reflective property. I know how I experience things, so I can craft hypotheses about how you experience things.

I don't have an answer to this question. I think it's one of the important questions. It's obviously one of the big questions, and it has always been one of the big questions. But it absolutely has relevance today in AI as we try to imagine whether we're going to have an unconscious, really intelligent AI or a conscious, really intelligent AI.

Fundamentally, that's a religious question in many ways. To understand it, you need to grapple with the history of religion.

Maybe I'll say 1 final thing about that. Religion has been around for thousands of years. One of the things I find fascinating about all these different traditions is that there is continuity of analysis, thought, and work.

We bootstrap on the work of the last few years. Religion has bootstrapped for longer than most intellectual threads. In the Judeo-Christian tradition, you have these debates in Judaism that lasted thousands of years. You had generation after generation after generation penetrating the same questions, and the same is true in Christianity and other traditions.

You can get a lot of depth, meaning, and understanding from that. These were the smartest people around back then, and maybe the smartest people around today are all working on AI.

One thing that's cool about AI is that everybody is focused on AI. But for a long time, the smartest people in the world were focused on these deep questions.

We'll see what happens next.