为什么 OpenAI 和 Anthropic 赢不了应用层｜Glean 创始人

Arvind, I’m so excited for this. We have a mutual friend in Mamoon, who says many wonderful things about you, and I think he’s one of the greatest ambassadors of our time. I’m really excited for this, so thank you for joining me.

Thank you for having me.

With entrepreneurs, you’re either thrilled by winning, and it’s that chase to win, or you’re terrified of losing, and it’s that fear of losing that inspires you. Which one are you?

That’s a good question. I would probably say the latter. I’m always worried about what can go wrong, and that keeps me up at night.

I love that. Only the paranoid survive. Has it always been that way?

Yeah, mostly. Yeah.

Even with the success you’ve had—it’s so interesting—Rubrik was a phenomenal success, a public company today, and you’re one of the co-founders.

Yeah.

It doesn’t change with time.

No, because every time you do a new company or start a new project, it’s sort of like starting from scratch, in my opinion. You have some good lessons from before, but it’s a new world and a new environment. Think about Glean: it’s fundamentally different from Rubrik in all ways possible.

Especially in the world of AI, you have to think that way because there’s a disruption every single day. If you start to focus more on building on what you’ve already built, that’s the winning mindset—you’ve won with something and want to double down on it. I don’t think that’s sufficient in this new AI world.

Can I ask, for those who don’t know, can you provide a 60-second summary of what Glean is and how you work?

Glean is an enterprise AI company. We started as a search company for businesses, helping an employee quickly find information they need that’s buried across 100 or 1,000 different systems inside their company. That was how we started: Google for your work life.

1. Are Enterprises Right to Fear Frontier Model Providers?

Over time, as AI models got better, it evolved into an AI platform. Today, the way to think about Glean is that, first, it’s a superset of ChatGPT, Claude, and Gemini, all of those combined into 1 product experience. It’s a coworker for your employees, and it’s connected to all of your company’s context—how work happens inside your company.

Mr. Alex K. from Palantir went on CNBC last week, and he said that the largest enterprises in the world were more skeptical than ever of frontier model providers. You work with some of the largest; you have incredible customers.

Yeah.

Do you agree with him? Are they more skeptical than ever?

Two things. One, they’re terrified of them, in the sense that every software company is worried about that: Will we be in business? Will the models eat it all? Similarly, enterprise leaders are also worried: Is their core IP—their data, their information, as well as their way of learning and their way of doing things—going to be subject to too much technological dependence on these model providers? That feeling is there for sure.

But I think what he said was that AI is not working in enterprises, and everybody’s afraid to actually say so.

Before we get to “AI not working,” because I think it’s probably one of the most important questions, but it’s a whole separate segment, do you think they’re right to be afraid of the frontier model providers eating their lunch or not?

Well, depending on the enterprise, yes. Look, if we’re talking about fundamentally changing how people work, and we’re saying that the majority of the work we do today is going to be done by an agent fully powered by one of these frontier model companies, then in some sense you’ve transferred a lot of your operations to these technology providers.

This is more than technology dependence. This is actual operational dependence on the companies that are running those agents for you. It’s actually interesting: If you think about how work happens over time, when you do a task for the first time, maybe you’ll document a process—what are the 10 steps you need to take to complete some piece of work? Then, over time, people start to optimize and tweak that process.

A lot of it never gets documented. Based on doing this work over and over again, you build all these learnings that you apply in real time to do this work in the future. All of that institutional learning is going to accumulate in the agent that’s doing that work. If you don’t have any control over running that agent yourself, if you don’t own the learning that it gains over the years, then you’re basically fully dependent on these AI companies to get your work done.

2. Who Owns the Institutional Learning That Agents Build?

So, it’s absolutely a fundamental question in front of enterprises today: How do they use these AI technologies but still retain control? All the compounding learnings that happen with AI belong to the enterprises.

Are you seeing enterprise customers move away from frontier model providers toward open source?

That’s something that’s happening now. I think we’re at a real inflection point with open source. Part of it is waiting on the open-source models to get better; the desire has been there for many years. There’s no enterprise we talk to that’s okay with saying, “I can get my work done with OpenAI or Anthropic, and I’m good.”

Everybody wants to make sure they’re in control of their destiny, that they get to use many of these models. Now, given that AI has become so expensive—people hear stories all the time about companies coming up with an annual budget for AI and running past that within a month or 2—

CFOs.

Yeah. That has really accelerated the desire for open source, coupled with the fact that we now have really good models in open source.

What do they care about? Do they care about cost? Do they care about ownership, in terms of their data staying on-prem and having visibility into the models? What is it?

I think right now the open-source drive is coming from the cost point of view. There are certain businesses, of course, that have requirements to keep all the inferencing workload within their own private data centers. When AI first arrived, companies were a lot more afraid of getting their data outside of their own control and model companies training on their data.

But that fear is no longer there. People believe that the model companies are going to be responsible and not train their models on enterprise data, so long as I’ve signed up for the right kind of contract. Right now, the drive is coming from cost.

3. Do We Need Sovereign AI Models?

In terms of where you sit in the landscape, every one of your investors that I spoke to said that I had to ask this question, which is the obvious question: Do you worry that Anthropic will do what they did to Figma, say, or what they’ve done with the legal space or what they’re doing with finance, and move into your space and cannibalize your business?

First of all, I think we should be careful in terms of what they’ve actually done for Figma, the legal space, or the finance space. They’re launching these sorts of vertical packs, but I think they’re quite shallow, in my opinion.

I don’t actually know of people who are moving their workload entirely from Figma—or, for that matter, from any other tool—to Anthropic. It’s net new, always, or I think it’s expanding the market. For example, in design, designers still use Figma, but non-designers are using Claude for design. That’s what we’re seeing: AI is making things simpler.

If people are not experts, or the primary users, of that particular tool, they can start to do some of that work with Claude.

So you don’t worry that they’ll all put emphasis on moving into enterprise and being that context layer?

They’re already doing it. Whether they’re doing it or not, we actually face that competition every day with enterprise customers. People will often ask us, “Claude can also connect with enterprise systems through MCP, so what’s different? What can Glean do that Claude cannot?” We have to go and explain what context really is and why it is actually complicated to build.

So we are competing. In fact, I would say that they probably started to compete with us before others. If you think about Claude Cowork as an application or Claude Desktop, the primary use case for that has always been question answering. That’s the largest application or use case for AI in the world today: information seeking and question answering.

How important do you think being first to market is?

It’s actually very advantageous, but it’s only a thing that helps you; it’s not going to carry you. For us, we get a lot of credit for being the first enterprise AI company in the world, the first ones to actually bring RAG into the enterprise, and the first ones to build conceptual semantic search. That gives us the brand and the right to compete in this market, even though now we’re much smaller compared to the giants that OpenAI and Anthropic have become.

It’s a huge sort of asset, but neither is it a requirement nor is it a savior.

How do you advise founders who are losing sleep at night, worried that the frontier model providers will come into their space?

Oh, right. I would say, absolutely, don’t worry about that. As a founder, you have to solve problems, not worry—that’s number 1. You have to always anticipate what they’re going to do and see their current capabilities. But for almost all other AI companies that are not doing frontier-model training, they should see the model companies as a huge asset, not a competition, in my opinion.

We believe that everything Anthropic is doing, everything OpenAI and Google are doing, as well as all the innovation that’s happening in open source, is great news for us. We don’t worry about that, and we don’t think of that as competition. In fact, they’ve allowed us to deliver a product that we could never have delivered without that help.

4. Are Frontier Models Commoditising?

Do you not think we’re seeing the ultimate commoditization of the model layer when you speak about Anthropic, OpenAI, the rise of the model layer, and the speed with which new models are coming out, especially from open-source Chinese providers?

So one thing is clear. Let’s talk about enterprise use cases.

Yeah.

5. Chinese Open Source Dominates Open Router

90% or greater of use cases can now be fully handled by many, many different models, including open-source models. So there’s definitely commoditization from that perspective. In fact, at Glean, that’s actually one of our core value adds to our customers: cost control. We will tell them, “As people complete their tasks on our platform, we pick the right model for you. If you’re okay with using open-source models, we’ll use them when we think it’s appropriate, when it’s going to generate a high-quality answer.”

What percentage of customers are not okay with open-source models?

This is actually so new. We only have—I would say—open source truly coming to within 3 months of frontier capabilities, and that has just happened literally a month back, or not even a month. I would say GLM 5.2, though, is the very first time where our own team, for example, feels comfortable that now we can run the majority of our workloads on that model.

We are yet to find what people are going to tell us. From a point of view of open source and using the model, everybody’s going to be fine. The question is going to be, are they okay with the Chinese model or not? That’s the only question here. It’s not open source versus closed source.

Why would they not be okay with a Chinese model when you look at the ownership, the ability to have it on-prem, and the fact that you’re not sharing anything back to China? Why would you not be?

I think it’s just comfort. It’s just, what if something goes wrong? There’s always paranoia and fear: what if there’s a backdoor? Some magic backdoor that we don’t even understand—then that could be a backdoor.

6. Have We Completely Mispriced the Frontier Model Landscape?

There are some concerns. There’s also the fact that if you use these models and it becomes a known thing, it could be used against you in some ways by your competitors and things like that. So, a variety of factors. But ultimately, it again boils down to who’s willing to be bold, because this is a new thing. Large enterprises have to make this move, and the early movers will make the move first. Then it’ll become a more normal thing.

I’m always doing this show to learn: if 90% of enterprise workflows can be done with open models, have we completely mispriced the frontier-model landscape? It’s a very different time.

I do feel like the model business on its own—forget open source for a minute—has plenty of competition even within the labs, and more and more companies are coming into that space. In that fierce competition, even in a 3-way race, I think you can actually get a good amount of pricing pressure. Now, of course, with open source, prices are an order of magnitude cheaper.

I actually heard rumors that OpenAI was going to drastically reduce its model prices in response to these developments, competition, and open source. I think the model business on its own is probably not as lucrative as everybody believes. But these companies now have a lot more things; they’re no longer model companies only.

Totally get that. But if they’re doing shallow things in those adjacencies, they’re not exactly going to generate $1 trillion of revenue, like Dario said.

If you think about it, first of all, these 2 labs are very fundamentally different businesses. OpenAI, of course, has an amazing consumer product, and Anthropic—the interesting thing that’s happening is that people are building on top of their platform.

When you think about Anthropic right now, there are a lot of folks who are developing automations and skills, and everybody’s creating these MCP servers to their internal systems, getting connected and connecting it all to Claude. So there’s an ecosystem actually being developed. You should very much consider them an application-level company, not just a model company.

If you were to make a guess, in 3 years’ time, what percentage of your workflows do you think will be through open source?

We’ve been telling customers that I believe the majority of enterprise workloads will actually be on open-source models in 3 years, for sure.

Yeah. Another competitive element that you face, forgetting the model providers, is actually Microsoft. Microsoft has made a phenomenal business on the back of creating a 70%-as-good product, but bundling it into a bundle for enterprises and then selling it with a nice sticker on it.

How do you think about the bundling pressure from Microsoft Copilot as a competitive threat?

Well, for us, they are one of our most significant competitors, and the bundling strategy actually works. You have to fight against that.

Luckily, there’s always been room for best-of-breed software, and our customers think of us exactly like that. If you’re trying to bring a great search product, if you’re trying to build a horizontal, comprehensive AI platform, they know that we do it better. So companies are willing to invest on top of that, as part of the bundled product suite from Microsoft.

But the other thing that is maybe making bundling not as effective a strategy anymore is the fact that AI is moving toward consumption-based models. Once you move toward consumption, there’s no inherent bundling advantage because, as a business, I can get 6 tools and let the users choose where they want to do their work. Wherever they do their work, I have to pay for that particular unit of work. So consumption can ultimately break that bundling strategy.

And respectfully, I don’t know if it does if you’re working with enterprise, because they will make you comply with an enterprise bundle. You’ll go through approval processes and sign-off processes internally for the largest enterprises in the world—your VWs or your Fords or your GEs or Tyson Foods. I always use them as random companies. But they’ll approve Microsoft as one vendor.

If they're suddenly having to approve 15 vendors, forgetting the pricing and the transactions, it creates a vendor-management problem that they didn't have before.

That is true. But I would say that if you go and talk to companies that have been on the other side of Microsoft's onslaught, most of them will talk about pricing as the main killer, because I think it's hard to compete with free.

Who's a fiercer competitor: Microsoft or the frontier models?

Good question. I think it's early to tell. But Microsoft is formidable. If you look at our experience as we go and prospect, we hear this answer more often: “We are a Microsoft customer, and we're already getting Copilot, so it doesn't make sense for us to consider you.”

We do hear that, and we hear it more often than we hear somebody say, “I've embraced one of the lab products, and therefore there's nothing else that I'm going to do.”

We mentioned Alex K.'s interview at the beginning, and I interrupted you and said, “Before we dive into ‘we're not getting value,’” because I think the second half of 2026 and 2027 is the year where everyone goes, “Hang on a minute.”

Yeah.

Is this spend generating output or a return on investment? How do we think about the return on investment that enterprises are getting? Is Alex K. right in saying everyone's going, “What the fuck? Where's my return?”

I would say that there are pockets of value realization today. Take customer support as a vertical. I think it was easy to measure productivity there. You could say that, in your company, a support agent resolves 10 cases a day, and now they're able to do 12 because of AI. You can see that in a very concrete measure of productivity increase.

7. Why AI Is Not Working in Enterprise

That's a use case where AI is actually pretty good, because a lot of the time spent by support teams is reading knowledge and then summarizing it to your customers in some ways. So there are definitely areas where there is clear value realization, and enterprises are feeling good.

Some others are more complex. I think the majority of AI spend right now is on coding, and coding as a practice has changed. Most developers now use AI to write code; they're not writing it by hand anymore. You can say that AI made a big impact, but are they shipping products faster? That's where we hear most companies saying, “No, the actual shipping speed of products has not increased,” even though coding speed has increased significantly, because coding is only a small part of shipping a product overall.

Has your shipping speed increased?

It's hard to actually measure. That's the challenge, because engineering productivity is one of the most difficult things to measure. It's the fuzziest of the jobs out there. If you look at some of the metrics, like lines of code written, of course we're writing way more lines of code now. But if you look at whether we're shipping features at a greater pace, yes, we are.

But that's also a result of having a larger team and a team with more tenure than it had before. Sometimes it's hard to tease those things apart. With that, what do we do as a company? We're saying, “Look, we're just going to keep investing.”

What percentage of Glean's code, say, do you think is written by AI now?

It's probably almost 100%. Nobody is actually writing the initial code by hand anymore. Maybe sometimes.

Maybe you've got an assistant in the corner. [Laughter]

Yeah. Almost all the code is being written with AI, but we enforce human reviews. You cannot generate tons of AI code and then just check it into the repositories. We're probably more conservative than most other companies.

There was, in fact, a discussion inside the company that, now AI can write so much code, the real bottleneck has shifted from the person who writes the code to the person who has to review it. There was a proposal to eliminate code reviews and just let the code get submitted directly into the repositories. Many companies are doing that.

If you have a stringent code-review process, it almost removes the point of having a faster code-development process.

Yeah, it's true. I think what it's doing right now is showing that we're still in the learning phase of using AI properly and effectively, and thinking about the long-term ramifications of it.

When you write code with AI, you can write a million lines of code, but it becomes incredibly hard to maintain, understand and manage over time.

Isn't that what AI does, though? You have AI that does refactoring, AI that does security and AI that does—

Yeah. The only thing is that it's not perfect right now. We have to make a trade-off. Right now, we're willing to pay the cost of reviewing the code.

We're still faster than before because the writing part is much faster now, and the person who writes the code is the one who does the first review.

When you say AI ROI is really a throughput problem, what does that mean?

The first thing we have to do is make sure that you're able to bring the right context to these AI agents. If you think about most enterprises today, the way they're rolling out AI is that they just throw it into the system and connect AI with all of their enterprise systems in a rudimentary manner using MCP servers.

8. The Agent Token Waste Problem

Now you're letting any piece of work that you're trying to do with AI allow the models to brute-force their way into figuring out and assembling the right raw materials they need to complete the task, and then doing it. In this mode, AI is super slow. It takes a lot of time just to assemble the basic information it needs to do the work.

It also becomes very costly, because most of the tokens are being burned just trying to assemble the right context for a given task. You're trying to use AI for things where it's not even good at or needed. To make AI really perform and deliver, you have to invest around it. You have to make sure that you provide it with the right context so that it can work faster and at a lower cost.

What does it mean to invest around it? And are we wrong as CEOs to be urging all of our team members to be trying to replace themselves with AI, even if it means that we're wasting tokens?

9. Should You Actually Try to Replace Yourself With AI?

I think it's a wrong goal, in my opinion, to say, “Hey, replace yourself with AI.” First of all, I think you're giving too much credit to AI. When you say that, it's just not ready right now. Give me the name of one job that you can replace with AI, for example. Do you think it can replace your EA?

Mine? No, but I'm a fucking diva. [Laughter]

For most people, I think it can do the majority. It can actually take care of a lot of things for any given role, but it cannot replace the final intangible.

No, but that can be a tipping point where, for a lot of people, if it does 90%, fine. You know what? You'll do that birthday present for your wife because it's once a year. It's not very often, and Claude isn't quite personal enough to know your wife's preferences for perfume. It's very close, but that role will get cannibalized.

I'm not sure, and I'll tell you why. I think you want to be performing your best in whatever you do, and I don't think you're going to take a 90% solution.

I'm not cost-constrained, being a dick. [Laughter]

Well, I mean, look, it's not about you not being cost-constrained. It's about having to be competitive in your work with others. Remember, they also have all the AI tools that you have, but if they also have a human on top, how are you going to compete with them?

How many people do you have now?

10. Glean Has 1,000 People - Will It Have 5,000 in 5 Years?

In our company, we're over 1,000 people now.

Over 1,000 people. How many do you think you'll have in 5 years' time?

Hopefully 5,000.

Wow, so you don't—

We're going to grow.

But that is very atypical. I sit with the biggest CEOs in the world, and every single one of them is shrinking teams. Every single one of them is saying—

I absolutely don't believe in it.

Why, though?

Why?

Well, I mean, think logically. Take 2 companies, Coca-Cola and Pepsi—2 companies that compete with each other. One company decides to shrink, and the other one still has a lot more people.

Both of them have full access to the same AI tools and technology. So now the question is: if you were trying to do the same amount of work and you believed you could do it with fewer people and therefore shrank, your competition could also do the same, but they chose not to do the same amount of work. They chose to elevate and build a 10x better product or produce 10 times more goods because they have more people. They’re going to be larger. They’re going to beat you.

But I don’t think more people makes for better products.

That’s a different thing.

I can cut headcount and then afford the best frontier models—the best technology for my 100x engineers—because I’ve reduced headcount. I think more people slow down everything.

But that’s not an AI argument. That argument has always been true.

Sure, but combined with the AI element, you’re able to ship more. If you’re able to ship more, I promise you—and you know this—when you have more people, they’ll just put up the barriers to get in the way of that product going out.

Well, look, even in the AI discussions we have right now, before that, post-COVID, many companies felt they were bloated. They cut down 15% or 20% of their staff, and every CEO came out and said that, as a result, they were moving 20% faster. A lot of companies came and talked about that.

That’s an argument that’s always there. At some point, teams get large and start to slow each other down. Humans do that. I also believe in that, but ultimately people are also your asset, and you have to be able to deploy them correctly in the right set of projects. I don’t think the world’s greatest companies are going to be companies with 100 people. Look at the model companies; same for them. Why are they hiring so aggressively?

Do you not think that the best people will want to work with the best technology, and we’ll see an increase in technology spend by the biggest companies in the world, from 8% to 12%, where it is today, to maybe 16% to 20%? Then, actually, you’ll see a reduction in headcount but an increase in technology spend. The best people will want to go where they have the best tools and equipment.

I’m not sure about that either, because I think technology is actually not supposed to increase in cost. First of all, do you admit that currently this technology is priced absurdly for what it delivers?

I think it totally depends on what it’s doing for you, so no, I don’t at all. For Cursor or any of the dev tools, I think it’s still dramatically underpriced. When you look at Mark B. spending $300 million on Anthropic, that’s 3.7% of developer salaries. I think that’s relatively small. I would say it’s absurdly expensive.

11. The $1M/Month Agent That Replaced 15 Engineers

I’ll give you an example. We had this really cool triage agent for engineering. We had a 15-person on-call team whose work was to triage every single production issue that happened—any system alerts or things that were going bad—and we built this agent that’s now taking care of 95% of those issues automatically for them. But even there, it’s doing that at a cost that’s questionable: is it actually more effective or more efficient than humans? We were spending $1 million a month on that particular agent.

A million a month?

Yeah.

Are you buying Cristiano Ronaldo? What are you doing?

No. A cost like that is quite expensive.

But sorry, can I go back? You said—and I discuss this a lot on the show, so you’re making me much smarter—you think that spending 3.8% of developer salaries on these tools is a lot. If you think that’s a lot, then these model providers are absolutely screwed.

Well, I think the point that I’m making is the 3.8% number actually doesn’t seem high at all when you look at it that way.

Yeah, but I also know that already, with open source, you can do the same amount of work for a tenth of the cost, right?

That’s number one. Number two, historically, as far as I can remember, we’ve never put technology cost and labor cost in the same sort of sentence ever before. This is the first time we’re actually hearing, “Hey, I would rather have fewer humans and more tokens.” It’s the first time, and I just feel like this is not how technology works. The models are supposed to get cheaper and cheaper. The tech is going to be more and more affordable.

But I’m so sorry. This is so funny for me because you’re the co-founder of Glean and Rubrik, so who the fuck am I but a podcaster? But this is exactly what technology is for. This is agents being proactive, having an opinion and making a decision. They should absolutely be included, or put in the same sentence as labor, because they are replacing the labor that we used to spend money on.

I think good technologies figure out how to make technology really, really cheap, and it’s going to happen here too. That’s my belief. You’re going to see it. You’re going to see inferencing costs come down by orders of magnitude.

I think we saw something bizarre in the last 6 to 9 months: every model actually increased its per-token price. If you go back 15 months, everybody thought the per-token price was going to keep falling, like it was before. We don’t know what happened here. This is also sort of unique.

They needed to prove that they were good businesses before they went public. That’s what happened. I can say things that you can’t.

Yeah. Yeah. But my bet is on AI getting much, much cheaper than what it is today.

If AI gets much, much cheaper than it is today, these already loss-making businesses—which prop up our entire global economy pretty much at this point—are very threatened.

Yeah. I mean, my take remains the same.

So, okay, it’s really interesting. You don’t expect an engineering team to get smaller in the future?

12. Per Person Productivity Will Go Up, But So Will the Bar to Compete

I think per-person productivity is going to shoot up, but so will the demands. To make the same amount of revenue, you have to produce a 10x better product in the future. Unfortunately, that’s the reality.

When you think about token spend internally, how did you sit down and think about it as a team, sitting with your CFO? How did you go through the decision of how to think about token budgeting?

Well, I think we did probably what most companies did, which is we didn’t do anything.

So I think we’re in this phase of letting people figure out what they can do with this tech.

And what did you see? People went crazy. People didn’t adopt it. What happened?

There’s a power law in our company and also at all of our customers. You’ll see some people who spend $10,000 or $15,000 in tokens every month, and then you have others who are spending $20.

One thing is interesting, though: everybody has embraced AI to some degree. Everybody’s using the basic tools. As I mentioned before, the number-one application or use case for AI today in the world is information-seeking and question-answering, and everybody’s doing that. You see everybody on our team, as well as at our customers, doing that. Everybody’s asking questions, and everybody’s getting some basic summarization and information synthesis going. But the advanced use cases are limited to about 5% of the employee base.

13. The AI Power Law Inside Companies

Is there anything you do as a leader to try to infuse AI as aggressively as possible? We had Nikesh Arora from Palo Alto Networks on. Every week, he has a leadership meeting where he’s like, “Show and tell,” and everyone needs to stand up and show something that they’ve done with AI that week—whether it replaces what they do, improves their job, whatever it is. Is there anything that you can do?

Yeah, that’s actually a really good idea. I’ve thought about doing that. We limited the token-maxing dashboards, and I always thought that was not the right idea—to just reward people who are consuming more tokens. I felt like we didn’t need to do that. We are a native AI company ourselves, and people are already educated enough. They’ll use AI when they need to.

But executives sharing a success story—we haven’t demanded it from every single executive every single week. We do have the showcase in our town hall. We’ll always ask people to share those wins. Every town hall has a section dedicated to the new AI agents that teams are using to work differently.

Can I ask about the executives and the people that you have?

I think recruiting has never been harder.

How hard is recruiting today, with some of the largest model providers, as we said, paying just enormous salaries that we haven’t seen before?

Yeah, I would actually say maybe the last 2 or 3 months.

Let's put that aside for a minute. I would say recruiting was actually getting easier for us compared to the SaaS peak. Why? Because companies haven't been growing their headcount. If you look at the largest employers of tech talent, many of them actually haven't been growing; many of them have been laying off continuously.

Take Meta, for example. Every year, there are significant layoffs, and I don't know if the overall headcount—my guess is that it's probably down from the peak of 2021 or 2022. So there was more talent available in the market than before. But if you start to talk about AI talent, ML talent, top people are sought after way more than ever before, and the pay scales have completely changed—not just from the model companies but even from startups, because startups are also—you are giving them too much money to compete for talent. Even startups these days pay a lot. We have to.

Yeah, we have to because their alternatives are so large, too.

And actually, if it costs $300,000, $400,000, or $500,000 for a great developer, well, the $2 million seed round just doesn't go anywhere.

But the founder building the team, even if they don't take a salary, if they're going to hire 4 people—

I need $6 million bucks.

That's right. Do you think founders should raise large seed rounds?

I think it's better. I always prefer to raise as much of a round as you can from the get-go.

Which round felt the most highly priced?

First, we never actually went out to raise, except for our first round of the company. We always had somebody come in, and it was a relationship that got built over some time and kind of became the de facto understanding that they were going to be the ones putting money in.

I would say our Series C probably felt the most expensive. We barely had any business—definitely sub-$2 million or $3 million, maybe $5 million; I don't remember exactly—but the valuation was north of $1 billion. That was extreme. I guess we take what we get.

That's incredible. Would you worry about scaling into that when you're doing it, or do you just keep your head down and think, "This is great: low dilution for a high price"?

The way we thought about it was that it was a statement to be made to prospective employees, more than anything else. We wanted the market to understand that we were building something special, and that kind of gave us that validation.

Do employees give a shit who your investors are?

Absolutely. Yeah.

A lot of founders are like, "The best people don't care. They're there for the mission." I'm always like, I promise you, if you have Kleiner, DST, or Sequoia, great candidates suddenly want to talk to you a lot more.

Yeah. Investor reputation directly impacts your reputation.

When you look today, what have you changed your mind on most in the last 12 months?

Personally, my style has been a little bit too disciplined to be the right strategy anymore. I get that feedback from my team that we're trying to be conservative and make sure our capital goes a long way. In that mindset, we may lose the land grab.

I'm feeling the pressure to change how I think about how we should be spending and how we should be investing. At the same time, I have this fundamental belief that a business is always built on discipline. You have to charge for the product; it has to generate value for customers. For every dollar that you invest in marketing, there has to be some good return from it. You cannot assume that you can just keep raising money to make up for all those things that were not there.

Do you agree with that when you have examples like Uber, which prove that a bad business model can turn good with scale?

Yeah, that's what I'm saying. That's the one where I feel that pressure—that perhaps my way of thinking is incorrect.

Do you think it is a land grab?

We are absolutely in a land grab. There is no question. Every single company in the world wants a product like ours today, and either we get in today or it's going to be 10 times harder to get in the future.

We spoke about job displacement. We had an interesting conversation around that. What job does not exist today that you think will be incredibly common in 3 to 5 years' time?

Composite roles will be very common. For example, somebody who can build a product—I don't know what to call them—but they can act like engineers, product managers, and designers.

Similarly, in go-to-market, somebody who can sell the product and is capable not only of doing the business negotiations but also of demoing the product and talking about use cases. Instead of having that segregation between account executives, solution engineers, and post-sales solution architects, I think we're going to see more and more generalization of roles away from specialization. In fact, I was trying to drive that very hard even in our own company.

I'm sorry—I mean this in a nice way—but composite roles go exactly against the idea of maintaining team size. If you have composite roles where you bring 4 different specialties into 1—

14. Which Roles Will Disappear First: Analysts, Recruiters & BI Teams

That is smaller teams.

It is. Yes.

But as I said, you have to do 10 times the work to get the same amount of revenue from your customers in the future. You have a much smaller team to deliver the same amount of work that you used to deliver before. We just are forced to do more.

Got you. What role do we have today that we will not have? What do we look at and go, "Oh my gosh, I can't believe we used to do that"?

A lot of analyst roles, or data analyst roles that are not business thinkers. They were given a task: "Hey, I need to see this data," and then they would go and build those specific dashboards and configure backend systems. I think that kind of work definitely goes away.

Business intelligence is just going to be very different. Business owners will directly be able to get answers to their questions. Business analysts, like data analysts, are one example.

Many HR roles will change as well. Sourcers, for example—a sourcer in recruiting is a role that I think is definitely going to get consumed into a full-cycle recruiting role.

15. Why Enterprises Are Shifting to Open Source

I do have to ask one final question. We're sitting here in Europe, and it brings about a question of sovereignty. The US and Europe, bluntly, have not come up to muster, so to speak, on open source. Do you think we will have a world of sovereign models? Given what we've seen in the last month or so, do we need to have sovereignty over our models?

The desire for sovereign models is strong. I would say it was probably stronger a year back compared to now. I feel like I'm hearing less of it. There was a period where every nation thought that it could build one, when AI was still in its early stages, but a lot of those nations figured out that this isn't going to be the way. They're okay with letting their enterprises within their own countries use OpenAI, Anthropic, or all the other models.

I'm not an expert. I don't know whether this trend is on the rise or on the decline a little bit.

I think it's unequivocally on the rise, given what we saw with the Trump administration banning Anthropic's latest models and this understanding from a lot of, especially, Europeans that we cannot rely on a US individual who could ban our access to intelligence.

Yeah, but where are the results from it?

That was a month ago. To expect a state-of-the-art model within 3 weeks would be tough.

Yeah, but even before that, it just hasn't happened. The only country in the world that has produced models outside the US is China, and then maybe a little bit in France.

Is that simply an incentive problem—the lack of an open-source community in the US, and why we don't have any US open source to any real degree or with any real substance?

No, you're right that we don't have open models, but it's not because open source as a movement, as a concept, is weak in the US. It's actually quite strong. If you think about models, they require a lot of upfront investment, which is not open-source-friendly in many ways.

A lot of open-source software has been skunkworks: developers get no funding associated with them, and they still get something built. They couldn't build models that way, and so that's why naturally this thing didn't work out. You need these techniques where super-high investment is not needed.

Do you worry, then, when you look at the state of things? I spent a lot of time on OpenRouter, and I see the model usage and traffic. Anthropic today was the first U.S. model; it was seventh. The first 6 were Chinese. Do we just get out of the bucket? Who cares that the CCP are funding the top 6 models?

The fact that you can actually run inference on those in that contained environment makes people feel comfortable, but I don't think the U.S. will feel absolutely okay with that trend. There's good work happening now to promote open-source and model development in the U.S. There are some models coming out.

The alternative is that Sam and OpenAI give 5% to Trump, and then he puts regulatory capture on Anthropic and OpenAI and puts taxes on open-source.

Well, I hope not. I doubt that's going to happen.

16. Sam Altman's 5% to Trump

Why would—I'm so sorry, I'm learning. Why else would Sam give them 5%? It's a quid pro quo: I need you; you need me.

Well, I mean, I guess I just believe more in the U.S. system. I don't think right now, by the way, you need to curb open-source; it's too far behind in the U.S.

You don't think Sam and Dario are sitting there going, “Oh, wow, we underestimated this, and this is a core threat to our business”?

They probably are thinking that, but I don't think they can fix that through regulation.

You don't think that Sam will be calling up Trump, who he has a direct line to, saying, “The CCP are funding your biggest—our biggest—competitors, and we cannot promise that there isn't a back door to Xi Jinping. You need to stop this, and I'll give you 5% for your troubles”?

Well, isn't the argument the other way around? Right now there are all these open-source models which are very good, and they're all built in China. The U.S. needs to build its own. The U.S. can't be seen as a country that doesn't innovate on technology, so it's actually paramount for the U.S. to build—

I think Sam will be saying it takes billions of dollars and years of time. “Trump, defend America and support OpenAI and Anthropic, and put barriers up to prevent Chinese open models from getting adoption—taxes, bans.”

Those, maybe, yes, but U.S. open-source models are going to have a lot of tailwinds. This is a known, accepted issue that every technologist in the Bay Area talks about. There are a lot of motivated parties that actually want to promote this, including NVIDIA, for example. They're putting a lot of investment into promoting the development of great open-source models in the U.S., and I hope they succeed.

17. Quick-Fire Round

Absolutely. [laughter] A multi-model world is important for all of us. Listen, I'm going to do a quick-fire round with you. I say a short statement, and you give me your immediate thoughts. Does that sound okay?

Okay. Yeah.

What's your biggest advice to someone studying computer science today?

It's fine to study it. Don't get too worried because of what other people are telling you.

Which legacy company has adopted AI the best, do you think?

Well, are you willing to call Google a legacy company?

Yeah.

Yeah, so Google probably rates higher than anybody else in terms of not only embracing AI internally but also launching products. But I guess they are AI companies, so it's kind of hard—it's unfair to put them in that category.

You start a new company, and you can only take 1 ambassador. Who do you take with you?

Well, I think I'll take 1 of our existing ones. We have great relationships with all of them.

Which one would you take?

I don't know. I won't answer that question. [snorts] I just don't have the answer, really. I have to think about it. I think it's probably circumstantial, depending on what I'm doing. Different people bring different strengths.

What would you most like to change about the startup ecosystem that we see today?

I actually do think that there is too much capital available today for startups, and it's sometimes creating failure paths for people. I think they're not getting what it takes to build a great company. I'll give you an example: a startup that has raised a seed round decides to pay $500,000 to an engineer, like you were saying before.

It's happening today, and the startup founder is okay with it, the investors are okay with it, but it's surely not a sustainable path to actually win. They're paying it while Google is not, and Google knows that they don't need to buy talent like that. So I think that is 1 thing that I feel: this overabundance of capital is getting startups to create structures which are not going to be sustainable for them.

Do you worry about the lack of exit options that are now becoming more and more real? What I mean by that is, honestly, if you don't have $1 billion in revenue today, it's hard to go public. Tech acquirers—your big companies—are very specific about what they want to buy. VCs are licking their wounds from having a portfolio that's full of markdowns.

It's a tough landscape.

Startups have never been easy. In fact, I would say in the last 25 years that I've seen, it's easier to build a startup and get a good exit from it these days than it used to be in the past. Startups are brutal. It's a brutal game.

What does no one know about being a founder and CEO from the outside that they should know?

That it's not a sexy job. It's actually one of the most stressful things, and you really have to be crazy.

I think they know that now. One thing for me is that you have to consistently be unhappy. You should never be happy, I think, as a CEO, because there's always something that needs doing or could be done better. Telling someone they will never be happy is something they're jarred by.

Yeah, that's a good one. This is a tough job all around, and I think oftentimes people who have not done it feel that there's a lot of glamour. They feel that this is going to make a lot of money, their life will be fantastic, and they're going to have a lot of respect. I think almost all of those things are irrelevant. You have to be truly mission-oriented to survive as a founder.

Did your style change with money? You've been successful before. I think founders are better and investors are better when they are already rich, if I'm being blunt. I think you make more rational, sound decisions that are not made with economic impatience.

I think for me, maybe not. But at the same time, I'm a man with minimal needs, and my needs were already met a long time back. So I guess I've definitely built these startups without that worry of, “Can I feed my family?” So, yeah, maybe that has helped me, but as I've seen more success, it hasn't changed me fundamentally.

You still have to have that drive. You have to work continuously. You have to work more than every other person in your company, lead by example and keep pushing, and you have to have this irrational need to make something big happen.

I so appreciate your time. I apologize for being robust in my pushback.

I think it was a different interview from a lot of interviews that you do.

It was more discursive, but I so appreciate the time. You've been fantastic, dude. Yeah.

Thank you.