Benchmark 的 AI 押注：Cerebras、Sierra、Legora、Fireworks、Starcloud、Gumloop……

We used to have golden rules or North Star metrics in investing. When you think about software, there were five or six things that mattered, and it was very legible and very spreadsheetable. Now all of the golden rules of the past that defined the spreadsheet investing era—

Mm.

—are all gone. In the new AI paradigm, you can have businesses that are well over $1 billion in revenue that haven't proven out their unit economics. They haven't proven out durable product differentiation. We have developers who are spending $3,000 per month each on Cloud Code. So it's like, wow, okay, that's $36,000 per developer.

When you think about the Anthropic $380 billion round, if Anthropic were to go public and get liquid at a $1.5 trillion valuation, I just don't know if people really understand or are ready for the impact that this much liquidity could have on the ecosystem.

Ev, welcome to Sorcery.

Thank you, Molly. It's great to be here.

I'm so happy. This is the first interview I've ever done with someone from Benchmark. You're one of the newer partners. I did an interview with Jack before he joined.

Mm.

But now we have you.

That's great. I'm excited to be the first.

So what's going on? I know you guys are just coming off the heels of your AGM.

Yeah.

You had a mass, collective reflection on the market today. How are the vibes?

1. AI Breaks The Risk Curve

Yeah, vibes—I mean, the vibes within Benchmark are definitely high. I think as a market, the whole venture-growth ecosystem is going through this phase of disorientation. Everyone sort of feels like down is up and up is down.

I think that, as we have our annual meeting with all of our partners, oftentimes that's when we try to introspect and figure out how we've been feeling about the market and the things we've been looking at. The one visualization, or the one framework, that I came up with that I think explains a lot of it is that, in the previous paradigm—especially in software—you had almost this inverse relationship between scale and risk, or scale and the risk of major impairment or a company going out of business.

That's because, historically, as you're scaling as a startup, you're sequentially de-risking different parts of your business. The first thing you de-risk, typically, is whether you can get product-market fit: Do people want to buy your product? Then you de-risk unit economics: Are you selling a product for what will end up being unit-economic-positive over time?

Then you de-risk things like the TAM of your market. Are you able to grow not just to $10 million of revenue, but to $100 million and maybe multiple billions of revenue? Then eventually, market leadership. If you don't de-risk those things while you're on the journey, you stop scaling. You just stop growing.

There's this nice, clean relationship between those 2 variables: If you're continuing to scale, it usually means you're becoming a less and less risky company. I think the thing that's changed massively in the new AI paradigm is that you can have businesses that are well over $1 billion in revenue that haven't proven out their unit economics. They haven't proven out durable product differentiation.

In many ways, it feels like the risk of impairment—or a company either going to zero, which is the worst-case scenario, or even just being devalued significantly from its last round—actually stays flat over time. Maybe there's even a weird positive correlation between scale and risk.

It's a very disorienting thing, because you're used to the fact that the bigger you are, the safer the company is, and it doesn't feel like that anymore. There are a lot of reasons for that that we can go into, but I think that's the reason why folks in the market, investors, our peers, and even us feel like something is different. The risk-reward paradigm is very different these days.

How do you evaluate that as an investor? We were kind of talking about this beforehand, but you say this is the death of spreadsheet investing.

Yeah.

Is that kind of what you mean? Is it categorical, or what's the premise there?

2. Spreadsheet Investing Is Dead

You know, we used to have all of these golden rules or North Star metrics in investing. There are some exceptions to this—hard tech has always been different, as have some of the capital-intensive consumer internet businesses of the past—but when you think about software, there were five or six things that mattered.

High gross margins were better than low gross margins. At least 70% or 80% was good; 90% was amazing. You wanted to be selling pure software, which meant you didn't have a ton of services or implementation load on your software, because that limited your ability to scale and brought down your gross margins.

Typically, these businesses were capital-light. There wasn't a lot of CapEx, and you had a lot of operating leverage on your R&D spend. R&D was relatively efficient over time. You also had really high gross retention, which meant that your customers, once they implemented your product, typically stayed around.

A lot of these software companies had gross retention rates above 90%. Over a year-long period, if you had 100 customers, fewer than 10 of them would leave. When you took all of those things together, the output was a capital-light business that, essentially at maturity, would generate really high rates of free cash flow in a way that grew secularly at multiples of GDP. That's sort of a mouthful.

Mm.

You can think about the whole reason why software companies traded at high revenue multiples, or high multiples of NTM revenue. This is really important for an asset class whose foundation is paying really, really high multiples of ARR. The whole reason why we're allowed—or have been allowed—to do that is because, eventually, these companies are extremely profitable, those profits are very durable, and the profits grow at a multiple of GDP for a very, very long time.

That was the setup for software, and it was very legible and very spreadsheetable. All those things fit really nicely in a spreadsheet. There were metrics like the Rule of 40—your growth rate plus your free cash flow margin—where sometimes you could literally abstract the quality of a software company into a single metric. Some investors would invest based just on a Rule of 40 score.

Now, if you think about all the most popular companies and categories, they're almost the inverse of every single one of those golden rules I just mentioned. There was a tweet that said, “FDE is the new PLG.” It's the most popular distribution strategy—

Thank you to Palantir.

That's right. The Palantirification of everything means that now you have all these sparkling implementation consultants. That used to be bad. You weren't supposed to have that. You were supposed to be selling pure software.

But that draws down gross margins. Now, if your gross margins are high, that's actually a bad thing, because AI inference costs a lot of money. If you have an AI product with high gross margins, that means no one's using your AI features.

There's this bizarro world where another popular idea right now is that, if you're going to be safe from the labs, you need to be training your own models, or at least post-training on the data that you're getting from your users. Now you want people to set up their own research labs within a software-app company in order to train their own models.

That's unbelievably CapEx-intensive, or capital-intensive, relative to the old era of software, when you didn't have any CapEx or any kind of capital outlays on GPUs or anything like that. All these things are bizarre, because all of the golden rules of the past that defined the spreadsheet investing era—

Mm.

—are all gone. The most popular companies and categories are almost the inverse of all these golden rules.

I think the reason why that's so confusing for people is that these golden rules, in a vacuum, are the first-principles building blocks of a high-quality company. At the end of the day, you need a company that's producing a lot of free cash flow for a very long period of time in order for it to be valued highly.

And when you look at the most popular companies, you say, “Well, from a first-principles basis, they just seem less attractive than SaaS.” You have to rethink a lot of how you invest.

I have so many secondary questions to this. The first one is: Who do you think has the best instinct for managing the scale and economics of AI in these companies? Is it the founders? Are you finding it within different firms or among other kinds of researchers or economists? This is new to everyone. How are you—

Yeah.

—making sense of it, and who do you think has the best handle?

3. AI Needs A New Taxonomy

It's a great question. It's really interesting because these companies are also extremely different from each other.

Yeah.

If you take a company in our portfolio like Fireworks, which we think is becoming the AI inference cloud, and you compare it to another incredible company that ostensibly is in the same category, like Crusoe, they're actually completely different companies with completely different business models and completely different economic models.

What I mean by that is that Crusoe is actually going and constructing data centers. They're acquiring power, land, and permits, and they're building data centers, oftentimes for hyperscale counterparties who become their customers. Sometimes they also have their own cloud product, where they can sell inference to startups or whatever.

Whereas Fireworks does not have its own data centers. It's actually leasing inference capacity and GPUs from partners. The thing that it's getting paid for is both the inference and what its software is doing on top of the inference to make it more cost-effective, reduce latency, and make it a better inference product altogether.

From the outside, you could say, “Oh, that's just 2 inference companies,” but they have completely different business models, capital intensities, and margin profiles. They're actually more different than they're alike.

I started my career at a software private equity firm called Vista Equity Partners, and the CEO, Robert Smith, would always tell us, “Software tastes like chicken, and that's why it's beautiful.” Every software company, when you look at the balance sheet and the P&L, has the same line items. The P&Ls look essentially the same at maturity. They're more alike than they're different, even if they're selling to completely different markets or offering completely different products.

That is not the case now. There are so many different business models that exist in AI. If you're an AI app creator, a foundation model lab, an inference platform, or a data center construction company, they're all over the place. It's hard to tell. The answer would actually be on a case-by-case basis, depending on the category and the company.

But I think there's a definitive advantage to investors and founders who have thought through the new taxonomy for an AI company. What it comes down to in my mind is the idea of P × Q × M. This is Econ 101: price times quantity times margin.

In SaaS, P was your ACV. What is the annual contract value of the contracts you sell to customers? Q was how many customers you had or that were in your TAM. M was gross margin, which was 70% to 90%.

Well, now in AI, if we take an AI app company, Q is probably the same. You're selling to the same people who would buy SaaS. M is almost definitively lower. For, I think, 99% of AI app companies, it's lower than 70%. But P can be immensely high. You have inference platforms that have 9-figure contracts with startups.

Mm-hmm.

There are very few SaaS companies that have 9-figure contracts with anyone, much less a startup. Now you have a lot of these AI companies with these unbelievably large contracts. I think understanding that taxonomy and what it means for the evolution of company quality, and how these things will look when they mature, is a very important dynamic that's still being figured out by everyone in the field.

So are you saying it's disorienting for making net-new investments, or also for understanding current portfolio companies and how they're growing and scaling in different ways, both marginally and in terms of their business models?

Definitely both. Thankfully, again, at Benchmark, the main thing we do well is partner with entrepreneurs extremely early. By doing that, you don't have to worry quite as much about a lot of these questions, because they have to do with things like what multiple the company is going to be worth when it IPOs or is sold to a company, and how it's going to use capital effectively at scale.

But when you're backing an entrepreneur at inception, at 50 Post, you've sort of already done your job. The company has already gotten to a level of scale and maturity that means you're probably looking pretty good. I think that's the case for a lot of the companies in our portfolio that are relevant to this conversation.

As they mature, and as we think about what categories going forward are going to have really big profit pools to go after, it is something that we care about. But I think the benefit we have is that great founders are always in style, whereas business models can go in and out of style, and things become popular and less popular.

You weren't supposed to touch hard tech, and now hard tech is apparently the only thing that's safe from the foundation model labs.

And consumer.

And consumer, because, yeah—

And sports.

—and sports. Yeah. Buying a baseball team is insulated from OpenAI and Anthropic, thankfully.

The nice thing is that at the center and core of all these companies, no matter the category, they all have amazing entrepreneurs driving them. I think the Benchmark model, as it's been since our founding, has been very entrepreneur-out versus theme-in.

Mm-hmm.

It would kind of suck to be the SaaS fund right now. I know a lot of people who have been the SaaS fund have reinvented themselves and said, “Oh, we're now the SaaS AI fund.” But that's a much tougher position to be in strategically than being the fund that's always said, “Let's back the greatest entrepreneurs, and then we'll figure out everything else along the way,” because those are the people who figure out those challenges and come out the other side even stronger.

Since it's not a thematic fund, but we're entering a new kind of state of the AI era where things are maturing, or—

Mm-hmm.

—Brad Gerstner thinks it's the age of inference. The companies that are going to be rewarded most now are downstream of that. Think about the AI agent economy, which you know well with Gumloop.

Mm.

Everything that's downstream of that, whether it's hyperscalers, clouds, power, or all the different kinds of connectors. Routers are now big.

Yeah.

Huge. People are going after those categories. How do you think about inference, too? And then maybe tie in Gumloop a bit, but this whole economy that's booming because of the proliferation of agents versus chats?

4. Inference Powers The Agent Economy

Yeah, 100%. The framework I have around inference right now is this: I was working with one of our portfolio companies that has an amazing mass of developer users and an amazing product, but they just hadn't figured out their business model yet.

I sat down with them and worked with them on how we were figuring out what the revenue model was going to be. The thing I said was, “It's sort of like you're on this riverbank and you have this bucket, and you want to fill the bucket up because it's the revenue bucket. But you're sitting on the riverbank, and over there there's a fucking waterfall.”

What you should do, instead of sitting on the riverbank wondering what to do, is walk under the waterfall. You don't know until you're under the waterfall whether there are holes in your bucket or how big the bucket is or anything like that. But the first thing you should do is get under the waterfall, and the waterfall is inference.

And just the absolute, unbelievable wave of revenue and demand for so many different businesses and business models that has come out of inference, I think, to your point, cannot be ignored. If you think about—again, we're huge investors in Fireworks—but if you think about the other inference platforms like fal, Baseten, and Modal, a lot of app companies that are using a usage-based model or an outcome-based model, all of these are different derivations of monetizing inference. And so when you hear about all these companies and all of these businesses that are going not 1 to 3 to 9 to 20 like they used to, but 1 to 20 to 100 or 1 to 30 to 300, all of that can be traced back to the enablement of a business model via inference.

Because the business model used to be, “We're going to charge some raw dollar amount times some amount of heads within your company,” and now it is, “Hey, we're going to charge some calculation of making a margin, essentially, on the inference that we're reselling you.” Some of those business models are a little bit thinner and are just actually reselling inference as a broker, and some are really abstracting it away. If you think about a Sierra that is outcome-based pricing for actual completed deflections of customer support queries and things like that, that's largely abstracted, but you're still monetizing this thing which is based on inference. And what that allows is that it just removes any sort of rate limiter that you have on your potential to grow, and it's the enabler to these unbelievable revenue growth rates that we're seeing time and time again with all these new AI companies.

What do you think is going to happen with the proliferation of agents?

I think it's interesting because, on one hand, I always cringe a little bit whenever a word or a term becomes over-product-marketed.

Yeah.

FDEs are certainly a version of this, and I think agents are absolutely a version of this.

What about AI?

Yeah. Well, I mean, AI is at least this umbrella term. But when private equity firms are telling all of their portfolio companies to say that they sell agents, that's when you know the term is cooked. We need to move on from this term onto something new or something different.

That said, it is still fundamentally the greatest product innovation and business model innovation that I think we've seen since the start of SaaS, since the cloud business model. And the reason for that is because it is the underpinning of how we move to this idea of selling work, or actually replicating what a human being is doing when they're trying to complete a task.

On the business model side, as I mentioned, it allows a buyer of software to move their mental framing from, “Oh, I buy a license,” to, “Oh, I buy intelligence on tap,” or a white-collar activity on tap, or some activity that creates economic output for my business on tap via an API or via a software product. And so it unlocks this new price-for-value equation for buyers that allows them to think about software and the budget that they have for software in a whole different way.

I think one of the things that we got insanely excited about when we were first tracking the rise of coding agents was when Cloud Code—even when it was still Cloud CLI, but when it became Cloud Code—we were talking with developers and companies within our portfolio, and they were like, “Yeah, we have developers that are spending $3,000 per month themselves, each, on Claude Code.” And so it's like, wow, okay, so that's $36,000 per developer.

Yeah.

Oftentimes in the SaaS era, if you had a $50,000 ACV, that was okay. But instead of having a $50,000 overall contract value with this customer, you were now getting that per developer, and it was still growing. And so it was just this idea that, wow, this could actually become not a $200,000 line item for the average company, but a $20 million line item for the average customer and average company.

Or for some.

Or for some, $500 million a month.

$500 million a month.

A month. That's a different era of technology. And so I think, even though I'm so annoyed at the whole agent thing because now it's just the only thing that you can see when you go online—

Yeah.

—and it's been marketed to death, it represents the most important product and technology shift, and then also just raw revenue and business model shift, we've ever seen probably in the history of technology. And I think that was one of the reasons—and we can go into Gumloop—but that's interesting. That was one of the reasons we also got really excited about Gumloop.

Yeah. Explain Gumloop.

Yeah. So Gumloop is an independent, third-party software platform that does a collaborative AI agent and AI automation canvas for enterprises. What that basically means, at a very simple level, is that we first saw it in coding and coding agents. With Opus 4.5, coding agents got to a place where developers felt like they could do a vast majority of their day-to-day job by offloading the work—not just tab-to-complete like we saw with Cursor, but actually offloading a bulk of their work to coding agents in Cloud Code, Codex, or Cognition, or any of these other products.

The thesis was that what we've seen in code is going to happen in most white-collar job functions, and eventually probably most blue-collar job functions as well. The average white-collar worker is going to be doing an immense amount to automate the rote, redundant work in their day-to-day via agents and via AI automations.

That's what Gumloop provides. It's a platform for enterprises, and every single employee within an organization—not just developers, not just salespeople, not just marketers—can create, iterate on, and collaborate on both very simple automations, like you did with Zapier back in the day, but then also full-on AI agents that they can build and use cross-functionally across the organization to do all sorts of things, either triggered by someone talking to them in Slack or Microsoft Teams, or running in the background without anyone having to trigger them at all.

Obviously, there are going to be—and there already are—great products like Cloud Cowork, and the news around OpenAI combining Codex and ChatGPT into a productivity product that are going to have big businesses in this general category as well. But we think that there's an immense amount of—

And this goes to what you were talking about with routers becoming a popular term as well in the agent space. We think it's really, really important that there's a third-party, independent vendor for 2 reasons. One is that the models are actually pretty jagged. The models at any given point are good at different things.

Gemini is the best multimodal model. Claude typically has the best coding model, although now a lot of people think GPT-5.5 is actually better than the latest Opus model at coding. And then you have open-source models, which typically have immense value for the cost that you're paying for the inference of those models. And so there are great models for all these different tasks, and you do need to have a third-party vendor that's sort of watching out for the customer and not just saying—

In some of these enterprises, you have employees checking the weather with Opus 4.8.

I know. Well, I’ve talked to a couple of companies about this already—how do you control spend on your tokens, and the token-maxing debate and all that kind of stuff? Because when I’m prompting Claude, whatever I’m using, it’s usually the dumbest thing. It’s a small fix. I could probably just go in myself, but the tool is magic, so I want to do it myself and make that small fix. But you do that over huge pieces of content or whatever—

And thousands of employees.

…and thousands of employees. It clearly aggregates very large bills and great revenue sources for some people.

That’s right. I’m asking Opus 4.8 what I should have for lunch, you know? That’s probably not the best use of our inference budget.

But it’s interesting. I know you guys are investors in Lagora.

Yeah.

I was at Harvey the other week, and—

Mm-hmm.

…and they were talking about the same thing. It was kind of similar to how text messaging was before.

Mm.

You would pay per text.

That’s right.

And now it’s happening within our chatbots—

Mm-hmm.

…and within the different kinds of prompts that we have. How are you seeing this? How do you think the business model of that will play out, and will it get more efficient?

Yeah. I think it all depends on 2 different things. I have something that I call—there are many other mom tests out there.

What’s a—

It might even be trademarked. I don’t know.

A mom test?

My personal AI mom test—

Okay.

…is—and not every mom is like this—but my mom lives in a rural suburban town in Colorado, and she’s not exactly an AI native.

Mm.

She doesn’t know how to reset her password on her iPhone, for example. She’s not the most tech-savvy person. And the thing I always ask is: For my mom, what is the amount of queries, or the amount of things, that she needs out of AI that can’t be done by a really, really, really cost-effective open-source model? I started asking this 2 years ago, and there was actually, “Oh, well, I don’t know.” Now it’s 100%. There’s nothing my mom actually asks of her AI products that needs to be done by a frontier or even a near-frontier model.

Mm-hmm.

And you can start layering that out. What about a high school student? What about this person or this person, or this professional in this area? Depending on your use case and who you are and what you’re trying to get out of AI, you might need the frontier. Again, I think what we’re seeing in the market right now is that people are willing to pay a huge premium to access frontier intelligence to this day.

But there’s a growing portion of tasks in the economy and within these enterprises where you just don’t need a frontier model, and oftentimes you don’t even need anything close to a frontier model. And so I think what you’re seeing—and Cognition published some work on this as well—is that they actually—I don’t know if they trained it; I think they post-trained an open-source model on tasks that they saw done within Cognition where there was low task complexity. It didn’t need frontier intelligence.

They were extremely popular things happening within Cognition. People were maybe just defaulting to using a frontier model for this thing, and they were like, “Wow, you’d actually save an immense amount of money if you took it away from the frontier, because it’s a very simple gated task that doesn’t need frontier intelligence, and you’re going to save 95% on that action or that query.”

Mm-hmm.

So I think, as the models get better and better, there’s a growing proportion of queries or actions, or whatever you want to call the use of inference that’s going on, that doesn’t need the frontier. That said, I think what’s also growing is usage for frontier intelligence. I don’t think we had unbelievable-quality coding models until Opus 4.5, until last winter, which is why you saw Anthropic’s revenue and Claude Code go so parabolic, because there was a genuine breakthrough in the usability of those models.

Yeah.

So it’s not that frontier intelligence isn’t growing, and the demand for it—and the ability and willingness to pay a premium for it—isn’t growing. It’s just that there are a lot of non-frontier tasks as well, and demand for those is also growing immensely.

My partner Eric Vishria had this really great tweet where a lot of people try to make this a very zero-sum thing. Is open source going to win, or are Anthropic and OpenAI going to win? And I forget all the things he went down, but it was basically: on-device inference, yes; open-source inference, yes; proprietary models, yes. It seems like there’s a lot of demand for all of this stuff, and the demand is all going parabolic.

So it’s not this zero-sum game, at least yet, in terms of either it’s all going to be open source or all going to be a frontier model from 1 of the 3 frontier providers. It turns out there’s use for all of it, and it’s all growing very, very quickly.

So I guess to push on that a little bit: to your point earlier this year and even to today, OpenAI has crazy amounts of revenue. Anthropic was last rumored and reported around 45 billion, and now people are speculating around 60 billion. Both are trying to go public, as are SpaceX and xAI. What happens once those models become efficient?

Yeah.

What do you think is going to happen to the scale? Do you think it’s going to continue? What do you think is going to happen?

5. Frontier Labs Face Open Source

Yeah. It’s the $1 trillion question—or the multi-trillion-dollar question if you combine SpaceX AI, OpenAI, and Anthropic into that equation.

And I think it depends on what the next 12 to many years look like. Do you believe we’re on a path to recursive self-improvement and this future of a bunch of geniuses in a data center? If you do have a bunch of geniuses in a data center, if the frontier goes that high, then you legitimately do have probably a lot of pricing power and a lot of ability to continue to grow and continue to monetize, and probably even reaccelerate growth if you’re a frontier lab.

On the other hand, if at any point it seems like capabilities are actually hitting an absolute ceiling, and distillation continues as it has historically, and open source gets 95% as good as wherever the ceiling of capabilities tops out, that’s a really scary situation for the frontier labs. I don’t think it’s a death knell for them because, again, most of the users of ChatGPT would use ChatGPT whether or not there was a GPT model in there or not. They like the product.

Yeah.

Most of them don’t even know. They wouldn’t be able to tell you if the model’s different. In our little bubble in San Francisco, people would be able to tell you, but the vast majority of the 900 million weekly active users who are—

They have no idea what 5.5 is.

No idea. They wouldn’t be able to tell you. If you swapped it out for 5.2 or something—

Mm-hmm.

…they wouldn’t be able to tell you. So I think that scenario wouldn’t kill them because people like the products, and both OpenAI and Anthropic have done a tremendous job of building amazing products on top of their models.

But it would certainly be a very, very different situation because open source would have a much greater impact, depressing their ability to have pricing power and charge a premium for the tokens they’re producing. Open-source inference also costs money. It’s not like open source means free. Someone still has to run the GPUs, build the data centers, and operate the data centers. It’s more about how much premium margin the frontier labs can create.

And I don’t think we’ll know until it’s very clear whether or not we actually have RSI—whether there are going to be geniuses, or God, in a data center, or whatever you want to call it—in which case we’ll have different problems to think about. Or if we do end up topping out on capabilities at some point in the next few years and distillation continues, I think it’s much harder to garner a premium margin if you’re a frontier-model company.

It becomes much harder.

6. AI Funding Rewrites Venture Capital

Given the scale and how fast things are running, another common question is: How do we fund this?

Mm-hmm.

There are different types of funding happening. I think we’re literally across from General Catalyst, and they have their CVF fund, and there are people who are going after debt for tokens—

Mm-hmm.

—and trying to create the right mix. But the undercurrent of all of this is that companies are raising money faster than ever before. One of the questions in our conversation beforehand was—

Yeah.

—whether we talk about the bubble. I don’t think the bubble is very interesting.

Yeah.

I’m most interested in what the biggest concerns are right now—

Mm.

—with that. Because, reading through the cracks of all of it, right—

Mm-hmm.

—it’s like, I think we all know when companies are running off the tracks—

Mm-hmm.

—and when they’re overfunded to an extent. But then again, there are total Hail Marys—

Yeah.

—where these companies are outperforming what you thought they were doing, and so they raise another round.

Mm-hmm.

And then 6 months after that, they raise another round, and then they just keep on getting more and more capital under them. At this point, the funding ecosystem has changed dramatically.

Mm-hmm.

The playbook is out the door.

Yeah.

With a firm like Benchmark, which is fundamentally focused on the early stages, how do you capture that value at the early stages and continue that?

Yeah. I think it’s something that we talk about a lot internally, because you’re right that even the funding market—it’s not just the change in the funding market and the venture-growth asset class. The trend that everyone talks about is companies staying private for longer.

Mm.

That’s been the prevailing trend that everyone talks about. You have—I mean, name your company. Obviously, some of these companies are now going public, like SpaceX. But before it was, “SpaceX is never going public, Stripe is never going public, and Databricks is never going public.” If this was 2005, all these companies would have been public for years. Now there’s just so much capital available for them in the private markets via these venture-growth platforms that they stay private essentially indefinitely.

I think a more recent change is that, as you’re mentioning, AI is just so different because, depending on what you want to do, oftentimes there are massive day-one costs.

Mm-hmm.

If you want to create a Neo lab and you have some incredible research direction that you want to pursue, but you need $2 billion of compute to see whether it’s going to work or not, that’s very different from Airbnb raising the $500K or whatever they raised at YC in their seed to see if Airbnb would have any product-market fit. It’s just a completely different equation and type of funding mechanism, and just a completely different capital market.

Also, I think what you’re seeing now, because of the staying-private-longer thing, is that you have these crazy situations where you almost see a rebirth of a company. Oftentimes, people think, “Oh, well, at the early stage, you get high multiples on money, but it’s higher risk. And then, at the late stage, you can get 2 to 3X, but it’s much lower risk.” I think the crazy thing is that, because these markets are so big in AI and elsewhere, and because these companies are staying private for much longer, you actually have situations where a late-stage company can have much higher upside than a Series C.

Mm-hmm.

Which is super weird.

So weird.

It’s so weird. My first investment when I was at Kleiner Perkins was SpaceX, and it was at an over-$100 billion valuation. Talking with peers and friends at the time, they were like, “Man, are you gunning for a 2.5X? How much upside could there possibly be if you’re getting in at a triple-digit-billion entry valuation?”

I think what ended up being true for SpaceX was that you had this thing called Starlink that had started a couple of years before that and had really started to scale. That was a rebirth of the company. If you look at the P&L today and in the S-1, the vast majority of the business is their consumer broadband business and B2B broadband business via Starlink. It’s not even a launch business.

I think the 2 things that we grapple with and think about a lot are: How do we continue to be the best venture capital firm that partners with entrepreneurs very early in their journeys and is their most meaningful partner when you have a much, much more capital-intensive early life cycle for some of these AI businesses? And when you have these businesses that actually go through these transformational moments at later stages, where it feels like even though they’re 15 years in, they’re still only 1% done with their journey?

I don’t have a great answer for that. I think we’re extremely flexible in what we do. We don’t say that we just do seed, or that we just do Series A. We partner with the very best entrepreneurs in the world when we think that we can deliver unbelievable upside for our LPs and be the most meaningful partner to them. Those are the constraints that we think about. The trends that we’re seeing both in AI and in capital markets writ large definitely expand our thinking in terms of what actual rounds and what types of situations are applicable to a firm like Benchmark.

There are definitely new types of funding vehicles, too. Private credit has really taken over where banks left off, and there are different kinds of funding to add into it. So how has venture capital, in the grand scheme of the alternative asset class, changed, and how has the role changed?

Yeah. I think this is where things get tricky when people are having these discussions: These firms can completely change what they are, and yet we still just use the term venture capital.

Whenever I write an occasional blog post or something, I always say “venture-growth,” because venture capital is something very particular. Even when I’m discussing something like Databricks, I say that should be called something different, and I’ll call it venture growth. That’s one way to do it.

I think a lot of these firms have just become alternative asset managers, and that is what they are. They have venture products, but they are not venture capital firms. That’s not me saying, “Oh, they’re not VC firms because they’re bad.” I think a lot of these firms are wonderful. But General Catalyst and Andreessen Horowitz are alternative asset managers because they have so many products.

Many of these places have—if you think about ICONIQ, ICONIQ has a great growth-stage venture practice, but it also has a high-net-worth wealth management practice. That’s a very different thing from just being a venture capital firm.

As these firms have evolved, one of the things that I think, as an industry, we haven’t done a good job of is evolving our nomenclature and the terminology we use to talk about these things. General Catalyst still has a venture product, and they also have a growth product, a debt product, a health assurance product, and a wealth management product because they’re an alternative asset manager.

I think that venture, in many ways, is still the same, but it’s a product now for many of these firms. It’s not the firms themselves.

We’ve definitely alluded to this a bunch during the conversation, but I’m really curious to get your perspective. The public markets are seeing the largest IPOs of all time—

Yeah.

—and it’s all apparently going to be happening this year. How do you think the markets are going to take that? Will it be absorbed? Will there be aftershocks? How will that actually affect this asset class?

Yeah. I think in a few ways. The first is, I did this very quick analysis because I thought, I know that using a single example with Anthropic...

Mm-hmm.

After the $380 billion round, I was like, “Man, I know this is so much bigger of a deal than we’ve ever seen in a growth round.” I also know that we’re not talking about it enough—just how different it is.

I did this analysis where I charted out, over the last 10 years, 4 of the best pre-IPO investments ever in terms of both their raw scale—they were big rounds—and the fact that the investors did really well in them. I think it was Slack, DoorDash, Snowflake, and Nubank. Typically, in those rounds, the total round size for the pre-IPO round was $500 million to $2 billion, and the investors over a 4-year period made from 2× to 5×. Everyone went home happy with that. That was a great outcome for all involved.

The Snowflake pre-IPO round, over a 4-year period, ended up being something like $500 million turning into $2.5 billion. That’s insane. That’s amazing. When you think about the Anthropic $30 billion round at a $380 billion valuation, if Anthropic was to go public and get liquid at a $1 trillion to $1.5 trillion valuation—which, you know, they just raised at about a $1 trillion valuation—I don’t think people would think that’s an ambitious estimate. I think a lot of people would have their estimates higher than that.

The gross return of their $30 billion round at a $380 billion valuation would be 35 times that of the Snowflake pre-IPO round. It’s 35 Snowflake pre-IPO rounds in a single round. I have friends—I know several people who have $3 billion to $4 billion invested in Anthropic—and it’s so hard to even talk about. These numbers are so big that it’s hard to comprehend, because 5 years ago, a growth fund was a billion dollars.

Yeah.

Until COVID brought all these growth fund sizes up, a normal growth fund was a billion dollars. Now there’s an individual company that these people have $4 billion invested in, and they might return 5 times their money in less than 5 years on $4 billion. It’s unbelievable.

At least SpaceX took 20-plus years to make everyone rich and put all this money into the ecosystem. I just don’t know if people really understand and are ready for the impact that this much liquidity could have on the ecosystem. I think the one place where people are definitely predicting it and are already seeing it is in the San Francisco housing market—

Oh my God. Yeah.

—where everything’s going for 2× the asking price at this point.

All cash.

All cash—

Or in equity.

—or in lab equity. But I think there are so many knock-on effects in terms of what those employees do, what they invest in, what causes or companies they then invest in. Do they start new companies? Do they stay at these labs? There’s just going to be a shock that’s going to impact every single aspect of life—both practical life in Silicon Valley, but also the lives of employees, investors, and founders in the ecosystem.

Oh, absolutely. I’ve done 2 episodes with Michelle Del Bono, who, to your point, at Andreessen Horowitz has multiple products. He runs their multifamily office for Mark and Ben and the principals there. So it’s the wealth-management division for them, and each time I talk to him, I’m asking because there’s going to be a huge wealth event—the most wealth creation probably ever—because of all these things happening simultaneously.

One of the biggest questions is, what do you do when 90% of your net worth, probably more at this rate, is in one position?

Right.

He has some really great answers to that.

So as we wrap up, I have a couple more questions, but one of the things that was super interesting when I was doing research on you all was how diverse your portfolio is, especially in the AI mix. You have StarCloud—data centers in space—named in SpaceX’s S-1. Then you have Cerebras, which just went public; Sierra; LangChain; Lagora; Fireworks, which we talked about a bit—

Mercur.

Vanis.

Vanis.

Mercur.

Mm-hmm.

Yeah, Mercor—interesting. I just listened to his podcast with Harry Stebbings, and I was really surprised by the positive things that are going on there after the data breach.

That’s right. He cleared a lot of air.

Good for him, honestly. HeyGen, Gumloop, which we talked about, and Exa. Because of this whole mix—and you guys have a concentrated strategy—how do you come up with a portfolio of this type? They’re all very mature. They’re all doing really well. What is the magic between all of that and the firm?

7. Benchmark Bets On Founders

When I was coming in—before I joined Benchmark—I was like, “Wow, they’ve done such an amazing job thematically.” They have the vertical AI winner in Lagora, the horizontal AI winner in Sierra, a data-infrastructure play in Lagora, a developer play in LangChain, and a prosumer play in HeyGen. You can keep going and keep going: orbital data centers in StarCloud.

Then I joined, and I realized, “Oh, wow.” There was literally no thought of, like, “This is the category that we’re going into.” It was truly what we talked about before: it was all founder-led. It was all based on the founder. Did we resonate with the idea that the founder was pitching? Yes. But there was zero thought to, “Oh, this is going to be a big category.” It was just, “This is an unbelievable entrepreneur who is going to put their energy into an idea that they’ve convinced us is really, really interesting.”

Even in some of those cases, some of these companies were pivots. There was no way that we were super smart to know that some of these companies were going to be in big categories, because they weren’t even the categories that we invested in. Again, I think that’s the benefit of Benchmark: we don’t invest in categories; we invest in people.

When Chetan invested in StarCloud, we closed the investment weeks before Elon started professing his love for and bullishness on orbital data centers.

You guys have the partner meeting. You’re like, “Wow, look at that.”

Oh, we definitely—

Okay.

I forget which platform it was on, but he did a long interview with somebody, and half the interview was on orbital data centers. We started sending it around our group chat. We were like, “Oh, man, this is about—”

This space is about to get really hot. And it was so ironic because, again, there's such a heated debate: Are orbital data centers the future? Are they going to be too hard to pull off?

We obviously discussed both sides when we were talking about the investment, but the investment was about the people. It was about Philip and the team that he had built. One, they already had a GPU working in space.

They had proven out their mettle, and they were the only company that had a GPU in space. There were already some things that we saw in their traction, but it was based on the people. I think when you focus and have your strategy around that, great entrepreneurs are always in style.

They always pull rabbits out of hats, and they end up being magnetized and drawn toward these really massive categories. They often end up pioneering them.

One element of performance that I think is very valuable is who you surround yourself with. You've been at so many legendary funds at this point, surrounded by the top investors over and over again. From your standpoint, who has been a mentor to you? Who do you keep as a guiding figure as you continue to evolve and grow in your career?

Yeah, I wake up every day and pinch myself because I've just been so unbelievably lucky and fortunate to, as you said, work with a lot of the people that have defined the history of venture capital and growth investing. Not only have they been extremely good at their jobs, but basically every single one of them has been a wonderful person. I feel like I learn something different from every single one of them.

If you take Napoleon Ta, who is a GP at Founders Fund and runs the growth investing practice there, I learned a ton about investing with him. But the most important thing that I learned from Napoleon was his focus on family and work. He had a very simple life.

He doesn't do podcasts. He doesn't network. He doesn't try to spend a bunch of time with other investors. He goes to work, works really, really hard, and then spends time with his family.

It's this beautiful, simple life, and he loves both spending time with his family and working with the Founders Fund team. It made me realize that and helped me prioritize my time and how I think about the important things in my life.

I think another one, currently, is Eric Vishria. I think he's just an incredible role model and leader within Benchmark. Obviously, when you look at his results—he's number 3 on the Midas List this year—he's the best of the best. He's a Mount Rushmore venture capitalist at this point.

But you spend an hour or 2 with him, however long, and he is just an extremely kind, hardworking, completely normal person. It's clear that he's unbelievably smart, but he's almost unassuming. He's not trying to prove anything to you. He's not trying to make you feel small or make you feel like he's the smartest person in the room.

He's worked really hard. He works really smart. He brings a kindness and a normalcy to his conversations and his work with his entrepreneurs and his partners in this room that motivates and inspires me to do more of the same.

There's not some alchemy to doing this job really, really well. If you work hard, know the direction that you need to go, are a great partner to entrepreneurs, and work really hard for them, that ultimately is the job. If you just do that and get some lucky breaks over your career, you can end up being one of the best VCs of all time.

Those are the 2 that come to mind, mostly because I think they've taught me more about lessons outside of the 9-to-5 than inside it. But every single person I've worked with, I've been very fortunate to do so because they've all taught me so much.

Amazing. It's a sophisticated answer compared to the typical growth investor who chooses Charlie Munger.

Yeah. Unfortunately, unlike David Satter, I've never had dinner with Charlie Munger. Otherwise, maybe it would be Charlie instead.

Oh, amazing. This is a great place to end. Thank you so much. I had so much fun talking about everything—AI and who knows what actually happened.

Exactly. That's the best part. We can look back at this in a year, and I've probably been wrong about everything.

What is an agent anymore? I don't know.

It'll be some new term.

Well, thank you so much, Ev.

Yeah. Thank you, Molly. It's been great.