20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO
Harry StebbingsErik Bernhardsson
- Modal has reached $1 billion ARR, but scaling compute keeps cash flow negative. Stebbings reports a $750 million Accel-led raise at $15.75 billion; Bernhardsson says Modal is profitable on a P&L basis while funding GPU capacity. After 15X growth in 12 months and continuing 30–40% monthly growth, he hopes revenue will be “way beyond” $4 billion in another year.
- Inference’s margin upside comes from selling a platform of products. Selling “tokens for dollars” creates easy switching and a “race to the bottom”; sandboxes, training, reinforcement learning and workflows give customers more reasons to stay. Bernhardsson thinks Modal could eventually reach 40%, 50% or 60% gross margins, while rejecting a return to 90% software margins because customer spending must support multiple infrastructure layers.
- Modal’s capacity aggregation turns unpredictable customer workloads into more manageable aggregate demand. It sources from 25 neoclouds, reserves capacity ahead and sells usage per GPU second, letting customers avoid multiyear commitments. With procurement lead times of three to four months, growth requires buying roughly two to three times projected capacity; hardware prices eventually fall “like gravity,” but Bernhardsson cannot say whether today’s increases last another year or five.
- Nvidia’s ecosystem matters more than CUDA itself. Bernhardsson calls CUDA’s moat “a little overstated,” pointing instead to Nvidia-first support across frameworks, kernels and inference engines. He expects specialized chipmakers and Nvidia both to grow, while admitting that 15 years of predicting the arrival of non-Nvidia accelerators have repeatedly proved him wrong.
- Bernhardsson rejects the claim that open-source inference has peaked. He thinks open and frontier-model usage can each grow “5X, 10X, 20X” over the next 12 months, supported by different requirements for customization, control and performance. He also sees an opportunity for a $100 billion specialized-model inference company and says he would have invested in the specialist discussed at $8.5 billion.
- Sandboxes demonstrate how infrastructure demand can arrive well after a product launches. Modal launched them in 2023, but demand accelerated in summer 2025 as models improved at generating code and whole applications; the business now runs at a $300 million annualized rate, roughly one-third of revenue. Bernhardsson hopes the newly launched training product could reach a $100 million run rate within six to 12 months.
- Reinforcement learning’s clearest advantage lies in tasks with verifiable, relatively short feedback loops. Coding and math fit; engineering decisions whose consequences emerge months later may expose limitations, although Bernhardsson repeatedly allows that he could be wrong. He remains very bullish on software engineering demand because greater productivity expands the need for software and shifts work toward understanding human requirements.
- He is almost certain there will be a 20–30% stock-market pullback. That does not change his view that the long-term opportunity is “so fucking big.” Separately, he says he would be nervous about companies dependent on continual equity financing; his operating rule is to raise every round “as if it’s your last round,” with a real path to commercialization and revenue.
1. Modal’s capital needs have shifted from operating burn to GPU capacity
Bernhardsson began coding around 1992, aged eight, but waited until 36–37 to found Modal, partly because he lacked confidence. His motivation remains building rather than an exit: “I don’t wanna sell this company ’cause I would just go and start Modal 2.” He rejects the “eating glass” description of entrepreneurship and hopes to work on Modal for 20, 30, 40 or 50 years.
Stebbings introduces Modal at $1 billion ARR, following a $750 million Accel-led financing at $15.75 billion. Bernhardsson distinguishes operating profitability from cash consumption: Modal is profitable on a P&L basis, but cash flow is “quite negative” because capacity requires upfront spending. Until roughly last year, monthly burn was only a few million dollars; capital now helps, although he resists calling it a moat.
Fundraising has relied on relationships rather than pitch decks. He had spoken with Rich and Gonzalo at Accel for at least a year before discussing terms. Earlier-stage founders should resist optimizing solely for price: VCs often want to invest more than a company needs, creating unnecessary dilution. Raising less at a lower valuation can preserve ownership and “keep the VCs hungry for the next round.”
2. Speed and personality matter before a startup earns structural defenses
Bernhardsson’s objection to casual moat language comes from a stricter, Seven Powers-style definition: “You never have moats in the early days.” His preferred analogy is Ben Thompson’s description of Snap’s gingerbread-man strategy: “They just run faster.” Stebbings offers his own investment criteria—top-1% teams, speed of product decision-making and brand—which Bernhardsson broadly accepts while remaining reluctant to call them moats.
Modal’s brand, in Bernhardsson’s telling, starts with developer experience and bottoms-up adoption. His Spotify background shaped a B2B product whose website, visual identity and first-use experience should feel distinctive. He thinks Modal has stronger community enthusiasm than Baseten and Fireworks, while stressing that he is “paranoid about competition” and closely follows what people say about those companies.
His memorable benchmark is the founder willing to wear a “fucking pink fur coat” onstage: personality cannot be copied simply by cloning features. Stebbings pushes back that Modal lacks a comparable founder-led following. Bernhardsson accepts the criticism; early growth involved extensive Twitter outreach asking people to try the product, with some of those users becoming angel investors. “You have to lean into what you’re good at.”
3. Inference becomes more valuable when it comes with adjacent tools
Bernhardsson calls inference “the new cloud,” but argues that inference alone is insufficient. Post-training, sandboxes, agentic tools, low-latency serving and batch inference belong together. Otherwise, providers are merely exchanging “tokens for dollars,” with little switching cost and a “race to the bottom.” He says the infrastructure supporting a suite of products is difficult to build, but explicitly adds that it is “not a moat.”
Stebbings asks whether industry gross margins of roughly 25–40% could reach 70%. Bernhardsson instead sees a possible 40%, 50% or 60% for Modal. His reasoning is “almost like topological”: a customer’s dollar passes through applications, models, infrastructure, cloud and Nvidia, with different layers taking margins. “The 90% software margin businesses are gone,” he argues, even though the market leaves substantial room to make money.
Specialization does not contradict breadth, he says when Stebbings challenges him. Inference is “10X or 100X bigger” than he expected, allowing distinct products for latency, throughput, batch processing, audio models and training. Modal can host multiple specialized products on one platform. He also expects unfamiliar technologies and new hardware that use fewer resources, rather than today’s serving architecture remaining fixed.
4. Pooling GPU supply and customer demand is Modal’s operating advantage
GPU procurement at Nvidia’s March GTC conference resembled “an open-air drug market, but GPUs instead of drugs,” Bernhardsson says: hundreds of vendors and constant capacity negotiations. Modal works with 25 neoclouds and handles that procurement for customers. His longer-term ambition is to aggregate more of the world’s compute while aggregating demand, then manage the two together efficiently.
Modal built its own file system, container runtime and scheduler rather than using Kubernetes and Docker. Bernhardsson says this lets it treat distributed compute as one pool. Individual customers have unpredictable bursts, but thousands of customers together produce smoother demand. The customer proposition is therefore not just access to GPUs: “You don’t even have to do capacity planning.”
One customer launched overnight on Modal, and “we weren’t even aware of it”: the platform automatically supplied the GPUs. Customers can scale up and down, paying per GPU second instead of committing to a fixed block for three years.
Procurement still requires forecasting. GPU lead times are about three to four months; Modal extrapolates aggregate demand using monthly growth rates that have been 30–40% for the preceding year. Bernhardsson says that means securing approximately two to three times projected capacity ahead of demand. Diversification across customers makes the aggregate easier to predict than any individual account.
5. Compute prices eventually decline, but the timing remains unknowable
Bernhardsson rejects the idea that capacity is simply unavailable: Nvidia is shipping enormous volumes, and buyers willing to pay more can obtain GPUs. Modal is “more long than we’re short GPUs,” reserving capacity for longer periods and renting it for shorter durations. He thinks higher underlying prices can be reflected in customer prices because willingness to pay also rises.
Asked what happens to H100 pricing, his answer is unusually direct: “Dude, I have no idea.” Hardware prices fall over the very long run—“it’s like gravity”—but shortages have pushed prices upward. They may continue rising for another year, two years or five years; he does not know. Eventual supply-demand convergence supplies no precise timetable for investors.
Stebbings argues that overbuilding is impossible within five years; Bernhardsson replies, “Probably not,” rather than making an unconditional forecast. Stebbings says Instinct allegedly does not yet charge or make revenue; Bernhardsson uses it as an example of potential future monetization. Companies can use a lot of GPUs before earning revenue, and that lag could become a constraint.
6. Nvidia’s ecosystem and neocloud financing require different capabilities
Specialized chipmakers do not automatically weaken Nvidia, in Bernhardsson’s framing. He expects Nvidia to keep growing while companies such as Cerebras and Etched also find markets. When Stebbings asks why Cerebras is down 30%, he offers no stock explanation: “I don’t know.” He invests passively in indexes rather than spending time following individual stocks.
CUDA itself is “a pretty annoying programming model”; the stronger defense is the surrounding ecosystem of optimized kernels, framework support and inference engines such as vLLM and SGLang. “Everything is NVIDIA first today.” Bernhardsson has predicted a breakthrough year for non-Nvidia accelerators for roughly 15 years and admits, “I’m always wrong.” Trainium, TPUs and AMD are improving, but he still thinks they are behind.
Neoclouds are both partners and potential competitors. He is skeptical about their moving into inference because Modal’s five-year engineering effort is difficult to replicate: millions of sandbox placements per minute, thousands of GPUs scaling and hundreds of thousands of concurrent tasks. His respect for CoreWeave and Nebius centers particularly on financial engineering—securing customer contracts, financing chips against them, delivering capacity and repaying debt over subsequent years.
7. Open models can expand alongside frontier providers
Stebbings presents Jason Lemkin’s bear case: cheaper frontier models and US enterprises’ reluctance to use Chinese models could mean peak open-source inference usage. Bernhardsson calls that “a fixed pie size mentality.” He thinks both categories can grow “5X, 10X, 20X over the next 12 months and beyond”; when Stebbings repeats the timeframe, he answers, “Yeah, for sure.”
Open models retain demand from customers who want fine-tuning, parameter control or particular latency and throughput characteristics—“It’s like Windows versus Linux.” Bernhardsson also says open-model providers are moving toward licensing arrangements that capture more value. In his reasoning, that revenue can finance better models, rather than open development necessarily losing to increasingly inexpensive proprietary alternatives.
He wants more European and US models alongside the Chinese labs he respects. Importantly, he explicitly retracts his initial implication that Europe lacks human capital: it has “an insane amount.” The weakness is aggregating talent inside one company and assembling large amounts of capital. He views Nvidia’s sponsorship of American open-model companies as a smart investment in its ecosystem.
Market size also supports narrower models. The classification-and-regression specialist discussed has, in his view, an exceptional brand and community; he would have invested at $8.5 billion. His broader bull case is that “there is an opportunity to create a $100 billion company with a lot of specialized models in the inference.” An exit or public listing are possibilities, not promised outcomes.
8. Sandboxes rewarded early infrastructure work and favorable timing
Modal’s serverless-function product launched in 2021 and found strong demand in 2023 through generative-AI inference; Stable Diffusion was not an anticipated killer application. Sandboxes launched in 2023 but remained relatively insignificant until summer 2025. Bernhardsson tentatively associates the inflection with “Sonnet 4.5, one of those,” as models became good at generating code and building entire applications.
Sandboxes now represent a $300 million run-rate business, about one-third of Modal’s revenue. Overall revenue grew 15X in 12 months, and Bernhardsson says monthly growth of 30–40% has not slowed. He hopes revenue will be “way beyond” $4 billion another year out. Stebbings’s cheap-valuation argument uses a hypothetical $4 billion revenue and $1.2 billion bottom line.
The resource-allocation mistake was underinvesting in open-source LLM inference. Modal initially regarded it as a commodity race to the bottom; DeepSeek changed the urgency. Bernhardsson recalls the reaction as “Fuck”—then staffing grew from one or two people to approximately 30 over 12 months. He now describes Modal as one of the top three providers.
Training has just launched, and he hopes it could reach a $100 million run rate in six to 12 months. Batch processing, reinforcement learning and other adjacencies are attractive, but Stebbings challenges whether breadth repeats the overextension Bernhardsson criticizes in neoclouds. He concedes the allocation problem: “You need to be extremely ruthless,” taking on only one or two directions at the appropriate time.
Bernhardsson says infrastructure requires trusting your gut about useful primitives: look at what engineers need, but do not follow every request. His approach is to build good building blocks that can “explode” when demand arrives. He also wants a broader customer experience, with more products and more FDEs and sales reps proactively helping customers—for example, by offering bio-related examples such as protein folding.
9. Reinforcement learning may encounter limits at longer feedback loops
Post-training works well where tasks have “fairly finite” feedback mechanisms, Bernhardsson argues. Code can run inside sandboxes—something labs already use Modal for—and math can be checked with Lean. Cybersecurity is another promising area; law may permit useful evaluation “to some extent.” The current focus reflects human tasks that have become reachable for agents, rather than proof that every task fits the same method.
His tentative boundary is time: agents will probably automate work they can complete within 24 hours or a week, while engineering often involves much longer feedback loops. He has “a feeling” reinforcement learning will encounter limitations there. But he preserves the counterargument: “I’ve been wrong for so long,” and perhaps it eventually works on tasks taking humans six months.
Software engineering, he argues, was never simply translating English into code. It means understanding requirements and mediating between human needs, processes and technical systems. Assembly, COBOL, C++, Python and now agents move that interface upward. Agents are the biggest change he has seen in 30 years, but historical productivity gains make him very bullish on the need for software engineering. Stebbings draws the conclusion that engineers will be fine for the next 10–20 years; Bernhardsson does not specify that horizon.
Modal spends roughly $500,000 monthly on coding agents across 150 engineers. Stebbings argues it should spend 10X more; Bernhardsson allows “maybe” and says engineers, including himself, already use agents heavily. On whether cheaper inference threatens provider revenue, his answer is that greater efficiency unlocks additional applications: “It just unlocks, like, five new use cases,” increasing total demand.
10. Owning intelligence is conditional, while standalone tools face platform pressure
Bernhardsson declines Stebbings’s invitation to say every company needs its own intelligence. Fine-tuning an open model can deliver strong performance and a stable foundation that will not unexpectedly change or disappear. But latency, cost, performance, hackability and adjacent services also matter. Modal supports proprietary-model businesses such as Runway as well as customers building on open models; ownership is one consideration among many.
For a 20-year-old seeking cash generation, he would consider selling coding environments to frontier labs: “There is so much money to be made.” His caution is financing structure, not present demand—“do you wanna get on the venture treadmill?” Long-term durability is uncertain. Stebbings argues that adding implementation, evaluation and monitoring can improve margins and longevity; Bernhardsson agrees with the direction of the point but still expects a tough long-run business.
Model routing can support a successful standalone business today, he thinks, but may eventually belong within a broader suite of evaluations, fine-tuning and performance controls. As an OpenRouter angel, he calls its outcome great yet would have advised continuing independently: rapid growth, revenue and brand outweighed the existence of “500 other competitors.” His long-term preference for platforms does not erase the immediate standalone opportunity.
11. Pullback risk, financing discipline and employee equity
Bernhardsson is “almost certain” a 20–30% stock-market pullback will arrive, followed by predictions that AI is finished. His long-term optimism survives that scenario. He does not identify an overheated category, but would avoid companies dependent on continual equity infusions: “Raise every round as if it’s your last round,” with a genuine route to commercialization and revenue.
He personally takes very little secondary liquidity because he already has enough money, but supports employee sales with guardrails, mentioning up to 10%. The recruitment purpose is concrete: establish that equity is “real money,” rather than merely paper value. A precedent of employee liquidity helps future candidates assign tangible value to compensation while preserving substantial incentives to keep building.
Competition for talent is extreme: Modal has lost exceptional new graduates to labs offering “20-plus million,” with no compensation period specified. Bernhardsson wants Modal around the 95th percentile of market pay, acknowledging that a few labs will pay more. His alternative pitch combines aggressive compensation, employee liquidity and what he believes is greater equity upside, rather than matching every lab offer outright.
Full transcript
I think there’s an opportunity to create a $100 billion company with a lot of specialized models in inference.
There’s going to be a pullback.
I’m okay with that, because I think the long-term opportunity is so fucking big. There’s just so much demand for it. It’s so large. I don’t want to sell this company, because I’d just go and start Modal 2. We’ve lost new grads to some of the labs for insane offers—talking 20-plus million. These are exceptional people. Exceptional.
This is 20VC with me, Harry Stebbings, and do you know what I wish I knew 12 months ago that I now know? The demand for inference, it's infinite, and we're seeing some of the biggest companies in the world be built in this space. Today, I'm thrilled to welcome one of those founders to the hot seat in the form of Eric, co-founder and CEO of Modal. Now, Modal just hit a billion in ARR. They just raised $750 million led by Accel, and they're priced at $15.75 billion today. I actually think it's a steal, by the way. But today, we dig into the inference market, what to come, what no one sees that everyone should see, and so much more. This was a phenomenal discussion with Eric. You have now arrived at your destination. Eric, listen, dude, it is so great to have you on the show. You know, I was sitting downstairs with you chatting before, and I was just like, you can feel when it's gonna be a good show because people are themselves. And that was an amazing conversation, and we're gonna have a lot of fun. So thank you for joining me.
It’s so exciting to be here. I’ve listened to so many episodes. It’s kind of surreal to be inside of it right now.
That’s really awesome. As I said to you—
Yeah.
—you could’ve texted me when you started the company, and maybe I would’ve been a little bit richer.
Yeah, but you probably didn’t know about me then.
No, I didn’t, but I would’ve welcomed investing in another Swede. I wanted to start with this: I was talking to Zach Frankel last night, and he was talking about your childhood and early indicators of success and exceptional ability. How did you fall in love with technology? And for me as an investor, what would’ve made me excited about you as a kid?
1. A Lifelong Love Of Coding
I started coding in 1992 or something like that. I was 8. I just loved it. I realized you could control computers. I had a Mac Plus, 1 megabyte of RAM or whatever, and I started making games. They were terrible, but I think back on my childhood and I’m so grateful. My friends were collecting hockey cards or whatever; that was their hobby. My hobby was programming, and now I make a lot of money from it. Isn’t that kind of crazy when you think about it? The fact that I can make a lot of money doing what’s been my hobby since I was 8 is kind of nuts. I’m just so grateful that I fell in love with technology and coding. Then I started doing programming competitions when I was 16 or 17, and they were so much fun.
Modal was your first company, or is your first company. Why did it take so long?
A consistent pattern in my life is that I never had quite enough confidence in myself. A lot of what was missing was just confidence, and a little bit was money, too. It wasn’t until I was 36 or 37 that I started Modal. I don’t think I ever necessarily wanted to start a company. I just saw this big opportunity in the market and thought to myself, “Why is no one else doing this? It’s such an obvious idea.” That led me to think, “Maybe I should start a company.” Five years later, here we are. It’s gone pretty well. In hindsight, maybe I should’ve started much earlier.
Do you wish you had?
I look at some of these Stanford kids, and they’re 21. I kind of feel like maybe they have too much confidence. They think they’re going to take over the world, and they’re 21. I think the ideal age is maybe 32 or something like that. You have a little more experience. You know how to navigate people, hire people, fire people—all these things. I don’t know if I would’ve been a good founder in my 20s.
Before we get into the market and everything beneath that, we were chatting downstairs. We often hear that founding companies is like eating glass. It’s hard; it’s painful. How do you feel about those two?
Dude, I hate that stuff. I don’t get it. You’re eating glass every day—why are you doing it, then? Do you expect some magic reward, some big cookie in the end? I never understood that mindset. A lot of people think, “I’m going to suffer pain, and then there’s going to be this big reward in the end.” I don’t think about it that way. I enjoy every day I go to work. Of course, there are hard times. You have to have hard conversations, maybe fire people, maybe do hard things. But in the grand scheme of things, such is life, and I enjoy life. I enjoy building a company, and I don’t think of it as, “One day we’re going to flip this company.” I hope to work on this for 20, 30, 40, 50 years. I’m having so much fun. I don’t want to sell this company, because I’d just go and start Modal 2.
You really wouldn’t?
No, I don’t. I don’t think money necessarily makes you happy. A little bit—it’s good to have some money. But beyond a point, what? I like building stuff. I love technology. I love coding. It’s so much fun.
You’re such a Swede.
Is that Swedish mentality? Maybe it is. And actually, do you know what—
Never has an American come on before and said it’s not about money.
I do care about money a little bit, but maybe you’re right. I do think in Swedish mentality, doing what you enjoy is kind of the highest form of prestige in a certain way. It could be music. It could be poetry, whatever. It almost doesn’t matter.
What do you do that you don’t enjoy?
Yeah.
Ugh.
I kind of hate fundraising, honestly. It’s fine. I enjoy it when it’s with good people, but it’s challenging. You have to say no to a lot of people. I guess that’s a good problem to have, but in general, it’s—
Oh, you have to say no to a lot of people? Oh, Erik, I’m sorry.
I know. That’s a horrible thing.
Oh, Harry, you can only have 10 million—
I know.
I know. It’s a tiny violin.
Not 50 million.
Yeah.
Oh, you’ve never done a deck, have you?
I’ve never made a pitch deck, yeah. Every round we’ve done so far, I’ve tried to get to know people well in advance of raising money. Most recently, we announced the D round yesterday with Accel. I’d been speaking to Rich and Gonzalo at Accel for at least a year, I think. At some point, we were like, “Actually, you know what? We could take some money. What are the terms? Let’s discuss terms.” Then it happened very quickly because we already had that relationship.
How do you think about price optimization? Is it a game of getting the highest price?
No, I don’t think so. Maybe in later stages, the money itself becomes a bit more of a commodity. But in the early days, for several reasons, there’s a misaligned incentive: VCs always want to put in way more money than a company needs, and you just take on a lot of dilution. It’s not good as a shareholder. The other thing is, I think you actually want to hold back. It’s often better to raise a lot less money at a lower valuation and keep the VCs hungry for the next round.
Is this not the generation where capital is the moat, though, and you need so much more money than you think? You said VCs want to give you more money than the company needs. I thought AI was finally the output exhaust for venture dollars that was needed.
I think there are times in the market when you’re capital-constrained. We haven’t been capital-constrained for most of the time we’ve been a company. We’ve been extremely capital-efficient. We burn only a few million a month. I mean, that’s a lot of money, and I have respect for it, but—
You burn a few million a month?
No, I mean, up until last year. Now we have to spend a lot of money on capacity.
How much do you burn a month now?
A lot. We’re actually profitable on a P&L basis, but cash flow is quite negative because we have to spend a lot of money on getting GPUs and building out capacity. That’s a different period for the company and a different way of thinking about capital. I don’t want to call capital a moat because people throw that word around to mean things that aren’t really moats. To me, “moat” has a very strict definition. But capital certainly helps at this point.
What’s the definition of a moat for you?
To me, you go read Seven Powers or one of those books—
I love Seven Powers.
It’s a good book. But I think people read it and get delusional. They’re like, “Yeah, we’re starting this company. It has all these moats.” You never have moats in the early days.
Google could do anything they wanted to.
Google could do—yeah. But what are the seven? I don’t even remember. It’s things like cornered resources, market power, whatever. None of those things happen until you’re 10 years in. There’s an old blog post by Ben Thompson I always come back to. It talks about Snap and the gingerbread man strategy. Do you know this one? The whole point is that Snap doesn’t have a moat. They just run faster. They’re the gingerbread man. I always come back to that. I think it’s true for almost every startup these days: you just have to move faster than everyone else.
I actually think the 3 things that matter when I invest today are, first, the team. Are they a top 1% team?
Yeah.
Second, the speed of product decision-making, very specifically. You look at Instinct today: the speed of feature shipping is really impressive. And then third is brand.
Yeah.
I think those are the 3 determinants of defensibility, to me, of a moat, but it’s a kind of package.
Yeah. Again, I don’t know if I would call it that, but I kind of agree with your point. I don’t want to call brand a moat, but I think the value of brand has gone up a lot.
Can I be a dick?
Yeah, please.
Do you think Modal has done well with brand?
I think pretty well. It’s coupled with the developer experience, the product, the website, the visual identity, and all these things. We just did an event in San Francisco, and I thought a lot about how we package it: the event, the visual look and feel, the vibes.
But do you think you have a bigger brand than Baseten and Fireworks?
I think we have really, really good vibes. I’m terminally online. People tweet way more about Modal. I think there’s a bit more of a community, and much better vibes around us. But I’m paranoid about those companies. I’m paranoid about competition, of course. I track it. I see everything that gets tweeted about our competitors, and this is important. Everyone wants to be part of the community. One thing people misunderstand is that I love TypeSafe with Jev.
Yeah.
Everyone wants to clone Jev. No one wants to get up on stage, wear a fucking pink fur coat, and talk about the Jevolution in an unhinged way. That’s what brand is to me. That guy is playing the game on a whole different level. I think that’s true of companies today. You have to think about that.
Can you explain why? Is it because he’s authentic and different?
I think so. You have to create a personality. You have to figure out what the personality of your company is, be consistent, and create a sense of movement and community. It’s something cool, and people want to use you because you’re one of the cool products out there. They’re like, “Have you tried Modal? It’s pretty cool.”
What has Modal done that Baseten and Fireworks haven’t to build that?
I think we started with the developer experience, so that was important. We had a very clear product-market fit and product-led growth, in the sense that people came in bottoms-up. They just sign up for the product. My consumer background from Spotify was very important in how I thought about building what’s now a B2B company selling tech. It’s a very boring category, but having personality and a bit of an edge is a big part of creating that movement, that sense of excitement, and the feeling that you’re different. I want people to feel that from the moment they go to our website, and the moment they try the product. That’s the magic moment.
We’ve spoken about 3 different providers. Inference as a market has many different speculators commenting on it. When you look at the next 5 years, how does the inference market play out?
First of all, I think inference is going to be an enormously large market. It’s basically the new cloud. But I don’t think inference in itself is the only thing you need. One thing I feel very strongly about is that it has to come with a lot of other products: post-training, sandboxes, agentic stuff, and all kinds of other types of inference, like low-latency inference and batch inference. We’ve been working on building that platform for the last 5 years. To me, that’s not a moat, but it’s a very hard product to build—the infrastructure.
Why do you need that breadth of product? Otherwise, you’re just in a commodity layer.
Otherwise, you’re just selling a commodity product. If you’re just selling tokens for dollars, there’s no switching cost. People can just switch to a different provider. It’s a race to the bottom. But if you couple it with all these other things—sandboxes, post-training, reinforcement learning, and all these things—you can create a whole set of tools. No one wants to buy a single-trick pony. People want a portfolio of different things. They want a whole suite of tools that lets them build amazing AI applications on top of it. The infrastructure underneath brings a lot of differentiation and staying power in the market, so it’s very hard to copy the infrastructure required to build all these products.
What does no one see about the inference market, or what do you not see commented on much, that’s very obvious to you?
One thing people don’t see about the inference market is that, in the long term, I think there’s so much gross-margin opportunity in this space because you can build all these things around it: sandboxes, agentic stuff, workflows, post-training, reinforcement learning.
If you do all those things, there’s actually so much opportunity to make money.
Is it a market that has software margins? We’re looking at margins now in the 25% to 40% range. That’s kind of traditional players. I don’t know if that’s yours, but that’s kind of where inference providers generally are. Do they go to 70% as you own more and more of the stack?
I think basically almost no company today has software margins anymore. What you used to see in Silicon Valley 10 years ago, with 90% margins or more, is basically gone. The reason is almost topological. You have this stack of vendors: applications, models, infrastructure, cloud, NVIDIA.
When a customer pays a dollar at the top of the stack, it has to flow through a lot of different layers. At every layer, there are going to be 20% to 30% gross margins. So I think there’s a lot of margin opportunity, and there’s a lot of opportunity to make a lot of money. The space is so big, but you’re also splitting it across a lot of layers in the stack.
The 90% software-margin businesses are gone. I think Modal has an opportunity long-term to get to 40%, 50%, 60% margins, but we’re never going to get to software margins.
What’s a wild prediction you have about the future of the inference market that you think many people would disagree with?
I think one prediction is that the inference market is going to split into a lot of different things, because the market got so big in the last—I don’t think anyone expected it, me included. I don’t think anyone expected it to be so big. It’s 10x or 100x bigger than I ever thought, and that means there’s going to be so much room for specialization: low latency, high latency, and various types of fine-tuning and post-training.
I think there are going to be some wacky new technologies that let you do it with way fewer resources, new hardware, all kinds of different things.
I’m really sorry. Doesn’t that go against what you just said about the need to own it all and have that breadth of product? If it’s so much larger that you need real specialization—
Specialization by product. I think it’s going to fragment into a lot of different products. We can build a lot of those products and host them all under Modal, and that’s very exciting to me, because companies are going to need a lot of different products.
For one product, they’re going to need low-latency inference. For another, they’re going to need batch inference. For another, they’re going to need sandboxes, which we have. For another, they’re going to do pre-training on their own models. Maybe they need certain specialized audio models or whatever. We do a lot of stuff besides LLM inference.
That creates a lot of opportunity to rethink what a cloud looks like in the era of AI, and to deconstruct the cloud and rebuild it for this new era of software.
Earlier, you said something about building capacity with GPUs.
Yeah.
What’s it like doing that today? Is it as crazy as reports suggest?
It is pretty crazy. I spent some time in March at GTC, NVIDIA’s big conference in San Jose. I would describe it as an open-air drug market, but with GPUs instead of drugs. You’re running around talking to 500 different cloud vendors, asking, “Do you have GPUs? Please get…” In the end, you do a lot of deals, and that’s a muscle we’ve had to develop.
Part of Modal’s value proposition is that our customers don’t have to think about that. We do the dirty work of doing all the GPU deals under the hood. We work with 25 different neoclouds. Right now, we control 0.1% or whatever—I don’t even know the number—of the world’s compute.
My crazy long-term goal is that, as we control more and more of the world’s compute, we can aggregate that capacity and aggregate demand on the other side, then manage that capacity and demand extremely efficiently. We think we can do that. We can fill a big gap.
What does being that aggregation layer allow you to do?
We built a lot of our infrastructure ourselves. We don’t use Kubernetes and Docker and things like that. We built our own file system, our own container runtime, our own scheduler, and basically our own way of managing capacity. That enables us to treat all of the world’s compute as one big pool of capacity.
We already have many thousands of paying customers, and over time that just grows and grows. Each of those customers has very variable requirements. Their workloads vary over the day, and they may have bursts. Individually, they’re very unpredictable. But if you pool all those customers together, you end up with smooth overall demand, which is something we can manage extremely efficiently.
One of Modal’s value propositions, besides being able to scale up with you, is that you don’t even have to do capacity planning. Just come to us, and we give you 100 GPUs. It’s amazing, going back to Typesafe, for instance. We’d talked to them a long time ago, before they launched, but they launched overnight on Modal, and we weren’t even aware of it. They just scaled up, we gave them a ton of GPUs, and it happened automatically.
That magic is something people really value with Modal: we can give them a lot of GPUs on very short notice. We can handle that kind of variability. If you come to us, you don’t have to lock up a rectangular block of GPUs for 3 years and know the exact number. It’s pay as you go. You scale up and down, and you pay per GPU-second.
How do you forecast GPU buying, then? You don’t want to have an insane excess of supply just sitting on your books.
Totally. GPU lead times right now are about 3 to 4 months, so we basically do a straight-line or exponential extrapolation. We look at demand in aggregate across all our customers. We have a very diversified set of customers, so you can look at aggregate demand pretty well and predict based on monthly growth rates, which have been about 30% to 40% month to month for the last 12 months.
You project that 3 to 4 months into the future. You end up with 2x to 3x, and you go out and buy that capacity and make sure to ramp up.
Can you just buy that capacity?
You can, yeah. I think there’s a bit of a misconception that we’re out of capacity. NVIDIA is shipping an insane number of GPUs every month. I don’t know how many hundred thousand Blackwells are getting installed. Prices are up, for sure. You have to pay more to get the same amount of GPUs. I don’t want to say it’s an efficient market, but if you’re willing to pay more than other people, you can get a lot of GPUs.
Is there a price at which it doesn’t make sense for you? Of course there is, but at what point does it stop making sense?
I don’t think so. In fact, I think we’re more long than we’re short on GPUs, because we can charge a higher price to our customers. There’s effectively a market price, and there are different dynamics because we lock up longer-term reservations and rent them out for slightly shorter durations. But if the underlying price goes up, people are also willing to pay more.
What happens to the price over time? H100s were meant to depreciate and go down.
Dude, I have no idea. Actually, it’s funny—
Well, dude, if you have no idea, what am I meant to know?
Did you see Jensen yesterday? He was in London, actually. Someone asked him this question, and he said, “Well, obviously, supply’s going to catch up with demand eventually, and then we’ll see what happens with pricing. It might happen in a year, it might happen in 5 years.” He said exactly the same thing I always tell people: hardware prices are like gravity. Hardware prices always go down in the very long run.
Right now, we’re in a supply shortage, and that means prices have gone up. They may continue to go up for another year, 2 years, 5 years—I don’t know—but eventually they will come down.
How early are we? I run Instinct for everything. I have over 100 automations on Instinct doing things that are absolutely wild, from buying concert tickets to planning shows to helping with medication for my mother. The amount of compute I’m using is Jevons paradox times 100.
Yeah.
If that’s true for many people in society—
Yeah.
Because it’s accessible within WhatsApp, how early are we?
I think we’re very early. I think you’re saying the same thing: Jevons paradox, whatever.
I think there’s so much latent demand for these things. Eventually, we’re going to start running up against these capital constraints. Interest rates are going up, and we just can’t build data centers fast enough. Maybe that’s good; I don’t know. In a way, it’s a guarantee against overbuilding, but I also don’t know if we can overbuild these things because there’s so much demand for them.
I don’t think you can overbuild in the next 5 years.
Probably not. I think what could happen is that people realize—Instinct is a good example. Allegedly, they don’t charge or make revenue yet, but clearly Instinct can make a ton of money in the future. That’s a dynamic in the market: a lot of companies aren’t making money yet, and they’re using a lot of GPUs. They just have to eventually start making money, but that lag could take a while, and we might run up against that.
2. Who Wins The AI Stack
You mentioned the specialization that comes within the inference market, and we spoke about NVIDIA there. We’re seeing more and more specialization within the chip market. Is NVIDIA’s business today stronger than ever, or weaker than ever, as you see the market fragment into very specialized players, from Cerebras to Etched to Olix[?] to Fractile, that weren’t there before?
I think you’re trying to phrase it as one versus the other. My view is that Eric Vishio was talking about the same thing: all of these things are going to grow. I think NVIDIA is going to keep growing and shipping amazing chips. I think Cerebras, Etched, and all these other companies are also going to ship specialized chips and grow. I think there are going to be markets for all of it.
So why is Cerebras down 30%?
I don’t know. I don’t really follow the stock market.
You don’t follow the stock market?
No. I’m mostly a passive, buy-the-index investor. I don’t feel like I have the energy to follow the stock market.
To what extent is CUDA the moat that everyone says it is for NVIDIA?
I think it’s a little overstated. I don’t think CUDA itself is the moat. CUDA is kind of annoying to work with; it’s a pretty annoying programming model. But I do think the NVIDIA ecosystem is. All the kernel optimizations and everything else mean that PyTorch always supports NVIDIA much better than anything else. The open-source engines like vLLM and SGLang support it much better, too. Everything is NVIDIA-first today, and that’s probably going to continue.
It’s funny: for the last 15 years, I’ve been saying, “Next year is the year of non-NVIDIA accelerators.” It never happens. I’m always wrong. NVIDIA has shown that it’s incredibly good at having staying power and dominating the market. We’ll see. There’s Trainium, TPUs, AMD—there’s a lot of other stuff that looks quite good, and they’re catching up, but I still think they’re quite behind.
To what extent are the neoclouds a friend—because they’re partnered with you in a lot of cases, and I’m sure, as you mentioned, I think you said 25—or a foe? I had the founder of Nebius on, and he was like, “We’re coming for the inference market. We need to own the full stack.” I don’t mean that against him; I totally understand it. But are they friends or foes?
I feel like everyone in startup land is a friend and a foe at the same time. They have this weird sort of relationship with everyone. I think some of the top neoclouds, Nebius included, are trying to move up the stack. I’m somewhat skeptical. I think it’s very hard to be good at your core business and also try to move into a different market at the same time.
Why would it be hard for them? What’s the bear case on them being able to move into the inference layer efficiently?
I’m biased, but we spent the last 5 years building an incredibly complex infrastructure stack. I think it’s very, very hard to replicate what we’ve built.
What’s the hardest thing?
All the code and everything around the file system, the container runtime, and the scheduler. Building a system that can place millions of sandboxes every minute, scale up and down many thousands of GPUs, and power hundreds of thousands of tasks running concurrently—these are very hard engineering problems. I’m an infra geek. I love that shit. It’s so much fun to work on, but these are hard problems.
Is Nebius a better business than CoreWeave?
I actually don’t know. I think they’re both executed extremely well. I have tremendous respect for what they’ve done. At the end of the day, the thing I have the most respect for is more on the financial engineering side. They’re very good at that kind of stuff.
Why do you say that?
I think a lot of the constraint for building a neocloud very quickly is on the financial side. You need to lock in a long-term customer contract, then raise a bunch of money against it, buy the chips, deliver the chips, and pay off the debt over the next few years. Doing that dance with the financials is an interesting game. CoreWeave has done it exceptionally well. It’s a capital game. It’s a financial game.
When you move a layer beneath them, we spoke about Jav[?], who obviously worked with Modal. We’ve seen incredible movement from the open-source community, which has been exciting.
Yeah.
But then you also see Haiku and Anthropic’s newest models being incredibly performant and incredibly cheap. What’s the benefit of open models if frontier models are almost as cheap and as efficient?
I think open source is always going to be around. There are always going to be great models coming out. Let’s also remember that open-source models are all moving into licensing models. They’re getting better at capturing the value of those models, and that means they’re going to have more funding to train better models.
I really hope there are European and US models, because I don’t think it’s a good situation that we have so many Chinese models. I have tremendous respect for all the Chinese labs. They’re extremely good at what they’re doing. But I think it’s good for the world if there’s more power spread out.
Going back to the question, I do think there are always going to be people who want to fine-tune, have more control over their models, and tweak the parameters—go low latency, high throughput, whatever it is. Open source is going to play a big role in that. It’s like Windows versus Linux.
In the broader AI space, do you think Europe will produce a competitive open-source model?
Yeah. I love Europe. I grew up in Sweden, but I do think Europe struggles with accumulating large amounts of capital. I think both of those things are things Europe struggles with. China has accumulated massive amounts of human capital, and the US is obviously very good at raising enormous amounts of money and training very large models. I don’t know if Europe has either of those.
Does it matter if NVIDIA owns the open-source model layer in the US, with first Poolside and now Reflection?
I think NVIDIA is very smart to focus on building up an ecosystem, and I think that includes propping up a lot of open-source models. They probably saw an opportunity to pick up open-source models. These companies are very resource-intensive, very CapEx-intensive. I think it’s smart of NVIDIA to buy some of those and effectively sponsor them in order to create American-built models.
By the way, I want to take back one thing I said. Europe does have an insane amount of human capital. I do think what Europe is very bad at is aggregating people and putting them all at the same company. That’s something I think the US is very good at.
And I think inspiring a mindset of ambition—
Yes. Yeah, yeah.
—and huge, huge scale. I think, to the point you made earlier, “I didn’t have the belief in myself,” I respectfully think that if you were in the US, you might have been surrounded by influences that told you to start a company earlier.
Yeah, yeah, yeah. I mean, it’s almost too much on the other side, to that point, but—
Oh, when you’re 4, you’re encouraged to start—
Yeah, yeah, exactly.
Leave school, drop out.
Yeah.
Kindergarten wasn’t worth it.
Exactly.
I totally agree with you. I want to talk about models, and Cohere in particular. Are we seeing the unbundling of models on a per-function basis, where Jav is obviously very good for its certain tasks?
I think so. It’s a function of what I talked about earlier: the market for inference is so large. When a market is $1 billion, you’re not going to have too many submarkets. But when a market is getting to $10 billion or $100 billion, you’re going to have a lot of different submarkets. Jav does classification and things like that, regression-type things, very well, and there’s a large enough market for that.
Okay. Welcome. It’s your first day as a partner at 20VC Fund, and we’re in the investment committee. Very kindly, we have the opportunity to invest in Jav. We’re all in on Jav’s new round.
All in.
What’s the bull case for Cohere being worth $100 billion?
I told you before, but I think the brand, the community, and the movement around everything they’re doing are incredible. To me, that’s a massive, massive asset in this new age.
That’s a $100 billion company?
I think there’s an opportunity to create a $100 billion company with a lot of specialized models in inference.
You would have done it, though, at $8.5 billion.
I would have done it at $8.5 billion. I think there’s a pretty big outcome here. It could be an exit, or it could be a public company. Who knows? But I do think there’s a lot of upside here.
Dude, I’d love to see him ringing the bell on the New York Stock Exchange.
Yeah.
Do you know what I mean?
Yeah. Fantastic guy. Love that guy.
Right? I mean, his charisma—you have to create a personality around your brand. You have to create a visual identity. It’s obviously not the visual identity I would have picked.
I hear you, but I think you should be louder, if I’m just honest, Erik.
Modal?
No. Well, yeah, but you.
Yeah. I mean, that’s possible, right?
I don’t think there’s a personality-led cult around Modal from your perspective.
I think that’s a fair point of criticism. I think I could be out there even more. I try to tweet. I’ve been kind of bad at blogging recently.
No one gives a shit about blogging.
Maybe you’re right. Maybe I just need to go more unhinged on Twitter. It’s founder-led marketing, basically. In the early days of Modal, that was a big part of our growth: me tweeting all the time about what we were building. I think that’s really important. Every company has to find its own go-to-market strategy.
In the early days, we spent a lot of time talking to people on Twitter, hitting people up: “Try the product. What do you think?” Half of them ended up angel investing. They’re now supporters of Modal. You have to lean into what you’re good at. In the early days, I did a lot of shitposting.
For every follower who followed me before 50,000, I would send them a personal thank-you message.
That’s amazing. You typed it out yourself?
Every night from 1:00 to 2:30 a.m., I used to send a personal message. I’d say, “I see that you’re from Sweden. I love the Grand Hotel. Let me know when you’re in London.” I’d send 50 to 100 a night.
That’s dedication. That’s amazing.
No, but it was the customer acquisition flywheel.
Yeah.
They would go into the office the next day and be like, “Wow, you know this guy? He sent me a message,” and, “You should listen to his show.”
Yeah.
“It’s pretty good.”
It’s smart. I love what you’ve been—
Modal.
—doing in terms of building up this media brand. It’s fantastic.
Neil Mehta has a great principle: “How can I 10X the customer experience?” And actually, let’s do it. Fuck it. Let’s play a game. How do I 10X the current Modal customer experience?
I think the biggest thing we need is way more products in Modal. We want to create a comprehensive experience where anything you need to do to build an AI application should be a product inside Modal. The other thing we can do is hire way more FDEs and way more sales reps. I’ll say we’ve been behind on that, but we should hire way more people who reach out to you personally and say, “Hey, I saw you get started on Modal. I saw you working in the bio space. How can we help you? Here are some bio examples—protein folding, whatever.” We work with a bunch of other similar companies. You seem skeptical.
No, I don’t at all. I’m actually just thinking: I like the product ambition, but as a European, I go, “Wow, that—”
We’re in the B2B space, so I think—
But do you worry about being stretched too thin?
Yeah.
To your point about why Nebius can’t do what you do so efficiently: you’re so good, and you’ve spent so much time on it. As you expand into all these specialties, there are players in each of them who’ve been doing it for years. I’m like, “Oof, Erik.”
I don’t think that’s true. I don’t think a lot of them are very good. But I sort of agree with your higher-level premise. If there’s anything that’s super critical and never goes away, it’s prioritization and resource allocation. You can’t boil the ocean and try to do everything at the same time. You need to be extremely ruthless. Having a multi-product strategy is a very hard balance. When do you branch out into a second or third product? It took us four years to see success with our second product—
What’s—
Sandboxes.
It took you four years to see success with your second product?
Sandboxes.
But you’re at a billion in revenue in 5 years.
Yeah. We launched Sandboxes in 2023 and didn’t see a lot of success in the early days. There was some success; it wasn’t horrible. But it was a relatively insignificant part of our revenue until the summer of 2025, when models like Sonnet 4.5, one of those, got really good at generating code and building whole apps. Then it just exploded.
Sandboxes is a $300 million run-rate business today, about a third of our revenue. It’s been a crazy journey. Now we’re thinking about what the third and fourth products will be. We just launched a training product, which I’m super excited about. We’re working on batch products, reinforcement learning, and all kinds of other stuff.
That evolution into becoming a platform is actually there. If you go back to the very early decks of Modal, I put some together just for my own thinking. It was all there. We had this market map and all these things. It ended up being very different products than we thought, but it’s so important to focus on a category and a set of solutions, not just be a single-widget vendor.
What do you know about the Sandboxes market today that you wish you’d known when you started and released the first product?
One of the benefits of being in infrastructure, but also one of the drawbacks in a weird way, is that you just have to trust your gut about what the important primitives are and build amazing building blocks. We try to look at what engineers need, but not listen too carefully, because they’ll ask for all kinds of random stuff. We distill it down to some of the building blocks.
We put out a serverless function product in 2021. It saw success in 2023 with GenAI inference. We put out a Sandbox product in 2023 because we saw a lot of people using LLMs to generate code. Then it saw success in 2025, like I said. A lot of how we think about building products has been: let’s put out good building blocks, and then they explode.
But to answer your question, it was reinforcement learning, vibe coding, background agents, and maybe some other stuff in the future. Those were things we didn’t anticipate. I think infrastructure is often like that. We didn’t anticipate Stable Diffusion when we started building our serverless function product. It just happened to be the killer app.
What’s your takeaway from that? The only thing that matters is that you’ve got to stay in the game long enough for your timing to be right?
Sometimes I think that’s right.
We’ve definitely been extremely lucky with timing in certain ways. We put out products right before Stable Diffusion and ChatGPT came out in 2022. We started the company in 2021, so we had really good timing. Those products created so much demand, and we were very well-positioned.
But some of it is also about infrastructure. I think it’s about building products that you anticipate will be useful in the future and that, from your own experience, are things you’ve always wanted to see. I think you have to trust your gut on that.
You said that about resource allocation and how important it is as you think about this breadth of products. Where did you allocate resources that, with the benefit of hindsight, you wish you hadn’t? How do you reflect on that?
I think we should have allocated more resources toward LLM inference earlier. We were a little bit on the sidelines for a while. We didn’t realize how big open-source inference would be, and we thought it would be a race to the bottom, just a commodity space. Then DeepSeek came out, and we were like, “Fuck. We’ve gotta move very fast.”
So we staffed up a team immediately. We went from 1 or 2 people working on this to 30 people 12 months later, and now we’re one of the top 3 providers.
Jason Lemkin, a dear friend of mine who I do a show with every Thursday, always says that we’ve now reached peak open-source LLM inference usage. We’re gonna see frontier providers provide incredibly cheap products, as they do now. US enterprises don’t wanna use Chinese models, and that will continue. So, with those 2 core prevailing winds, frontier providers will actually gain ground, and we’ll see a reduction or plateauing of the open-source LLM inference market.
I think that’s a little bit of a fixed-pie-size mentality. With all due respect, I do think both spaces will grow 5×, 10×, 20× over the next 12 months and beyond. It’s hard to argue which one’s gonna be bigger than the other, but I think both will grow by orders of magnitude over the next few years.
5×, 10×, 20× over the next 12 months?
Yeah, for sure.
You just hit $1 billion in revenue.
Yeah.
Well done.
Yeah, thank you.
Does that feel pretty cool?
It feels pretty cool. I don’t know. I’m this Swedish guy who grew up in a rental. For me, these numbers are absurd, but it feels great.
Yeah. I love you.
It’s so much money.
You’re so likable and cool. I wasn’t expecting that.
I don’t know. It’s a lot of zeros. How many zeros are in a billion?
It’s a lot. Your wealth manager will tell you. When you look at that, where will you be in 12 months?
I don’t know. What I can say is we grew 15× in the last 12 months, so you can extrapolate that. At some point, markets just get very big, and they start to slow down a little bit. But with the growth rate right now, the growth rates aren’t slowing down. We’re growing 30% to 40% a month.
But do you think you’ll be at $4 billion?
I hope to be way beyond that.
With the greatest of respect, $15 billion then feels like quite a cheap price. If I’m an investor, this is my ballpark: if you think you’re gonna be at $4 billion, say, dropping, I don’t know, $1.2 billion to the bottom line, then they actually only paid, what, 14× end-of-year-next-year bottom line, 3× top line.
Yeah, it’s kind of cheap.
That feels pretty cheap.
I think it’s cheap. I wouldn’t do this if I didn’t think we could grow and become a generational company. I think there’s so much upside. This is one of the largest market opportunities that’s ever existed. The demand for compute and the demand for AI applications are just gonna be enormous. They’re all gonna need a place to run, and I hope a lot of them are gonna run on Modal.
What revenue line do you not have today, or is tiny, that you think will be very significant in 3 years’ time, 5 years’ time?
That’s a good question. I think training is something we just launched, and I think that could be very large. I hope that could get to a $100 million run rate in the next 6 to 12 months. Post-training, similarly. I’m working on batch jobs.
I’m also very excited about a lot of traditional data warehousing stuff: databases, workflow orchestration, feature stores, vector databases. There are so many parts of this, and they all fit together very nicely. That, to me, is what I’m so excited about. Once you start building in this space, there are so many natural adjacencies.
The problem with me is I get excited about so many things. I have no lack of ideas. At Modal, we have so many ideas for things we wanna do. I would love to get into continuous integration. That’s a very interesting space. Agent harnesses. Maybe we should train models. There are 5,000 directions, and my ADD brain wants to run in all of them.
You obviously have to be ruthless with prioritization and only take on 1 or 2 things at the right time, but there are so many different directions we can go.
It’s amazing that your wife calls you for decisions. She must get a very incoherent answer.
I know, I know.
Like, “It is, it’s this, is this.”
Yeah, yeah.
“I just wanted some curtains done.”
Yeah. With the curtains, I just pick one of them. Have you ever seen The Simpsons, where Rainier Wolfcastle becomes the president?
Yeah.
They’re like, “Which one do you wanna pick, one to seven?” He’s like, “I was born to lead, not to read.” “Number 4.”
3. What AI Changes Next
Yeah. Can I ask: we’ve seen this real shift in emphasis, but also in resource allocation, from pre-training to post-training. How do you feel about that, and do you think that’s the new normal?
I don’t know. I think post-training works because there are a lot of tasks people do in real life that have fairly finite feedback mechanisms. Coding is a great example. You can do coding with reinforcement learning pretty well because you can run code in sandboxes, and a lot of labs are running Modal Sandboxes for that exact purpose.
Cybersecurity is another area. Math, obviously, with Lean. We’ve seen some fantastic proofs from OpenAI the other day. Law, maybe, to some extent, where you can do evals. I think there’s such incredible focus on post-training right now because you have all these tasks that traditionally only humans can do but are actually well within reach for agents and LLMs.
But I don’t know if it extends beyond that. There are a lot of tasks out there with much longer feedback loops, and I don’t know if post-training is quite within reach for those yet. A lot of people ask me, “What’s gonna happen with software engineers in the long term?” I think we’re probably gonna automate everything an agent can do within 24 hours or a week or something like that.
But engineers work on a longer timescale, and they have longer feedback loops. I have a feeling reinforcement learning is gonna start to run into some limitations around that. But I’ve been wrong for so long, so maybe I’ll be wrong and it just keeps going, and eventually we can do reinforcement learning on things that take 6 months for a human. I don’t know.
What does the day look like in 3 years for an engineer?
I’m extremely bullish on software engineering. I’ve been coding for 30 years, and I’ve seen so many changes with programming languages, the cloud, and so many different frameworks. Every single time we make programmers more efficient, it just increases demand for software.
There are so many parts of the world that desperately need software, so I’m very, very bullish on software and the need for software engineering. One of the biggest misconceptions from the outside is that being a developer is about translating from English to code. That has never been the job of software engineers. It’s about understanding the requirements in the first place and being an interface between the humans you’re building for—because you’re building for humans at the end of the day—and figuring out what that means in terms of processes and workflows.
That continuously moves up in the stack. Back in the ’60s, we used to write assembly and punch cards, and then we moved to COBOL, then to C++, then to Python, and now we’re using agents. It’s definitely the biggest thing I’ve seen in the last 30 years. But I do think we have many, many analogs and parallels throughout the history of software and computers.
To me, that makes me feel like software engineers are going to be fine for the next 10 or 20 years. How much money do you spend on coding agents a month at Modal?
I think half a million or something like that.
How many engineers do you have?
150.
Dude, that’s so worrying, you know?
You want to—should be token-maxing more?
Dude, there should be 10x that.
Yeah, maybe. I don’t know.
Brandon McCool said they spend 3x their human salaries on inference.
Yeah, we’re not quite there yet, but our software engineers are extremely agent-heavy, including myself. I write code with agents all the time.
But I think the big question, if you’re looking at an Anthropic IPO—and don’t worry, I’m not ringing up Bloomberg and pulling you in to be like, “What do you think of the Anthropic IPO?”
Great.
I hate those ones. You know where it’s like—
I think it’s great.
Yeah, it’s great. It’s—
Wonderful.
No one in AI is like, “Oh, not a fan.”
Yeah, yeah, yeah.
Everyone is like—
I’m against it.
“Go, go.” But I do think the mother lode, so to speak, is Claude Code. It’s $60 billion of revenue. It largely dictates the $2 trillion price. You look at Marc Benioff, who says that they spend 5% of engineers’ salaries—$300 million—on Claude Code today. Now, if that comes way down as efficiencies improve, that’s a challenge. Or will that go up as it eats more and more of engineers’ workflows?
Yeah, it’s just going to eat more. We’ve seen this so many times: you make something more efficient, it makes something more powerful, it unlocks five new use cases, and the demand just goes up in the end. That’s just the history of the last 50 years of computers and software. I wouldn’t be here if I weren’t such an optimist. That’s why I started a company. I’m the biggest optimist there is in the world about software and computers. I just love this shit, and I think it’s going to keep growing.
The data market is an interesting one. We have so many players—
Yeah.
—that have scaled so fast.
Yeah, it’s incredible. If I were 20 and just wanted to make a lot of money, I would go out and start an environment company and sell data to the frontier labs. There is so much money to be made in that space, it’s kind of incredible.
Let’s go a layer deeper. What would you do next? Would you specialize in a specific market?
The coding space is where you see the most demand right now. If you can sell code environments to the labs, you can make a lot of money. There’s a big space in that.
Still?
Still. Yeah, yeah. I meet startups all the time. I’ve angel-invested in a bunch of them. They’re incredible businesses. My one piece of advice is: you make a lot of money today, but is this a venture-funded business? Do you want to get on the venture treadmill? I’m not sure. I think these are companies that are going to make a lot of money today. I don’t know if they’re going to make a lot of money in the very long run, but who cares? Take the money. It’s great money. It’s a fantastic business today, and that’s all that matters.
I think there are data marketplaces, and then there’s data and services. When you look at the leading providers, they started as the data acquirers, and then just sold you the data. That was kind of it. I think they’ve really gone full life cycle now, into deployment, implementation, evaluation, and monitoring, with a real ability to hand-hold enterprises—
Yeah.
—similar to the FDEs—
Yeah.
—where they’ll really be that partner to ensure full implementation. That’s how you expand margin and have some form of business in the long term.
For sure. I think it’s going to be a tough business in the very long run, but today you can have extremely good cash flow selling these datasets.
Can I ask: when we think about that kind of implementation, working so closely with partners, you’ve said—and Lynn has said too from Fireworks—that owning your own intelligence is important. Can you unpack what this looks like? The jury’s still out on specialized models for companies in this way.
Yeah.
You know, we have Harvey, who says it’s fantastic and has been amazing, and then you have Legora, who say, “Ridiculous. The underlying models move so fast; it’s the wrong strategy.”
For some stuff, it’s good to own your own intelligence. For other stuff, it’s not. I don’t think it matters for everyone. We have a lot of customers who run proprietary models on Modal, and they use a proprietary model for their customers. I don’t think it makes sense for everyone to own their own intelligence.
Where does it make sense for you to, versus not?
I think for certain use cases, you can take an open-source model and fine-tune it, and you can get really good performance. The nice thing is that you know what the fine-tuned model is based on. It’s not going to change; they’re not going to take the model away from you. It’s nice to have that consistency and the ability to continuously tweak it without the underlying model changing, which has happened in some cases with the frontier labs. That’s nice. I support that, but I think there are many other things that matter to people. Owning your intelligence is just one thing out of many. Latency matters, cost matters, performance matters. Hackability matters. Having a platform with a lot of adjacent services matters, too. So, owning your own intelligence—yeah, I’m supportive, but there are many other things.
Isn’t that kind of going against your business?
No, I don’t think so. I—
I mean, I love the honesty, but it would be in your favor to say, “Yes, every company is going to need it, and we’re the partner to help you.” I kind of teed you up there, and you—
Well, I mean, not really. If you think about Typesafe with Jav, it’s a highly proprietary model. I want people to use Jav[?]. It runs on Modal. We have many customers with proprietary models that they’re hosting on Modal and serving to their customers, and I’m very much in favor of that, too. I think the beauty of Modal has always been that it’s a general-purpose platform, and we support so many different things. To me, that’s the beauty: we can help you if you want to use open-source models. There are many customers who build incredible AI applications on top of open-source models, fine-tune them, use reinforcement learning, and have adjacent sandbox use cases or training. But there are a lot of companies that have built highly successful businesses—Soon or Runway, Typesafe, and many others—running their own proprietary models on Modal, and their customers don’t care. I think that puts us in a really good spot.
I want to ask 2 more questions before I do a quickfire.
Okay.
Model routing, as the dear Americans say—is that a defensible business?
What are you saying? Do you get routing?
Routing.
I love that.
Well, we are, actually. The original language—America just took ours and bastardized it.
Yeah.
But my question to you: is that a defensible segment of this value chain, or is it a commoditized product that everyone just builds themselves?
I think it’s possible to build a successful business in model routing. You can do that today as a standalone product. I wonder if, long term, it has to be part of a more comprehensive platform. At the end of the day, I think what these companies want is evaluations, ways to fine-tune, and more control over performance—latency, throughput, all these things. It’s one of those things that makes sense today. These are very good businesses. I invested in OpenRouter personally, and they had a great outcome. I actually think they could have done even better on a standalone basis. But in the very long term, it feels like one of those categories that should be part of a larger, more comprehensive set of tools. That’s my view on everything in AI.
Congratulations, you’re on the OpenRouter board, and an acquisition offer comes in. Please make the case for why they would have done that.
Reject. Yeah.
Yeah, please make the case for why.
I think OpenRouter—the growth rate and the revenue, and the brand, frankly. They have a really strong brand. There are 500 other competitors, but the space is growing so fast and there’s so much revenue, I would have told them to keep going.
Final one. I think a really telling question is: What do you believe does not work in favor of your business? For example, I think seed investing today is kind of dead and bullshit. This idea of 2 on 10, or slowly building for 18 months.
Well, it’s kind of bad if you want to make money, right? But it’s great if you want to develop a relationship with founders.
Sure, but that goes against what I do. I’m traditionally—
You’re against having a relationship with founders?
No, no, no. I’m saying I think the death of seed investing—the 2 on 10 and building for 18 months—is real.
Yeah, no, I understand what you’re saying.
So what would yours be?
I don’t know. One thought I have is that this AI stuff is growing so fast. We’re here, and we’re going to go to here, and then there’s going to be a pullback, maybe 20% or 30% in the stock market. Everyone’s going to freak out, and people are going to say it’s the end of the world. I’m almost certain it’s going to happen, but I’m so bullish because then we’re going to keep going. So that would be one thing: there’s going to be a pullback. I’m okay with that because I think the long-term opportunity is so fucking big. It’s so large.
Do you take secondaries?
Very little.
Why?
I personally don’t need a lot of money. I’m pretty happy with life, and I have enough. I actually support it for my employees, though. It’s quite important as a founder that you make the equity worth as much as possible. One way to do that, and to tell future candidates, “No, this isn’t just paper money; this is real money,” is to have a precedent of doing employee secondaries. Then you can point to the last round and say we let employees sell this much, up to 10%, or whatever. I think you have to have guardrails. You don’t want employees to sell too much. You want them to still have a lot of skin in the game and be very incentivized to make the company successful. But I think it’s a way to create a very liquid, hard currency out of the equity.
I totally agree. And as you think about a talent marketplace that’s ever more competitive—
Absolutely.
—it’s very—
It’s extremely competitive out there.
Is it nuts?
It’s nuts. The offers we see for new grads going to some of the labs are insane—talking $20-plus million. These are exceptional people. Exceptional.
What makes an exceptional new grad?
Just fucking sharp people. I know it when I see it. They’re extremely hungry and hardworking—
And where do they go?
They know everything about technology. They go to some of the big labs.
Is that it? They go to OpenAI and Anthropic?
Yeah, the big labs. They pay a lot.
But is that it? Is that where the talent comes from?
There are a few others, but mainly those people are known for paying a lot of money.
And you won’t pay that much?
We pay a lot. We can be very aggressive. We want to be out there paying at the 95th percentile in the market, and there are a few companies I know that are probably going to pay a little bit more. I think we have way more upside in the equity. Going back to what I said about secondaries, I think it’s important for that reason, too, to create real tangible value in the equity that makes our offers seem very strong. Plus, we have a lot of upside. You said it: our stock seems cheap. I kind of agree. There’s a lot of upside in our stock, so that enables us not necessarily to match the labs, but to create a compelling sales pitch for why our comp package is basically better than the labs. We’re willing to go very far to get the best talent.
Okay, we’re going to do a quick-fire round, because otherwise I could talk to you all day. You have a great angel portfolio. You mentioned OpenRouter and Lovable before. You can put all your money into only one.
I love TurboPuffer. I know Simon really well. He’s a Danish guy, so we have the Scandinavian connection. Just a fantastic entrepreneur. I just love the business.
How big can that business be?
That’s a good question. It is somewhat specialized. They’re doing vector databases. I think it could be a very large category. If I ran that company, I’d have 500 different ideas for how to expand into adjacent areas around databases, data warehouses, and other types of query systems. It could be a large category.
What have you changed your mind on most in the last 12 months?
Probably the magnitude of the sandbox space. I think it’s kind of insane. I look at the revenue, and it just keeps going up. I’m like, “Well, clearly it has to stop right now.” I think I’m doing some sort of reverse psychology on the revenue, because it just keeps going up. I’m like, “What is going on?” It just keeps growing, and I’ve been wrong so many times—in a good way. It’s incredible how large that space has gotten.
What do you think is the most overheated category within AI today?
Oh, boy. I don’t know. I don’t think anything is overheated. I will say, one of the things I’m a little nervous about is companies that are dependent on a continual infusion of equity capital. I would not invest in them, and there are a lot of them. I would be nervous about anything where, if there’s a venture capital pullback, they’ll struggle to raise more money. I think it’s important to raise every round as if it’s your last round, and to have a real path toward commercialization and revenue.
You’ve got 2 kids.
2 kids.
Biggest lesson on being a good dad and being a good CEO?
Sometimes I feel like it’s kind of similar.
Ah.
Do you mean in terms of how to balance it, or just in terms of being a good dad?
No, what does it take to be a good dad?
Setting them up for success in life. I always think about the right amount to push the kids. I’m not a tiger dad, but I think you need to push the kids a little bit to get them exposure. I don’t want to create a situation where my kids come back to me and say, “Hey, Dad, I wish you’d made me do X,” or, “I wish you’d made me do Y.” But I also don’t want them to come back when they’re 20 and say, “Hey, Dad, why did you make us do X?” or, “Why did you make us do Y?” So, a little bit of pushing. I make them play piano. I don’t know why, but I think it’s a good way to learn something that I hope they’ll appreciate when they’re 20 or 30. They don’t like it that much, but I think stuff like that is good. You’ve got to push them a little bit, but not too much. My parents were very lax. I discovered math and programming, and they were like, “Erik, why are you doing so much math? You should go out and play like the other kids.” In a way, I wish they’d pushed me even harder, but I pushed myself so hard anyway, so it doesn’t matter. It’s that right amount of pushing.
How could you push yourself so hard? You’re actually from a nation where you’re told it’s bad to stand out. I’m forgetting that word, but I had a Swedish girlfriend.
Yeah.
Uh—
Jante, the Jante Law.
Yeah, where it’s bad to stand out.
Yeah, don’t think you’re special.
Yeah, why—
I don’t know if that’s true. I think there are a lot of misconceptions about Scandinavian culture. I do think there’s a lot of respect for people who follow their dreams and do things, whether that’s music, writing books, making furniture or art, or being an entrepreneur. I think self-actualization and following your dreams are highly regarded. That’s the flip side of having a very lax culture: you kind of encourage people to follow their dreams. My dream was doing math and programming. I love that shit, so I did that.
Unless you’re from Norway, where a tax structure exists that basically makes it impossible to start a company.
Yeah, Norway has some issues.
Yeah.
They also barely work.
Yeah, this is very true. But we do have a sovereign wealth fund.
Yeah.
So thank—
They are rich, yeah.
—thank God for that.
Maybe it’s the first post-economic society. I don’t know.
Or literally almost.
In multiple meanings of the word.
Final one for you. What are you most excited for? It can be anything. I look at drug discovery for chronic conditions. My mom's got MS.
Actually, drug discovery is one of the applications I think is incredibly cool. I can't wait to take AI-generated drugs. I'm old, my skin is getting bad, and every time I drink a beer, I feel horrible the day after. Give me all the AI-generated drugs I can take. I want to live long. I want to live a healthy life.
I think so. But with the FDA and other things, it's going to take a while. I think there's just so much incredible progress today in AI-aided drug design. There are so many cool things. So many things.
Dude, I've so enjoyed this show in a way that I honestly wasn't expecting you to be this fun, open, charismatic. You've been amazing, so thank you so much.
Thank you so much for having me. This was a lot of fun.