20VC: Modal’s Erik Bernhardsson on inference and RL’s limits

Summary

• 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.

Deep dive

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.