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

Erik Bernhardsson says he began coding around age eight on a Mac Plus with one megabyte of RAM, making terrible games before competing in programming contests at 16 or 17. Programming was his hobby while friends collected hockey cards, and he is grateful it became his career. He waited until 36 or 37 to found Modal because he lacked confidence and money, then saw what seemed like an obvious market opportunity and asked why nobody else was building it. He thinks an ideal founder age may be around 32: old enough to know how to hire, fire, and navigate people, but not so established that confidence becomes excessive.

Erik rejects the startup cliché that founders must “eat glass” every day for a future reward. He enjoys going to work and treats hard conversations, including firing people, as normal parts of life rather than suffering to be exchanged for a payoff. He does not plan to sell Modal: if it were sold, he says he would simply start “Modal 2.” Money matters somewhat, but he believes happiness comes more from building technology and doing work he enjoys. He also dislikes fundraising, though Modal has avoided pitch decks by building investor relationships over roughly a year before each round, including its latest Accel-led financing.

Erik argues that early companies should often raise less money at a lower valuation rather than accept excessive capital and dilution; keeping investors hungry can help future rounds. Modal was capital-efficient for most of its history, burning only a few million dollars monthly, but now has negative cash flow because it is buying GPUs and expanding capacity while remaining profitable on a P&L basis. He rejects calling early-stage advantages “moats,” saying durable powers such as cornered resources usually appear only after roughly ten years. Instead, startups must move faster. Harry identifies team quality, product-decision speed, and brand as modern defensibility; Erik agrees that brand has become more valuable.

Erik says Modal built its brand through developer experience, product-led growth, consumer-style presentation, and a distinct personality rather than a conventional enterprise image. He contrasts Modal’s “good vibes” and active online community with competitors Base10 and Fireworks, while admitting he could be more publicly unhinged like TypeSafe’s Jev, whose pink fur coat and “Jevolution” persona create a movement. He sees inference as potentially as large as the cloud, but only if combined with post-training, sandboxes, reinforcement learning, agentic tools, and both low-latency and batch inference. Selling tokens alone would be a commodity race to the bottom; a broad platform creates switching costs and differentiation.

Erik predicts inference will divide into many specialized markets because demand is already 10 to 100 times larger than he expected: low latency, batch workloads, fine-tuning, post-training, specialized audio, and new hardware will all have room to grow. He says this does not contradict Modal’s broad-platform strategy because specialization will occur by product, with different customers needing different combinations of inference, sandboxes, training, and other tools. He expects Modal to help “deconstruct the cloud and rebuild it” for AI. He estimates Modal could eventually reach 40–60% margins, but says the 90% software margins of older SaaS businesses are largely gone because value is split across applications, models, infrastructure, cloud, and NVIDIA.

Erik describes NVIDIA’s GTC conference as an “open-air drug market” where Modal negotiates with hundreds of cloud vendors for GPUs. Modal works with about 25 neoclouds and hides that procurement work from customers, offering pay-as-you-go access rather than requiring multi-year rectangular blocks of capacity. Its own infrastructure replaces Kubernetes and Docker with a custom filesystem, container runtime, scheduler, and capacity-management system that treats global compute as one pool. Thousands of customers’ individually unpredictable workloads become smoother in aggregate. Modal forecasts demand three to four months ahead, using roughly 30–40% monthly growth and buying two to three times projected capacity; Erik says GPUs are available if a buyer pays the market price.

Erik expects NVIDIA and specialized chipmakers such as Cerebras, AMD, Etched, and others to grow simultaneously rather than one eliminating the rest. He thinks CUDA itself is an overstated moat because it is an awkward, low-level programming model; the stronger advantage is NVIDIA’s ecosystem of kernel optimizations and superior support across PyTorch, vLLM, and SGLang. He has predicted non-NVIDIA accelerators would break through for 15 years and has repeatedly been wrong, though he sees TPUs, Trainium, and AMD improving. Neoclouds are both partners and competitors. Erik is skeptical that companies such as Nebius can easily reproduce Modal’s five-year infrastructure stack, while he respects Nebius and CoreWeave most for their financial engineering and ability to finance GPU purchases against long-term contracts.

Erik expects open-source models to remain important because developers want to fine-tune, control, and optimize models for latency and throughput; licensing changes may also let open-model companies capture more value and fund better training. He wants more European and US models alongside Chinese systems, but says Europe struggles to concentrate human and financial capital at the scale required, unlike the US and China. He believes inference will unbundle into specialized models by function, making a $100 billion company possible through many verticalized models. He cites Jev’s brand, community, and movement as a major asset and admits Modal has not created an equally personality-led cult, though its early growth came from his own Twitter outreach and personal messages to followers.

Erik says Modal’s biggest resource-allocation mistake was underinvesting in LLM inference because the team initially expected open-source inference to become a commodity. DeepSeek changed that judgment, prompting a team to grow from one or two people to about 30 within 12 months; Modal is now among the top three providers. He rejects the idea that open-source inference has peaked, predicting both open and frontier-model markets could grow 5x, 10x, or 20x. Modal launched Sandboxes in 2023 and saw limited early traction, but code-generating models made the product take off in summer 2025. Sandboxes now represent about $300 million in run-rate revenue, roughly one-third of Modal’s business.

Erik says reinforcement learning works best where feedback is finite and quickly verifiable, including coding in sandboxes, cybersecurity, and mathematics with Lean. He expects limitations when tasks have feedback loops lasting months, though he admits repeated technological progress has made him wrong before. He remains strongly bullish on software engineering: making developers more efficient historically increases demand for software, and engineering is about understanding human requirements and workflows, not merely translating English into code. Modal spends about $500,000 monthly on coding agents for 150 engineers, who are already highly agent-oriented. Erik also sees a major opportunity in selling data environments and services to frontier labs, especially for coding, while warning that such companies may be excellent cash businesses without becoming enduring venture-scale platforms.

Erik says proprietary models make sense for some customers because fine-tuning an open model provides stable behavior and control, but latency, cost, performance, hackability, and adjacent services matter just as much. He sees model routing as a viable standalone business today, citing OpenRouter’s growth and brand, but thinks it will eventually belong inside a broader platform offering evaluation, fine-tuning, and performance controls. Modal’s next opportunities include training, post-training, batch jobs, databases, vector stores, workflow orchestration, and feature stores; Erik hopes training can reach $100 million in run rate within 6–12 months. He supports employee secondaries with guardrails and says exceptional new graduates can receive offers above $20 million from major labs, while Modal competes through high-percentile compensation and greater equity upside.

Erik expects an AI-market pullback, perhaps a 20–30% stock-market decline that causes widespread panic, but says he is comfortable with it because the long-term demand for inference and compute is enormous. Modal grew 50 times in the last year and is currently growing 30–40% month over month; he hopes to be far beyond $4 billion in revenue within 12 months and considers the company’s roughly $15.75 billion valuation inexpensive relative to the opportunity. He changed his mind most dramatically on the size of the sandbox market. In the closing rapid fire, he backed TurboPuffer as his favorite angel investment, urged companies to raise each round as if it were their last, and expressed excitement about AI-assisted drug discovery and healthier, longer lives.