Modal CEO Erik Bernhardsson on Inference, Compute and Building an AI Platform

Harry Stebbings introduces Modal co-founder and CEO Erik Bernhardsson for a conversation about the inference market, AI infrastructure and the company’s growth. Figures about revenue, funding and valuation in the discussion are the speakers’ descriptions.

Bernhardsson says he started programming at eight on a Macintosh and made games for fun. He remains grateful that a childhood hobby became his work, and recalls enjoying programming competitions as a teenager.

He started Modal in his thirties after seeing what he considered an obvious market opportunity. Looking back, he thinks earlier confidence might have helped, though he is unsure he would have been a good founder in his twenties.

He rejects the idea that founding a company must feel like ‘eating glass’ every day for a reward at the end. There are hard conversations and difficult periods, he says, but he enjoys building and hopes to work on Modal for many years.

Stebbings relates a similar question about selling his own business: what would he do the next day? Bernhardsson says that if he sold Modal he would likely start a similar company, because he enjoys the work itself.

Fundraising is less enjoyable for him. He says he has not made a conventional pitch deck for the rounds discussed, preferring to build relationships with investors well before a financing and then discuss terms when the company needs capital.

Asked whether founders should maximize valuation, he says early companies can take more money and dilution than they need. He distinguishes that from Modal’s current period, in which securing GPU capacity requires much more cash.

He describes the company as profitable on a profit-and-loss basis but cash-flow negative because of capacity investment. Capital helps, he says, though he is cautious about using the word ‘moat’ for every advantage.

Bernhardsson argues that early startups rarely possess durable moats in a strict sense. He returns instead to an image of a company staying ahead by moving quickly, while Stebbings emphasizes team, product-decision speed and brand.

Both see brand as increasingly valuable. Bernhardsson connects Modal’s identity to developer experience, website design, community and consistency, while saying he remains attentive to competitors such as Baseten and Fireworks.

He says Modal began with a product-led approach: developers could sign up and experience the product directly. His consumer-software background influenced the desire to make infrastructure feel distinctive from the first visit.

Bernhardsson expects inference to become a very large market, comparable in importance to a new cloud layer. He argues that customers also need post-training, sandboxes, agent tools, low-latency serving and batch processing.

If a provider sells only tokens for dollars, he says, customers can switch easily and pricing can become a race to the bottom. A broader set of tools and the infrastructure connecting them can provide more differentiation.

He sees room for gross margin across those adjacent services, but is skeptical that most infrastructure companies will achieve the exceptionally high margins associated with some older software businesses.

His explanation is that a customer’s spending passes through several layers—application, model, infrastructure, cloud and hardware—each with its own economics. He suggests Modal could eventually improve its margins while remaining below traditional software levels.

His prediction is that inference will fragment into many specialized products as the market expands. Specialization by product does not prevent a platform from offering multiple products, he says; different applications need different combinations of latency, batch processing and training.

He describes sourcing GPUs as a frenetic market in which Modal works with many specialized cloud providers. The company’s goal is to spare customers those negotiations and aggregate capacity on one side with demand on the other.

Modal has built its own scheduling, container runtime and filesystem components, Bernhardsson says. Pooling many customers’ variable workloads can smooth aggregate demand, making it possible to offer elastic capacity instead of asking each customer to reserve a fixed block for years.

With GPU lead times of several months, the company projects aggregate demand using recent growth and obtains capacity ahead of it. He says that prices have risen, but believes customers willing to pay enough can still acquire GPUs.

Asked where prices go next, he says he does not know when supply will catch up. His expectation is that hardware prices decline in the very long run, even though shortages can keep them elevated for years.

Both speakers believe AI-compute demand remains early. Bernhardsson notes that some products consume significant compute before generating revenue, so the timing between investment and sustainable revenue remains an open question.

On specialized chips, he expects NVIDIA and alternative suppliers to grow rather than assuming one must eliminate the other. He calls CUDA’s specific programming model less important as a moat than the wider NVIDIA-first ecosystem of frameworks, kernels and inference engines.

Specialized cloud providers are both partners and potential competitors as some move into inference. Bernhardsson says he is biased, but believes reproducing Modal’s scheduler, runtime and large-scale orchestration would be difficult.

He also respects the financial execution required to build a GPU cloud: secure customer commitments, finance equipment, deliver capacity and service debt. That capital-intensive sequence is a different challenge from building software.

Stebbings asks why open models matter if proprietary frontier models become cheaper. Bernhardsson expects continuing demand for fine-tuning, control and the ability to optimize specific latency or throughput requirements.

He says licensing can help model developers capture value and fund future work. He wants strong open-model ecosystems in Europe and the United States as well as China, while expressing respect for the technical strength of Chinese labs.

He worries that Europe can struggle to concentrate talent and capital at the scale needed for frontier model training. He later clarifies that Europe has abundant talent; his concern is aggregating it within a single organization.

Bernhardsson views NVIDIA’s support for open-model developers as an ecosystem strategy. The companies are capital intensive, he says, and investment can help sustain models built in the United States.

He also expects more specialized models as markets grow. He cites Jev as an example focused on tasks such as classification and regression, and speculates that a portfolio of specialized models could support a very large business.

The discussion returns to brand. Bernhardsson says early Modal growth included direct conversations with developers online; each company, he argues, needs a way of communicating that fits its own product and audience.

Asked how to improve the customer experience, he wants more products within Modal and more people who can help customers directly. Stebbings asks whether expansion risks spreading the company too thin.

Bernhardsson agrees that prioritization and resource allocation are hard. A multi-product strategy requires choosing when to launch another product rather than attempting everything at once.

Modal launched its sandbox product in 2023, he says, but its major growth came around 2025 as models improved at generating code. He describes it as a roughly $300 million run-rate business at the time of the interview.

Infrastructure products can take years to find a decisive use case. Modal’s earlier serverless functions found demand in generative inference, while sandboxes later benefited from coding tools, reinforcement learning and background agents that the company had not fully anticipated.

He credits both timing and a willingness to build useful primitives before the eventual use case is obvious. He says the company benefited from beginning before later generative-AI demand accelerated.

With hindsight, Bernhardsson says Modal underinvested in open-model inference. After DeepSeek attracted attention, the company expanded the team substantially, having previously worried that inference would be mainly a commodity.

Stebbings presents a counterargument that open-model inference may plateau as proprietary models become inexpensive and some enterprises avoid Chinese models. Bernhardsson calls that a fixed-pie view and forecasts substantial growth for both categories; his numbers are predictions, not results.

The speakers discuss Modal reaching a billion-dollar revenue run rate. Bernhardsson says recent growth has been rapid but does not know exactly where it will settle; his optimistic comments about future revenue are forecasts rather than guarantees.

Among possible new revenue streams, he names training, post-training and batch jobs. He also sees adjacent opportunities in data systems and orchestration, while acknowledging that the company must select only a few priorities at a time.

Post-training benefits from tasks with relatively clear feedback, he says. Coding can be tested in sandboxes, while mathematics and some security tasks also provide ways to evaluate results.

He is less certain how quickly the same methods will work for tasks whose feedback loops span months. He notes that he has been wrong about AI progress before and leaves open the possibility that such limitations will move.

Bernhardsson remains optimistic about software engineering. He argues that making programmers more productive has historically enabled more software, and that the job includes understanding human needs and workflows, not simply translating English into code.

He says Modal spends roughly $500,000 per month on coding agents for around 150 engineers. He expects more use cases to emerge as tools improve, though future spending and prices are uncertain.

He sees strong current demand for data and environments used by frontier labs, especially coding environments. At the same time, he is unsure whether every profitable provider of such data is necessarily suited to venture funding or durable long-term economics.

Asked whether every company should own its models, he says it depends on the use case. A fine-tuned open model can give a customer consistency and control, but latency, price, performance and surrounding services also matter.

Modal supports both customers using open models and companies serving proprietary models on its infrastructure. Bernhardsson presents that flexibility as part of the platform strategy rather than insisting that one model choice suits everyone.

On model routing, he says a standalone company can succeed today, though over the long term routing may fit naturally into a broader collection of evaluation, fine-tuning and performance tools. He discloses that he invested in OpenRouter.

Asked for a view that could work against his business, he anticipates an eventual AI-market pullback even though he remains optimistic about long-term demand. This is his market expectation, not a prediction with a known date or magnitude.

He says he personally takes little secondary liquidity but supports limited employee share sales with guardrails. In his view, a precedent for liquidity can make equity more tangible to employees while preserving incentives.

Competition for AI talent is intense. He says Modal has lost exceptional new graduates to labs offering very high compensation and believes its own mix of pay, equity and growth potential must remain compelling.

He names sandbox demand as an area where he underestimated the market. He is wary of companies that depend on repeated equity funding without a credible path to revenue, particularly if capital markets tighten.

The conversation closes with reflections on parenting, Scandinavian culture and interests beyond work. Bernhardsson says he is excited about AI-assisted drug discovery while acknowledging that development and regulatory processes take time.