Erik Bernhardsson from Modal Labs
Modal founder Erik Bernhardsson explains how a cloud built for bursty, compute-heavy Python workloads found its wedge in generative AI — and why the fastest route to adoption is letting developers feel the product's magic within minutes.
Erik Bernhardsson founded Modal after engineering and leadership roles that included Spotify and Better.com. Modal gives data teams a Python-native way to run models, GPU workloads, batch jobs and other demanding computation without managing the underlying infrastructure.
We talk about Modal's development loop, its unexpected generative-AI wedge, bottom-up adoption, why onboarding and documentation are part of the core product, the limits of early-stage analytics, and Erik's decision to spend the company's first years unusually focused on engineering.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
What kinds of workloads is Modal designed to run?
Modal is cloud infrastructure for data teams and data products. It handles jobs such as model inference, speech and video processing, computational biotech, 3D rendering and web scraping. The common thread is work that stretches traditional serverless systems: it may need GPUs, run for a long time, fan out sharply or require a complex environment.
What does the developer workflow look like?
Data teams have varied infrastructure needs but overwhelmingly use Python, so we made Modal an end-to-end Python SDK. You describe the compute environment alongside the application code, then deploy or run it ephemerally from the command line. Because we built our own container infrastructure and filesystem, code on a laptop can begin executing in the cloud in about a second without a separate Docker build-and-push loop.
Who adopted Modal first?
I originally imagined a general-purpose replacement for Kubernetes and tried selling into enterprises, but that was a difficult replacement sale. Generative AI created a new cohort of engineers with no existing infrastructure to displace. They simply wanted to run Stable Diffusion, fine-tuning, audio, music or video workloads, and that clean slate became the wedge behind most of our growth.
How did people discover the product?
Almost all discovery was inbound. I had written a blog and maintained a presence on Twitter, so founder-led posts and conversations brought people toward Modal without much spending on sales or marketing. The early audience skewed toward indie hackers and young startups, with enterprise adoption beginning later.
Why optimize so aggressively for a quick first run?
Modal can be hard to explain, and the product sells itself better when somebody feels it. We therefore try to move a visitor from the site to installing one Python library and running cloud code within a few minutes. The absence of a complex configuration process is not just onboarding polish; it demonstrates the different developer experience we are selling.
How did you get from running a 300-person organization back to founder mode?
After managing hundreds of people, I wanted to write code for a year and went into full hacker mode for roughly two years. I explored web scraping, workflow scheduling and vector databases before focusing on the compute layer. The platform vision stayed fairly stable, but its killer use case did not become clear until generative AI supplied product-market fit.
What should DevTools founders get right about docs and onboarding?
Treat account creation, setup and the first successful run as the core product. Documentation should begin with useful code rather than a wall of invented concepts. Developers often structure docs as if decomposing a system for a computer, but a human needs a mental model. SDK and API design have the same requirement: you are writing code for a person as well as a machine.
How much can an early startup learn from product analytics?
Data is a tool, not an obligation. We use it for operational questions such as latency and utilization, but our customer count was too small to read onboarding conversion changes with confidence. At that scale, you need qualitative judgment in the same way a small company chooses a shade of blue rather than running Google's enormous experiment.
How do you turn customer feedback into the right product decision?
Talk to users early and maintain a feedback loop, even before the product has meaningful production adoption. Then look beneath the feature they request to the problem creating it. People asked to run infrastructure locally because cloud iteration was slow; solving cloud startup time instead removed the need for a second environment and produced a simpler system.
What founder advice best reflects how you built Modal?
The more time I spend on the core product and validate it with real users, the easier everything else becomes. Until shortly before this conversation, all 12 people at Modal were software engineers. I spent about two and a half years building before putting serious energy into marketing. That will not transfer perfectly to every company, but relentless product focus worked for us.