Building AI-Native Developer Infrastructure | Erik Berhnardsson, CEO of Modal
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
This episode features Eric Bernhardsen, co-founder and CEO of Modal, discussing his journey building AI infrastructure from the ground up. Eric shares his experience going from recommendation systems at Spotify to creating a new cloud provider optimized for AI workloads. (00:00) The conversation covers the great GPU shortage, lessons from hypergrowth, and why traditional infrastructure wasn't built for AI applications that require fast feedback loops and dynamic scaling. (08:00)
- Main themes: Building AI-native infrastructure, the evolution from traditional cloud to AI-optimized platforms, entrepreneurship lessons from Spotify's early days, and navigating the AI infrastructure landscape
Speakers
Eric Bernhardsen
Co-founder and CEO of Modal, Eric previously spent seven years at Spotify as the 40th employee, where he built the music recommendation system and led a team focused on machine learning infrastructure. Before Modal, he served as CTO at Better.com for five years, scaling the tech team from one to 300 people. Eric started his coding journey at age 8 on a Macintosh Plus and competed internationally in programming competitions, representing Sweden at the International Olympiad in Informatics.
Turner Novak
Host of The PL and founder of Banana Capital. Turner conducts in-depth interviews with founders and investors, exploring the world's greatest startup stories across over 100 episodes.
Key Takeaways
Build Infrastructure for Fast Feedback Loops
Traditional infrastructure works well for backend applications but breaks down for AI development where you need rapid iteration cycles. (07:30) Eric observed that 90% of AI work involves building infrastructure rather than applications, creating friction that slows development. Modal's approach enables developers to write code and run it in the cloud immediately, collapsing the traditional development-to-deployment cycle. This isn't just about convenience - it's about enabling the kind of rapid experimentation that AI development demands.
Focus on Problem Discovery, Not Solution Suggestions
When working with early customers, take their pain points extremely seriously but almost ignore their suggested solutions. (09:09) Eric learned that customers often propose specific features when the real need is much broader. By going several steps back to understand the underlying problem, companies can build more general, valuable solutions rather than narrow fixes. This approach prevents you from building a "ridiculous MVP" based on every customer's specific requests.
Choose Hard Problems That Take Time
Eric intentionally picked a problem that would take 1-2 years to solve, believing this creates natural moats. (26:08) While accelerators push for immediate product-market fit, some startups benefit from solving genuinely difficult technical challenges that competitors can't easily replicate overnight. This approach requires more upfront investment and patience but can lead to more defensible businesses in the long run.
Optimize for Carbon-Silicon Balance
The relationship between human costs (carbon) and compute costs (silicon) determines optimal development strategies. (54:57) When GPUs are more expensive than engineers, teams over-optimize for compute efficiency at the expense of developer productivity. As compute becomes cheaper relative to human talent, the winning strategy shifts to tools that make engineers more productive, even if they're less computationally efficient.
Start Companies Later in Your Career
Eric advocates for more people starting companies in their 30s and 40s rather than their 20s. (67:58) By age 37, he had developed crucial skills in hiring, product building, and team management from his time at Spotify and Better.com. This experience made many startup challenges easier to navigate compared to learning everything on the job. The key is having enough savings or exit proceeds to take the financial risk while leveraging accumulated expertise.
Statistics & Facts
- Modal aggregates compute capacity across 80 different regions to ensure GPU availability for customers. (14:23) This multi-region approach helps solve capacity constraints that single cloud providers face.
- Eric estimates there's approximately $500 million in combined annual revenue across AI coding platforms today. (80:19) This represents massive growth in a market that barely existed a year prior, demonstrating real commercial demand for AI applications.
- CO2 levels in offices regularly hit 2,000-3,000 PPM, while airplane cabins can reach 2,500-3,000 PPM during takeoff. (96:51) These levels significantly impact cognitive performance compared to optimal outdoor air levels of 300-400 PPM.
