Overview of 20VC with Modal CEO Erik Bernhardsson
In this episode, Harry Stebbings speaks with Erik Bernhardsson, co-founder and CEO of Modal, about the exploding inference market, the economics of compute, why AI infrastructure is becoming a massive platform opportunity, and how Modal is positioning itself beyond raw inference into a broader AI cloud stack. Erik is extremely bullish on demand, arguing that inference is still early, that specialization will create many submarkets, and that the winners will be platforms offering a full suite of products rather than “token-selling” commodity infrastructure. The conversation also covers GPU scarcity, NVIDIA’s staying power, open-source vs frontier models, post-training and RL, brand-building, and what it takes to build a durable AI company.
Key Takeaways
-
Inference is still in a very early, massive expansion phase
- Erik believes inference will become an “enormously large market,” effectively the new cloud layer.
- He expects demand to keep growing by orders of magnitude, not plateau.
- His core thesis: the market is large enough for both frontier and open-source inference to grow substantially.
-
The winning companies will be broad platforms, not narrow commodity providers
- Modal’s strategy is to build a suite around inference:
- low-latency inference
- batch inference
- sandboxes
- post-training / reinforcement learning
- training
- agentic workflows
- His view: if a company only sells “tokens for dollars,” it risks becoming a commoditized compression layer.
- Modal’s strategy is to build a suite around inference:
-
Compute is becoming a financial and operational game
- GPU sourcing is still highly competitive and messy.
- Modal treats compute as an aggregation problem: pooling demand across many customers lets them manage capacity more efficiently.
- Erik frames the business as part infrastructure, part financial engineering, part supply-chain optimization.
-
Open-source models still matter
- He does not believe frontier models will fully crowd out open-source models.
- Open-source is valuable for control, fine-tuning, latency, throughput, and consistency.
- He also thinks it matters strategically that the US and Europe have strong model ecosystems, not just China and the frontier labs.
-
Post-training is important, but likely not universal
- Erik is highly bullish on RL/post-training for domains with clear feedback loops, like coding, math, and cybersecurity.
- He’s less certain it will work equally well for tasks with long feedback cycles, like large-scale engineering or business workflows.
- He sees it as powerful, but not a complete replacement for pretraining.
-
Brand, speed, and team matter more than traditional “moats”
- He is skeptical of moat talk in early-stage startups.
- He emphasized:
- top-tier team
- fast product decision-making
- brand/community
- In his view, startup defensibility often comes from moving faster and building stronger product energy than competitors.
Modal’s Strategy and Product Expansion
Why Modal is Broadening the Product Set
Erik argues that inference alone is too narrow to create enduring value. Modal is building adjacent products to become the AI infrastructure layer for many workloads, not just serving model requests.
Areas Modal Is Expanding Into
- training
- post-training / RL
- sandboxes
- batch jobs
- agent harnesses
- workflow orchestration
- data infrastructure adjacent products
Why That Matters
- It improves retention and switching costs.
- It makes Modal more than a “commodity inference provider.”
- It lets the company capture more of the value stack as AI applications become more complex.
Compute Economics and GPU Market Reality
GPU Supply Is Still Tight
Erik described the GPU market as chaotic and highly competitive, with lead times around 3–4 months and ongoing price pressure.
How Modal Manages It
- It buys capacity ahead of demand based on aggregated customer growth.
- It works with a large set of NeoClouds.
- It builds much of its own infra rather than relying fully on Kubernetes/Docker-style defaults.
Long-Term View
- He believes compute supply will eventually normalize, but not necessarily soon.
- For now, demand is so strong that the market can keep absorbing more GPUs.
- He thinks hardware prices will fall in the very long run, but near-term pricing can stay elevated.
NVIDIA, Custom Chips, and the Infrastructure Stack
On NVIDIA
Erik does not think NVIDIA is weakening because specialized chips are emerging. His view is that both NVIDIA and specialized players can grow.
On CUDA
- He thinks CUDA’s moat is somewhat overstated.
- The real advantage is the broader NVIDIA ecosystem:
- kernel optimizations
- tooling support
- framework compatibility
- default support from open-source engines
On New Chip Players
- He expects a fragmented market with multiple winners.
- He doesn’t see a single winner-take-all outcome.
Open Source vs Frontier Models
His View
- Open source will remain important.
- Frontier labs will still be powerful because they can move fast and monetize.
- Open models matter for:
- customization
- control
- fine-tuning
- stable behavior
- specialized use cases
Strategic Note
Erik is supportive of more US- and Europe-based model development, partly for geopolitical balance and partly because customers want choice.
Post-Training, RL, and the Future of Software
Why Post-Training Is Taking Off
Erik points to tasks with clear verification loops:
- coding
- math
- cybersecurity
- some law workflows
His Bigger Point
More capable tools tend to increase demand, not reduce it. He sees AI as another major wave of software abstraction:
- assembly → COBOL → C++ → Python → agents
On Software Engineers
He is extremely bullish:
- software demand keeps expanding when tools get better
- engineers are more than “English to code” translators
- the real work is understanding requirements and building systems for humans
Company Building, Talent, and Culture
Founder Timing and Confidence
Erik says he started Modal later than he might have liked because he lacked confidence earlier. He believes many founders start too early with too much confidence, while his own case was the opposite.
Why He Thinks Modal Has Worked
- strong developer experience
- bottom-up product adoption
- clear product-led growth
- strong brand and community energy
Talent Market
- He says the competition for elite AI talent is intense.
- Labs are reportedly offering enormous packages to new grads.
- Modal aims to stay competitive via equity upside, strong culture, and long-term mission.
On Employee Secondaries
He supports them in moderation:
- useful for making equity feel like real currency
- helps recruiting
- but should be limited so employees still have skin in the game
Brand as a Real Advantage
Erik strongly believes brand now matters more than many founders admit.
What Brand Means to Him
- distinctive personality
- consistent identity
- community feeling
- visible “movement” around the company
Modal’s Brand Thesis
- build a product people want to talk about
- make the developer experience feel cool and different
- lean into personality and online presence
He also acknowledges he could do even more on this front.
Notable Predictions and Opinions
Predictions
- Inference will fragment into many specialized submarkets.
- Modal can become a much broader platform than just an inference provider.
- Open-source and frontier inference will both grow significantly.
- Post-training/RL will be a major force, especially in code-heavy domains.
- AI demand may keep rising even if there is a market pullback.
Opinions That Stand Out
- “Moats” are often overused in early-stage startup discourse.
- Brand is increasingly important in developer and AI infrastructure markets.
- Many AI businesses are still underestimating how much demand will emerge.
- The long-term market is large enough that multiple categories can win at once.
Final Thought
Erik’s worldview is fundamentally optimistic: AI infrastructure is not a narrow commodity race, but the foundation for an expanding ecosystem of products, workloads, and workflows. Modal’s bet is that the real winner in inference won’t just serve models — it will become the cloud platform where AI applications are built, tuned, deployed, and scaled.