Modal CEO Erik Bernhardsson says the GPU market is far from efficient, with compute too heterogeneous to trade like a fungible exchange contract.
Bernhardsson says Modal's core value is aggregating demand across 25 clouds so it can manage scarce capacity more efficiently, while reliability separates providers.
Sandboxes now make up about a third of Modal's revenue, and Bernhardsson expects GPU rental prices to rise as Anthropic and OpenAI keep buying capacity.
Modal's CEO on why platforms beat tokens in AI infrastructure
Modal sells infrastructure for AI workloads, and its bet is that a full platform outlasts any single product. Commodity token serving will race to the bottom, while sandboxes, training and inference together build lasting advantage. Compute stays a messy, non-fungible market, so aggregating demand and running reliable capacity is the real work.
Takeaways
The GPU market is chaotic and inefficient. Compute is heterogeneous, so it cannot trade like a fungible contract on an exchange. Modal does two or three deals a week across 25 clouds. Its value is aggregating demand, and reliability is the main difference between providers.
Sandboxes grew into a third of revenue. Launched in 2023 for running model-generated code in isolation, Sandboxes took off in 2025 with a standalone run rate of 300m dollars. Inference is another third. Total run rate rose from about 2m at the Series A pitch to over 1bn.
Durable inference value is a platform, not tokens. Commodity token serving, which OpenRouter is commoditizing, will be a race to the bottom. Lasting advantage comes from offering sandboxes, reinforcement learning, fine-tuning, batch processing and deployment together. Customers include Cognition, Ramp, Suno, Runway, Decagon, DoorDash and 1X.
Building its own stack drives performance. Modal built its own container runtime and scheduler instead of using Kubernetes and Docker, and uses gVisor and now KVM. Snapshotting loads large models onto GPUs in seconds. A team of 20 to 30 tunes inference, and Modal contributes to vLLM and SGLang.
GPU prices should rise, and neoclouds are exposed. Neoclouds are fine but undifferentiated businesses with high leverage and expensive capital. GPU commitments needing about 30 percent upfront create implicit leverage for neolabs, and not all will succeed. Anthropic and OpenAI keep buying, so rental prices should rise.
Agents are now both builders and users. About 95 percent of merged code at Modal is written by agents to some extent. Modal tracks how coding assistants recommend it, much like search optimization. Agents favor frameworks in their training data, which gives software more staying power.
Ruthless prioritization is the CEO's hardest job. With about 150 engineers, choosing what to build matters most. Sandboxes were an asymmetric bet, cheap to prototype, then rewritten about five times in 2025 to handle scale. Bernhardsson says Modal was slow to chase LLM inference until DeepSeek.
Modal's next bets
Low-latency inference that behaves like a CDN.
Mid-sized model training and reinforcement learning.
Tap a name to go there · tap a line to see why they are linked
On this episode 4
Show 1
VMs that run full operating systems.
Automatic GPU switching under the hood to raise utilization, alongside a ramp of Blackwell GPUs while many H100s keep running older and smaller workloads.
Bernhardsson predicts Modal will look like a conglomerate in years ten to fifteen.