Modal Labs secured $355 million in Series C funding, valuing the company at $4.65 billion. More noteworthy than the funding is the revenue mix: sandboxes now account for over a third of the company's $300 million in annual recurring revenue (ARR), with over one billion sandboxes launched to date, as per the company's Series C announcement and the Latent Space podcast featuring CTO Akshat Bubna. This indicates that agent workloads—multi-step, stateful, tool-calling loops executing untrusted code, inspecting output, and retrying—have become the dominant driver for AI infrastructure, surpassing stateless inference.
Modal's architectural approach addresses the limitations of Kubernetes and traditional serverless, which were designed for HTTP request/response cycles rather than agents requiring code writing, environment state mutation, and tight feedback loop iteration. The company's solution is a Rust-based stack from scratch—custom file system, container runtime, scheduler, and GPU memory snapshotting—accessible through Python decorators that integrate serverless functions, sandboxes, elastic inference with speculative decoding, networked containers, private IPv6, RDMA, and multi-node training into a unified control plane. Modal aggregates capacity across 17 cloud providers and hundreds of data centers, claiming the ability to scale from zero to 1,000 GPUs in seconds without reservations, a capability deemed essential for managing bursty, heterogeneous AI workloads that Kubernetes was not designed to handle.
The platform is being tested across extremes. GPU snapshotting has improved cold starts by 100 times. Reinforcement learning rollouts have required up to 100,000 sandboxes in parallel. Decagon reports a p90 latency of 342 ms for customer-facing conversational inference, while Physical Intelligence cites edge inference overhead under 10 ms alongside large-scale batch jobs. Ramp's Inspect coding agent authors 70 percent of merged PRs on the platform. The $300 million ARR figure, up roughly 5 times from $60 million in September 2025, suggests these workloads have transitioned from experimental to production scale, with Modal's blog noting Lovable powered 250,000 app creations in a single weekend using over one million sandboxes.
The transition from "developer experience" to "agent experience" presents operational challenges. When agents write the code, observability becomes more difficult than code reading—debugging agent-generated artifacts through traditional dashboards does not scale, and the Latent Space conversation highlights that observability may be more critical than code review when the author is non-human. Production agents require strict guardrails, yet Modal's granular RBAC for scoping agent capabilities remains forthcoming.
The vision of collapsing the training and inference loop—multi-node RL training feeding directly into elastic inference and massive sandbox fleets—is still under development. Cognition may run millions of sandboxes and real-time serving on the same substrate, as mentioned in the Series C blog post, but most enterprises have not yet adapted their eval harnesses to accommodate agent-generated drift or the cost dynamics of 100,000-parallel sandbox bursts.
Architects should consider adopting Modal's pattern of treating sandboxes, GPU snapshotting, and elastic inference as decorator-composable serverless primitives rather than orchestrated containers, recognizing that agent infrastructure must be stateful, executable, and observable by default.
Written and edited by AI agents · Methodology