Modal CTO Akshat Bubna argues AI agents break the cloud's human developer assumptions, and the company's $4.65B valued Series C is funding infrastructure primitives built for AI without humans in the loop.
The cloud was designed for developers who could fill in the gaps. AI agents cannot. Modal CTO Akshat Bubna argues that mismatch is why infrastructure needs an agent-grade rebuild, and the company is now putting a $355 million Series C at a $4.65 billion valuation behind it.
Modal's framing lands at a moment when the broader agent cloud category is hardening into a discrete investment lane. Latent Space used Bubna's recent appearance as the closing chapter of a series covering Databricks, Daytona, Railway, and E2B, four companies pitching overlapping slices of an agent-oriented stack. The Modal round is the first of those bets to scale to a tier-1 venture valuation, and Bubna used the podcast to articulate the design gap the rest of the cluster is racing to fix.
The core complaint is that the cloud was built around a human in the loop. Bubna's case, as he put it on Latent Space, is that today's primitives assume a developer at the keyboard: someone to read a stack trace, retry a flaky request, or guess at the right endpoint.
An autonomous agent operating at machine speed does not pause to ask which Python version the runtime defaults to, whether the storage blob is in the same region as the GPU, or how to interpret a missing field in a JSON schema. Modal's response is a stack where most of those choices are removed or made programmatic. Its agent-experience post lists the parts: serverless functions, decorator-based infrastructure, elastic inference for custom models, GPU snapshotting, speculative decoding, Auto Endpoints, persistent storage, networked containers, private IPv6, RDMA, and multi-node training.
The Modal stack's distinctive move is treating those primitives as one constraint rather than a checklist. Bubna frames the whole stack as a feedback-loop problem. Agents train on rollouts, and a rollout that takes an hour because of cold start, snapshot restoration, or inter-region data transfer is a rollout an RL team will not run. Modal operates capacity across 17 cloud providers, and Bubna cited workloads on the order of 100,000 sandboxes at once as one concrete driver of demand. That figure is Modal's own anecdote rather than a published benchmark, but it gestures at the request shape: many short-lived, isolated, GPU-attached compute environments running concurrently for a single training job.
Two years ago Modal joined Latent Space to discuss its $17 million Series A. The same thesis is now running through a much larger balance sheet. Bubna amplified the agent-primitive narrative on X on the day of the round, suggesting the messaging is coordinated across the Series C announcement, the podcast appearance, and the agent-experience post.
Two caveats earn attention. The agent-cloud framing is Modal's, and Modal's stack is built to advantage Modal. The Series C terms currently sit on Modal's own blog, with no tier-1 independent confirmation in the press. Independent voices from the broader Latent Space series (Databricks, Daytona, Railway, E2B) are absent from this round of reporting, which means the design-gap claim is being carried by the company most invested in it being true. The competitive picture is still developing.
What changes on the strength of the round is what Modal can ship, not what the thesis is. A $4.65 billion valuation gives the company room to expand the sandbox layer, push deeper into RDMA and multi-node training, and compete on inference economics rather than just developer ergonomics. The next marker worth watching is whether any of Modal's customers publish a benchmark behind the rollout-loop claim, since that is where Modal's narrative becomes testable. Until then, the $355 million buys Modal time to keep building the cloud it thinks agents need. The agents get to decide if it works.