Modal Raises $750M at a $15.75B Valuation and Reaches $1B in ARR

Modal, a New York-based AI infrastructure provider, has raised $750M in a Series D round at a $15.75B valuation. Accel led the round. The company's annual recurring revenue (ARR) has passed $1B, and platform usage has grown more than 15x over the past year. In the five months since its previous round, Modal has tripled both its workload and its valuation. It has moved from the ranks of infrastructure unicorns into decacorn territory.

Sandboxes That Outgrew the Rest of the Product

Sandboxes became the main growth driver. Modal launched them in 2023, when it noticed that people had started generating code with LLMs at scale. That code has to be executed safely somewhere, and an isolated environment turned out to be exactly what agents were missing. Three years later, Sandboxes is a standalone product with a $300M run rate.

The scale is substantial. The platform has launched about 3B sandboxes and can hold up to a million running concurrently. Use cases vary: background agents, reinforcement learning, vibe-coding platforms and more. In his post about the round, co-founder and CEO Erik Bernhardsson frames it as the market catching up to an idea Modal built before the agent boom.

At the AIE NYC conference in July, CTO Akshat described the company's plans for agent workloads:

• millions of concurrent sandboxes;
• secure, elastic launch of RDMA clusters for distributed training;
• training of modern speculators for inference;
• a new sandbox runtime.

The approach is clear. Modal is building more than GPU rental: it is building an execution layer for code written by models. Sandboxes, training and inference run on the same foundation, so a single customer can grow across all three.

Neighboring Players Are Growing Just as Fast

Modal is not alone. According to DealBook, other inference providers report comparable figures:

• Together — $1.2B in annualized bookings;
• Fireworks — above $1B in ARR;
• Baseten — revenue tripled to a $600M annualized run rate;
• Modal — $1B in ARR and a $15.75B valuation.

At the same time, agent sandboxes are becoming a niche of their own. Several players have shown their own solutions:

• Perplexity SPACE: handles millions of executions per day with an error rate below 0.0002%, and ties requests to executions through short-lived signed identifiers;
• Factory OpenCompute: a self-hosted sandbox orchestrator that keeps agents inside the customer's VPC;
• Prime Sandboxes: filesystem checkpoints let teams branch RL rollouts from intermediate states.

Each addresses a different pain point: security, data control or training reproducibility. Modal wins on versatility and speed of scaling.

What Teams Building Agents Should Do

For developers, the main takeaway is that agent code execution has become a separate layer of the stack, with its own metrics: startup time, number of parallel environments, isolation and cost. When choosing a platform, it is worth checking up front whether it can handle not dozens but thousands of concurrent sandboxes. For regulated industries, options like Factory OpenCompute matter more, since code and data stay within the customer's perimeter. Teams working on reinforcement learning should look at checkpointing and state-forking mechanisms like those in Prime Sandboxes.

For the market as a whole, the numbers point to one conclusion: money is flowing not only to training large models, but also to the infrastructure that makes them work. If the valuation, tripled in five months, holds, the next question will be the profit margins of such services as competition grows from cloud giants and specialized players.