Modal Raises $355 Million at a $4.65 Billion Valuation

Modal, building high-performance AI infrastructure, led by Erik Bernhardsson and Akshat Bubna, has raised $355 million at a $4.65B post-money valuation led by General Catalyst and Redpoint, with Menlo Ventures, Bain Capital Ventures, and Accel joining as new investors. All the existing major investors participated as well, doubling down on their conviction in Modal.

Modal was founded on the belief that traditional cloud infrastructure was not designed for AI workloads.

The company is building an AI-native cloud platform with infrastructure primitives for applications such as low-latency inference, dynamic agent runtimes, reinforcement learning, and large-scale batch processing.

The shift from frontier APIs to model ownership is accelerating as teams, from companies like DoorDash to AI-native startups such as Reducto, fine-tune models on proprietary data, run reinforcement learning, and optimize inference for their own latency, throughput, and cost needs.

With production-ready open-weight models (e.g., DeepSeek, Qwen) and mature inference engines like vLLM and SGLang, companies can now fully own and serve their models without sacrificing capability.

“Modal powers both reinforcement learning infrastructure and production inference, from large-scale sandboxes to real-time serving on a single platform,” said Scott Wu, CEO of Cognition.

“Decagon achieved a p90 latency of 342ms, well below the sub-second threshold required for natural customer conversations, delivering enterprise-scale speed and reliability,” said the Decagon research team.

Agents need better execution environments. Modal built Sandboxes as an isolated runtime for untrusted code, anticipating the rise of AI-generated execution at scale.

Over the past year, agent adoption has surged across use cases like merchant automation, coding agents, and large-scale RL workloads, leading to over 1 billion sandboxes launched on the platform.

“Sandboxes are a key building block for reinforcement learning,” said Yash Patil, CEO of Applied Compute. “Modal stood out for its flexibility, performance, and reliability in building complex environments.”

“As we scale agentic commerce for local businesses, we need a highly efficient path to production with full control, scale, and reliability,” said Andy Fang, CTO of DoorDash. “We’re excited to evaluate Claude Managed Agents for this next step, building on our AI infrastructure with Modal.”

The shape of AI workloads keeps expanding, requiring flexible infrastructure for inference, training, and orchestration at scale.

Modal provides a general compute platform built around elastic compute, safe isolation, and programmatic control, enabling diverse applications, from real-time robotics inference and drug discovery pipelines to large-scale generative music systems, all using the same core primitives.

“We use Modal to run edge inference with under 10ms overhead and large-scale batch jobs, giving us strong performance, power, and flexibility,” said Brian Ichter, Co-founder of Physical Intelligence.

“It’s not just time savings, it’s the mental overhead that disappears,” said Kevin Wu, ML Researcher at Chai Discovery. “With Modal, we can scale functions with a few decorators, and they just work.”

Modal has spent five years building a full-stack compute and storage system, enabling breakthroughs like 100x faster cold starts via GPU snapshotting, elastic low-latency global inference, and scaling from zero to 1,000 GPUs in seconds through a distributed multi–data center capacity pool.

Owning the full stack allows the company to continuously improve performance and developer experience over time.

Modal is expanding its platform around three core areas.

First, it is pushing low-latency inference at scale, improving serving primitives, observability, and contributing to the open inference stack through work on tools like Flash Attention, vLLM, and SGLang.

Second, it is working to collapse the training and inference loop by strengthening support for reinforcement learning, multi-node training, elastic inference, and sandboxes, making the full model lifecycle more accessible.

Finally, it is building the compute layer for agents by scaling its Sandbox infrastructure and adding features like granular RBAC, enabling safe, large-scale agentic execution.