Company Spotlight: Modal and the Infrastructure AI Runs On

A researcher working on Modal can keep changing the reward function while another team operates the cluster underneath the experiment. That division of work runs through the New York company's proposition: give AI builders control over their applications and models, with a managed platform handling the infrastructure needed to execute them.

Founded in 2021 by CEO Erik Bernhardsson and CTO Akshat Bubna, Modal supplies cloud infrastructure for training, inference, batch processing and agent execution. Its customers span application developers, research organizations and enterprise AI teams. The company has built its own underlying systems while making important developer tools publicly inspectable, joining an opinionated operating platform to code that customers can work with directly.

Modal's October 14, 2026 Series D brought $750M at a $15.75B post-money valuation, led by Accel. That financing belongs inside a broader company story about who operates the growing collection of environments, model servers and training jobs behind AI products, and which decisions the customer continues to own.

How Modal connects training code with managed compute

Modal's reinforcement-learning engineering account breaks a training loop into connected jobs: updating a model, generating attempts through inference and executing actions in isolated environments. Each part has its own infrastructure demands. Coordinating them matters because an expensive training resource can sit idle while another part of the experiment catches up.

The company introduced Modal Dojo to reduce the setup around those workflows. The current open-source library supplies training configuration and tutorials, letting teams work with models, rewards and environments. A researcher can inspect and change the code that defines the experiment while using Modal's infrastructure underneath it. The library's public availability is directly observable; claims about an individual workload's performance still need to be evaluated against that workload.

Modal's official SDK repository makes another part of the relationship visible. Python supports application deployment and the command-line interface, while JavaScript, TypeScript and Go support Sandboxes, deployed functions and selected platform resources. Those interfaces give developers ways to make compute part of application code. The distinction is commercially useful: customers can retain responsibility for what the software does while paying a supplier to carry much of the operational machinery required to run it, with SDK capabilities varying by language.

The company beneath the developer interface

Bernhardsson and Bubna built Modal around making iteration easier for developers. The company's account of its platform identifies a custom file system, container runtime, scheduler and image builder. Those are substantial systems to operate behind an interface designed to make a compute request feel routine, and they explain why Modal's engineering work extends well below model APIs.

That foundation serves several kinds of application. Functions execute code on configured resources; Sandboxes create isolated environments for arbitrary code; model-serving and training capabilities connect the platform to different stages of an AI product. The customer can assemble those building blocks around a workload, while Modal manages the fleet and the runtime. Product breadth therefore increases the number of operating relationships the company has to support.

Modal has also brought another infrastructure team into the business. Its July 2025 Jamsocket acquisition announcement described products for collaborative backends, document synchronization and stateful Python execution, with functionality intended to migrate into Modal. That historical announcement establishes the acquisition and its rationale; it does not establish that every migration has finished. The connection is still instructive: applications using AI often need state and interaction around the model, giving a compute platform reasons to understand more of the surrounding software rather than treating inference as an isolated request.

Why Modal's customer breadth matters

Modal names customers across very different application categories, including Suno in music, Chai Discovery in drug discovery and Physical Intelligence in robotics. It also describes work with companies building agents and enterprise applications. These are company-published accounts, but the range illustrates the intended reuse of a compute platform: storage, execution and scaling can be shared needs even when the products built above them have little in common.

The economics are growing alongside that scope. Modal says it recently crossed $1B in annualized run rate, with Sandboxes at a $300M pace, and now reports over $1.2B raised in total. Annualizing a current rate does not turn it into a completed year's recognized revenue. The figures describe the company's reported operating scale without establishing margins, retention or the concentration of spending among its customers.

The latest $750M Series D was led by Accel, with Kleiner Perkins and existing investors participating. Funding gives Modal resources to continue developing the platform, while its customer mix makes the assignment unusually broad. Research experiments, agent tasks and production serving can demand different combinations of responsiveness, isolation and capacity. A supplier serving all three must make those differences manageable without expecting every customer to build a dedicated infrastructure organization.

What customers still own when they use Modal

A managed execution platform changes the boundary of an engineering team's work. Modal's security documentation describes a shared-responsibility model in which Modal operates and patches the platform while customers manage identities, secrets, application code, dependencies and the endpoints they expose. That boundary becomes particularly consequential when an AI agent can generate and run code.

Modal documents isolation primitives and controls for untrusted workloads, but the customer still chooses guardrails such as resource limits, timeouts and network restrictions. The documentation also identifies customer responsibilities for backups and application failover. These details belong in a product evaluation because an environment's ability to execute a task and an organization's authority to permit that task are related operating decisions.

Pricing creates another part of the relationship. Modal's current pricing page combines usage-based compute with plan features and charges that depend on resources and configuration. A team evaluating the service needs to understand the full workload, including attached storage, networking and requirements that influence placement or execution behavior. The relevant comparison is the cost and effort of operating that application under its actual conditions. A simple function call can be a useful interface while the engineering organization still keeps a detailed view of what happens behind it.

What Modal's hiring and public tools reveal

Modal's current careers listings cover systems, storage, SDK development, inference, post-training research, security and customer-facing work. Roles in solutions architecture, account management and operations appear alongside low-level engineering positions. The company reports a team of 200+ across New York, San Francisco and Stockholm; its office and job listings also include London.

The range of roles is a useful market signal without becoming a recruiting slogan. Developers adopting the platform need reliable systems and clear interfaces, while organizations putting those workloads into production need help with integration, support and commercial decisions. The staffing requirements follow the work Modal is asking customers to entrust to it. Current listings show the functions being recruited, rather than proving how quickly those positions will be filled or how much revenue they will produce.

The public SDKs, examples and training library give builders a way to examine that proposition before a sales conversation. They also let the community learn from runnable interfaces and contribute around the tools it uses. For an operator assessing Modal, the product documentation and open roles can be read together: one describes the capabilities available today, while the other shows where people are needed to keep those capabilities useful as a researcher's experiment becomes somebody else's production application.