Modal Labs

Modal Labs, usually called Modal, sells cloud computing for AI software without requiring customers to manage servers. Its pitch is that running a model on a GPU should feel like calling a Python function. A developer writes a normal Python file, adds a decorator such as @app.function with arguments that name the container image and GPU type, and runs a command. Modal packages the code into a container, finds a machine with the right hardware, runs the function, and shuts the container down when it is no longer needed. Its documentation states that configuration lives entirely in code, with no YAML files, and that users do not need their own Kubernetes cluster, Docker setup or AWS account.

Under the hood the workflow has several stages. First, the developer defines an image, a description of the container environment such as a base Python version and the packages to install. Second, Modal builds that image once and caches it. Third, when the function is called, from a laptop, a web endpoint, a schedule or another function, Modal's scheduler places a container on a worker machine with the requested CPU, memory and GPU. Fourth, the container runs any setup code, such as loading model weights, and then processes inputs. If requests pile up, the autoscaler adds containers; when they stop, idle containers shut down, by default after 60 seconds. This pattern is called serverless because the user never provisions or keeps a server running. Modal says it pools capacity across multiple clouds, more than 20 according to its October 2026 post, and chooses where to run each workload based on availability and cost.

The company was founded in 2021 by Erik Bernhardsson, its CEO, and Akshat Bubna, its CTO, according to TechCrunch. Bernhardsson spent more than 15 years building data teams, including at Spotify and as CTO of the mortgage lender Better.com; Bubna studied math and computer science at MIT and was an early engineer at Scale AI. Modal left beta and announced general availability on October 10, 2023, alongside a $16 million Series A led by Redpoint Ventures, bringing total funding to $23 million. Early customers named then included Ramp and Substack.

Growth accelerated with demand for AI inference. On September 29, 2025, Modal announced an $87 million Series B led by Lux Capital at a $1.1 billion post-money valuation, saying it served thousands of customers and ran inference on thousands of GPUs. TechCrunch reported a $355 million round at a $4.65 billion valuation in May 2026, and that Modal had told Reuters its annualized revenue had passed $300 million. On October 14, 2026, Modal announced a $750 million Series D led by Accel at a $15.75 billion post-money valuation, stating that it had crossed $1 billion in annualized run rate, that usage had tripled since its previous round, and that it employed more than 200 people across New York, San Francisco and Stockholm.

Modal's products fall into four groups. Inference covers serving large language models, image and video generators and custom models behind web endpoints, with support for open-source serving engines such as vLLM and SGLang. Sandboxes are isolated, short-lived environments where AI agents can execute code they wrote; Modal said in October 2026 that Sandboxes alone was a $300 million run-rate product and that close to 3 billion had been created. Batch runs large parallel jobs such as transcription. Training provides multi-node GPU clusters connected with RDMA networking, a high-speed link that lets GPUs on different machines exchange data directly, without a long-term capacity commitment.

Compared with alternatives, Modal sits between raw infrastructure and fully managed model APIs. Renting GPU virtual machines by the hour gives full control but leaves the customer to install drivers, manage scaling and pay for idle time. Running Kubernetes gives flexibility but demands an operations team. Model API providers serve popular models through a fixed interface but offer less control over custom code. Modal's approach suits teams that want to run their own code and models with little infrastructure work. TechCrunch's 2026 coverage places it in a crowded inference market alongside Baseten, Fireworks and Fal, and notes that margins across the sector are thin because GPU compute is expensive to acquire or lease.

The trade-offs are concrete. Cold starts, the delay when a new container has to boot and load a model before serving, remain the main latency cost of serverless GPUs. Modal's documentation says containers boot in about a second, but loading tens of gigabytes of model weights can take much longer unless weights are baked into the image or stored on a Modal Volume, and keeping containers warm with settings like min_containers means paying for idle GPUs. Code is written against Modal's Python SDK, so moving to another platform requires changes. Per-second pricing is convenient for bursty workloads, but for a GPU that is busy around the clock, reserved capacity elsewhere can cost less. Capacity for the newest GPUs depends on Modal's supply across its cloud partners.

Pricing is published per second. As listed on Modal's pricing page in October 2026, an Nvidia H100 costs $0.001097 per second, about $3.95 per hour, an A100 80 GB costs $0.000694 per second, an L4 costs $0.000222 per second and a T4 costs $0.000164 per second, with CPU billed at $0.0000131 per physical core per second and memory at $0.00000222 per GiB per second. The free Starter plan includes $30 of compute credit per month for up to three seats, and the Team plan costs $250 per month with $100 of included credit.