Replicate

Replicate is a cloud platform that lets developers run, fine-tune and deploy machine learning models through one API without managing GPU infrastructure. Users call a hosted endpoint, the platform scales containerised model workers up or down, and billing is per second of compute used. Since December 2025 it has operated as a brand of Cloudflare.^[1][2]

History

Replicate was founded in 2019 by Ben Firshman, who previously led open-source product work at Docker, and Andreas Jansson, a machine learning engineer formerly of the BBC. Its foundational open-source tool, Cog, packages a model and its dependencies into a Docker container, configures the GPU software stack and generates a validated inference API directly from Python input and output definitions.^[3]

The company went through Y Combinator and raised a $5.3 million seed, a $12.5 million Series A and a $40 million Series B led by Andreessen Horowitz in December 2023 — which included NVIDIA's NVentures arm, Sequoia Capital, Heavybit and Y Combinator — bringing total funding to roughly $58 million. At the time it reported about 2 million sign-ups and 30,000 paying customers.^[3][4]

On 17 November 2025 Cloudflare announced its acquisition of Replicate; the deal closed on 1 December 2025 for a total cash consideration of $57.4 million, disclosed in Cloudflare's annual filing. Replicate continues as a distinct brand inside Cloudflare's AI product line.^[2][4]

Technology

The platform's core value is the removal of inference plumbing. A model published with Cog gets a reproducible container, an auto-generated API and automatic horizontal scaling; Replicate hosts a catalogue of more than 50,000 containerised models spanning image and video generation, speech, music, restoration and language models. Fine-tuning jobs, serverless GPU endpoints and version-pinned model URLs round out the workflow, so a prototype can move from a notebook to a production endpoint without an MLOps team.^[1][3]

Why It Matters for AI

Replicate democratised model-as-a-service: before platforms like it, running an open research model meant provisioning GPUs, resolving CUDA dependencies and writing serving code by hand. Its acquisition by Cloudflare for $57.4 million also signals consolidation in the AI infrastructure layer — model hosting is being bundled into connectivity clouds alongside compute, storage and networking rather than surviving as a standalone business.^[2]

>See Also

• Model Serving
• Hugging Face
• Together AI
• vLLM
• MLOps