Replicate

Replicate runs machine learning models behind a single API, mostly other people's open-source models packaged with its own open-source container format, Cog (Apache-2.0, github.com/replicate/cog, 9,459 stars, actively merged as of August 2026). There are two billing shapes: public/shared models are billed per second of processing or per token/output unit (e.g. Claude 3.7 Sonnet at $3/million input tokens; FLUX 1.1 Pro at $0.04/image), while private deployments are billed for the wall-clock time dedicated hardware is online — setup and idle time included, not just active inference. Hardware rates run from $0.000025/sec for a small CPU instance up to $0.0122/sec for 8x H100 GPUs, confirmed against the pricing page's embedded rate table. There's no free tier or trial credit — you need billing set up to run anything beyond the public playground.

Cold-starting a model on Replicate — pulling the container image to a compute node before it can serve a request — has been a long-running complaint. In a 2024 Hacker News thread, Replicate's own founder wrote: "Yeah, our cold boots suck," and described mitigations in progress (faster fine-tune boots, keeping models warm via deployments, more efficient weight loading, distributing images as individual files). Those mitigations shipped over the following year, but the underlying trade-off remains: unlike a dedicated GPU box you keep warm yourself, Replicate's autoscaled cluster has to place your model somewhere before it can run, and that placement cost shows up as latency on the first request after idle.

Cloudflare completed its acquisition of Replicate on December 1, 2025, for total purchase consideration of $57.4 million ($44.4M cash net of acquired cash, $9.5M in holdbacks, $3.5M in assumed liabilities), per Cloudflare's Form 10-Q. That is well below the $350M valuation Replicate reached in its June 2023 Series B (Andreessen Horowitz-led, with Sequoia and Y Combinator participating), though a private valuation and an acquisition price aren't directly comparable and no reason for the gap has been disclosed. Cloudflare and Replicate both say the API, pricing, and existing workflows continue unchanged, and Replicate keeps operating as a distinct brand for now, with 50,000+ of its models being folded into Cloudflare Workers AI over time. Hacker News reaction to the announcement was mixed — skepticism about the lack of disclosed deal terms, some users noting they'd already outgrown Replicate's Cog abstraction for GPU work they self-host, and general wariness about Cloudflare's growing footprint.

The draw is breadth: Cloudflare's own announcement cites 50,000+ production-ready models on Replicate, spanning image, video, audio, and language models from many different maintainers, plus the ability to package and fine-tune your own via Cog. Cog itself is Apache-2.0 and works outside Replicate (you can build and run Cog containers on your own infrastructure), which caps lock-in on the packaging format even if you'd still need to re-plumb billing and scaling yourself. We could not verify a breakdown of how many of the 50,000+ are actively maintained versus stale community uploads — that figure is Cloudflare's, unverified independently.

Replicate publishes a public status page (replicatestatus.com, run on Incident.io) covering API, inference/training, website, registry, official models, and support separately. At the time of this check it showed all systems operational with no incidents visible in the recent history window we could load; we did not pull a full 90-day incident log, so we can't state an uptime percentage or incident count — the page itself doesn't publish a rolling uptime figure either.