Koyeb

Filed by the Vibe Code Weekly desk · Updated Nov 12, 2026

The story

Koyeb is serverless infrastructure for code that already exists. It deploys containers, models and AI agent sandboxes onto CPUs, GPUs and accelerators, scales them automatically and bills by the second. The company has announced it is joining Mistral AI.

AI coding tools have made it cheap to produce working code, which moves more of the effort to deployment. Koyeb covers that runtime layer. A developer installs the CLI, runs a single deploy command or pushes to Git, and the platform builds the project and serves it at a koyeb.app address. A custom domain can be attached afterwards. Workloads run inside secure VMs on bare metal servers. The platform scales them out to hundreds of servers and back down to zero, and teams can choose one location or all of more than 50.

The company describes its focus as AI-heavy work. Its stated reason is that generative AI and agentic workloads changed what cloud infrastructure has to do. Hardware comes from AMD, Intel and Nvidia, and Koyeb has added Tenstorrent accelerators. Beyond web services and APIs, Koyeb pitches inference endpoints, MCP servers and isolated sandboxes where agents can execute AI-generated code. One-click templates cover serving stacks such as vLLM, TGI and Ollama, alongside framework starters and databases.

Koyeb is led by Yann Léger (CEO), Edouard Bonlieu (CPO) and Bastien Chatelard (CTO). Koyeb's own account says the idea followed 10 years of building and operating large-scale cloud platforms with Edouard and Bastien. Investors include Serena, Isai, Samsung Next, MongoDB, Acequia Capital, Kima Ventures and Plug & Play Ventures. For builders, the practical result is one place to host an AI-generated app, the model behind it, a Postgres database and agent execution environments, all under per-second billing.

AI coding tools make working code cheap to produce, which moves the hard part to deployment. Koyeb runs that layer: a single CLI command or a Git push serves an app, workloads scale to zero, and compute is billed by the second. The company has announced it is joining Mistral AI.

The good

One-command CLI deploys and Git-based deploys, with Docker and popular languages supported without modification. GPU lineup ranging from single L4 and L40S cards to eight-way H100 and H200 machines. Agent sandboxes and inference sit on the same platform as ordinary web services. Per-second billing paired with configurable autoscaling limits. Startup credits of up to $30k and reserved-capacity discounts of up to 50%.

The catch

Paid plans charge a monthly fee and bill compute on top. Uptime SLAs start at the Scale tier; SSO, RBAC and audit trails are Enterprise only. Tenstorrent N300s instances need an access request. Serverless Postgres runs in only three locations.

Who should read on

Builders who have generated an app with AI tools and need it running without server administration. Teams serving their own models as autoscaling inference endpoints on GPUs. Agent developers who need isolated environments to execute model-written code. Workloads with idle periods that benefit from scaling to zero.

Who can skip it

Anyone looking for an editor or assistant that writes code; Koyeb only runs it. Projects that need managed Postgres in regions other than the US East Coast, Germany or Singapore.

The price

Compute is billed per second. Plans: Pro $29/mo +compute ($10 included compute, 10 users, 100 services, e-mail and chat support); Scale $299/mo +compute ($100 included compute, 50 users, 1000 services, AWS Regions, 99.9% uptime SLA); Enterprise Custom +compute, starting at $1000/mo (99.99% uptime SLA, SSO, RBAC, audit trail). GPU instances range from RTX-4000-SFF-ADA at $0.50 /hr to 8x H200 at $24.0 /hr, for example L40S $1.20 /hr, A100 $1.60 /hr and H100 $2.50 /hr. Serverless Postgres ranges from Free 5h at $0 /mo to 3XLarge at $1.28 /hr ($952.32 /mo), with storage at $0.50/mo per GB. Early-stage startups can get up to $30k in credits.

Reader mail