Northflank

Northflank is a full-stack deployment platform founded in London in 2019 by CEO Will Stewart and CTO Fredrik Brix. The platform solves a specific problem that haunts growing engineering teams: Kubernetes is the right architecture for production workloads, but it requires substantial DevOps expertise to run. Northflank builds a developer-friendly abstraction over Kubernetes so teams can deploy containerized apps, managed databases, scheduled jobs, GPU inference services, and AI code execution sandboxes without authoring YAML. By November 2024, the company had raised $22.3 million in funding, tripled revenue year-over-year, and was handling more than 1 million container deployments per month across 70,000+ developers in production.

The platform covers the full application stack: Git-triggered CI/CD pipelines, ephemeral preview environments per pull request, managed PostgreSQL, MySQL, MongoDB, Redis, and RabbitMQ, GPU workloads on NVIDIA H100s and B200s, and microVM-backed sandboxes for secure AI code execution. Its standout capability is Bring Your Own Cloud (BYOC): teams deploy through Northflank's UI and API, but compute runs inside their own AWS, GCP, Azure, Oracle Cloud, or Civo account. Northflank currently serves 2,000+ startups and enterprises and has processed over 130 billion requests.

What Northflank actually does in May 2026

Northflank operates as a control plane that sits between your development workflow and whichever cloud infrastructure you're running on. You push code to Git, Northflank builds a container image, runs it as a long-lived service or job, provisions the database it needs, and exposes it behind a load balancer with TLS. That's the baseline. The interesting parts are what happen around that baseline.

Preview environments are one of the most practically useful features: when a developer opens a pull request, Northflank automatically spins up a complete copy of the application stack, with its own database, environment variables, and routing, so reviewers can test the actual change in an environment identical to production. No more "it works on my machine" gaps between review and merge. Templates allow you to define the full topology of a project as code, with draft templates functioning like pull request workflows for infrastructure changes.

GPU support covers the AI inference and training cases that have become core for many product teams. Available hardware runs from A100 40GB at $1.42/hr up to H100 at $2.74/hr and B200 at $5.87/hr, with spot instance optimization scanning AWS, GCP, and Azure for cheapest available capacity automatically. Fractional GPU allocation is supported, so you are not forced to rent a full GPU for a lightly loaded inference endpoint. A pgvector-equipped PostgreSQL instance is a one-click add for teams building RAG pipelines, running adjacent to the inference service in the same project.

The sandbox capability is specifically relevant for AI coding tools and agentic applications. Northflank runs microVM-backed sandboxes using Kata Containers (with Cloud Hypervisor) and gVisor, selectable per workload depending on security requirements. Kata provides hardware-level VM isolation; gVisor provides syscall interception with lower overhead. Cold start time is approximately 200ms. This is the infrastructure that products like Writer and Sentry use to run untrusted, multi-tenant code at scale. Northflank reports processing 2 million microVMs monthly across these workloads. For teams building AI coding assistants that need to execute user-submitted code safely, this capability is production-tested rather than experimental.

On November 11, 2024, Northflank announced $22.3 million in financing: a $16 million Series A led by Bain Capital Ventures and a $6.3 million seed round led by Vertex Ventures US, with participation from Kindred Ventures, Pebblebed, Tapestry VC, and Uncorrelated Ventures. Alongside the announcement, Northflank revealed revenue had tripled year-over-year and outlined a roadmap including ARM hardware support, local Docker caching for faster builds, expanded BYOC to additional cloud providers, and a self-deployable enterprise control plane for on-premises deployments.

"Northflank is way easier than assembling multiple tools. It's the ideal platform to deploy containers in our cloud account, avoiding the complexity of big cloud and Kubernetes. It's more powerful and flexible than traditional PaaS, all within our VPC." - David Cramer, Co-Founder and CPO at Sentry, November 2024

Where Northflank sits versus Modal and Render

Northflank occupies a different part of the market than its most-mentioned comparisons, and the differences are architectural rather than cosmetic.

Modal is a Python-centric serverless compute platform built specifically for AI and data science workflows. The core difference is scope and flexibility. Modal uses gVisor exclusively for isolation; Northflank supports both Kata Containers and gVisor with per-workload selection. Modal requires all container images to be defined through its Python SDK, which locks you into their image-building pipeline. Northflank accepts any OCI image from any registry without modification. Modal is managed-only, with no BYOC option and SDK-limited session durations that restrict long-running agent loops. Northflank supports unlimited session durations and full BYOC on AWS, GCP, Azure, and OCI. The pricing difference is significant: Modal's H100 runs approximately $7.25/hr when CPU, RAM, and storage are summed; Northflank's H100 is $2.74/hr all-inclusive. The tradeoff is that Modal is simpler for a Python AI engineer who just wants to decorate a function and get a scalable endpoint, while Northflank requires more setup but gives you a complete infrastructure stack alongside the GPU compute. Teams that only need serverless AI function execution and don't care about app hosting, databases, or BYOC will find Modal simpler. Teams building complete AI products with multiple components will find Northflank more coherent. For specific serverless inference comparisons, Modal remains the specialist for pure Python ML workloads.

Render is the closest competitor in the general PaaS category, but the gap has widened significantly in the past two years. Render has no GPU support at all. Northflank supports H100, B200, A100, L40S, and more. Render has no BYOC support; everything runs on Render's own infrastructure, which creates data residency and compliance complications for enterprise customers. Render's free tier puts applications to sleep after inactivity; Northflank's Sandbox tier keeps services always-on. Render requires paid plan upgrades for static IPs; Northflank includes these across all tiers. On managed databases, Render supports PostgreSQL and Redis natively but requires manual setup for MongoDB and other engines. Northflank manages PostgreSQL, MySQL, MongoDB, Redis, and RabbitMQ as first-class platform features. The one area where Render consistently wins is simplicity for beginners: its interface is less dense, the path from GitHub to deployed app is shorter, and the community resources are larger. If the primary requirement is deploying a web app or API without GPU, complex job workflows, or enterprise compliance needs, Render's lower friction is a real advantage. But for teams that will eventually need GPU support or need to keep data inside their own cloud account, starting on Render means a migration later. Developers often compare Northflank to RunPod for pure GPU workloads, though RunPod lacks Northflank's full-stack app hosting and BYOC capability.

"Anyone familiar with Kubernetes should feel right at home." - eschatology, Hacker News, February 2023

Railway and Fly.io round out the comparison. Railway is the most beginner-friendly option with the fastest path from repository to running app, but has no GPU support and no BYOC. Fly.io supports more infrastructure control with anycast routing across its own network, but has no BYOC and no managed GPU. Both are worth considering for pure app hosting without AI workloads; for teams that expect to need GPU or compliance controls, Northflank is the more durable choice.

What the daily workflow reality looks like

A typical Northflank project for an AI product team looks like this: a Python inference service runs on an H100 GPU service, a Node.js API sits in front of it handling authentication and routing, a PostgreSQL instance with pgvector handles embeddings storage, a Redis instance handles session caching, and a scheduled job runs nightly data processing. All of this lives in one Northflank project with unified logs, metrics, and rollback controls. CI/CD pipelines trigger on Git push. Preview environments spin up automatically for each pull request so the inference changes can be tested before merging.

The BYOC path adds one layer: instead of Northflank-managed infrastructure, the team imports their existing GKE, EKS, or AKS cluster. Northflank's control plane manages deployments, secrets, databases, and pipelines while all compute stays inside the customer's VPC. This is the Sentry architecture, addressing the data sovereignty requirements that prevent most enterprises from adopting a fully managed PaaS.

Build times have been a recurring complaint. Northflank historically didn't support local Docker layer caching, meaning builds re-pulled base images from scratch on each run. The November 2024 funding announcement listed faster builds with local caching as a roadmap priority. Partial caching improvements were shipped through early 2025, but teams with large images still report longer build times than on platforms with mature caching.

For AI sandbox requirements, Northflank pairs secure microVM isolation with full-stack infrastructure in the same platform. The same project that runs your LLM inference service also handles sandbox execution for code generation outputs, avoiding a separate integration with providers like Replicate. For GPU-focused research workflows, Lambda Labs and Together AI serve different segments of that market.

Who Northflank is built for

Northflank serves three audiences well. The first is the team migrating off Heroku that has grown beyond basic hosting and needs GPU support, better CI/CD, and more infrastructure control, without hiring a full Kubernetes engineering team. The full-stack nature of the platform means one tool covers what previously required stitching together a deployment platform, a database host, a CI/CD system, and a GPU cloud.

The second is the AI startup that needs to deploy open-source models (Llama 4, Deepseek, custom fine-tunes) alongside the traditional web infrastructure that serves those models to users. Running both the inference service and the product application in the same Northflank project, on the same BYOC cluster if required, simplifies operations significantly versus maintaining two separate infrastructure configurations.

The third is the enterprise team with compliance requirements (SOC 2, GDPR data residency, internal security reviews) that cannot use a managed SaaS cloud where data leaves their VPC. Northflank's BYOC model specifically addresses this: the company gets Northflank's developer experience, but the actual data and compute never leave the customer's own AWS or GCP account.

What Northflank is not

Northflank is not the right choice for developers who want the absolute simplest deployment experience. If your requirement is "connect GitHub, click deploy, get a URL," Railway and Render have shorter paths to that outcome and better beginner documentation. Northflank's interface exposes more infrastructure concepts, which is valuable when you need them but adds friction when you don't.

It is not the choice for pure GPU compute at minimum cost. If you need the cheapest possible H100 access and can manage your own container setup, RunPod or Vast.ai will undercut Northflank on price. At $2.74/hr for H100 all-inclusive, Northflank is reasonable, but the premium covers the platform layer rather than the raw GPU hour.

Northflank is not a managed ML platform. There is no built-in experiment tracking, dataset versioning, or model registry. Teams that need MLOps capabilities run Weights and Biases or MLflow on top of Northflank's infrastructure. The platform handles deployment and execution; it does not manage the experiment lifecycle that produces the models being deployed.