Northflank Pricing 2026: What You Actually Pay

Northflank sells itself as a self-serve platform for running containers, databases, and jobs without wiring up Kubernetes yourself. The pricing page gives you a free Sandbox, a wall of nf-compute plan names, and an hourly rate card. What it does not give you is the one number that decides your bill: the cost of the memory you allocate and then never touch. Here is the whole meter, decoded, with three real 30-day bills and the single lever that moves the total.

What does Northflank actually charge for?

Strip away the plan names and Northflank is a straight resource meter. You pay for the vCPU and the memory you allocate to a service, for every hour that service is deployed, plus disk and egress. Rates below are from Northflank's pricing page (2026).

Two things fall straight out of that card. First, one vCPU-hour costs exactly the same as two GB-hours: the CPU-to-memory price ratio is 2 to 1. Second, there is no request meter, no bandwidth-in meter, no per-seat compute charge. The box you allocate is the bill.

Are the nf-compute plans discounts? No.

The pricing page lists named presets, nf-compute-50, nf-compute-100-2, and so on, which reads like a plan ladder where bigger tiers get a bulk break. They do not. Every preset is the raw hourly meter multiplied by 720 hours, the number of hours in a 30-day month. You can rebuild each published price from the rate card exactly.

So nf-compute-100-2 at $24 a month is nothing more than (1 x $0.01667) + (2 x $0.00833), times 720. Picking a preset versus setting a custom size changes nothing about the price. That is good news: it means you can size a service to the exact vCPU and GB it needs and pay the metered rate, with no penalty for going off-menu.

What does a real Northflank bill look like?

Three worked 30-day bills, each built from the rate card above. Every service is assumed to run the full month.

Small production SaaS, one web tier with two replicas, a worker, and a database:

Growing SaaS, same shape but the team bumped every web replica to 4 GB "to be safe" and scaled to three replicas:

GPU batch inference, an H100 running six hours a day plus a small always-on API:

Notice what is missing from all three: requests, traffic, and bandwidth-in never show up as a meaningful line. On Northflank the size of the box, held for the month, is the whole story.

Why is memory the line item that gets you?

Northflank prices the resources you request, not the resources your process actually uses. This is the standard Kubernetes requests-and-limits model : you declare a vCPU and memory allocation, the scheduler reserves it, and Northflank bills the reservation for as long as the service is deployed. A running service does not scale its bill down to what the process touches. If you ask for 4 GB and use 1.5 GB, you pay for 4 GB.

That makes idle memory the quiet default overspend. At $0.00833 per GB-hour, one spare gigabyte on an always-on service costs $0.00833 x 720 = $6.00 a month, every month, for doing nothing.

Look back at the Growing SaaS bill. Those three web replicas request 4 GB each but the app's working set is about 1.5 GB. Right-size the request to 2 GB and each replica drops from $36 to $24: $12 saved per replica, $36 a month across three, with zero code change and no throughput impact. That is the one lever on Northflank. Not switching regions, not haggling a plan, just matching the requested box to the working set.

Northflank vs Railway: which meter is cheaper?

The closest comparison is Railway , which meters the same two resources but on a different clock (rates 2026):

Per allocated resource-hour Northflank is roughly 40% cheaper for the same box. But Railway bills per second of real usage, so a service that idles most of the day, or a preview environment that only wakes for a demo, can total less on Railway even at the higher rate. The rule that falls out: an always-on production tier leans Northflank, where the flat allocation is cheaper; a bursty, dev, or preview workload leans Railway, where the per-second meter only charges for the minutes it runs. If your workloads sit closer to bare metal, the same math extends to our Fly.io pricing breakdown and the Railway teardown.

What about GPUs?

GPU is where Northflank's flat hourly rate gets genuinely interesting, because reserved cloud GPUs are expensive and the meter here is simple (rates 2026):

The trap is the same one as memory, scaled up. An H100 attached to an always-on service runs 720 hours a month: 720 x $2.74 = $1,972.80, whether or not a single request hits it. Run the same H100 as a scheduled job six hours a day and it is $493.20. That is a $1,479.60 monthly gap created entirely by leaving a GPU switched on. For inference you rarely want a GPU pinned to an always-on service; you want it on a job or a queue that spins it up and lets it go.

The lever that matters

Northflank's pricing is honest and easy to model once you see past the preset names: a per-hour meter on the vCPU and memory you allocate, 720 hours a month, disk and egress on top. There is no clever discount to chase and no usage cliff to fear. There is only the size of the box.

Math check: on Northflank you are not paying for traffic or requests, you are paying for the size of the box, every hour it is switched on. Size the box to the working set, switch off the ones nobody uses, and run GPUs as jobs instead of always-on services. Do those three things and the bill is boring, which on an infra invoice is exactly what you want.