Serverless

Endpoints that scale to zero.

Ship a container, get an HTTPS endpoint. Workers start when requests queue and stop when traffic drops, and every GPU-second in between is on the invoice in euros.

Create an endpoint Talk to an engineer

The endpoint view

Every worker, in one screen.

Status, region and uptime per worker, the request rate, and how many workers are billable right now. Min, max and idle timeout sit next to them.

console.nexpod.ai/serverless/llm-eu

live

Serverless /llm-eu ReadyManage

Workers

4of 20 max

In progress

14requests

Queued

0requests

p95 latency

241ms · 1h

OverviewMetricsWorkersLogs

WorkerStatus

w-3e9c51av14EU-AMS-11× H200 · 141 GBRunning2h 14m
w-8b1d07fv14EU-AMS-11× H200 · 141 GBRunning2h 14m
w-c42a9e3v14EU-FRA-11× H200 · 141 GBRunning1h 02m
w-19f0b6dv14EU-FRA-11× H200 · 141 GBRunning41m 09s
w-7d55e20v14EU-AMS-11× H200 · 141 GBIdle · warm12m 30s
w-a0c3f81v14EU-FRA-11× H200 · 141 GBBooting4s

Per GPU-second · H200 €0.001115 billable workers

Dockerfile

1FROM nexpod/pytorch:2.9-cuda13
2WORKDIR /app
3COPY requirements.txt handler.py ./
4RUN pip install -r requirements.txt
5ENV MODEL_DIR=/models
6CMD ["python", "handler.py"]

Bring your own image

Any registry, any base. Start from a Nexpod CUDA image or your own.

vol · model-weights420 GB

/models · EU-AMS-1ro

w-3e9cmounted

w-8b1dmounted

w-c42amounted

w-19f0mounted

Weights on a shared volume

Mount one network volume on every worker. No re-download per boot.

RTT from Berlinms

EU-FRA-1 Frankfurt9
EU-AMS-1 Amsterdam14
US-NYC-1 New York88

Requests route to the nearest region with a warm worker.

Close to your users

Requests go to the nearest warm region.

In plain words

What is Nexpod Serverless?

A managed queue in front of GPU workers. You supply a container with a request handler; Nexpod runs as many copies as the traffic needs, on B300, B200, H200 or H100, and stops them when it quiets down.

Nexpod Serverless at a glance

GPU families: B300 (288 GB HBM3e) · B200 (180 GB HBM3e) · H200 (141 GB HBM3e) · H100 (80 GB HBM3)
Billing: Per GPU-second, from worker start to stop, rounded up
Price per GPU-second: €0.00083 (H100) to €0.00167 (B300)
Scaling: From zero to your max workers on queue depth; warm minimum optional
Regions: EU-AMS-1 · EU-FRA-1 · US-NYC-1
Invoicing: EUR, VAT on its own line

Rates from the Nexpod rate card · excl. VAT

How it works

From image to endpoint.

Four steps, all from one CLI or the console. Nothing to provision, nothing to patch.

01
01 · seconds
Create the endpoint
Point at an image, pick a GPU family and set min and max workers.
02
02 · automatic
Let it scale
Workers follow queue depth. Idle ones stop after a timeout you choose.
03
03 · live
Watch it run
Logs, traces, GPU memory and error rates per worker, in one view.
04
04 · no downtime
Ship the next release
Deploy v15 while v14 drains. Roll back with one command.

nexctl — endpoints

$ nexctl endpoints create llm-eu \

--image ghcr.io/acme/llm:14 \

--gpu h200 --min 0 --max 20

→ release v14 · volume model-weights mounted

✓ https://llm-eu.run.nexpod.ai

$ nexctl endpoints scale llm-eu --min 2

✓ 2 warm workers · eu-ams-1, eu-fra-1

$ nexctl endpoints logs llm-eu -f

12:04:31 w-3e9c51a 200 · 182 ms

12:04:31 w-8b1d07f 200 · 176 ms

Features

Inference that scales on its own.

The parts that usually take a platform team: warm pools, fast starts, any runtime and releases from Git.

Workers · 24 h running warm booting

16:00 · peak

running12

warm2

00:0006:0012:0018:0024:00

Warm workers on standby

Keep a floor of warm workers for instant answers. Everything above it scales with traffic and back down.

Time to first responselog scale

Cold · image pull38 s
first deploy of a new image
Cached image6.2 s
image already on the node
Warm worker180 ms
min workers above zero

Starts measured in seconds

Cached images and volume-backed weights cut a cold start to seconds. Warm workers answer in milliseconds.

Runtimes · tested images

Python 3.12
Node.js 22
Go 1.23
Rust 1.82
C++ 20
vLLM
TensorRT
ONNX Runtime

nexpod/ imageB300B200H200H100

pytorch:2.9-cuda13✓✓✓✓
pytorch:2.8-cuda12.8✓✓✓✓
jax:0.7-cuda12.8✓✓✓✓
pytorch:2.4-cuda12.4––✓✓

Any runtime you can containerise

Python, Node.js, Go, Rust or C++, with the framework and CUDA version you pin.

releases · llm-eutag v15 · built 1m 52s

v153f9a1c2rolling 6
v14a07e4d1draining 3
v13c51b9e0rollback

12:04:18w-3e9c51av14 → v15health 200
12:04:31w-8b1d07fv14 → v15health 200
12:04:47w-c42a9e3v14 → v15warming

$ nexctl endpoints rollback llm-eu --to v14

Releases straight from Git

Push a tag, Nexpod builds the image and rolls workers over one by one. The previous release stays one command away.

Steady traffic

Busy all day? Pin it to a server.

When an endpoint keeps eight GPUs near full, a dedicated Cloud Server is cheaper per request. Keep the same image and handler, move the traffic.

See Cloud Servers Compare rates

94%

avg · peak 99%

idleGPU utilisation · 7 daysfull

Use cases

What runs well behind an endpoint.

Anything request-shaped: traffic that comes in bursts, jobs that should cost nothing while nobody is calling.

llm-eu · 24 h req/min workers

peak 13 workers

00:00 · 0 workers12:0024:00

Model serving

Text, image and speech models behind one HTTPS endpoint, scaled by queue depth.

job ft-1042 · 8× H200€22.77

train · 3 epochs

queued: 4s
boot: 6s
train: 41m 52s
save adapter: 44s

workers 0 → 8 → 0billed 42m 42s

Fine-tuning jobs

Submit a training job as a request. Workers exist only for the length of the run.

agents-eu · queue depth2 warm · max 12

−60 min scale up above 8now

Agent backends

Many short model and tool calls per session, with warm workers for the first token.

batch-0917 · 2,400 items6m 10s

queue

2,400

items in

24 workers during the runthen 0 · €0.00 idle

Batch and rendering

Fan a batch out to dozens of workers, then let the endpoint fall back to zero.

Pricing

Pay per GPU-second.

A worker is billed from start to stop, rounded up to the next second. Idle endpoints at zero workers cost nothing.

Create an endpoint All pricing

288GB HBM3e
B300Newest
Most memory per GPU, for the largest models at long context.
€0.00167/GPU·s
180GB HBM3e
B200
High throughput for large models and heavy batches.
€0.00139/GPU·s
141GB HBM3e
H200
141 GB fits a 70B-class model on one GPU.
€0.00111/GPU·s
80GB HBM3
H100
Lowest rate. Suits 7B to 34B models and embeddings.
€0.00083/GPU·s

Example · one traffic burstH200

3 workers × 95 s each = 285 GPU-seconds

285 × €0.00111 · then scale to zero, €0.00 idle

€0.32

Smallest unit we bill

€0.00083

One H100 GPU for one second. Endpoints at zero workers bill nothing at all.

0 ms500 ms1 s · billed

€0.00083 × 3,600 s = €3.00 per H100 GPU-hour

FAQ

Serverless, answered.

How endpoints start, scale, ship and fail, in short answers.

How is Serverless different from a Cloud Server?

A Cloud Server is a whole machine you keep running. A Serverless endpoint is a queue in front of workers: Nexpod starts GPU workers when requests arrive and stops them when the queue is empty, so you pay only while one is up.

Which languages and frameworks can I run?

Anything that runs in a Linux container with a GPU. Python is the common case, with PyTorch, JAX or ONNX Runtime, but handlers in Rust, Go, Node.js or C++ work the same way.

How do you keep cold starts short?

Images are cached on the nodes that served them before, weights load from a network volume in the same region, and you can keep a minimum number of warm workers so the first request never waits for a boot.

How do releases and rollbacks work?

Each deploy creates a numbered release. Workers move to the new release as they finish their current request. Rolling back is one command, or one click in the console.

Can an endpoint react to webhooks or queues?

Yes. Call it synchronously over HTTPS, submit async jobs and poll, or register a webhook that receives the result. Each request carries its own ID for tracing.

What do I see when something goes wrong?

Per-worker logs, request traces with timings, GPU memory and utilisation, and alerts on queue depth or error rate. Metrics export to your own Prometheus as well.

Customers

Endpoints for teams with bursty traffic.

Product launches, nightly batches and agent sessions all share one pattern: spikes, then quiet. Serverless bills for the spikes.

Placeholder

Placeholder quote. An AI app team says, in one or two plain sentences, what changed after moving production traffic to Nexpod.

Customer nameEngineering lead, AI app

Customer 01
Customer 02
Customer 03
Customer 04
Customer 05
Customer 06
Customer 07
Customer 08

Customer logosPlaceholder

Start now

Deploy an endpoint before lunch.

Push an image, set min and max workers, call the URL. The first GPU-second is the first thing you pay for.

Create an endpoint Book a demo