Backed by Y Combinator

Your GPU data center.Sell it as your own AI cloud.

NeoChamber is the cloud software for GPU data centers. You own the land, the power and the GPUs. It turns them into an AI cloud under your name: GPUs to rent, models to call, and Chami to help run it.

Launch your cloudRenting GPUs instead?

Backers and tools

Backed by Y Combinator. Built on tools you already run.

  • Y CombinatorWinter 2026 batch
  • NVIDIAGPU health and faults, from the driver
  • DockerModel servers run in containers
  • vLLMServes the models
  • OpenSSHCertificates, not passwords
  • OpenAI APICompatible: same calls, your models

Tested and attested by Prescient Assurance

  • SOC 2 Type IAttested
  • SOC 2 Type IIAttested

How a GPU cloud is made

Four parts make a GPU cloud. You own three of them.

  1. 01 · Land

    It starts with land. Green, open, near the grid.

    Room to build, and a line in from the power company.

  2. 02 · Power

    Then the power. Enough for every rack.

    Steady power on site, for GPUs that run day and night.

  3. 03 · GPUs

    Then the GPUs. Racked and running.

    The costly part, and the part you have already done.

  4. 04 · Software

    Then the software that makes it an AI cloud. That is NeoChamber.

    Customers rent your GPUs and call your models, under your name. You get paid.

    See the eight jobs it does

Why NeoChamber

A GPU cloud is eight jobs. NeoChamber does all eight.

Pick one to see it work in the console.

MachinesBring it online
Watch the fleetHealth, faults and logs, live.
One console8jobs, from rack to invoice
Your brandMake it yours
CapacitySell capacity
Sell tokensModels, priced per million tokens.
CustomersWhat customers get
MoneyGet paid
HelpRun it with help

The product

Pick a job. Watch it work.

Every panel is the console’s own parts, with demonstration data. It plays by itself if you would rather watch.

One command

Every machine joins the same way: one command, run as root, with a token that is shown once.

Add servers

Empty fleet to rentable capacity in about ten minutes

curl -fsSL https://get.neochamber.io | sh
reading hardware…
8× NVIDIA H100 80GB HBM3
joined as gpu-sfo-01 · San Francisco

The tour

One week at a data center. Racked Monday, selling Friday.

Northwind’s week, from racked machines to a first customer. All of it is demonstration data.

The narration shows here as the film plays.

Read the narration

Every GPU data center already has the hard part: the power, the cooling, and the cards themselves.

What’s missing is the software that turns those racks into a cloud business. That’s NeoChamber.

Meet Northwind: two data centers, in San Francisco and Dallas, and a hundred and twenty-eight GPUs. This is their week.

Monday. The servers are racked. Each one joins with a single command as root: the token is shown once, and the agent reports its own hardware.

By the afternoon both sites are on the map, sized by the GPUs behind them.

Tuesday. Every machine, with its site, state, GPUs live, utilization, power and last heartbeat, sampled every fifteen seconds.

Open one and the whole machine is there: per-GPU memory, temperature and power, the agent’s health checks, ECC and Xid counters, and a log search that highlights Xid faults.

Every agent reports its version on each heartbeat, so a site stuck between releases is one rollout from fixed.

Every console has one box. Type a machine, a customer, or a page, and it opens.

Wednesday. Twenty-two ready-made themes, or their own colours, previewed live and then published to every screen.

And one switch moves every screen a customer sees: the console they sign in to, and the storefront they arrived from.

Their storefront goes live on their own domain, with their GPUs, their rates, and the three ways to rent.

And the storefront a stranger lands on is painted from the same switch.

They price each GPU three ways: on demand; spot, which they can take back with two minutes’ notice; and reserved terms from one month to five years.

Thursday. The same machines can sell more than hours: models from the library, with their licences and the GPUs they are tested on.

The library is long, so it narrows: by size, by what the model is good at, by maker, by the cards they already own.

Every model opens: what it is for, how much memory it needs, and how long a prompt it will take.

Nothing runs until its licence is accepted, and the page says plainly what accepting one means.

Pick a free GPU and deploy. The agent pulls the weights, loads them, and the page counts the wait.

One model to a GPU. Accept the licence, deploy it, and the page counts the start: the download first, then the load.

They set what a million tokens costs, in and out, with a cheaper rate for prompts the server already has, and how fast anyone may call.

A customer can have their own rate, or their own limits, without touching the public list.

Planned work gets a window, so a model that stops for it reads as maintenance and not as a fault.

And on the machine itself: which card holds which model, and how many tokens a minute it is writing.

Friday. Verity AI signs in to Northwind Cloud: their machines, their utilization, their spend.

They reserve an H100 node for three months. The rate is fixed at signing, and the end date is on the row before they click.

One click makes an SSH key in the browser. Only the public half ever leaves the tab, and the command is ready to paste.

Or they skip the terminal: Notebooks opens JupyterLab on the rented machine, tunnelled over SSH.

Eight H100s, ready to train.

The models are there for them too, at Northwind’s prices, with an API key of their own.

They make their own keys, each one scoped to the models it is allowed to call.

They can try one in the browser before they write any code, tools and all.

Then the same call from their own program: an OpenAI-compatible address, their key, and the model’s name.

They watch their own GPUs, and a live log tail, for their machines and nobody else’s.

Machines meter from the moment they are ready, rounded up to the minute. Tokens land on the same bill.

Invoices draft themselves from what was rented and what was called. Publishing freezes the figures and gives the invoice its number.

Back in the data center: which machines are leased, which are ready, and which need a hand, with the reason and the next step.

Support threads carry the logs, and every reply is signed by Northwind’s team.

Idle machines earn on the open market, and come back to Northwind’s own pool whenever they choose. A rental already running always finishes.

Rentals, the market and tokens land in one total, for any period, with the machines that earned it.

Rentals, the market, and tokens, in one total, split by where it came from.

And every morning: what needs attention, machines online, live rentals, customers, and revenue this month.

Racked on Monday. Selling on Friday.

NeoChamber turns your data center into your own GPU cloud. Pure revenue share. Start at neochamber dot io.

Pricing

Nothing up front. A share of what it earns.

NeoChamber

The cloud software for GPU data centers.

Backed by Y Combinator

No licence, no setup fee and no minimum.

$0upfront

  • 5% of GPU services revenue
  • Your brand on every screen
  • Idle machines still earn
  • 1command to join a machine
  • 15sbetween live readings from every GPU
  • 3ways to price a GPU: on demand, spot, reserved
  • 22ready-made themes, or your own colours

Photos from Pexels: site, land, field.

Questions

Asked before, answered here.

Sorted by what you are deciding. Every answer is something the product does today.

Five per cent of your GPU services revenue, idle-capacity resale included. No licence, no per-seat fee, and nothing to pay on a machine that earns nothing.

Get started

Two ways in. Sell GPUs, or rent them.

Each side has its own console and its own guide.

Operator

You own the metal.

Sell your machines as a cloud under your own name. Your customers never see ours.

  • Your brand, your prices, your storefront
  • Managed or sovereign
  • One command per machine
  • Idle machines earn on the open market

5% of GPU services revenue. No licence, no per-seat fees.

Set up your cloud

Customer

You need the compute.

Rent whole machines from a provider you can see into. What you rent is yours alone until you hand it back.

  • Live figures from every GPU you pay for
  • Isolation tested on every release
  • Rent by GPU model from what is free
  • Hand machines back the hour you are done

Your machines, your numbers, nobody else’s.

Create an account

Questions? Email us.

Write to the founders, or ask for a 45-minute call to scope your fleet and launch date.

Email usfounders@usechamber.io