# How to deploy a B300 cluster, step by step

Nine steps and six counterparties before the first job runs. We line up all of them for a flat $25K, not a percentage of your project.

## Six counterparties, or one

### On your own

- Server vendor
- NVIDIA compliance
- Network vendor
- Data center
- Freight and insurer
- Installer

You deal with each of them directly.

- Six contracts, six contacts, six timelines.
- One missing form or a site without power holds everything.
- An infrastructure engineer, about $150K a year ([American Compute](https://www.amcompute.com/blog/tco-own-vs-rent-gpu-clusters)).

### With us

- Server vendor
- NVIDIA compliance
- Network vendor
- Data center
- Freight and insurer
- Installer

American Compute deals with each of them for you.

- One team lines up every party, in your name.
- No preferred vendors. You get the best price on the market right now.
- A flat $25K for 1-7 servers, not a percentage of the project.

## Nine steps from quote to first job

### Before the order

1. Size it. Servers, network, storage and power. Who you deal with: Your team. Ballpark: [Bill of materials builder](https://managedgpus.com/guides/how-to-build-a-gpu-cluster).
2. Get quotes. Entity details, a credit application and NVIDIA’s end-use forms come before any firm quote. Who you deal with: Vendors, NVIDIA compliance. Ballpark: Best price on the market now.

### Order to delivery

1. Order. Weeks if in stock, five to six months from the factory. Who you deal with: Server vendor. Ballpark: 20-50% deposit.
2. Sign a data center. High-density colocation, about 15 kW per server, usually on a 3-year contract. Who you deal with: Data center. Ballpark: $240-325 per kW a month.
3. Ship and insure. Freight from the vendor, insured from the factory door. Who you deal with: Freight and insurer. Ballpark: $2,500 per server.

### Bring-up and run

1. Rack and network. Power, racking and an 800G network of switches, optics and cables. Who you deal with: Installer, network vendor. Ballpark: $50-90K per server.
2. Burn in. Stress-test every GPU and the links between servers before you accept them. Who you deal with: Installer. Ballpark: [NVIDIA DCGM](https://docs.nvidia.com/datacenter/dcgm/latest/user-guide/dcgm-diagnostics.html), [NCCL tests](https://github.com/NVIDIA/nccl-tests).
3. Hand over. Drivers, CUDA, and Kubernetes or Slurm. Who you deal with: Your team. Ballpark: [NVIDIA GPU Operator](https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/index.html).
4. Run it. Monitoring, repairs and warranty claims. Who you deal with: Server vendor. Ballpark: For the life of the servers.

## Where B300 projects stall

- **An order held for a form.** Vendors won’t ship until NVIDIA’s end-use forms check out.
- **Servers with nowhere to go.** Hardware that arrives before the site has power sits in storage, already paid for.
- **Lead times that slip.** The cheapest factory orders can take five to six months.
- **Bad GPUs found late.** Skip burn-in, and failures show up after you’ve accepted the servers.

## What four B300 servers cost

- Upfront. $2.6M. - Servers $2.2M - Networking $220K - Shipping and our setup $35K - Install, storage and colo setup $109K.
- Each month. $24K. - Data center $17K - Management $1K - Parts, bandwidth and insurance $6K.
- Our fee. $25K. - Flat, for 1-7 servers. 8 or more quoted - Not a commission or a percentage of the project - 1% of this build alone would be $26K.

Ballpark, October 2026, at $550K per server. Colocation at $280 per kW a month. After parts, bandwidth and insurance; before taxes and financing.

## Plan your B300 deployment.

- A call with our CTO about what you need
- Ballpark costs and timelines for your budget
- A straight answer if it’s not a fit

Prefer email? [hello@amcompute.com](mailto:hello@amcompute.com)

## Deployment questions

**How long does it take to deploy a B300 cluster?**

Smaller deployments usually go live in 60-90 days. Servers that ship within weeks cost more, and the cheapest factory orders can take five to six months.

**Can I deploy GPU servers myself?**

Yes, if you can line up the server vendor, NVIDIA’s compliance check, a network vendor, a data center, freight and insurance, and an installer, then handle repairs for years. One missing form or a site without power can hold the project for weeks.

**What paperwork do vendors ask for?**

The buying entity’s details, a credit application or a deposit, and NVIDIA’s end-use forms, which name the buyer, the site, and what the GPUs are for.

**Do I need InfiniBand?**

Not for one server. A cluster needs a fast network between servers, either InfiniBand or 800G Ethernet with RoCE, at about $50-90K per server.

**What do you charge to deploy it for me?**

A flat $25K setup for 1-7 servers, half at signing and half when live, quoted for 8 or more. Management from $1K a month per 4 servers. It is not a commission or a percentage of project cost. On a $2.6M four-server build, 1% alone would be $26K. We have no preferred vendors, so you get the best price on the market right now and pay the vendor directly.

## Want B300s without the headache? We’ll run them.

[Plan a deployment](https://managedgpus.com/guides/deploy-a-b300-cluster#discuss)

Source: https://managedgpus.com/guides/deploy-a-b300-cluster
