# Owning a GB300 NVL72 rack

72 Blackwell Ultra GPUs and 36 Grace CPUs in one liquid-cooled rack. Here’s what owning one takes, and when HGX B300 servers make more sense.

## GB300 NVL72 at a glance

- **GPUs:** 72x NVIDIA B300 (Blackwell Ultra) linked as one system
- **CPUs:** 36x NVIDIA Grace
- **GPU memory:** About 21 TB per rack
- **Cooling:** Liquid only
- **Power:** About 130-140 kW per rack
- **Ballpark price:** $8-8.6M per rack, with networking and installation

## Is a rack right for you?

It fits the largest models, where 72 GPUs need to work as one. For most inference, [HGX B300 servers](https://www.amcompute.com/managedgpus/b300) use the same GPU, run air-cooled, and can be bought one at a time.

Vera Rubin, NVIDIA’s next generation, comes as racks and as 8-GPU servers. Early supply goes to the largest buyers. Tell us if you want to be in line.

## What it takes to own one

- **A liquid-cooled site.** Few data centers can cool racks this dense, and their space goes fast.
- **Power and floor.** About 130-140 kW per rack, and a floor that can hold it.
- **Lead time.** Orders placed now arrive in about six months.
- **Compliance.** Expect a thorough review of the buyer, the site, and the use.

## We manage. You own.

1. **Scope it.** Tell us your budget, timeline, and power. You get a plan and ballpark costs. One call.
2. **We procure it.** No preferred vendors. We source servers, networking, and data center space at today’s best price, in your name. 60-90 days typically.
3. **We run it.** Operations and maintenance (O&M), software, and customers, on an annual contract. Ongoing.

## What it costs

Flat fees, not revenue-share.

### Paid to vendors, in your name

- **Hardware.** Servers, networking, shipping, and installation, with nothing added.
- **Colocation.** Space, power, and cooling, usually on a 3-year contract.
- **Parts, bandwidth, and insurance.** Spare parts, internet, and property insurance on the hardware.

### Paid to us, flat

- **Setup fee.** A flat $25K for 1-7 servers, quoted for 8 or more.
- **Management fee.** From $1K a month per 4 servers.

Want a customer for the GPUs? We charge a small percentage of the contract, only if it signs.

## Tell us what you’re planning.

- 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)

## NVIDIA GB300 NVL72 racks: questions

**How much does a GB300 NVL72 rack cost?**

Ballpark $8-8.6M per rack, including networking and installation. The data center comes on top.

**Can it be air-cooled?**

No. It needs liquid cooling and a data center built for it.

**How long until delivery?**

Orders placed now arrive in about six months. The largest buyers are served first.

**GB300 NVL72 or HGX B300?**

Same GPU. The NVL72 links 72 of them for the largest workloads. HGX B300 links eight per server and fits far more data centers.

## Further reading

- [Liquid vs air cooling for GPU servers](https://www.amcompute.com/blog/liquid-vs-air-cooling)
- [The power budget of an AI data center](https://www.amcompute.com/blog/the-power-budget-of-an-ai-data-center)
- [HGX, DGX, MGX: NVIDIA’s server platforms](https://www.amcompute.com/blog/hgx-dgx-mgx)

## More hardware and hosting

- [NVIDIA B300 servers](https://www.amcompute.com/managedgpus/b300) - Ballpark pricing, power, and lead times for HGX B300 servers.
- [NVIDIA B200 servers](https://www.amcompute.com/managedgpus/b200) - Ballpark pricing, power, and availability for HGX B200 servers.
- [NVIDIA H200 servers](https://www.amcompute.com/managedgpus/h200) - Ballpark pricing and availability for HGX H200 servers.
- [NVIDIA RTX PRO 6000 servers](https://www.amcompute.com/managedgpus/rtx-pro-6000) - Ballpark pricing and power for RTX PRO 6000 servers, the lower-cost way in.
- [GPU colocation](https://www.amcompute.com/managedgpus/gpu-colocation) - A U.S. data center for the GPU servers you own.
- [GPU colocation directory](https://www.amcompute.com/managedgpus/colocation) - High-density and liquid-cooled data centers by metro: power, kW per rack and liquid cooling, each with a source.
- [GPU server vendors](https://www.amcompute.com/managedgpus/gpu-server-vendors) - Who sells GPU servers worldwide, from the public partner lists of NVIDIA, Supermicro, Dell, HPE, Lenovo, Cisco, Gigabyte, ASUS and AMD, with server models by maker.

## Who it’s for

- [AI startups](https://www.amcompute.com/managedgpus/ai-startups) - Own your GPU servers instead of renting cloud GPUs.
- [Universities and research labs](https://www.amcompute.com/managedgpus/research) - A private GPU cluster in your institution’s name, run for you.
- [Companies and enterprises](https://www.amcompute.com/managedgpus/enterprise) - Run AI models on GPU servers your company owns, hosted and run for you.
- [Family offices and funds](https://www.amcompute.com/managedgpus/family-offices) - Own GPU servers in your own entity and rent them to AI companies.
- [Individual investors](https://www.amcompute.com/managedgpus/individual-investors) - Own one or a few GPU servers through your LLC, rented out and run for you.

## Planning a rack? We’ll tell you what’s realistic.

[Book a call](https://www.amcompute.com/managedgpus/gb300-nvl72#discuss)

Source: https://www.amcompute.com/managedgpus/gb300-nvl72
