Ethernet
B200e
$6.75M
- 64 × NVIDIA B200, 8 servers
- High-speed Ethernet between servers
- Good fit for inference and fine-tuning
20 ft micro datacenter · 64 NVIDIA Blackwell GPUs
Own your compute.
Buy it, set it on power you already have, and run your training and inference on your own terms. No reservation, no renewal, no price change in the middle of a run.
What it is
A micro datacenter you can buy.
Each TWR box is a standard 20 ft ISO container with 64 NVIDIA Blackwell GPUs: eight servers in two racks of four, eight GPUs each. Roughly half the box is enclosed IT. The other half is an open-frame dry cooler in the same frame, so the site never supplies water.
20 ft ISOTurnkey container on a concrete pad. Ships and sets like any 20 ft box.
64 GPUs8 servers, 2 racks × 4, 8 GPUs each. NVIDIA B200 or B300.
In-frame coolerDry cooler built in. No site water. Air-cooled servers by default; liquid options available.
480 V three-phaseAbout 200–275 A spare for a full box, a pad, and a licensed electrician.
Why own
Owning wins over years, if you can host it.
Over a long run, owning usually costs less than shared cloud capacity on a reservation that ends or gets repriced. You pay once, you control the schedule and the data, and the box is still yours when the term would have ended.
Most teams don’t have spare 480 V three-phase, a pad for a 20 ft box, or the people to look after it. That’s fine. Rent a dedicated TWR box instead: white-glove service, with management and operating costs rolled in.
- ControlYour hardware, your network, your models and data. Schedule training and inference however you like.
- Clear costOne purchase up front. After that, your ongoing cost is power, network, and upkeep.
- An assetEquipment you can depreciate, move, or resell when your plans change.
- SpeedIf you already have power and a pad, the box sets without waiting on a new datacenter build.
Buy vs. reserve
See when owning pays off.
Pick a box and enter the reservation you’d otherwise sign. The chart compares the one-time list price of the box with what a reservation for the same 64 GPUs costs over time.
Assumes 64 GPUs reserved at that rate, 24/7, for the full term: the same capacity as one box.
Break-even—
Reservation, 3 yrs—
Buying saves—
Buy (one-time list price)Reservation (cumulative)
Buy line is the box’s list price, paid once. After the term, the box is still yours to run, move, or resell. Reservation cost = 64 GPUs × rate × 8,760 hrs/yr × years.
Prefer to reserve instead of buy?
Rent a dedicated 64- or 128-GPU box on a 3–5 year reservation. Its own internet, white-glove service, and all management and operating costs rolled in. Example: B300c, 36 months, $5.50 / GPU-hr, 30% up front.
Boxes
Four boxes. 64 GPUs each.
Ethernet boxes suit inference and most fine-tuning. Cluster boxes add InfiniBand and central NVMe storage for multi-node training. List prices, USD, per box.
Cluster
B200c
$7.5M
- 64 × NVIDIA B200, 8 servers
- InfiniBand fabric and central NVMe storage
- Built for multi-node training. UPS options available
Ethernet
B300e
$8.25M
- 64 × NVIDIA B300 (Blackwell Ultra), 8 servers
- High-speed Ethernet between servers
- More memory per GPU for larger models
Cluster
B300c
$9.0M
- 64 × NVIDIA B300 (Blackwell Ultra), 8 servers
- InfiniBand fabric and central NVMe storage
- The full training box. UPS options available
List prices per box. Final terms in a signed purchase agreement. Compare Ethernet and Cluster →
Need more than 64 GPUs?
Paired 20 ft boxes, 128-GPU and larger builds, and 40 ft configurations available.
