Build a GPU cluster: bill of materials and cost
The servers set most of the bill. Power sets where the cluster can live and when it starts. Price both in the builder.
Upfront for 16 B300 servers (128 GPUs)
From $10.2M
$80K per GPU, all in
- Power at full load
- 269.9 kW
- Data center a month
- $76K
- Racks
- 9
- GPU network
- 4 leaf + 2 spine, 128 ports of 800G
- Optics and cables
- 512 optical modules, 256 cables
- Other switches
- 2 front-end, 2 management, 1 firewall
- Shared storage
- 256 TB usable on 2 nodes
- Power split
- 248 kW servers, 18.3 kW network, 3.6 kW storage and management
- Racks
- 8 for GPUs (2 per rack), 1 for network and storage
- Monthly running cost
- $76K data center + $4K our management fee
| Line | Qty | Total |
|---|---|---|
| GPU servers | ||
| B300 servers, 8 GPUs eachOur October 2026 ballpark | 16 | $8.80M |
| Network | ||
| Back-end switches, Spectrum-4 SN5600 (800G)4 leaf + 2 spine, non-blocking | 6 | $390K |
| Optical modules, 800G4 per GPU; a twin-port module counts once | 512 | $461K |
| Fiber and patch cables | 256 | $31K |
| Front-end Ethernet switches (storage, in-band)Deployed in pairs | 2 | $60K |
| 1G management switches | 2 | $10K |
| Firewall | 1 | $25K |
| Storage and management | ||
| Shared storage, 256 TB usable2 TB per GPU, 2 nodes | 256 | $256K |
| Management nodes (scheduling, monitoring) | 2 | $45K |
| Shipping and colocation setup | ||
| Shipping | 16 | $40K |
| Colocation setup, one timeAbout $250 per kW, once | $67K | |
| Setup | ||
| Managed GPUs setup feeCustom above 7 servers; shown as the floor | $25K | |
| Total | $10.2M | |
Ballpark, October 2026. Not included: sales tax, 3-year support where a vendor doesn't bundle it, insurance, financing, and integrator installation. Server prices are our October 2026 ballparks; network, storage and management lines follow the assumptions below the builder. Not a quote.
The servers are about 86% of the bill
Servers that ship within weeks cost the most. Factory orders cost less but can take five to six months. The power column decides where each one can go. Figures are per 8-GPU server; the share is the 16-server B300 example in the builder.
| GPU | Ballpark price | Power | GPU network | Best for |
|---|---|---|---|---|
| B300 | $550-650K | 15.5 kW | 8 x 800G | Large training, long contracts |
| B200 | $450-500K | 14 kW | 8 x 400G | Training and inference |
| H200 | $340K | 10.2 kW | 8 x 400G | Inference on Hopper software |
| RTX PRO 6000 | $100-150K | 9 kW | 4 x 400G | Inference and small budgets |
Power is what we size for; RTX PRO 6000 at Supermicro’s full-load rating (Supermicro). Four network ports on RTX PRO 6000 is our assumption.
The network is the line budgets miss
Every GPU gets its own port, and every link needs an optic at each end. In the example the network is $977K, about $61K a server. NVIDIA’s 64-server B300 Ethernet unit uses 16 leaf and 8 spine switches (NVIDIA). One 144-port InfiniBand switch holds every GPU of up to 18 B300 servers, so small clusters need no spine. What the builder assumes:
- GPU network
- One port per GPU. One switch if every port fits; otherwise a non-blocking leaf-spine, half of each leaf facing up
- Ethernet switch
- 64 ports of 800G or 128 of 400G, about $65K
- InfiniBand switch
- 144 ports of 800G at about $130K, or 64 of 400G at about $31K
- Optics
- 4 per GPU on B300 Ethernet, 2.5 on B300 InfiniBand, 2 at 400G, half with one switch; $900 to $2,400 each
- Storage
- 2 TB per GPU for training, 0.5 TB for inference, $1,000 per usable TB
- Management nodes
- One per 8 servers, at least 2
- Colocation
- $240-325 per kW a month, power included; setup about $250 per kW once
Ballpark, October 2026. Not a quote.
Power decides where it lives and when it starts
Two B300 servers fill a high-density rack. NVIDIA rates its DGX B300 at 14.5 kW (NVIDIA), and the example draws about 269.9 kW. More than 80% of data centers have no racks above 30 kW (Uptime Institute), so plan on two B300 servers to a 30-40 kW rack.
Space can take as long as the servers. A site with that power takes about four weeks to six months to get, and it bills from its start date. Sign it to open when the servers land. Compare sites in the GPU colocation directory.
Free PDF · October 2026 · Not legal advice
Plan your cluster with us.
- 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
GPU cluster questions
- How much does it cost to build a GPU cluster?
- On our October 2026 ballparks, 16 B300 servers (128 GPUs) with an Ethernet GPU network, storage and setup come to about $10.21M, or $80K per GPU. Hosting adds about $76K a month.
- How much power does a B300 server use?
- NVIDIA rates its DGX B300 at 14.5 kW. We size colocation at 15.5 kW per 8-GPU B300 server.
- Do I need InfiniBand?
- No. InfiniBand and Ethernet with RoCE both carry GPU-to-GPU traffic, on the same NVIDIA network cards. InfiniBand needs less tuning; Ethernet costs less per port and comes from several switch makers.