Building a distributed datacenter for the future of AI.
Big datacenters are waiting years for permits and power. We containerize GPU capacity to run on excess power at thousands of small industrial sites around the world. The first container is being built now.
What it is
Your property. Your power. Our box. Our GPU.
TWR sets a self-contained 20 ft container on a parking pad. Inside are up to 12 liquid-cooled NVIDIA HGX B300 servers, 8 GPUs each, 96 GPUs in all. Power and cooling stay in the box. No water needed. The building stays a building. You lease the pad and the power. We operate the servers.
What the property needs is spare 480 volt three-phase power, about 275 amps for a full box, a concrete pad, and a business fiber line of about 10 gigabits per second. A site with less power can run fewer servers in the same box. A licensed electrician makes the connection. You do not have to run the computers, and the site does not supply water.
Why
GPU capacity is waiting on sites that take years.
New hyperscale power, substations, and generation take years, so GPU capacity is stuck waiting on big sites.
A 20 ft self-contained container on an industrial parking pad is the way through. Each box holds up to 12 liquid-cooled NVIDIA HGX B300 servers, 96 GPUs. The property brings the pad, 480 V, and fiber. There is no site water. Many of those small sites run as one distributed compute cloud.
Time to get it
Demand vs the grid, 2025
Published ranges, not a TWR forecast. New transmission is 5–15 years (IEA). US projects that reached operation in 2025 spent a median of over 5 years in the interconnection queue (Berkeley Lab). Substation transformers are quoted past 160 weeks (Wood Mackenzie, 2026). Data-center electricity use rose 17% in 2025, against 3% for global electricity (IEA). About 20% of planned data-center projects are at risk of delay if grid bottlenecks stay (IEA).
Hardware
NVIDIA HGX B300, liquid cooled
- 8Blackwell Ultra GPUs
- 2.1 TBGPU memory
- 144 PFFP4 Tensor Core
- 72 PFFP8 Tensor Core
- 14.4 TB/sFifth-generation NVLink, total
- 1.6 TB/sNetworking, 8 × ConnectX-8 at up to 800 Gb/s
- ~14 kWPower per server (DGX B300)
Per server, from NVIDIA. The FP4 and FP8 figures include sparsity. Dense FP4 is 108 petaflops and dense FP8 is 36. The ConnectX-8 and power figures are from NVIDIA’s DGX B300 specs. Each server is liquid cooled and slides out on its own, so a tech can swap one without touching the rest. A full box holds 12 servers, 96 GPUs.
GPU reservations
Reserve NVIDIA HGX B300 capacity.
The first container is being built now. The next GPUs come online in 4 to 8 weeks. Reserve by the GPU or by the server, 8 GPUs per server, for 2, 3, or 4 years.
NVIDIA HGX B300
8-GPU server · USD per GPU-hour
| Term | 30% upfront | 50% upfront |
|---|---|---|
| 2 years | $6.46 | $6.15 |
| 3 years | $4.38 | $4.16 |
| 4 years | $3.48 | $3.26 |
Longer terms and more upfront lower the rate. Pricing is confirmed when the order is signed and is subject to hardware availability. Next GPUs come online in 4 to 8 weeks.
Property owners
List your property.
We pay $30 per square foot per year for a 400 sq ft pad. That is $12,000 a year. TWR brings the box, connects it, and runs it.
- Power480 V three-phase with about 275 amps to spare
- FiberBusiness fiber can reach the site
- SpaceRoom for a 20 ft container pad, 400 sq ft
- WaterNone needed
Investors
Interested in investing in a site?
Contact us.
This is not an offer to sell securities.