1Legion has 32 NVIDIA HGX B200 servers coming into service in the United States in January 2027, and allocation is open now. This post sets out what the servers hold, which workloads they suit, how they are rented and what happens before a cluster is handed over. The full specification is on the NVIDIA HGX B200 page.
What the 32 servers hold
Each HGX B200 server carries eight NVIDIA Blackwell GPUs on one baseboard. Every GPU has 180 GB of HBM3e and 7.7 TB/s of memory bandwidth, for 1,440 GB per server. Fifth-generation NVLink connects the eight GPUs at 1.8 TB/s each, 14.4 TB/s across the server.
Across 32 servers, that adds up to:
- 256 Blackwell GPUs
- 46 TB of GPU memory
- 2,304 PFLOPS of dense FP4, at 72 PFLOPS per server
- 9.5 PFLOPS of FP64, at 296 TFLOPS per server
The per-server figures come from NVIDIA's DGX B200 specifications and the HGX B200 product documentation listed on our server page.
What the HGX B200 is for
The HGX B200 suits teams that want Blackwell for training and inference and still run double-precision work.
- Training and fine-tuning of models that fit in 180 GB per GPU. A 70B-parameter model in FP8 takes about 70 GB of weights, which leaves about 110 GB on one GPU for KV cache and batching. Our team worked through that arithmetic across four GPUs in 70 GB of weights, four GPUs.
- Inference in FP4, FP8 and INT8.
- Scientific computing and simulation that needs FP64 next to AI. At 296 TFLOPS of FP64 per server, the HGX B200 carries about 30 times the FP64 of an HGX B300 board.
If memory per GPU is the limit, the NVIDIA HGX B300 carries 270 GB per GPU.
How the servers are rented
Customers rent whole 8-GPU servers, never split between customers, or a cluster on one fabric dedicated to them. Contracts run 36, 48 or 60 months. Every proposal is written, starts from our delivery specification and names the site from our data center locations.
Our how to buy page walks through each step, from the onboarding forms to signing.
What happens before handover
Every cluster passes NVIDIA's hardware validation tests, HW-1 to HW-5, before handover: NCCL throughput, every InfiniBand or RoCE rail, GPUDirect RDMA and storage, NVLink and NVSwitch, and a 24-hour soak under a representative inference load. A node that fails is quarantined, repaired or replaced, and tested again.
The customer receives the results in an acceptance report and keeps it as the cluster's baseline. Five tests before handover explains what each test checks, and the burn-in and acceptance page has the full procedure. We do not offer trials: the burn-in and the acceptance report take their place.
"A January ready-for-service date gives teams time to plan a training or inference roadmap around hardware they know will be there," said David Vargas, CEO of 1Legion. "Every cluster is built to our published specification and burned in before handover, so customers know exactly what they are signing for."
How to request a quote
Share the hardware, the number of servers, the term and the start date, and you receive a written proposal. Request a quote or talk to an engineer.