StorageReview has published the first throughput data for two GB300 DGX Stations cabled back to back, and the gains on models that overflow one GPU's 252GB of HBM3e run far beyond 2x. The result makes the deskside pair a real alternative to waiting for a GPU server slot in a data centre, with caveats worth reading first.
What was announced
- Each GB300 DGX Station pairs a 72-core Grace CPU with a Blackwell Ultra GPU carrying 252GB of HBM3e at 7.1TB/s, plus 496GB of LPDDR5X at 396GB/s over NVLink-C2C. A ConnectX-8 SuperNIC exposes two 400GbE QSFP112 ports, up to 800Gb/s in total.
- StorageReview linked an ASUS ExpertCenter Pro ET900N G3 and an HP ZGX Fury AI Station with two 400G direct attach copper cables, one RoCEv2 rail per port, following NVIDIA's published clustering playbook.
- Across the two links, NCCL broadcast, reduce and one-way send/receive reached about 98GB/s against the 100GB/s the pair of cables can carry. All-reduce landed at about 92GB/s and all-to-all at about 80GB/s. GPUDirect RDMA reached 392Gb/s per rail, and an 8-byte RDMA write averaged 1.46 microseconds.
- On GLM-5.2, a roughly 433GB NVFP4 checkpoint that forces about 215GB of expert weights into Grace memory on one Station, output rose from 139 tokens per second on one Station to 1,525 on the pair at 32 concurrent streams, about 11x. The pair reached 3,141 tokens per second at 256 streams.
- On MiniMax-M3 with 8,192-token prompts, a single Station dropped to 103 output tokens per second at eight streams as its KV cache ran out, while the pair delivered 736 at the same point and kept scaling to 1,801 at 256 streams.
- For models that fit on one Station, StorageReview ran a full copy on each and split prefill from decode. GLM-5.3 Flash went from 828 to 2,721 output tokens per second at 32 streams on short prompts, and from 1,219 to 2,448 on long prompts.
- The ASUS unit is specified with a single 1,600W Titanium ATX power supply. StorageReview notes each data point is a single pass and that its time with the pair was limited to inference testing.
The ETON view
The headline multiples need reading carefully. A second Station does not make the GPU 11 times faster. What it does is move a model that was spilling half its weights into slower CPU memory back into HBM, and the gain is mostly the removal of that penalty. On models that already fit in one Station, the benefit drops to the more ordinary 2x to 3x that comes from splitting prefill and decode across two boxes. So the first question for any buyer is not which Station, it is how large the model you actually intend to serve is, at what precision, and how much KV cache your prompt lengths will need.
The more interesting point is where this sits in the estate. Two towers on roughly 1.6kW supplies each, cabled with two DACs, can now serve models that a year ago would have meant an HGX server, a liquid or high-density air rack and a colocation contract. For UK teams stuck behind 6 to 12 kW colo racks or a waiting list for AI-ready capacity, that is a genuine option for a department-sized inference service or a secure on-site deployment. The trade-offs are just as real: no redundant power, no hot-swap, no rack management, a lab-style network with no switch in it, and a ceiling of two GPUs before you need a proper fabric anyway.
Our view is to treat the Station pair as a bridge, not an architecture. It works well for pilots, data that cannot leave the building, and teams waiting on a rack. Once usage grows past one department, the economics usually move back to a rack server with four or eight GPUs where power, support and lifecycle sit inside infrastructure you already run. We can price GB300 Stations from the different OEMs alongside HGX and PCIe GPU servers, source the 400G cabling and optics, and quote buyback for the hardware either route replaces.
Related infrastructure
Category: AI & GPU · Vendor: NVIDIA, ASUS, HP · Technology: GB300 Grace Blackwell Ultra, DGX Station, ConnectX-8, 400G Ethernet, RoCEv2 · Last verified: 06 Oct 2026 · ~3 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
