NVIDIA GB300 Reaches the Desk as ASICs Set to Overtake GPUs and ABF Tightens
Liquid-cooled DGX Station-class workstations are delivering a full Grace Blackwell stack to individual engineers, while DIGITIMES expects custom ASIC shipments to pass GPU shipments in 2027 with ABF substrate capacity emerging as the new constraint.
Gigabyte's W775-V10-L01 is an NVIDIA DGX Station GB300-based system built for desk-side deployment, pairing a 72-core NVIDIA Grace Neoverse V2 CPU with a Blackwell Ultra GPU carrying 252GB of HBM3E at 7.1TB/s. The tower draws up to 1.6kW through an 80 PLUS Platinum supply, needs a 20-amp circuit, weighs about 29kg and is liquid-cooled, with every side of the chassis except the decorative front acting as an intake or exhaust. The system effectively requires its own dedicated circuit.
Networking is dominated by two QSFP112 ports fed by NVIDIA's ConnectX-8 SuperNIC at 400Gbps each, alongside a 10GbE port that also routes through the SuperNIC rather than the CPU. Storage comes from two PCIe Gen6 x4 and two PCIe Gen5 x4 M.2 slots. ServeTheHome measured Qwen3.6-35B-A3B Flash NVFP4 running at more than 26,000 tokens per second on the system, equivalent to over 2.25 billion tokens a day.
At the other end of the market, DIGITIMES reports the high-end cloud AI accelerator segment will see a structural shift in 2027, with total ASIC shipments surpassing GPU shipments for the first time as output rises from Google, Amazon and Huawei. ABF substrate bottlenecks are shifting accordingly, because the large, complex packages used by custom accelerators and high-end server processors consume far more substrate area per unit than conventional designs.
The practical implication for buyers is that the constraint on AI hardware delivery is moving away from the accelerator itself and toward the substrate, memory and power infrastructure needed to complete a deployable system. That shift is visible even at the desk-side end: delivering a full Grace Blackwell stack to one engineer now requires a dedicated circuit, a liquid cooling loop and a 29kg tower, and the same input constraints that bind a rack are starting to bind a workstation.