The power supply inside the rack design
An engineer designing an AI rack has work beyond selecting processors. Delivering electricity in a form those processors can use is part of the hardware too. In TechNews’s October 24 report, up to 73 million dollars in proposed support for Vertical Semiconductor addresses that task: developing vertical gallium nitride, or GaN, power semiconductors. The company plans to improve conversion efficiency while shrinking power-supply systems and demonstrating the technology in a 2.5 MW-class data-center rack. The development target sits in the power hardware beside the computing equipment.[1]
The physical constraint here is the electricity consumed during conversion and the space occupied by the equipment doing that work. The department’s account, as reported by TechNews, connects power semiconductors’ conversion, control and delivery functions with efficiency and reduced thermal-loss objectives. My reading is that the power supply belongs at the center of a high-power rack’s design. Research into a smaller supply allocates resources to an engineering task distinct from adding processors.[1]
What does 2.5 MW describe?
The 2.5 MW scale of Vertical’s planned demonstration describes the rack’s power class. It supplies neither a quantity of useful computation nor the conversion-loss fraction or power supply’s volume. The report names higher efficiency and a smaller system as separate objectives. Electricity losses and space requirements impose different design burdens. Lower losses and smaller hardware can affect different operating expenses, but this report supplies no basis for calculating a percentage electricity saving or additional computing capacity.[1]
Better conversion alone also leaves the facility’s overall constraint unsettled. Reducing a rack’s electricity requirement and supplying a more power-hungry rack are different outcomes. Vertical’s stated objective addresses advanced computing equipment’s high power requirements. A different rack design or a workload requiring less power could instead change the demand placed on this component. I therefore locate the research’s potential value in conversion suited to the rack, rather than in the superiority of any particular processor.[1]
From research budgets to operating hardware
The proposed total of up to 389 million dollars is not all allocated to AI racks. TechNews reports ceilings of 175 million dollars for Helion and 85 million dollars for Commonwealth Fusion Systems for fusion-related semiconductor and electronics development, and 56 million dollars for Sibelco’s quartz purification and finishing. These are not orders for four components destined for the same power supply. Adding research allocations for different applications and treating the sum as power or computing capacity available to a data center changes the quantity being counted.[1]
All four proposals also remain letters of intent awaiting final approval. According to the report, the government intends to require minority equity without control or voting rights. That condition describes the planned structure for public research support; it provides no hardware delivery date. For Vertical, the concrete engineering objective is demonstrating vertical GaN conversion at the stated rack scale. For an operator buying racks, the hardware volume and losses involved in doing the same power-conversion work at that demonstration are more useful than the funding ceiling.[1]