High-Voltage Server Power Supply Engineering

The Hidden Megawatt: Why We Are Engineering Our Own Gridlock ⚡🏗️ Everyone is racing to secure the next grid connection 🔌 Almost nobody is asking how much compute is being left on the table with the connection they already have. I call it the Hidden Megawatt. In a traditional data center power chain, electricity is converted about five times between the grid and the GPU. Each stage adds loss, heat, cost, and failure risk. By the time power reaches the chip, roughly 5 to 7 percent of total facility capacity is already gone, burned as heat before a single token is produced ♨️ At 100 MW, that equals an entire row of GPU racks you paid for but never use 🖥️ At gigawatt scale, it becomes a full building of stranded compute capacity 🏢 This would matter less if new grid capacity were easy to obtain. It is not. In major hubs, large connections can take many years to secure ⏳ While the industry waits for new megawatts, existing megawatts quietly disappear inside legacy conversion chains. This is an architecture problem 🧠 An 800 VDC architecture cuts the conversion chain down to a minimal path ⚡ Converting medium voltage AC directly to high voltage DC and distributing DC natively turns more incoming watts into usable compute instead of heat. The ecosystem is already shifting 🚀 Next generation AI racks, high density power shelves, and 800 VDC reference designs are entering deployment now. The most valuable megawatt is often not the next one you are trying to connect. It is the one you can recover inside your existing facility 💡 No permits 📄 No queue 🚦 No multi year wait ⏱️ Just better power architecture ⚡ #DataCenterDesign #800VDC #AIInfrastructure #EnergyEfficiency #DataCenterDC #DataCenterSST #SolarEdgeSST #DataCenter800VDC

I'm excited to share the recent publication of my latest patent application, which anchors the next-generation power topology of the datacenter. Designed around the realities of AI-scale compute, not inherited enterprise infrastructure. US 2025/0239852 A1 – Disaggregated Solid State Power Distribution System This architecture doesn’t just eliminate the centralized UPS, it eliminates the entire logic of conventional power distribution, replacing it with rack/row-level solid-state conversion, flexible power redundancy, and dynamic fault tolerance. No fixed pathways, no stranded capex, no latency between failure and response. NVIDIA is already signaling this shift with their move to 800Vdc rack architectures. This patent defines the structural framework for the facility that makes that viable; not as a spec sheet, but as a deployable system. Power can’t be static if the compute it supports is dynamic. USPTO Link: https://lnkd.in/gqfQjXh6 Patent Publication No.: US-2025/0239852-A1 #AIinfrastructure #solidstatepower #800Vdc #nvidia #datacenterdesign #nextgencompute #SST

Data centers are rapidly becoming a major driver for DC power distribution and DC microgrids. Hyperscale facilities consume 20 to 100MW each, wit most of that power ultimately delivered at ~1V at the point-of-load for xPUs and memory. To improve efficiency and to reduce distribution cost, designers push voltage levels upward, with the conversion to 1V as close to the silicon as possible. Hereby, the power conversion system architecture is of crucial importance! Across the industry, facility-scale DC distribution is converging on either +/-400VDC (Google, Meta, Microsoft) or 800VDC (NVidia). In a future architectures, these DC buses will likely be fed from the medium voltage AC grid via solid-state transformers (SSTs). Today, the power delivery from 800V to 1V is envisioned to move from 800 V → 48 V → 12 V → 1 V. A rack-level conversion from 800V to 48V, is followed by a tray- or GPU card-level conversion from 48V to 12V. The final conversion from 12V to 1V happens on the GPU card, as close to the silicon as possible. Exact voltages may vary a bit. This structure evolved from traditional AC-fed architectures. However, it has two big drawbacks: it still requires substantial copper at rack- and tray-level and it has multiple conversion stages. Both add loss and cost. Skipping a stage—for example, jumping from 800 V directly to ~12 V—sounds attractive, but creates challenges for converter semiconductors and magnetics. A multi-module series architecture may be more promising! On the high-voltage side, modules connect in series, naturally dividing the input bus (e.g., 800 V into ~100 V segments). Each module converts directly to 12V, a much more favorable design point for both semiconductors and magnetics. These modules can be integrated directly on the GPU board, minimizing the amount of copper needed to transport power within a rack. A series architecture taps into low-voltage power devices which are more more efficient and more reliable than high voltage ones. Moreover, power converter transformer ratios are less extreme which simplifies magnetics. On the flip side, a series architecture requires a more complex communication and control. But embedded digital control, and high-speed communication are becoming inexpensive, making the control challenge solvable. Power system design is ultimately about managing the “conservation of misery”. Design challenges remain, but you can choose where the burden sits. The arrival of smart, all-digital power modules unlocks new possibilities to redistribute that burden more intelligently. #DC, #800V, #microgrids, #datacenters, #nvidia

As our need for computing resources continues to increase servers are consuming more and more power. Racks are expected to eventually consume a megawatt of power which requires rethinking the data center power architecture. This movement is envisioning DC distribution at 400 or 800V DC that is converted from AC at the point of connection with the grid and extends to the server. Two architectures are emerging: NVIDIA’s monopolar 800VDC end-to-end architecture and the OCP’s Mt. Diablo ±400VDC specification; This design change allows for several levels of power conversion to be eliminated improving energy efficiency and reducing the footprint needed for electrical infrastructure. To achieve this vision requires facility level AC to DC conversion devices. Rising to meet this challenge are solid state transformers. Solid state transformers or SSTs are power electronic devices that converts AC to AC or AC to DC. The transformer has been a fundamental component of the grid since its earliest days. They are simple magnetic devices that transform voltage. An SST replaces that with a rectifier, converter, and inverter. For data center applications SSTs can convert AC to 400V or 800V DC to supply the DC power distribution system. They can also be paired with battery backup units or BBUs to enable the facility to meet ride-through requirements and manage demand fluctuations that can induce forced oscillations. Numerous players are emerging to manufacturer SSTs. There are the new entrants like Heron Power and DG Matrix along with existing companies like Solar Edge. With SSTs we see a new implementation of one of the most fundamental components on the grid. Transformers are simple magnetic devices but as SSTs are power electronic devices it is unknown if they will last as long as conventional transformers. They also have implications for the power grid that need to be studied. We have had numerous challenges with power electronic generation such as IBRs that we don’t want with power electronic transformers. We learned that we need proper dynamic and EMT models for these generators, and will need the same for these SSTs. We also need to see what their protection and fault behavior is and study how that interacts with the grid. Finally we need to ensure that grid codes are updated as the transformer moves from a passive device to an actively controlled power electronic device.

The air-cooling era of the data center is officially over. A traditional enterprise server rack runs at about 8kW. We are now entering the era of the 1MW rack: A concentration of power and heat that fundamentally breaks the laws of traditional infrastructure. Average power density per rack doubled from 8kW to 17kW in twenty-four months and is projected to hit 30kW by 2027. But that is just the beginning. The individual GPU levels are scaling from 150W to 1,500W, with some projections exceeding 10,000W per unit. This requires a transition from 100kW IT racks to 1MW configurations. How do you deliver and dissipate 1MW in a single IT rack without catastrophic failure? To reach this density, the industry is moving toward disaggregated systems and high-voltage direct current (HVDC) architectures. By bringing 480V AC into the rack and converting it to +/- 400V or even 800V DC, you can distribute power directly to the IT gear through HVDC bus bars. This configuration, based on the OCP Diablo spec, allows for 110kW to 180kW power shelves interspersed with battery or capacitive backup units. The thermodynamic reality is even steeper. As rack power increases 20x, we see a parallel 20x increase in heat generation. This requires in-row Coolant Distribution Units (CDUs) with a 1.8MW capacity, operating at flow rates of 1.5 liters per minute per kilowatt. Liquid cooling is a mandatory physical constraint to prevent clock-speed throttling across clusters of 72 GPUs acting as a single logical chip. But where does a 1GW site find this level of stable, always-on power? Traditional grids cannot handle the 1GW "flat load" profile of a massive AI cluster. In France, our nuclear baseload provides the sovereign, carbon-free energy needed for 99.9% frequency stability. If the grid frequency deviates, you risk losing an entire training epoch during checkpointing. Most companies simply cannot build this vertical integration.