Nvidia’s reference design for megawatt-class AI racks asks for two things at once: 800 VDC power distribution and storage that works across timescales from milliseconds to minutes, with supercapacitors named for the fastest layer. A memorandum of understanding announced by Infineon Technologies and Skeleton Technologies in early September sets out a framework to explore both, from the grid connection to the processor.
The memorandum maps directly onto that two-part requirement. Its first track covers next-generation solid-state transformers that convert medium-voltage AC to high-voltage DC, combining Infineon’s CoolSiC power semiconductors with Skeleton’s power conversion systems and supercapacitor technology. Its second track covers gallium nitride-based peak-shaving sidecars that pair Skeleton’s supercapacitors with Infineon’s CoolGaN devices. Both companies describe the aim as higher power density, so that more compute fits within the same physical footprint. The agreement is a non-binding framework for joint exploration, which leaves room to define products and priorities as the work develops, and the companies have said they will explore and develop architectures for the entire power chain from grid to core.
The need the sidecar track targets is well documented. Nvidia’s October 2025 technical post says rack power in synchronized AI training can swing from roughly 30% to 100% of load and back in milliseconds, and it calls for multi-timescale storage: high-power capacitors and supercapacitors close to the racks to absorb fast spikes and fill brief valleys, and facility-level batteries at the interconnection for slower shifts. Nvidia’s July 2025 post on GB300 notes that generation resources can take from one minute to 90 minutes to respond to a sudden ramp. It also shows the value of storage in practice: capacitors that fill about half the volume of each GB300 NVL72 power shelf, storing 65 joules per GPU, cut peak grid demand by 30% in a Megatron training test. A sidecar built on supercapacitors and fast-switching GaN devices could extend that kind of smoothing beyond the shelf, and Schneider Electric’s work with Nvidia on an 800 VDC sidecar with modular energy storage shows the concept is gaining traction across the ecosystem.
Infineon brings an established position in the 800 VDC architecture. It has been developing power systems for it with Nvidia since May 2025 and is among the silicon providers Nvidia lists for the design. Its recent solid-state transformer work spans two other partners: a collaboration with SolarEdge on a modular 2 to 5 MW SST block designed to convert 13.8 to 34.5 kV directly to 800 to 1,500 VDC at more than 99% efficiency, and the supply of silicon carbide devices to Eaton for its MVSST 2.0 platform, which Eaton describes as one of the first medium-voltage SST platforms to achieve IEC certification. The Skeleton memorandum extends that presence from conversion into fast storage, the other half of Nvidia’s blueprint.
Skeleton brings manufacturing scale and a customer base in the storage layer. It opened its 220 million euro Leipzig plant in November 2025, designed for up to 12 million supercapacitor cells a year, and says it supplies Siemens, General Electric and Hitachi Energy for grid applications as well as US hyperscalers for AI data centers. Tech Funding News reports that it delivered millions of cells to a major American hyperscaler this year. The company has raised 392 million euros in venture funding, including a 33 million euro first close in May, ahead of a planned US listing in 2027. Its GrapheneGPU system, which the company says completed validation under hyperscaler power profiles, is described by Skeleton as capable of cutting AI energy consumption by up to 45%, lowering peak power connection requirements by 44% and raising GPU computing output by up to 40%; those are company figures, and the memorandum itself quantifies none of them.
Integration between storage and solid-state transformers is already a recognized design direction. In July, Skeleton announced a technical and commercial collaboration with DG Matrix to integrate its GrapheneCBU800 and GrapheneBBU800 storage systems into DG Matrix’s Interport CM medium-voltage SST skid and its Interport CL sidecar for GPU pulse loads. The Infineon memorandum adds a semiconductor partner to that picture, with silicon carbide for the high-voltage conversion stages and gallium nitride for fast peak-shaving.
The timing leaves room for the possibilities to take shape. Nvidia says full-scale production of 800 VDC data centers will coincide with its Kyber rack systems in 2027, and it attributes up to 5% better end-to-end efficiency, 45% less copper and up to 30% lower total cost of ownership to the architecture, against baselines of 54 V rack distribution, 415 VAC distribution and an unstated one, respectively. Third-party projections cited in the DG Matrix and Skeleton announcement from SemiAnalysis point to roughly 39 GW of incremental AI data center capacity on 800 VDC by 2030 and a solid-state transformer market of about 13 billion dollars. The memorandum sets no schedule, but the 2027 Kyber window gives the partners a natural reference point as they move from exploring architectures toward named products, prototypes, and customer testing. The non-binding structure also lets each company keep building on its existing relationships while this one develops: Infineon with SolarEdge and Eaton, Skeleton with DG Matrix.
Add EnergyNews.biz as a preferred source on Google
