ABB launches Infinitus, a source-to-rack DC portfolio for AI data centers

ABB announced Infinitus on Monday, a portfolio of direct-current equipment aimed at AI data centers that the company describes as the industry's first source-to-rack DC offering — covering the path from the medium-voltage grid connection down to the server rack. Nothing is available today. ABB says a number of the main components will be produced over the next 12 months, and that the first DC-native facilities using the complete portfolio are expected to be installed and operational within two to three years.

What the portfolio contains

At the center is what ABB calls its Infinitus solid-state transformer technology, which it says lets sites run with a smaller footprint and higher energy efficiency than conventional conversion chains. Around it the company groups five building blocks:

• Medium Voltage Powertrains — power from the grid or on-site generation
• DC Sources and Power Quality — AC-to-DC conversion moved out of the white space
• DC Power Distribution — higher power density in a smaller footprint
• DC Power Protection — protection for operators, AI servers and hardware from DC overcurrents
• Cooling Optimization — DC-connected variable speed drives and motors

The number driving it: 1 MW per rack

ABB's case rests on rack power. Today's racks draw almost 200 kW, the company says, while next-generation AI chips are expected to pull 1 MW per rack and beyond. At 800 VDC, ABB argues, the footprint gets more compact while redundant conversion steps disappear and the heat losses of conventional AC systems are cut. The company cites the International Energy Agency's expectation that global data center electricity demand more than doubles by 2030, and adds its own estimate that roughly 25-40% of new capacity installed in 2030 could use DC distribution — an ABB forecast, not an independent one.

Where the 5% comes from

The efficiency figure ABB leads with — energy efficiency gains of more than 5% — comes from a report it produced with BCG. Applied to a 500 MW data center, ABB says a 5% gain means 25 MW more power available for revenue-generating IT load and more than US$300 million in potential additional revenue per year. That revenue figure assumes the recovered capacity is sold, and the underlying study is one ABB co-authored, so it reads as a vendor estimate rather than an independent benchmark. The same 5%, ABB adds, would save energy over a year equivalent to powering Washington, D.C. for almost a full week.

Infinitus is the first portfolio providing the building blocks for any data center architecture – either DC native or hybrid AC / DC – to power next-generation AI chips efficiently.

Not ABB alone

The launch extends a collaboration with NVIDIA that the two first announced in October 2025, covering 800 VDC power architectures for gigawatt-scale AI facilities. NVIDIA has been pushing the same transition alongside Google and Microsoft through the Open Compute Project. In an August 2026 post, NVIDIA said more than 80 equipment manufacturers and infrastructure companies were already building products to the specification, with an MGX-compatible 800 VDC power rack arriving in the second half of 2026 and a centralized row power center — a power station for a full rack row — expected in 2027. NVIDIA's 2025 technical write-up named ABB among its data center power systems partners and made the same underlying argument: at megawatt racks, 54 VDC in-rack distribution demands copper volumes and carries resistive losses that stop making sense.

What the announcement does not say

• No pricing for any component.
• No named customer, design win or committed deployment.
• No per-product availability dates beyond "a number of main components" in 12 months and complete DC-native sites in two to three years.
• "Industry's first" and "the only major electrification supplier" with an IEC-certified solid-state circuit breaker and a static medium-voltage UPS in production are ABB's own characterizations, not independently verified here.

For anyone scoping AI capacity, the usable content is the timeline. ABB is telling the market that DC-native sites are a two-to-three-year build rather than a 2026 procurement decision — which puts its schedule roughly alongside NVIDIA's, where the 800 VDC rack lands this half and the row-level product a year after.