The growth of AI compute is forcing a rethink of how electricity is distributed inside data centers. A new report from ABB and Boston Consulting Group (BCG) argues that 800-volt direct current (800 VDC) distribution is emerging as the reference architecture for next-generation AI racks, while alternating current will keep dominating the external grid. The shift, then, isn’t about replacing one technology with another, but combining both wherever each works best — the same power-density problem behind ABB’s broader alliance with NVIDIA to build gigawatt-scale AI data centers.
Direct current and data centers in 30 seconds
- ABB and BCG rank AI data centers among the first sectors where DC already makes economic sense.
- The report identifies 800 VDC as the emerging architecture for high-density AI racks.
- In its 100 MW campus model, DC distribution improves grid-to-rack efficiency by 5%.
- That gain would free up an extra 5-6 MW of IT capacity.
- The main obstacle is no longer technical, but a lack of common standards and skilled professionals.
The report’s thesis starts from a contradiction that’s becoming more visible in modern energy infrastructure. Solar panels, batteries, electric vehicles and much of computing electronics run internally on direct current (DC), while the grids that connect them are still built mainly around alternating current (AC).
For decades, that hardly posed a problem. The losses from successive conversions were manageable, and AC infrastructure was fully standardized.
The arrival of AI racks drawing hundreds of kilowatts changes the scale of the problem.
ABB and BCG calculate that every conversion avoided can start to represent a meaningful amount of energy, space and electrical capacity once multiplied across facilities of tens or hundreds of megawatts. Their report, The Strategic Case for Hybrid AC/DC Power, analyzes data centers, industry, commercial buildings and homes, and concludes that the economic case is currently clearest in the first two.
600 kW racks, and the road to 1 MW
The starting point is density.
The report uses NVIDIA’s published roadmap to note that some racks for advanced AI systems could reach around 600 kW by 2027 and later approach 1 MW per rack.
Those figures are far removed from the historical densities of conventional data centers.
The problem isn’t just getting enough power from the grid. That electricity has to be converted, distributed throughout the building and delivered to the accelerators with reasonable losses, cabling and protection equipment.
According to ABB and BCG’s analysis, above roughly 200 kW per rack traditional AC distribution starts becoming increasingly impractical. For densities near a megawatt, the report considers 800 VDC to be the technically better-suited alternative.
The difference can be summarized like this:
| Architecture | Expected role |
|---|---|
| High-voltage DC | Long-distance transport of large amounts of electricity |
| Regional AC distribution | Remains the backbone of the grid |
| AC/DC conversion | Concentrated at the facility’s entry point |
| Internal 800 VDC distribution | Powers high-density IT infrastructure |
| Batteries | More direct connection to the DC bus |
| AI rack | Receives high-voltage direct current |
The report doesn’t anticipate a fully DC grid, then.
Its proposal is hybrid: AC remains the backbone of regional distribution, while certain consumers use DC behind the meter.
The diagram included by BCG and ABB shows exactly that split: the AC grid reaches the data center, passes through the transformer and a centralized AC/DC conversion stage, and from there batteries and IT loads share DC distribution.
This approach reduces the number of intermediate conversions the energy needs before reaching the hardware.
A 5% efficiency gain can mean several extra megawatts of GPUs
In a home computer, saving a few percentage points on electrical conversions can look like a modest improvement.
In a 100 MW campus, the picture changes.
BCG uses as its reference a data center of that size located in Northern Virginia, one of the world’s largest concentrations of digital infrastructure and also a region where available grid capacity has become a constraint on new projects.
In that scenario, the study estimates that centralizing conversion and using DC distribution can improve grid-to-rack efficiency by around 5%.
That 5% would have a particularly important consequence: it would free up between 5 and 6 extra MW for IT equipment without expanding the connection granted by the utility.
In the AI era, that detail transforms the economic case.
| BCG’s estimate | Reference campus |
|---|---|
| Data center power | 100 MW |
| Grid-to-rack efficiency gain | ~5% |
| Extra capacity for IT | 5-6 MW |
| Space freed up | ~3,000 m² |
| Estimated additional revenue | $92 million/year |
| Estimated additional operating profit | $35 million/year |
| Initial CAPEX reduction | $3.5-6 million |
These figures are estimates from a model built by BCG, not guaranteed results for any given data center.
Their value lies in showing why electrical efficiency takes on a different meaning once infrastructure is constrained by grid connection capacity.
If an operator has only 100 MW authorized, cutting losses doesn’t just mean a slightly lower power bill. It can also mean fitting more servers or accelerators within that same 100 MW.
That leads to one of the study’s central ideas: for AI infrastructure, direct current can be a capacity decision before it’s an energy-savings decision.
Fewer conversions, less copper, more room for compute
The report attributes part of the savings to the physical properties of electrical distribution itself.
At higher voltages, a DC system can carry a given amount of power using a smaller conductor cross-section than an equivalent AC architecture. BCG notes that, in certain configurations, copper use can drop by up to 45%.
That reduction has economic effects, but physical ones too.
Electrical cabling takes up space, adds weight, and needs supports, busbars, protection gear and associated distribution systems. As a rack’s power grows, these elements take up an increasingly large share of the data center’s design.
Current architectures can also use external power modules or sidecars placed next to the rack.
BCG estimates that centralized 800 VDC distribution could free up around 3,000 square meters in its hypothetical 100 MW campus, roughly a fifth of the space a sidecar-based architecture would require.
That space can then be used to install more compute capacity.
The calculation helps explain why companies designing large AI systems are reconsidering a part of the electrical infrastructure that had barely changed for years.
Batteries fit more naturally on a direct-current bus
Batteries are another key component.
A battery stores energy as direct current. When connected to predominantly AC infrastructure, several conversion stages are needed to charge it and later use that energy.
A DC-bus-based architecture can simplify that chain.
ABB and BCG propose batteries connected directly to the 800V bus through a simpler conversion stage. Beyond providing backup during outages, those batteries can absorb spikes and temporarily reduce demand pulled from the grid.
The report considers this design can provide several minutes of autonomy.
That matters because rising rack density is shrinking some of the electrical margins available to respond to outages.
BCG notes that a conventional rack may have roughly 20 milliseconds of margin, while a rack near 1 MW can shrink that window to just a few milliseconds.
By bringing storage and power electronics closer to the load, designers are trying to react faster to these fluctuations.
Data centers could consume 945 TWh by 2030
The pressure isn’t only coming from inside the rack.
The report uses International Energy Agency (IEA) estimates showing global data center electricity consumption rising from around 415 TWh in 2024 to 945 TWh in 2030, a trajectory that lines up with separate forecasts showing the data center electricity equipment market itself surpassing $50 billion.
That’s more than double in six years.
At the same time, securing new grid connections can take years.
BCG cites waits that can reach around seven years in Northern Virginia and up to ten years in some of Europe’s largest markets.
When the grid is the bottleneck, having an empty building or enough financing doesn’t guarantee more compute can actually be installed.
That’s why the report considers extracting more usable capacity from an existing electrical connection could become a competitive factor.
DC distribution doesn’t create additional electricity. It reduces losses and the intermediate infrastructure needed so that a larger share of contracted power actually ends up feeding servers.
Not every data center needs to switch to 800 VDC
The report also draws an important distinction.
The transition doesn’t affect every facility equally.
Large hyperscalers building AI clusters with racks drawing hundreds of kilowatts have more reasons to change their electrical architecture.
An enterprise data center hosting conventional servers, storage or low-density cloud workloads can keep using AC for years.
Colocation operators and so-called neoclouds sit in an intermediate position. Some buildings were designed for much lower densities but now need to accommodate AI infrastructure.
According to BCG’s model, replacing certain traditional low-voltage uninterruptible power systems with centralized conversion and 800V storage could reduce capital spending by up to $35 million in a 100 MW facility and cut the annual energy bill by up to 9%.
Again, these results come from BCG’s reference case studies and aren’t a universal forecast.
Actual returns will depend on electricity prices, the architecture being replaced, IT density, system design and how much of the recovered capacity can be monetized.
The biggest problem may be standards
If power electronics make these facilities possible to build, a less visible difficulty remains: getting equipment from different vendors to work safely under common standards.
ABB and BCG identify fragmented standards as one of the two main real barriers to DC adoption.
DC standards exist for different voltages and applications, but they still don’t have the level of uniformity reached by many AC technologies.
The report cites initiatives such as Current/OS, the Open Direct Current Alliance, and the 800 VDC specifications being developed within the Open Compute Project.
It also notes that some current DC protection systems can clear faults in microseconds using solid-state switches or hybrid designs.
The challenge is bringing those technologies to pricing, industrial scale and interoperability levels comparable to mature AC infrastructure.
The second barrier is human.
Designing a DC facility requires specific expertise in protection, fault coordination, power electronics, cross-vendor integration and commissioning.
ABB and BCG cite a 2025 survey of German manufacturers in which 69% identified the lack of standards as the main obstacle, while 38% pointed to a shortage of technical specialists.
Direct current also wants into factories and buildings
Although data centers take up much of the report, ABB and BCG extend the analysis to other sectors.
A modern factory already contains many elements that run internally on direct current. Variable-frequency drives used by robots, conveyors and machinery first convert the AC drawn from the grid into DC before regenerating the signal motors need.
The study proposes sharing a DC bus among those systems.
For a reference 10 MW industrial plant, the model estimates initial savings of roughly $1.2 million in investment and around $500,000 a year in operating costs.
Large commercial buildings are the next step.
Solar panels, batteries, LED lighting, computing electronics, EV charging and some HVAC systems already use direct current internally.
BCG studies a 10,000-square-meter office building with annual consumption near 1 GWh. In that case, it calculates around $46,000 in upfront savings and roughly $20,000 a year, equivalent to about 7% of the electricity bill.
For homes, the result is still different.
The report considers that, for now, the economics of a hybrid DC home don’t offset the upfront extra cost. Its reference case would save around $300 a year, not enough on its own to justify the investment.
That’s why it places data centers and industry first chronologically, then large commercial buildings, leaving housing for a more distant horizon.
AC isn’t going away: the grid will be hybrid
Perhaps the most important conclusion from ABB and BCG’s work is precisely that direct current’s comeback doesn’t amount to another “war of currents.”
AC retains very important advantages.
Passive voltage transformation via transformers, already-deployed infrastructure, its protection systems and more than a century of standardization make it unlikely to disappear from regional grids.
DC has different advantages.
High-capacity, long-distance transmission via HVDC is already established for interconnections and offshore wind farms. Now the novelty sits at the other end of the grid, behind the meter.
The framework ABB and BCG lay out splits into three layers:
Long-distance transmission: growing room for HVDC where it makes sense.
Regional distribution: predominantly AC.
High-density facilities: growth of internal DC networks connected to the grid through a central converter.
The expansion of artificial intelligence could accelerate that last shift.
For years, the conversation around data centers has focused on processors, GPUs, cooling and grid access. Racks approaching a megawatt now add another piece to the puzzle: how to carry every one of those watts from the substation to the accelerator without wasting capacity along the way.
If density forecasts hold up, 800 VDC could end up being one of the least visible — and, at the same time, most important — technologies of the next generation of AI infrastructure.
via: ABB report: The Strategic Case for Hybrid AC/DC Power
