How Data Centers Are Preparing Their Power Systems For AI
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: How Data Centers Are Preparing Their Power Systems For AI on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Data center operators are changing power designs as AI racks require much more electricity: approaches include 800-volt direct current, on-site batteries and new generation. Vantage reserved more than a gigawatt of manufacturing capacity for a solid-state transformer, while NiSource and Microsoft signed a long-term power deal that includes generation and battery storage. Grid connections, equipment delivery times and who pays for infrastructure remain major constraints.

Data center operators are redesigning how power reaches AI servers, with recent moves including a transformer-capacity reservation by Vantage and a long-term electricity deal between NiSource and Microsoft. The developments address rising rack power demands through higher-voltage direct current, new generation and batteries, as grid access and electrical equipment become constraints on new capacity.

On October 21, data center operator Vantage reserved more than one gigawatt of factory capacity for a solid-state transformer from Heron Power, according to the source report. The transformer is designed to convert grid electricity directly into 800-volt direct current (DC). The reservation is for manufacturing capacity; it does not, on its own, establish that a corresponding amount of equipment has been delivered or installed.

On October 16, Indiana utility NiSource signed a 20-year power agreement with Microsoft, adding about 1,100 megawatts of generation and a 600-megawatt battery, the report says. The deal illustrates a separate but related response: arranging long-term electricity supply and storage alongside data center growth. The source does not specify the project’s exact operating schedule or how its costs will be allocated.

The engineering shift reflects denser computing equipment. The International Energy Agency (IEA) says AI server power density rose 11-fold between 2020 and 2025 and is projected to quadruple again by 2027. NVIDIA’s 800-volt design proposal would convert power near a facility’s perimeter and distribute DC to racks. The company says this could use 45% less copper than 415-volt AC distribution and improve end-to-end efficiency by up to 5% compared with current 54-volt systems. These are NVIDIA’s projections, not independently verified measurements; production with its Kyber systems is planned for 2027.

At a glance
reportWhen: Developments reported October 16–26, 20…
The developmentRecent infrastructure deals and power-system proposals show data center operators adapting electricity supply and distribution to the rising demands of AI computing.
Thorsten Meyer AI
Power architecture

Infographic 01 · From grid to rack

AI racks near a megawatt are outgrowing the AC power chain.

Why the industry is moving to 800-volt direct current: fewer conversions, less copper, batteries on the main path. The numbers are vendor figures with different baselines, so they cannot be added up.

Today: many conversions

Typical AC chain from the grid to the rack

  1. Medium-voltage AC from the grid
  2. Transformer steps it down to low-voltage AC
  3. UPS converts to DC for the batteries and back to AC
  4. Distribution units feed the racks
  5. Every server power supply converts AC to DC again

Each stage loses energy, and almost all of the loss becomes heat the cooling system must remove.

800 V DC: convert once, outside

NVIDIA reference design, full-scale production planned with Kyber racks in 2027

  1. 13.8 kV AC from the grid
  2. Converted once to 800 V DC at the edge of the site, batteries on the same DC bus
  3. 800 V DC busway to the racks
  4. Stepped down inside the rack, for 1 MW racks and beyond

Bridge for existing halls: a “sidecar” cabinet next to the rack converts AC to 800 V DC, while the old chain stays in place.

Power lost per MW delivered to the rack

kW, SMA white-paper model vendor model

  • AC chain, conversion at the racktraditional design~80 kW
  • 800 V DC from outside the buildingmedium voltage converted once~20 kW

What NVIDIA says 800 V DC changes

NVIDIA’s engineering claims claim

−45%copper, vs 415 V AC distribution
+85%power through the same conductor size
≤5%better end-to-end efficiency vs 54 V racks
200 kgcopper busbar for one 1 MW rack at 54 V

Sources: NVIDIA technical blog (20 May 2025, updated 31 Jul 2025); SMA America, sponsored content in Utility Dive (26 Oct 2026). Rack density: IEA, Key Questions on Energy and AI (Apr 2026) (AI server power density 11× from 2020 to 2025, a further 4× by 2027).
Powered by Thorsten Meyer AI

Power Becomes a Data Center Bottleneck

The shift matters because AI capacity depends not only on chips and buildings but also on whether enough electricity can reach servers reliably. Higher rack loads put pressure on the full electrical chain: grid connections, transformers, power distribution, backup systems and cooling. A design that reduces conversion losses or copper requirements could help operators use space and electrical infrastructure more effectively, though the size of any real-world gains will depend on implementation.

Storage could also help manage rapid changes in data center demand. The IEA says batteries may become important as AI workloads create large power swings, and estimates that data centers could have 20–25 gigawatts of battery capacity by 2030. It says storage could become a grid asset if incentives support that role. Batteries may help smooth demand or respond to disturbances, but they do not replace the need for adequate generation and grid connections.

The consequences extend beyond operators. The source report says U.S. household electricity averaged 18.21 cents per kilowatt-hour in 2026, compared with 17.30 cents in 2025, and cites a 41% increase in wholesale prices in the PJM region. Those figures do not establish that data centers caused the increases. The distribution of infrastructure costs—and whether customers or developers pay for new plants and wires—depends on specific contracts and regulatory decisions.

Products Worth Considering

Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W

Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W

Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black

Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black

As an affiliate, we earn on qualifying purchases.

From AC Distribution to 800-Volt DC

Most data centers have historically received alternating current (AC), then stepped it down and converted it through several stages before servers use DC internally. Each conversion can lose energy as heat, which also adds to cooling demand. Conventional designs were developed for lower-power racks; AI systems are pushing power density far beyond the 10- to 20-kilowatt racks cited in the report.

The IEA projects global data center electricity use will rise from 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030, roughly 3% of global demand. Berkeley Lab estimated that U.S. data centers used 4.4% of national electricity in 2023 and could account for 6.7% to 12% by 2028. The latter is a range of projected outcomes, not a confirmed share.

Moving distribution to higher-voltage DC could reduce current for a given amount of power and potentially cut conductor needs. NVIDIA says a one-megawatt rack supplied at 54 volts could require up to 200 kilograms of copper busbar. A separate Utility Dive article published October 26, sponsored by inverter maker SMA America, describes a “sidecar” conversion cabinet that could connect 800-volt racks to existing AC halls. Its model estimates losses of about 80 kilowatts in a traditional one-megawatt AC chain versus about 20 kilowatts with conversion to 800-volt DC outside the building. That comparison is a vendor-sponsored model, not an independent field result.

“Leaving no room for compute”

— NVIDIA

Costs and Performance Still Need Proof

Several headline efficiency and materials figures come from companies selling or promoting the relevant systems. Independent measurements at operating sites are not provided for NVIDIA’s projected efficiency and copper savings, or for SMA’s modeled conversion-loss comparison. The results may vary with facility design, workload, equipment and how much of an existing power chain is replaced.

The Vantage reservation does not confirm delivery dates, deployment locations or the number of transformers that will be installed. The source also gives limited detail about the NiSource-Microsoft deal’s project timeline, ownership and cost allocation. More broadly, it is not yet clear how quickly utilities can provide new connections or how the expenses of added generation and transmission will be divided among data center operators, other businesses and households.

Deployment and Grid Access Ahead

The next milestones will be equipment production, project construction and evidence from operating systems. NVIDIA says full-scale production of its 800-volt design is planned alongside Kyber systems in 2027. Vantage’s manufacturing-capacity reservation and the NiSource-Microsoft agreement will need to be followed by details on installation and operation before their effects on data center capacity can be assessed.

Grid access may remain the longer-term constraint. The IEA says transformers can take two to three years to arrive and grid connections can take five to ten years in many places. Berkeley Lab reported that U.S. interconnection queues held about 2,300 gigawatts of proposed generation and storage at the end of 2024, while only 13% of earlier requests had been built. These figures describe proposed projects and historical completion, not guaranteed future supply. New contracts, regulatory decisions and utility construction will help determine which planned data centers can secure power and who pays for the supporting infrastructure.

Key Questions

What does 800-volt DC change for a data center?

It changes how electricity is distributed inside a facility. A higher-voltage DC system could reduce conversion stages and the amount of conductor needed, but the savings cited so far include company claims and vendor models that require real-world validation.

Why are batteries being added to data center power plans?

Batteries can store electricity and may help smooth rapid changes in AI workloads or respond to grid disturbances. The IEA estimates data centers could have 20–25 gigawatts of battery capacity by 2030, but that is a projection, not installed capacity today.

What are the Vantage and NiSource deals?

Vantage reserved more than one gigawatt of manufacturing capacity for Heron Power solid-state transformers designed for 800-volt DC. NiSource and Microsoft signed a 20-year agreement adding about 1,100 megawatts of generation and a 600-megawatt battery, according to the source report.

Will these projects lower household electricity bills?

The available information does not establish that. The effect on bills depends on how new generation, grid upgrades and other costs are assigned, along with local electricity-market conditions. The cited price figures alone do not show that data centers caused household increases.

Source: ThorstenMeyerAI.com

You May Also Like

Synaptics Surges In Global Coverage

Synaptics experiences a significant increase in media mentions, with GDELT recording 27 mentions in a recent window, indicating heightened global attention.

Enhancing Student Skills Through AI And Critical Thinking Education

OpenAI publishes a report claiming students improve answer quality and reasoning when AI tools like ChatGPT are paired with explicit critical-thinking training.

How To Effectively Measure Benchmark Improvements In Speech Recognition AI

New tests from Hugging Face reveal that leading open-source speech recognition models may overstate real-world performance, highlighting measurement challenges.

Stenvrik: News as Geography

Stenvrik launches a live news globe organizing stories by city, offering a new spatial approach to news consumption with minimal costs and strategic insights.