
Why is the line between a data center’s electrical backbone and its IT room inverting? Explore how 800-volt direct current (VDC), side power racks, and hyperscaler-written specifications are making engineering talent the scarcest input in AI infrastructure.
In our last piece, we followed AI’s power bottleneck from the grid to the rack and left one question open: who captures the incremental content per megawatt? The common answer is every electrical supplier, since all of them are selling into record backlogs. However, that answer misses what is changing inside the building.
The industry divides a data center into two domains, “Grey Space” and “White Space.” Grey Space is the electrical and mechanical backbone: transformers, switchgear, backup power, and chillers. White Space is the IT room, together with the power and cooling hardware that sits next to the chip. AI’s power density has forced that boundary to shift: Google expects machine learning to require more than 500 kilowatts per IT rack before 2030, and equipment designed for far lighter loads cannot meet that demand. The boundary is not just blurring, it’s inverting. Functions are moving out of the Grey Space toward the rack, and some are disappearing altogether. The scarce input is shifting from capital to engineering bandwidth: the number of teams able to design this equipment.
To see why, let us start with what is physically moving inside the data center.
What Moves, What Disappears, What Stays
The inversion begins with physics: at AI rack densities, the old way of delivering power runs out of room. Our earlier piece covered why 800 VDC matters: roughly 15 times less current than low-voltage distribution, and far less copper. Here is what it changes inside the building.
Power is voltage times current, and the heat conductor’s output rises with the square of the current. At one megawatt, a 54-volt rack bus would carry about 18,500 amps; at 800 volts, about 1,250 amps. Beyond a few hundred kilowatts per rack, low-voltage distribution runs out of copper and space before it runs out of power.
Two designs now compete to deliver 800 volts. Bipolar ±400V uses three conductors: a positive rail, a negative rail, and a midpoint, so the system carries 800 volts across the outer rails but never more than 400 volts to ground. That lets designers use components built for electric vehicles, which Google cites as a reason for choosing it. Native 800V uses two conductors at a full 800 volts to ground, which simplifies cabling and is NVIDIA’s preference. They are not mutually exclusive: one side of a power rack can be built to output either.
Either of these designs changes where Alternating Current (AC) is converted to DC. Today’s first step is a side power rack, a cabinet next to the compute racks that converts the facility’s 480-volt AC locally. The Open Compute Project (OCP), the industry body where Google, Microsoft, and NVIDIA publish these specifications, calls it “the fastest path to 800 VDC capability in today’s AI factories.” The longer-term design uses solid-state transformers (SSTs), power-electronics replacements for the iron-and-copper transformer, to convert medium-voltage grid power directly to 800 VDC for the entire data hall. SSTs are at the first-product stage. Google and its partners say the published ±400V specification covers current deployments while the native-800V version “will be coming later.” They have given no date. The figure below traces the three steps, from today’s AC chain to direct conversion.
Three steps from AC chain to 800 VDC
Each step removes conversion stages between the grid and the chip. Steps are sequential, not dated.
1. Today: AC chain
- Medium-voltage AC
- Transformer and switchgear
- UPS (AC to DC to AC)
- AC PDU
- Rack power supply (AC to 54 V DC)
2. Side power rack
- Medium-voltage AC
- Transformer and switchgear
- Side power rack converts AC to DC
- ±400 V or 0 to 800 V DC output
- Compute rack
3. Direct to 800 VDC
- Medium-voltage AC
- Solid-state transformer: medium-voltage AC straight to 800 VDC
- 800 VDC distribution
- Compute rack (DC-DC conversion only)
We classify the following steps as ‘Moving toward the rack’: AC-to-DC conversion, the backup-power function (shifting into battery units beside the rack), and thermal hardware such as coolant distribution units and cold plates. Then the steps that are ‘Disappearing on new builds’: AC power distribution units, centralized AC backup power, and the main distribution boards inside the data hall, with NVIDIA’s reference designs also letting many sites skip mechanical chillers. Finally, the items that are ‘Staying at the facility’ are: medium-voltage switchgear and transformers at the grid connection, long-duration generators, large-scale batteries, and centralized chillers for sites that maintain a secondary water loop. This elimination applies only to DC-native new builds. OCP stresses that 800 VDC “can coexist” with existing AC systems, so existing sites mostly add side racks and keep their AC busways.
Where the grey space goes in AI data centers
Functions of the building’s electrical and mechanical backbone, sorted by what happens to them on new AI builds.
Moves toward the rack
- AC-to-DC conversionPower supplies and side power racks take over the conversion stage once handled at the facility.
- Backup power functionBattery backup units sit at or inside the rack.
- Coolant distributionCoolant distribution units and cold plates move next to the chip.
Disappears on new builds
- AC power distribution units
- Centralized AC backup power
- Main distribution boards in the data hall
- Many mechanical chillersThrough NVIDIA reference designs.
Stays at the facility
- Medium-voltage switchgear and transformers
- Long-duration generators
- Large batteries
- Centralized chillersOn secondary loops.
The equipment that leaves the grey space ceases to be a catalog item because it must now be designed around the chip it feeds. That changes how it is bought.
Why Catalogs Stop Working
The equipment moving into the rack can no longer be ordered; it must be engineered. For decades, specifying data center power resembled ordering from a menu. Loads were steady, conversion followed a standard multi-stage AC path, and the building blocks – 480-volt distribution and 48-volt DC inside the rack (a standard inherited from telephone networks) – were mature and sold from catalogs. The work was selection and sizing. That is also why earlier standardization was achievable: it packaged known technology to cut labor, and prefabricated modular builds can cut timelines by up to 60%.
Rack-level 800V is closer to working with the chef. The power system has to be designed around the chip’s behavior, and five problems arrive at once: power stages built on silicon carbide and gallium nitride devices that switch at high speed at 800 volts; heat removal designed together with the chip’s load swings; battery backup that responds in milliseconds; protection against DC arcs, for which standards are still being written; and testing the server, rack, and cluster as one system. Validating parts one by one no longer proves the system works. Everything has to be validated together and hence designed together. Out goes the cookie-cutter model of the past. This naturally requires more engineering resources to design these products.
If engineering depth is the scarce input, it should show up in what suppliers spend on it, and it does. The first chart below shows R&D as a share of sales over 2016–25. Delta, a power-electronics specialist, spent 8.7%; most diversified electrical peers spent 3% to 5%. Those peer totals include large non-data-center businesses (by analyst estimates, data centers account for about 30% of Schneider’s sales and about 5% of Siemens’ sales), so the gap is likely wider.
R&D as a % of Sales
Source: Bloomberg
The second chart shows the same decade in absolute dollars – Delta has spent as much in R&D dollars as the largest electrical components companies despite being smaller. Keep in mind that Delta is a power-electronics specialist. Most other peers in the chart are diversified capital-goods groups, so their dollar R&D includes business lines unrelated to data centers. That means the R&D dollar figures for everyone else correspond to business lines that are not relevant to Data Centers. And even the incumbents have been highlighting this renewed focus on new product expertise – Vertiv’s management has said that success with solid-state transformers will depend on in-house power-electronics expertise, not transformer-making experience.
R&D 2016-2025 ($ Million)
Source: Bloomberg
Capital can buy a factory or a company, but it cannot quickly buy a design culture that works alongside a chip designer. The next question is, who sets the design that those teams build toward?
Who Writes the Spec
The customers, not the suppliers, now write the architecture. Google, Meta, and Microsoft co-authored Mt. Diablo, the ±400V specification. Google, Microsoft, and NVIDIA developed the solid-state transformer specification, released as version 0.3 in mid-2026, and a low-voltage DC white paper in March. More than 80 companies are developing compatible equipment.
Their motives are economic. Hyperscalers optimize for cost per token and per watt, run custom chips alongside NVIDIA’s, and want options beyond any one vendor’s reference design. We think that makes them white-space architects: they set the power architecture around their own silicon and buy components to it, capturing design economics once held by the chip vendor and by the electrical incumbents.
Publishing the spec does not publish the know-how. Supplier lists from these programs and supply-chain reports show a small group of rack-power specialists on most Hyperscaler programs, though supplier shares are not disclosed. Business models also differ: some specialists sell only components, while others also sell full power racks, which some customers may view as competition. Lite-On told investors in July that its 800V power rack, initially built for an unnamed ASIC customer, would enter small-volume production in November if qualification succeeds and scale in the first quarter of 2027.
Timing remains open. Google calls its ±400V sidecar the “first embodiment,” with higher-voltage DC to come “longer term.” As of mid-October, Google has published no date for native 800V.
So, the customer sets the target, and engineering bandwidth decides which suppliers can hit it. But specifications are also how industries commoditize, so does that undo the argument?
What Could Close the Gap
The strongest objection to our thesis is that the specifications exist to commoditize. OCP says that when three of the largest operators publish common requirements, “the supply chain can build once and sell everywhere,” so suppliers need not “navigate conflicting proprietary specifications from different customers.” Past transitions suggest commoditization takes years: x86 servers took about thirteen years to reach dominance, and modular UPS systems shifted value toward service and deployment scale once they standardized. If 800V follows that path, the advantage could move toward suppliers with scale and global service networks.
Our answer to this pushback is timing. Broker estimates of 800V’s share of new deployments run from about 10% to 40% in 2027 and from about 20% to more than 90% by 2030, depending on whether ±400V counts and what is measured. That spread is itself the point: if people cannot agree even on what is being counted, the industry has not settled on a dominant design that suppliers can simply build to. Commoditization also requires certified safety standards, supply-chain scale, and Hyperscaler alignment on a single architecture, and none of the three is yet in place. The table below shows how far apart the broker estimates are.
Brokers disagree on how fast 800 VDC arrives
Estimated 800 VDC share by year. The figures use different bases, so they should not be compared row to row.
| Source | 2027 | 2028 | 2030 | What the share measures |
|---|---|---|---|---|
| Citi | 16% | 58% | 79% | New capacity additions |
| HSBC | 38% | n/a | 78% | New capacity, single-bus 800 V core |
| Huatai | 40% | 76% | 93% | New builds plus retrofits, ±400 V included |
| UBS | n/a | n/a | 58% | New deployments |
| TD Cowen | 10% | 30% | n/a | 600 kW+ racks only |
| Daiwa | n/a | 50% | n/a | HVDC adoption |
| ABB / Berenberg | n/a | n/a | 25% to 45% | Total market |
| Deutsche Bank | n/a | n/a | 20% | New data centers, measured in GW |
Incumbents hold cushions meanwhile. Vertiv earns about a 58% gross margin on services versus about 36% on products; services account for about 20% of its sales, roughly three-quarters of which are recurring, backed by more than 5,300 field engineers. A components-only specialist has no equivalent annuity. Incumbents also expect new profit pools in medium-voltage rectification, DC protection, and solid-state transformers, but orders are still at prototype scale: Eaton has 10 committed SST prototype orders, per Redburn; ABB expects its first meaningful 800V orders in the first half of 2027; and Vertiv’s systems are in customer testing. We found no public Hyperscaler confirmation of an incumbent as an 800V co-design partner, though NVIDIA names several as reference-design partners.
China may offer an early look at how this shakes out. Domestic accelerators deliver far less compute per watt than NVIDIA’s, so Chinese operators must pack more power into each rack and face the density problem sooner. Domestic power vendors are winning qualification with Chinese Hyperscalers on speed and local responsiveness, while Delta is the global leader in white space power equipment, holding an estimated 6% to 10% of China’s UPS and HVDC markets. If that pattern holds, the first designs will come from teams that sit closest to the customer, not from the largest catalog. We confine our argument to the U.S. and global Hyperscaler programs, but China is where the pattern should show first.
Until standards, scale, and a single agreed-upon architecture arrive, design capacity, not capital, sets the pace. We think that could make things interesting.
Five Things We’re Watching
Each of the five markers below tests the thesis:
- If DC protection and arc-flash standards are finalized, and how quickly certified solid-state transformers reach volume.
- Whether dated native-800V roadmaps follow the specifications.
- Solid-state transformer orders beyond prototypes.
- Whether any Hyperscaler publicly confirms an incumbent as a co-design partner.
- If specialists’ R&D intensity holds as designs standardize.
If these markers break our way, the value in AI power will go to the teams that can design it, not to the companies that can simply supply it. For investors, the question is no longer who can supply the power, but who can design it?
More Insights
When the Machines Sell: Amplified Market Fragility and Why It Could Favor the Patient, Unlevered Investor
Thornburg Income Builder Opportunities Trust Announces Distribution
Gradually, Then Suddenly: What Q3 Revealed About Bond Markets
Rates up, Credit Steady: Municipal Yields Hit 15-Year Highs
Why Advisors Should Consider an International Approach to Income
