ENVIZNThe Value of What Sits Between the Grid and the GPU
How AI is changing the economics of electrical infrastructure, and why the move toward DC distribution may be about more than efficiency.
More of a fixed utility allocation reaches productive compute. In the illustrative model, one recovered megawatt contributes $16 million over five years before the capital needed to use it.
For decades, progress in computing has been measured in processors, memory, and software. The infrastructure beneath them improved alongside, adding electrical capacity, better cooling, and more sophisticated distribution, but it was valued as a supporting system.
Artificial intelligence is starting to change that relationship. How much computation a facility can add no longer depends only on the processor. It depends on how much electricity can reach that processor, how effectively its heat can be removed, and how reliably everything around it holds up. As AI systems grow more demanding, the value of the supporting infrastructure becomes harder to separate from the value of the compute it serves.
The work now underway on 800 VDC distribution, solid-state circuit breakers, and semiconductor-based power conversion is an early signal of where that relationship may be heading. Each technology matters on its own terms, and together they suggest a change in valuation: electrical equipment may increasingly be priced by what it enables and protects downstream, alongside its own cost and performance.
When supporting infrastructure becomes the constraint
In a data center, the computing equipment is the productive asset. It runs the calculations and delivers the services that earn revenue. The electrical infrastructure exists to support it, and that division of labor has pushed manufacturers to improve components one at a time: more efficient transformers, more reliable breakers, better UPS systems, more capable distribution equipment.
Those improvements still matter, but a better component does not necessarily improve the economics of the whole facility, and that gap is the subject of this essay.
Take two data centers with identical utility power allocations. The first uses a conventional architecture, with several conversion stages between the utility connection and the servers. The second uses a redesigned architecture that cuts conversion losses, consolidates some distribution functions, and frees physical space for compute. Both receive the same amount of electricity. The second may deliver more of it to productive hardware.
For the operator, the useful measure is how much computing capacity a fixed purchase of power can support. A system that costs more up front can still produce the better outcome if it enables more computation, reduces downtime, or accepts future hardware without extensive rework. Evaluated that way, the electrical system is paid for partly by the computation it makes possible.
Why the industry is looking toward DC
The movement toward DC distribution did not start with AI: solar arrays generate DC, battery storage holds it, electric vehicles run on high-voltage DC packs, and semiconductor-based computing ultimately consumes regulated DC. Each of these industries has pushed power electronics, protection, and system design toward handling larger DC loads, and AI adds a powerful new incentive to keep going.
In May 2025, NVIDIA proposed an 800 VDC architecture for AI factories built around megawatt-class racks, with full-scale production timed to its Kyber rack systems in 2027. By August 2026, Google, Microsoft, and NVIDIA were working through the Open Compute Project to standardize 800 VDC as an open power architecture, covering performance requirements, power quality, system interfaces, and safety certification. NVIDIA counts more than 80 equipment and infrastructure companies building products to the specification.
The motivation is relatively simple. A conventional data center distributes power as AC and converts it to DC near the load. The path can include medium-voltage distribution, transformers, UPS systems, low-voltage distribution equipment, rack-level power supplies, and further DC/DC conversion next to the processor. Each stage does necessary work, and each adds some combination of electrical losses, equipment, heat, and maintenance.
An 800 VDC architecture moves conversion upstream and carries DC across more of the facility before stepping it down for the processor. NVIDIA estimates the approach could cut copper requirements by 45 percent compared with 415 VAC distribution and remove rack-level AC/DC conversion, freeing space for compute. Those are projected architecture-level benefits, not guaranteed results at any given site.
The GPUs do not run at 800 VDC. Voltage conversion, protection, isolation, redundancy, and energy storage all remain necessary. The opportunity is to shorten the chain of equipment between the utility connection and the processor.
Nor is this an argument for abandoning AC. The utility grid is overwhelmingly AC, and conventional architectures remain proven across countless applications. The OCP collaborators describe 800 VDC as an additional deployment option alongside existing AC infrastructure. The practical question is where in a facility AC should end and DC should begin, and for AI infrastructure the economics of that boundary may be shifting.
Solid-state protection preserves value downstream
Higher-voltage DC distribution brings a harder protection problem. An AC breaker benefits from the current naturally passing through zero 100 or 120 times a second, which gives an arc a moment to extinguish. DC offers no such moment, so fault interruption grows more demanding as distribution voltages and power densities rise.
Solid-state circuit breakers take a different approach. Instead of relying mainly on mechanical contacts, they interrupt current with power semiconductors that switch within microseconds. ABB’s SACE Infinitus, rated for 1,000 VDC and 2.5 kA, is an existing example; ABB puts its interruption time at roughly 20 to 50 microseconds. In September 2026, SolarEdge and Infineon extended their collaboration to develop solid-state breakers for 800 VDC AI data centers, built on Infineon’s silicon carbide JFETs and aimed squarely at high-speed DC fault interruption.
The value of that speed comes mostly from the equipment it protects. A protection device earns its cost through the damage, downtime, and disruption it prevents, so if better protection limits the fault energy that reaches expensive computing equipment, its premium can be justified by the value of the assets behind it.
The claim should stay narrow. A solid-state breaker does not reduce the heat a GPU produces, and it does not shield equipment from every electrical disturbance. What it can do, when properly selected and coordinated, is cut off certain faults before their energy causes more extensive damage downstream. The more valuable the equipment behind a breaker, the more that interruption is worth.
That makes protection part of how the facility preserves its productive capacity, on top of meeting electrical code. Standards are where that becomes concrete. UL 489I, published in October 2025, sets requirements for solid-state and hybrid circuit breakers rated up to 1,000 VAC and 1,500 VDC, including an integral air gap for galvanic isolation. UL 891, meanwhile, covers switchboards rated 1,000 V nominal or less. The two operate at different levels: qualifying a protection device does not automatically qualify the assembly built around it.
For electrical manufacturers, the work therefore extends beyond a faster breaker to proving that new protection can be integrated safely and reliably into complete distribution systems. For operators, the question is whether that added sophistication creates enough value to justify it.
The electrical room becomes an extension of the rack
The transition also changes the relationship between electrical and computing infrastructure. Some of the complexity that used to sit beside the servers may move upstream. Centralized conversion can reduce the need for rack-level power supplies, freeing space and removing conversion heat from the rack.
The equipment delivering that power gets more sophisticated in turn. Solid-state transformers bring semiconductor conversion into the distribution path. Solid-state breakers carry power electronics, sensing, and digital control. Higher-power semiconductor systems bring their own thermal load. ABB is candid about the trade: it describes higher on-state losses, caused by the voltage drop across the semiconductor, as a historical hurdle for solid-state breakers. Infinitus uses reverse-blocking IGCTs to cut conduction losses by about 70 percent relative to an equivalent IGBT, and it still relies on liquid-cooled cold plates to remove the heat.
Reducing conversion losses near the rack does not make losses disappear. Some heat and complexity may simply relocate to upstream power equipment, and whether the total cooling burden falls depends on the complete electrical and thermal design. The electrical room may become more sophisticated even as some equipment categories consolidate.
There is a broader convergence here. The computing industry is investing more and more in what surrounds the processor: advanced cooling, high-bandwidth interconnects, power delivery, and rack-scale integration. Electrical infrastructure, meanwhile, is absorbing semiconductors, embedded controllers, software, sensing, and thermal management. The two industries are approaching some of the same engineering problems from opposite directions, and the sophistication is arriving because the computation downstream is valuable enough to pay for it. How much value would actually justify the investment is the question the next section tries to put numbers on.
What is an additional megawatt of compute worth?
Consider a hypothetical AI data center with a 100 MW utility allocation. Under its existing architecture, 90 MW reaches productive computing equipment and 10 MW goes to electrical losses, cooling, and other facility loads. Now suppose an alternative architecture reduces supporting power enough to make one more megawatt available to compute, after accounting for the cooling and overhead that the additional hardware brings with it.
The allocation has not changed. The facility now delivers 91 MW to compute instead of 90, assuming it has the cooling capacity, space, and compatible infrastructure to host the extra hardware. That megawatt may be worth far more than the electricity the distribution system saved. An entirely hypothetical model shows why.
| Variable | Assumption |
|---|---|
| Additional productive compute capacity | 1 MW |
| Annual economic contribution per MW at full utilization | $4 million |
| Expected commercial utilization | 80% |
| Evaluation period | 5 years |
| Additional compute hardware investment | $8 million |
| Electrical infrastructure cost premium | $4 million |
These figures are illustrative assumptions, not estimates of the economics of any actual NVIDIA-powered data center.
The annual contribution is what remains after direct computing operating expenses, including the electricity to run the added compute, but before the incremental capital. On those assumptions, the added capacity contributes $4 million × 80% × 5 years, or $16 million. The operator spent $12 million on computing hardware and electrical infrastructure to get it. Before financing, taxes, or the time value of money, that leaves $4 million over five years.
Productive capacity is one dimension of the infrastructure’s value. Two others, protection and longevity, change the total.
Protection
Suppose better fault protection also reduces the expected annual cost of equipment damage and operational disruption by $300,000, or $1.5 million over five years. That benefit has to be calculated separately. If it represents avoided losses on the facility’s existing equipment, it can sit beside the new megawatt’s contribution. The same lost revenue cannot be counted in both.
Longevity and operating cost
Suppose the redesign also avoids a $2 million electrical retrofit that would otherwise be needed to host a future generation of hardware. That is a separate benefit, but only if its value is not already embedded in the capacity calculation. Against all of this sit new costs: semiconductor cooling, maintenance, and specialized equipment support. Assume those run $200,000 a year, or $1 million over five years.
| Economic impact | Five-year value |
|---|---|
| Incremental productive compute contribution | +$16 million |
| Avoided fault-related losses | +$1.5 million |
| Avoided infrastructure replacement costs | +$2 million |
| Additional compute hardware investment | −$8 million |
| Electrical infrastructure cost premium | −$4 million |
| Additional infrastructure operating expenses | −$1 million |
| Illustrative net economic benefit | +$6.5 million |
The calculation is deliberately simple. A complete investment model would discount future cash flows, account for financing and taxes, consider the residual value of equipment, and weigh the risks of deployment and utilization. The result could change substantially if compute becomes less profitable, utilization falls short, or the new infrastructure proves more expensive to maintain.
That uncertainty is why the framework matters. Some advanced electrical architectures will earn a positive return and some will not, and in either case the answer depends on more than a comparison of equipment prices.
The broader calculation
Physical space, electrical efficiency, and equipment reliability are inputs to this equation, not separate line items. If freed rack space enables more compute, that benefit already sits in the productive compute term; pricing the recovered space again counts it twice. Likewise, if an efficiency gain lets more compute run within a fixed allocation, the operator cannot also book the redirected electricity as a lower power bill.
The model predicts nothing about any particular facility. It is a way to see why supporting infrastructure can gain value even as its individual components get more expensive.
The economics of the transition
This framework changes how technologies like solid-state breakers and 800 VDC distribution should be judged. A solid-state breaker may cost more than its mechanical equivalent. It may add semiconductor losses, cooling requirements, electronics, and different maintenance considerations. Measured only against the purchase price of a conventional breaker, it can look hard to justify. If it enables an architecture that better protects valuable computing equipment or supports higher-power DC distribution, the comparison changes.
The same reasoning applies to solid-state transformers and centralized conversion. A more expensive conversion system can be justified if it replaces several conventional stages, reduces losses, frees space, or allows more compute within an existing utility allocation.
None of these benefits is automatic. A solid-state breaker does not remove conversion stages on its own. An 800 VDC architecture does not make every data center more profitable. More sophisticated equipment brings new reliability, certification, and maintenance challenges. The evaluation has to happen at the system level.
That creates different incentives for each participant. For computing manufacturers, better electrical architecture removes limits on future power density and system design. For operators, it makes existing allocations, floor space, and capital more productive. For electrical manufacturers, it opens new equipment categories and a way to hold position as customer requirements shift.
Demand for better breakers and transformers is part of this, but the larger driver is that the value of downstream equipment makes upstream improvements increasingly consequential.
Deployment time and equipment lifetime
Deployment complicates the transition. An advanced DC architecture may deliver better efficiency and more compute over a facility’s life. A conventional AC design may let the same facility start operating sooner, because its equipment, engineering practice, supply chains, and installation methods are already established. When demand for compute is strong, an earlier commissioning date can be worth a great deal.
So there is a trade between immediate deployment and long-term productive capacity. A new electrical system may eventually enable more compute, but that advantage has to be weighed against any additional time needed to engineer, procure, certify, and commission it.
NVIDIA’s roadmap reflects the problem. It stages the transition: power racks designed to work with existing AC infrastructure in the second half of 2026, row-level 800 VDC power centers in 2027, and eventually facility-scale units that convert grid power to 800 VDC in a single step. The sequencing recognizes that operators cannot necessarily afford to abandon what they have while a new architecture matures.
The value of electrical infrastructure therefore includes how quickly capacity becomes productive, not only how much capacity it enables. A facility that earns revenue sooner can produce a better investment outcome than one that runs more efficiently but reaches commercial operation substantially later. Which wins depends on the operator’s circumstances, cost of capital, expected demand, and investment horizon.
The infrastructure that outlives the processor
Equipment lifetime raises a second timing question. A data center is built around infrastructure expected to serve several generations of computing equipment. Processors, memory, and networking can change dramatically within a facility’s operating life. The electrical system is expected to last much longer.
That creates an unusual relationship. An architecture designed around today’s requirements can become the constraint on tomorrow’s hardware while every one of its components still works, leaving a building with decades of physical life whose economic usefulness is capped by what it was originally designed to accommodate.
This is another reason to consider more adaptable power architectures. Nobody can predict every future generation of hardware, but a flexible design lowers the odds that supporting infrastructure becomes the reason a facility cannot host it. Solid-state protection, modular conversion, and higher-voltage DC distribution may help, though every architecture keeps physical limits and introduces its own dependencies.
In practice, future-proofing means lengthening the period in which a facility can economically support the technology it was built to serve. Some equipment will still need replacing along the way.
Rethinking the value of infrastructure
For most of computing’s history, better processors were the visible engine of progress. Faster chips, larger memory systems, and more capable software expanded what computers could do. The infrastructure around them was essential, and comparatively easy to overlook.
AI is making it harder to overlook. Installing more GPUs, deploying new hardware generations, operating equipment reliably, and making productive use of a limited electrical allocation all depend increasingly on decisions made across the facility.
The way electrical infrastructure is evaluated may change as a result. The cheapest breaker, the distribution design with the fewest components, and the smallest electrical room can each turn out to be the costlier choice once their effect on the rest of the system is counted.
Conventional electrical economics have not become irrelevant. Efficiency, reliability, maintainability, and capital cost remain fundamental. What changes is the scale of the downstream opportunity. The move toward DC distribution may ultimately be judged mainly by the computing capacity it makes possible, and only secondarily by the equipment it replaces.
As AI raises the value of the load, the electrical industry, long judged on how reliably it delivers power, may increasingly be judged on how much productive compute that power supports.
As the value of compute keeps rising, the open question is whether operators and manufacturers will evaluate electrical systems by the productive compute they enable and keep in service over a facility’s life.
In the illustrative model, one more megawatt reaching compute under the same 100 MW allocation and electricity bill leaves $4 million over five years, before financing, taxes, or the time value of money.
All insights