AI Data Centers Abandon Copper Cables for Controllability: How NVIDIA's 800V DC Architecture Changes Power Supply Control

AI Data Centers: Abandoning Copper Cables, Seeking Controllability. Source: Semiconductor Horizons. Published: 2026-11-02 15:32. Server data centers generate posters. High-voltage direct current solutions are expected to be deployed at scale in 2027–2028, which is when 800-volt DC data centers enter the mass production stage.

Server power supply was once a mature, solved problem. For decades, ordinary server racks drew 10 to 15 kilowatts through 48-volt power distribution buses, and to this day the vast majority of non-AI racks still use this approach. The accompanying control circuits are mature and not particularly noteworthy. But artificial intelligence has completely overturned this situation. Today's AI racks already consume about 130 kilowatts, and next-generation rack designs aim directly at one megawatt. This transformation is not driven by a single factor but is the result of multiple changes stacking up simultaneously: computing density keeps climbing, distribution voltage is rising by more than an order of magnitude, wide-bandgap power devices switch far faster than silicon devices, accelerator loads fluctuate violently within extremely short periods, and at the same time requirements for critical infrastructure in terms of efficiency, safety, and trustworthiness are steadily tightening.

All these changes ultimately converge on the same core: the digital controllers used at every stage of power conversion between the grid and the processor. And they all follow the same trend: copper cables and bulky passive components are gradually exiting the rack, replaced by intelligent control.

Eric Chou, Director of STMicroelectronics' PDSA Power and Energy Technology Center (PECC) in China, said, "Everyone is focused on power devices evolving toward larger specifications and higher switching speeds, but few notice what is being discarded inside the rack: copper, large-capacity passive components, and design margins. What replaces them is control capability, and the competitive landscape of AI power supply over the next decade will be determined by control capability."

NVIDIA's 800-volt DC architecture is designed to scale single-rack power consumption from 100 kilowatts to more than one megawatt, with the first megawatt-class racks expected to be deployed in 2027–2028. The 54-volt busbars currently used inside AI racks can no longer support this roadmap. According to NVIDIA's estimates, if 54-volt distribution continued to be used in a one-megawatt rack, the copper busbars alone would require as much as 200 kilograms; a one-gigawatt data center would use as much as 200,000 kilograms of copper. Adopting 800-volt DC distribution can greatly reduce copper usage, and under the same estimates, end-to-end power supply efficiency can improve by up to 5%.

Along with the voltage upgrade, the entire power supply architecture changes as well. In the past, power conversion units and IT equipment were placed in the same rack and powered by AC. Mainstream new solutions now separate the two: an independent power rack transmits high-voltage DC power to the IT rack. Cloud providers are promoting standardization of this approach through the Diablo specification released by the Open Compute Project (OCP) in October 2025. Cooling solutions are evolving in parallel, and new-generation rack-level AI systems come standard with liquid cooling at the design stage.

Each of these changes reduces copper and transmission losses inside the rack while concentrating the core pressure on the remaining power conversion stages — that is, the conversion units that step 800 volts down stage by stage to the fractions of a volt required by the processor.

Why 800 volts changes the difficulty of control

Raising the voltage is an effective way to improve efficiency, but it places higher demands on control. Every switching action must withstand an 800-volt potential difference, so switching must be faster and waveforms cleaner. The fault tolerance margin at 800 volts is far smaller than at 54 volts, protection mechanisms must be more rigorous, and even scenarios where maintenance personnel replace modules while energized must be covered. At the same time, the conversion chain is being streamlined. To reduce losses, the industry is merging and reducing the number of conversion stages; the remaining stages are fewer, each carries a higher load, and design margins are smaller. Converters can already step down directly from 800 volts to the low-voltage bus near the processor, and this approach has been validated in 800-volt DC racks.

The load side also brings additional challenges: AI accelerator power consumption fluctuates violently at the millisecond level with computing load. For this reason, the 800-volt DC architecture is paired with energy storage units to smooth out such power fluctuations. Each energy storage, conversion, and output unit is an independent closed-loop control link, and all links must work in coordination to ensure that the rack presents a stable and clean power supply externally.

Controllers face entirely new requirements

The switching devices inside conversion units increasingly use wide-bandgap devices: gallium nitride (GaN) and silicon carbide (SiC). Compared with traditional silicon devices, they switch faster and have lower losses. GaN switching frequencies can approach 1 megahertz, and GaN converters based on 800-volt inputs have already exceeded 98% efficiency. Higher switching frequencies can also shrink the size of accompanying magnetic components and capacitors, which is the underlying reason bulky passive components can be streamlined. But whether the high-frequency advantage can be realized depends on whether the controller can keep up. At 1 megahertz, the entire switching period is only 1 microsecond, and every level transition requires nanosecond-level precision; a deviation of just 1 nanosecond will cause efficiency loss or device stress.

Multiple pressures are stacking up, redefining what "control" means in power conversion. For decades, the controller's core tasks were only two: hold the output voltage steady, and shut down the circuit on a fault. In AI racks, however, the controller must simultaneously accomplish several highly demanding tasks and run stably for years: the control response must be fast and highly precise; the control algorithms are more complex, and to fully exploit the performance of wide-bandgap devices, a large amount of computation is needed in every switching cycle—once the closed-loop response lags, the consequences can be even worse than taking no action; the protection mechanisms must be robust, since the time left for software to handle an 800-volt fault is almost zero, so the protection logic must be hardened in hardware; the endurance must be strong, operating continuously for the long term right next to the hottest devices in the rack; and it must be secure and trustworthy—since power supply is critical infrastructure, the controller needs to verify firmware authenticity, accept only trusted signed updates, and adapt to security requirements far beyond the service life of current cryptographic systems. Therefore, adapting to the post-quantum cryptography standards finalized by NIST in 2024 has become a newly added requirement today.

How to implement these capabilities in chips is an engineering detail, to be expanded on in a companion technical article. For decision-makers on power supply solutions, the core conclusion is simple: capabilities that only high-end controllers once had have now become the baseline threshold for every conversion stage between the grid and the processor.

Power conversion becomes a system-level decision

Digital power conversion units are no longer just a matter of selecting discrete components, but a system-level solution decision. Switches, drivers, magnetic components, and controllers need to be selected and defined as a unified whole, because the ceiling of any one component is the ceiling of the entire solution: no matter how fast the transistor, if the closed-loop control cannot keep up, its performance cannot be realized; and no matter how good the control loop, it cannot rescue a switching device with a poorly designed driver.

This is changing the industry's procurement and development models. Compute operators and system partners increasingly tend to evaluate power conversion units as a whole, and suppliers need to provide matching complete solutions rather than scattered components. All the 800-volt projects we participate in show this trend, and the impact of this change has already gone beyond any single supplier.

Future development directions

Rack compute density will continue to rise. High-voltage direct current solutions are expected to be deployed at scale in 2027–2028, which is when 800-volt DC data centers enter mass production. This also paves the way for solid-state transformers: power electronics solutions will replace the traditional line-frequency transformers between the grid and the DC bus. The switching speed of wide-bandgap devices will continue to increase. Each round of iteration follows the same logic: passive components with high fault tolerance are removed from the chain and replaced by active, controllable solutions.

The core dividing line for next-generation AI infrastructure is not how much power a rack can take in, but how precisely we can control that power once it is delivered. In this transformation, the winners will be the vendors that can simultaneously provide power devices, driver chips, and intelligent control solutions.

This article is reproduced from media reports or online platforms and represents the author's personal stance or views. Our reproduction is solely for sharing and does not mean we endorse or agree with it. If the source is incorrectly attributed or your lawful rights and interests are infringed, please contact customer service promptly, and as a neutral platform service provider we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

Record-breaking! 2.75GHz gallium nitride micro-LED unveiled, targeting low-power data centers 2026-11-02 16:18

Broadcom becomes a big winner in Anthropic's IPO 2026-11-02 16:15

Backed by the two memory giants' hugely profitable earnings reports, South Korea unveils its largest-ever AI semiconductor budget 2026-11-02 16:11