NVIDIA’s rumored 800V power supply architecture may be pushed to 2028? Delta Electronics says it will begin small-scale production in Q1 2027

NVIDIA’s originally planned 800V high-voltage DC (HVDC) power-supply architecture, set to be deployed in 2027, reportedly has been delayed. The market research firm SemiAnalysis has warned that large-scale commissioning could be pushed to 2028 or even later. Supply-chain sources indicate the bottleneck is “verification” rather than the technology itself. Initial testing found potential short-circuit risks, and U.S. customers requested additional testing, moving the schedule back. Delta (DELT) is expected to start small-volume production by the end of Q1 2027, then ramp up gradually in Q2. Foreign investors, meanwhile, believe HVDC remains the long-term power direction for AI data centers, and next year the penetration rate of 800V products could reach about 20%.

(Background recap: NVIDIA’s 800V power revolution won’t be delayed! Confirmed with Delta and ABB partners: Q3 on-time volume production)
(Background addition: Dell teams up with NVIDIA to launch “fully liquid-cooled” AI servers! First with the Vera Rubin architecture; a single rack cabinet with 144 GPU cards delivers off-the-charts compute)

Table of contents

Toggle

• Delay is stuck in verification, not because the technology can’t be done
• What problem does 800V HVDC really solve
• Will individual users use it?

In early July, supply-chain contacts were still confidently saying NVIDIA’s 800V high-voltage power supply “won’t be delayed” and would “go into mass production on time.” Less than half a month later, the wind changed. Market research firm SemiAnalysis stepped in to warn that this architecture, originally scheduled to land in 2027, may be pushed to 2028 for large-scale rollout, or even later.

But moving the timeline back doesn’t mean the technology can’t be built. Supply-chain sources say it clearly: the bottleneck is “verification,” not “technology.”

Delay is stuck in verification, not because the technology can’t be done

The issue lies in the system verification phase. During initial tests, the engineering team found potential short-circuit risk. U.S. customers immediately demanded additional testing. A series of redesigns plus the customer certification process has pushed back the shipment schedules of upstream suppliers across the board.

Whether this counts as good news or bad news depends on which perspective you take. Once high-voltage DC experiences a short circuit, the consequences are far worse than with low-voltage power. Customers would rather run a few more rounds of validation than give the green light lightly. What’s being held back isn’t that it can’t be built—it’s that they’re not willing to risk lives on the gamble.

Delta’s proposed schedule is: small-volume production by the end of Q1 2027, then gradually ramp up shipments in Q2. Foreign investors are not as pessimistic. Most believe HVDC is still the long-term power direction for AI data centers, and next year the penetration rate of 800V-related products could reach around 20%. In other words, what’s delayed is the timetable—not the direction.

What problem does 800V high-voltage DC really solve

To understand this architecture, you first need to know how much power an AI rack consumes today.

In NVIDIA’s next-generation Rubin Ultra generation, the power draw of a single rack cabinet is nearing 1MW (one megawatt)—roughly the same order as thousands of households running air conditioners at the same time. Traditional data centers use 54V DC or 415V/480V AC power. To feed a monster like this, you have to pack multiple layers of power-conversion equipment inside the rack cabinet, and also pull in heavy copper busbars weighing 200 kilograms. The more copper busbars you add, the less space is left for the compute chips.

The solution of 800V high-voltage DC is straightforward: convert mains power into 800-volt high-voltage DC outside the rack cabinet, send it directly into the cabinet, then use DC-DC conversion to step it down and power the GPUs. Raise the voltage, and each copper line can deliver 150% more power. NVIDIA has calculated that this approach saves about 45% of copper, eliminates up to 4 additional power-conversion stages, improves end-to-end efficiency by more than 5%, and reduces overall cost of ownership by roughly 30%.

The same logic applies as to why high-voltage power transmission towers use high voltage. The higher the voltage, the smaller the current-related losses, making it more suitable for long-distance, high-power transmission. What AI lacks isn’t just chips—it’s the entire infrastructure that can deliver stable power to the chips.

Will individual users use it?

First, the conclusion: this 800V HVDC is the power backbone for data centers and AI factories. Your home wall outlets won’t turn into this in the near term. Residential electricity will still be low-voltage AC at 110V or 220V. In between are several layers of conversion, so ordinary users won’t directly encounter the 800V high-voltage backbone line inside a datacenter.

But the concept of “800V high voltage” is something individual users have already encountered—electric vehicles. Porsche Taycan and Hyundai’s E-GMP 800V high-voltage platforms rely on the same logic of raising voltage and reducing losses, enabling ultra-fast charging to fill the battery to 80% within a dozen minutes.

So data-center HVDC and the 800V in EVs are two different applications, but they follow the same direction. Your next electric vehicle may use 800 volts earlier than your home’s outlets.