Tech BusinessChips & Semiconductors
AI accelerator power delivery moves closer to chips
Semiconductor Engineering reports that AI chips are driving a shift to vertical power delivery, placing voltage regulators nearer to processors to cut losses and improve control.

Key takeaways
- AI accelerators are pushing current demand to 1,000 amps and beyond at low voltages.
- Designers are moving voltage regulators closer to processors through vertical and backside power delivery.
- OSATs, foundries, and substrate suppliers must co-design power delivery, packaging, and thermal performance.
- Power loss and voltage droop can limit frequency and efficiency in data centre chips.
In this article
Power delivery becomes a design bottleneck
According to Semiconductor Engineering, AI accelerator power delivery is becoming a major design challenge for data centre chips. The report, published on 19 November 2026, says AI accelerators and graphics processors are drawing large currents at low voltages. This is forcing designers to move voltage regulation closer to the processor. The details come from one report and are not independently confirmed.
The problem is not only more power. It is also where that power is delivered. A power delivery network, or PDN, carries electricity from the board to the chip. As current rises, resistance causes voltage loss, known as IR drop. That loss can reduce performance and make chips less efficient.
The report gives an example. A 600W GPU running near 0.7 V can require 1,000 amps or more. At that level, even small resistance in the PDN can create significant voltage droop.
Voltage droop is a sudden fall in supply voltage. Multi-chiplet designs add to the pressure. They increase current density and signal congestion inside the package.
Vertical power delivery and backside designs
One response is vertical power delivery. Designers are moving voltage regulator modules away from the board edge. They place them under the board, inside the package, or inside the silicon interposer.
An interposer is a silicon layer used inside advanced packages. Placing regulation nearer the processor shortens the path for electricity and reduces loss.
Backside power delivery is another approach. It places power connections on the back of the wafer, rather than the front. This approach started about 10 years ago with backside power delivery on the wafer backside.
Semiconductor Engineering says Intel Foundry's 18A was first to use backside wafer processing in advanced high-volume production. Samsung and TSMC plan to use backside power delivery in future chip generations, starting next year. The report also names ASE, Marvell and Empower as companies working on closer voltage regulation.
Compact inductors and embedded capacitors can sit inside packages or interposers. These parts help store and smooth electrical energy. They allow DC-DC converters, which change voltage, to sit closer to the load.
Packaging and thermal teams must work together

These changes affect many companies. Semiconductor Engineering says power delivery is now a system-level design challenge. It involves the power delivery network, interposer, power management chips, integrated voltage regulators, capacitors, memory, and interconnects.
Integrated voltage regulators control voltage close to the chip. OSATs, which work on package-level design and validation, are being brought in earlier. Alex Kizeev of Amkor says OSATs work at the meeting point of chip, substrate, interposer, passive parts, and board.
His team models the package power network and combines it with chip power data from the customer. Good outcomes need a power budget split between chip, package, and board. Thermal design is also affected because backside power delivery changes how heat escapes.
Why it matters for AI performance
The shift is driven by efficiency. Lipen Yuan of TSMC spoke at a 2024 IEDM short course. He said energy efficiency will be the top performance measure for AI platforms.
Faster voltage control can help chips respond to sudden demand changes. Nir Sever of proteanTecs says faster control loops can cut response time from milliseconds to microseconds.
The changes also affect performance limits. Synopsys says AI accelerators are nearing currents of about 1,000 amps. At such currents, small resistance can reduce voltage margin.
That can affect maximum frequency and minimum operating voltage. CP Hung of ASE says AI systems are putting heavy pressure on board-to-chip power delivery. He says the last stage of voltage regulation matters more.
For readers, the main point is that AI chip design is no longer only about transistors. Power delivery, packaging, and cooling are now central to performance. Companies that can co-design these areas may gain an advantage in data centre AI systems.
Sources
- AI Accelerators Drive Changes In Power Delivery NetworksSemiconductor Engineering
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.


