The Power Delivery Bottleneck in AI Data Centers: Where and How to Invest

Damnang · Damnang’s Substack · August 04, 2026 at 01:37 · ⏱ 5 min read  | Read on Substack ↗
Summary
The article argues that the AI data center power bottleneck is fundamentally an internal power-delivery problem: GPUs demand over 1kW at under 1V, which creates extremely high current and efficiency losses unless power conversion and regulation are improved. It frames MPWR, Vicor, Murata, and Samsung Electro-Mechanics as the companies closest to this bottleneck, supporting a positive but mostly thematic lens on power-delivery component makers. No explicit buy or sell recommendations are made in the provided text.
  • •A single GPU accelerator draws roughly 1kW, while a rack runs in the hundreds of kilowatts today and is heading toward 1MW; campuses are measured in gigawatts.
  • •A modern data center GPU runs at under one volt while pulling more than 1,000 watts, pushing current into thousands of amps versus around 200 amps for a typical home.
  • •Power losses scale with the square of current, so doubling current quadruples loss; this makes short, thick power-delivery paths critical.
  • •Transient load spikes cause voltage sag because distant power sources cannot respond instantly, forcing chips to throttle and underperform.
  • •OpenAI's Stargate originally committed to 10GW by 2029, but by April 2026 had already passed that and added more than 3GW in the prior 90 days.
  • •Meta has said it will build tens of gigawatts of AI data center capacity this decade, while US grid interconnection delays can take years.
Read time 5 min
Length 5,657 chars
Category finance
Ideas
Damnang Substack author, Damnang’s Substack
The article explicitly names MPWR as one of the four companies 'closest to' the AI power delivery bottleneck, while arguing that solving high-current, low-voltage delivery is essential to getting prom
The article explicitly names MPWR as one of the four companies 'closest to' the AI power delivery bottleneck, while arguing that solving high-current, low-voltage delivery is essential to getting promised GPU performance. Risk: The excerpt provides no product-level data or revenue estimates for MPWR's AI power delivery exposure.
MPWR
Damnang Substack author, Damnang’s Substack
Murata is named as one of the four companies closest to the power delivery bottleneck; its power inductors, MLCCs, and power modules are core components in DC-DC conversion and noise filtering for AI
Murata is named as one of the four companies closest to the power delivery bottleneck; its power inductors, MLCCs, and power modules are core components in DC-DC conversion and noise filtering for AI racks. Risk: Power delivery may be a small slice of Murata's overall electronics business; no specific revenue catalyst is given.
6981.T
Damnang Substack author, Damnang’s Substack
Vicor is named among the four companies closest to the power delivery bottleneck; its high-density modular power converters directly address the article's central problem of delivering >1kW to sub-1V
Vicor is named among the four companies closest to the power delivery bottleneck; its high-density modular power converters directly address the article's central problem of delivering >1kW to sub-1V GPUs without resistive loss or voltage sag. Risk: Customer concentration in a few AI platforms; the article does not quantify design-win economics.
VICR
Damnang Substack author, Damnang’s Substack
Samsung Electro-Mechanics is named among the four companies closest to the bottleneck; it supplies high-end substrates and power-related components relevant to efficient current delivery for AI accele
Samsung Electro-Mechanics is named among the four companies closest to the bottleneck; it supplies high-end substrates and power-related components relevant to efficient current delivery for AI accelerators. Risk: The article provides no product-specific details or comparative data, making the bullish read thematic rather than company-specific.
009150.KS
More from Damnang’s Substack
What Would Stop This Selloff?
Jul 29, 00:31
After CXMT’s Listing, Do the Memory Three Have Further to Fall?
Jul 27, 06:19
I Met with Semiconductor Experts in Korea
Jul 25, 13:11
The Open-Source AI Era: Will Hyperscalers Keep Buying Chips?
Jul 19, 18:57
Optical Investing, Act Two: Three Signals From AEHR’s Earnings
Jul 15, 06:48

This newsletter, published August 04, 2026, features Damnang discussing MPWR, 6981.T, VICR, 009150.KS. 4 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Damnang  · Tickers: MPWR, 6981.T, VICR, 009150.KS