惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

大猫的无限游戏
大猫的无限游戏
阮一峰的网络日志
阮一峰的网络日志
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
WordPress大学
WordPress大学
博客园 - 叶小钗
Hugging Face - Blog
Hugging Face - Blog
博客园 - 司徒正美
Last Week in AI
Last Week in AI
罗磊的独立博客
量子位
Jina AI
Jina AI
T
Tailwind CSS Blog
Apple Machine Learning Research
Apple Machine Learning Research
IT之家
IT之家
美团技术团队
雷峰网
雷峰网
爱范儿
爱范儿
S
SegmentFault 最新的问题
小众软件
小众软件
月光博客
月光博客
酷 壳 – CoolShell
酷 壳 – CoolShell
人人都是产品经理
人人都是产品经理
The Cloudflare Blog
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报

IEEE Spectrum

Why Some Coders Now Reach for GLM 5.2 Before Frontier AI Models New Soft Exoskeleton Outperforms Most Hip Assist Devices We’re Squandering LEDs’ Potential to Save Our Night Skies Balcony Solar Is Sneaking Onto Grids Before the Rules Are Ready Cortisol Could Be the Next Frontier for Wearables Million P Bit Machine Pushes Probabilistic Computing to New Scale Would You Let This Humanoid Robot Do Your Laparoscopic Surgery? How a Spinning Drone Exploits Your Eyes to Become Nearly Invisible Why Indonesia’s Fisheries Future Hinges On Data Integrity and Trust Stable Jobs Can Hide the Riskiest Move In Your Tech Career Inside ELIZA’s Source Code and Its Multiple Personalities Tiny Puerto Rican Island Tests Hydrogen to Slash Sky High Power Bills AI Turns DNA Into Tiny Dogs and Mona Lisa Nanostructures How Darth Vader Taught Me Card Counting and AI Security Got Weird The Memory in Your Thumb Drive Could Fix AI's Big Problem The AI Arms Race in Technical Interviews Is Escalating Inside Nokia’s Race to Catch the iPhone and Android Wave Quantum Sensor Sniffs Out Radio Signals in 3D Two New Wheelchairs Reveal What “Smart” Really Means Today Video Friday: A World Cup for Robots Japan Pulls Off One of the Closest Asteroid Flybys Ever How Cheap Ground Robots Are Rewriting Frontline Warfare in Ukraine Nvidia’s NVLink Fusion Quietly Pushes Optics Inside the Rack Large Tabular Models Excel Where LLMs Fail Are Battery PoweredTrailers the Shortcut to Cleaner Long Haul Freight? The Hidden Overthinking Flaw That Could Drag AI Services Down Stacking Chips Sideways Gives AI More Memory There Independent Labs Crack Google Brain Inspired Camera Sensor Learns to See and Gently Forget
Inside the Race to Tame AI’s Wild Power Swings
https://www.facebook.com/48576411181 · 2026-07-16 · via IEEE Spectrum

Data center developers and hyperscalers are experimenting with 800 volt direct current (800 VDC) inside their largest data centers. It enables them to pack more compute power in much smaller spaces to run the latest AI applications. 800 VDC also reduces waste heat from electric wires, lowers cable weight, and eliminates equipment needed in multiple AC to DC to AC conversions.

The Large Load Working Group (LLWG) of the Electric Reliability Council of Texas (ERCOT) is evaluating how to accommodate several 800 VDC facility designs. Data center power infrastructure provider Dimaag, for example, presented in April a way to protect the grid from AI’s massive load variations, which can switch from almost nothing to hundreds of megawatts and back again several times per second. DiMaag’s system also enables data centers to stay connected to the grid, or “ride through,” during short drops in voltage. This ability is known as Low Voltage Ride Through (LVRT).

“Grid infrastructure wasn’t built for the large load oscillations from AI data centers and large load disconnections due to voltage dips,” says Sadha Kameswaran, Vice President of Business Development at Dimaag.ai.

Dimaag’s solution comprises batteries and control software that work with devices that switch electricity from AC to DC called rectifiers. This system, located between the grid and an AI data center, isolates DC load fluctuations from the grid and delivers LVRT support that can respond to demand shifts in real time at scale. Further, it lowers the number of AC to DC conversions required. Hyperscale data center developers are now evaluating this approach, Kameswaran says.

A power system study conducted by Electric Power Engineers (EPE) on behalf of Dimaag evaluated the technology in a simulated data center environment with regard to load smoothing, LVRT compliance, provision of backup power, and demand response. EPE concluded that this approach was an “effective solution for addressing the power quality and stability challenges inherent in modern data center operations.” The Dimaag module was found to be effective in buffering extreme DC-side power fluctuations and delivering LVRT response.

“Voltage ride through is key to maintaining balance between power generation and load,” said Woody Rickerson, COO of ERCOT during a panel at the Data Center World conference.

Alternative Power Electronics to Protect Grids, Consumers

Companies are developing other ways of helping large loads integrate into the grid. The National Laboratory of the Rockies (NLR), for example, has tested ON.energy’s AI Uninterruptible Power Supply against ERCOT’s large electronic load (LEL) ride-through requirements. Another alternative from Ramboll acts as a shock absorber for the grid, helping utilities and grid operators to add more renewable energy sources while lowering risk, preventing outages, raising resilience, and lowering development costs.

One of Ramboll’s innovations is to address reactive power, a component of electricity that does not perform useful work but is essential in maintaining electrical and magnetic fields and controlling voltage levels. Ramboll have developed a Static Synchronous Compensator (STATCOM) that adds more controls than traditional STATCOMs that provide damping capabilities to detect and neutralize any voltage oscillations by absorbing energy from the oscillations and converting it into useable power.

This work to address data center load variability comes at a time of mounting public and governmental pressure. The Governor of Texas just ordered the Public Utilities Commission (PUC) and ERCOT to safeguard residential and small business ratepayers from added costs due to the expansion of AI data centers and to require that data centers fund any new electric infrastructure their operations require.

That was followed by the PUC approving ERCOT’s proposal on how to connect large electricity users such as AI data centers while protecting the reliability of the grid via provisions for LVRT and load smoothing. As well as incurring the costs for additional substations, transmission, and power electronics equipment for their new facilities, large load customers are encouraged to agree to curtail power usage at times of local transmission constraints. This comes at a time when ERCOT has 438 GW of large load interconnection requests, almost all from AI data centers.

“Texas is experiencing an energy transformation unlike anything we have seen before,” said ERCOT President and CEO Pablo Vegas in a press release. “This new process represents a fundamental shift in how ERCOT manages the significant growth of large load interconnection, providing a structured, transparent path forward that protects reliability for Texans while supporting the state’s continued economic growth.”

Nvidia, AI hyperscale developers, and their supply chain partners are pressing ahead with their plans for 800 VDC data centers. NVIDIA envisions that 800 VDC facilities will begin appearing in 2027 to support its forthcoming GPU design. However, these organizations are moving at a pace that is far in advance of the power industry.

ERCOT is known as one of the most progressive grid operators in the nation. Texas has more solar energy, more wind power, and more battery energy storage systems (BESS) than any other state, but it is far from ready for 800 VDC. The LLWG continues to evaluate 800 VDC.

For now, ERCOT continues to review a variety of 800 VDC but no decisions are imminent on the preferred approach.

“There are a lot of solutions out there if you are listening for them as help can come from many areas,” said Rickerson.