Here’s what’s on our plate today:

• 🧪 Why electricity may matter more than chips in the next AI race.

• 📰 Microsoft's Scout assistant, Gemini, plans your trips, and Trump's AI executive order.

• 🛠️ Three tools worth trying: Electricity Maps, Ember Climate Data, and Modal.

• 🗳️ Poll: Who's better positioned in the AI energy race?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Chips aren’t the ceiling: Running AI at scale is closer to operating a steel mill than a website. A single data center can draw as much electricity as 100k homes, and the largest facilities can consume up to 2M homes’ worth of electricity, making energy supply the real bottleneck.

• America’s grid gap: Two decades of flat electricity demand left U.S. utilities unprepared. Interconnection queues are backlogged, transformers are scarce, and local opposition has already blocked or delayed over $60B in planned data center investment.

• China’s overcorrection: Beijing’s East Data, West Computing program was rolled out quickly and widely, but many facilities operate at only 20-30% capacity due to latency and misaligned demand, leaving hundreds of completed sites without customers.

• The stakes: Both superpowers are stuck in mirror-image traps. Whoever figures out how to match infrastructure buildout with actual demand, not just projected demand, holds the real structural edge in AI.

Why electricity may matter more than chips in the next AI race

Over the past couple of years, most coverage of the contest between the United States and China over artificial intelligence has focused on semiconductors. This has made the entire conversation revolve around questions like who is building the fastest chips, who can buy or hoard the most of them, whether Washington’s export controls are slowing Beijing down, or just pushing it to improvise.

However, while these questions are real, they are no longer the whole picture, because what may end up mattering most over the next ten years does not come off a production line in Taiwan. It comes from a power plant, travels down a high-voltage line, and arrives at a warehouse full of computers that cannot do anything at all without an enormous, uninterrupted supply of it.

Running an AI model, whether you are training a new one or serving an existing one to millions of users, is closer to operating a steel mill than a website. A single data center can draw as much electricity as 100k homes, and the largest hyperscale facilities can use as much as 2M, according to the International Energy Agency. Every time the models get bigger, and the number of people using them grows, the appetite for power grows with them. A country that can deliver that power cheaply, reliably, and on a timeline the industry demands ends up with an edge that is far stickier than a lead in chip design, because it lives inside an entire energy system that took decades to build and cannot be copied in a quarter.

A grid built for a different era

As of 2026, the United States is walking into the AI race with a particular kind of debt. American electricity demand barely moved for 20 years, and now AI is pushing it upward at several times the rate anyone anticipated. Utilities make their bets on horizons measured in decades; they were not braced for demand that shows up in two or three years, and the result is a mismatch between what AI companies are asking for and what the system can physically hand them. The shortfall goes beyond the need for money or technology. It is the accumulated cost of two decades spent building almost nothing.

In practice, that looks like a queue in which, to plug a new data center into the grid, a developer files an interconnection request, which triggers a long sequence of capacity studies, engineering reviews, and eventually a physical connection, all of it designed for a slower world. As a result, even completed projects cannot be turned on. According to a Bloomberg report, two completed data centers in Silicon Valley sat idle because the transformers needed to power them were unavailable. Those transformers, the bulky custom-built units that manage how power moves across the grid, have seen their wait times climb steadily since 2023.

Even if the grid buildout can manage the queue, there is also local political opposition to data centers that has now blocked or delayed more than $60B in planned U.S. data center investment, much of it fueled by the fact that wholesale electricity near these clusters has climbed as much as 267% over five years, with the cost landing on local ratepayers.

A July 2025 executive order tried to fast-track federal permitting, but the parts that actually slow projects down, the interconnection queue and local land-use fights, mostly sit outside what an executive order can touch.

China’s coordinated bet

From across the Pacific, China appears to have solved the exact problem the U.S. is stuck on. Electricity is cheap and getting cheaper, and that cheap power has become the centerpiece of Beijing’s pitch.

NVIDIA's chief executive, Jensen Huang, told an audience in late 2025 that standing up an AI data center in the United States takes roughly three years from breaking ground, while in China, he said, they can put up a hospital in a weekend. None of this is cultural mystique. It is what 20 years of relentless grid investment buys you: a system limber enough to swallow a sudden new load instead of choking on it.

The clearest expression of the strategy is a program called “East Data, West Computing,” launched in 2022 and run jointly by four central agencies, among them the National Development and Reform Commission and the National Energy Administration. The idea is to push new data centers into China’s thinly populated western provinces, where land is cheap, the climate does some of the cooling for free, and vast wind and solar farms sit nearby. In 2025 alone, China added more than 430 gigawatts of wind and solar capacity, more than half of the world’s total that year. That is a real advantage, and a lasting one, over a rival that spent the same stretch adding next to nothing.

The limit of speed

The tempting conclusion is that China has already won the energy round. However, that is the surface reading, and the reporting points somewhere messier.

While China developed a speedy system, all that speed produced a different failure: China built faster than anyone could fill the buildings, and it built them in the wrong places.

Many of those western facilities ended up running at only 20 to 30% of capacity, in part because of latency, the small lag that creeps in when data has to travel hundreds of miles between where it is processed and where it is actually used. For the live, responsive AI services that make up much of the near-term business, that lag is a dealbreaker.

MIT Technology Review found that China announced more than 500 data center projects across 2023 and 2024, with at least 150 finished sites unable to find customers in a market that had moved on.

To counter this, in November 2025, Beijing began offering electricity discounts of up to 50% to data centers running domestic chips, billed as an industrial policy but read by many as a quiet admission that demand never materialized at the scale the buildout had assumed.

So the two superpowers have walked into mirror images of the same trap. America built too little for too long, and now it cannot connect new supply to demand quickly enough. China built fast and wide without lining up capacity with where it was needed, and is now holding a lot of expensive infrastructure it cannot fully switch on. One side is stuck behind its own institutions. The other is sitting on a stranded concrete slab.

The invisible race

For the past several years, the AI race has been framed as a contest over chips, models, and export controls. Those are the parts that are easy to see and easy to measure. What matters increasingly, however, is the infrastructure beneath them.

Power lines, substations, transformers, transmission networks, and permitting systems rarely attract attention. Yet they increasingly determine how much AI capacity can actually be deployed. The United States and China have run into opposite versions of the same problem. One struggles to quickly connect to new demand. The other built aggressively only to discover that some of its infrastructure sits far from where demand exists.

That is why the next phase of the AI race may be decided less by who builds the best model and more by who can align their infrastructure with the speed of technological progress. The headlines may remain focused on the chip lab. The outcome could be determined in a much less visible place.

Quick Bits, No Fluff

• Microsoft's Scout assistant: Microsoft launched Scout, an OpenClaw-inspired personal assistant, further advancing in the agentic AI race with a consumer-facing helper.

• Gemini plans your trips: Google's new Gemini Spark agent can now plan entire trips autonomously, handling research, booking logic, and itinerary planning in a single flow.

• Trump's AI executive order: A new Trump executive order on AI aims to reshape federal policy, signaling another shift in how Washington plans to govern the technology.

Thursday Poll

3 Things Worth Trying

• Electricity Maps: Live map of grid electricity and carbon intensity by region, a fascinating way to see where the cheap, clean power actually is.

• Ember Climate Data: Open dashboards on global wind, solar, and grid capacity, useful for understanding the buildout numbers behind the AI power race.

• Modal: Serverless GPU platform that lets you run AI workloads without managing infrastructure, a hands-on way to feel how compute gets provisioned on demand.

The Toolkit

• Assembly AI: Speech-to-text API that handles transcription, speaker detection, and audio intelligence for production apps. 

• Chroma: Open-source vector database built for AI apps, fast to set up and easy to scale for RAG and embeddings. 

• Continue: Open-source AI code assistant that plugs into VS Code and JetBrains with full control over models and context.

Rate This Edition