The US Grid Can't Handle AI, and China Just Proved It Doesn't Need Us Either

The US Grid Can't Handle AI, and China Just Proved It Doesn't Need Us Either

Two stories landed on the same day this week. One from BloombergNEF, one from Bloomberg's China desk. They're about the same topic — AI infrastructure — and they arrive at opposite conclusions that somehow both scare the hell out of me as someone who runs actual servers for a living.

Let me break them down side by side, because together they tell you more about where this industry is headed than any single earnings call or analyst note ever could.

US Power Grid vs China's 1GW All-Domestic Data Center

New York / Beijing — July 21, 2026 — Start with the report that should've been front-page news everywhere but probably wasn't: BloombergNEF dropped its latest US power demand forecast on Tuesday, and the numbers are staggering.

The BloombergNEF Report That Should Scare Every Data Center Operator

Here's the headline: data centers in the US will consume 20% of the nation's electricity by 2035. That's four times today's 5.9%. Not "could consume." Will consume. BloombergNEF's new estimate is 83% higher than what the same consultancy predicted just seven months ago, in December. Let that sink in — they raised their own forecast by nearly double in under a year.

EPRI, the electrical industry nonprofit, has already more than doubled its 2024 estimate. S&P's forecast rose by more than a third between October and April. Every single forecaster in this space keeps raising their numbers because actual buildout is outpacing every model they built.

The details matter. BloombergNEF expects US data center capacity to hit nearly 200 gigawatts over the next decade. Nearly half of that will go to AI training and inference. By 2033, the US will host 64% of the world's AI chips by power demand.

Now here's where it gets real — the grid. In PJM, the interconnection that runs from Virginia to Illinois, 34% of all electricity will go to data centers. In ERCOT, covering most of Texas, 22%. PJM has already struggled so badly with connection requests that it paused accepting new applications for four years. Electricity prices in PJM are up 76% over the past year. One utility, American Electric Power, has threatened to pull out of the interconnection entirely.

Data centers represented 38% of all charges in PJM's most recent capacity auction. Think about what that means when the auction clears at prices driven by data center demand: every other electricity consumer in the region — hospitals, schools, factories, homes — is paying rates set by AI compute.

And that's just the US. BloombergNEF says that by 2033, if AI adoption continues along an aggressive trajectory, data centers will create 1,935 terawatt-hours of new electricity demand globally. That's nearly as much as the entire country of India.

While We're Fighting Over Watts, China Just Flipped the Switch on 1 Gigawatt of Domestic Silicon

Meanwhile, on the other side of the Pacific, Z.AI — the Chinese lab formerly known as Zhipu, the team behind the GLM model family — just completed a 1-gigawatt data center that runs entirely on Chinese-made chips. No Nvidia. No AMD. No Western silicon at all.

This is a direct product of US export controls. The Biden administration's chip sanctions were designed to choke off China's access to advanced Nvidia GPUs. The idea was that without H100s and B200s, Chinese AI development would stall. Instead, the controls created exactly the kind of pressure that forces a country to build its own supply chain.

Z.AI's data center is that supply chain in action. One gigawatt is not small — that's a hyperscale facility by any standard. And it's running on Chinese chips, likely Huawei Ascend 910-series or similar domestic alternatives. The facility is already partially operational, per Bloomberg's sources.

Here's what this tells us: China has now demonstrated it can build and operate a 1GW-class AI data center without a single Nvidia chip inside. The performance per watt won't match Nvidia's best — Huawei's Ascend line is probably 2-3 generations behind — but the scale is real. When you can throw a gigawatt of domestic silicon at a problem, you don't need best-in-class per-chip performance. You just brute-force it with volume.

And they're not stopping at one. Z.AI is reportedly planning additional facilities. Other Chinese AI labs — Moonshot, Baidu, Alibaba — are watching and learning from the Z.AI playbook.

The Conspicuous Absence — Same Day, Two Different Worlds

Notice what neither story talks about. The BloombergNEF report doesn't mention export controls or Chinese competition. The Z.AI story doesn't mention US power constraints. These two narratives exist in separate universes, but they're describing the same phenomenon: AI infrastructure is the most capital-intensive buildout in the history of computing, and both superpowers are hitting different walls.

The US has the chips but can't power them fast enough. The grid can't keep up, PJM is backlogged for years, prices are spiking, communities are revolting. We have all the Nvidia silicon in the world and nowhere to plug it in.

China has the power — or at least the ability to build gigawatt-scale facilities — but has been cut off from the best chips. So they built their own. They're behind on performance per watt but ahead on deployment speed because they don't have to deal with the same permitting, community opposition, or grid interconnection delays.

Both sides have a bottleneck. It's just different bottlenecks.

What These Two Stories Together Tell Us About the AI Infrastructure Thesis

The market has been pricing AI infrastructure as a one-sided bet: hyperscalers build, everyone profits. But these two stories together reveal a much more complicated picture.

First, the US power grid is the single biggest constraint on AI growth in the West, and nobody's talking about it enough. The BloombergNEF numbers are not a prediction — they're a warning. If the grid can't deliver 20% of US electricity to data centers by 2035, then AI growth in the US caps out well below current market expectations. That's not a maybe. That's arithmetic.

Second, export controls are forcing China to build an independent AI hardware ecosystem whether we like it or not. Z.AI's 1GW facility proves the strategy works at scale. Yes, Chinese chips are behind. But they're catching up faster than the制裁 regime anticipated, because necessity is the mother of invention.

Third, the AI infrastructure buildout is now a geopolitical contest, not just a market one. The country that solves its bottleneck first — whether that's the US fixing its grid or China closing the chip performance gap — gets first-mover advantage in the next wave of AI development.

What This Actually Means for Independent Hosting Providers

First — US grid constraints are your competitive moat. Every independent hosting provider reading this should understand that the grid interconnection crisis is the single best argument for colocation and dedicated server hosting that isn't in a hyperscaler data center. When PJM can't connect new 100MW facilities for four years, your existing 5MW or 10MW deployment in a well-connected facility is suddenly worth more than any new hyperscaler build.

Second — Watch equipment lead times like a hawk. The Z.AI story tells us that Chinese chip manufacturing capacity is being absorbed by domestic AI infrastructure. That means less Chinese silicon for the global market, which means higher prices for DRAM, SSDs, and networking gear that uses Chinese fabs. If you're planning a hardware refresh in the next 12 months, order now.

Third — The power price picture is about to get worse before it gets better. With PJM prices up 76% year-over-year and data centers representing 38% of capacity auction charges, electricity costs for existing facilities in PJM territories are going to keep climbing. If you colocate in Virginia, Northern Virginia, or any PJM-connected region, lock in your power rates for as long as your contract allows.

Fourth — Diversify your infrastructure geography. The BloombergNEF report confirms that data center demand is concentrating in PJM and ERCOT — exactly the regions where grid constraints are worst. If you're running hosting out of a single facility in Ashburn or Dallas, that's a concentration risk you need to hedge. Look at secondary markets: the Midwest, the Mountain West, even Canada. The grid pressure is going to push data center development into regions with surplus power, and independent providers who get there first will have the advantage.

The Structural Reality — Both Sides of the Pacific Are Betting Big, and Both Have Problems

The BloombergNEF report and the Z.AI story are not competing narratives. They're two sides of the same coin. The US has the silicon advantage but is losing the power race. China has the power and deployment speed but is catching up on chips. Neither side has a clean path to AI dominance, and both are spending insane amounts of capital trying to solve their respective bottlenecks.

For the market as a whole, this means the AI infrastructure boom is not a straight line up. It's going to hit speed bumps — some physical (power), some geopolitical (chips), some regulatory (moratoriums, community opposition). The idea that we just build our way out of these constraints with enough capex is not supported by the data. You can't build a gigawatt of compute without a gigawatt of available power and a functioning supply chain.

And the market is starting to price this risk in. The same day these two stories broke, you saw it in the bond market — the divergence I wrote about in the Oracle credit analysis. Lenders are getting selective. Junk-rated data center debt at 10%+ spreads tells you the market sees risk that the earnings calls aren't mentioning.

The Bottom Line

Here's where I land after reading both stories on the same day. The AI infrastructure buildout is happening at a scale and speed that the existing power grid and supply chain were never designed to support. That doesn't mean it stops — it means it reshapes around constraints. New facilities go where power is available, not where demand is highest. Chips get sourced from whoever can make them, not whoever makes the best ones. And hosting providers who understand these constraints and plan around them will be the ones still standing when the music stops.

The US doesn't have infinite power, and China doesn't have infinite chips. Both problems are solvable, but not overnight. And the gap between "building the stuff" and "having the stuff actually work at scale" is where the real money — and the real risk — lives.

Plan for the world where power is the bottleneck and chips are the leverage. Because according to the data, that's the world we're already living in.

-- Allan Ali, Founder