What China Understands About AI That the US Doesn’t

Kind: captions Language: en The AI revolution is supposed to be happening in the cloud, a battle fought with code and software. It's not. Right now, Microsoft and other tech giants are quietly locking up America's nuclear energy capacity to train the next generation of AI models. They're pulling in massive amounts of power off of the public grid, and as demand explodes, regular consumers are competing for what's left. The result? A private AI energy race is emerging behind the scenes, and you are picking up the bill. Chapter 1: The $500 AI tax. PJM runs the power grid for 67 million Americans across 13 states, and once a year it holds a massive auction to lock in electricity 3 years before anyone uses it. In June 2025, that auction cleared almost $270 per megawatt day. The year before, just $29. In only 12 months, prices exploded 833%. That single auction sent total household costs surging from $2.2 billion to $14.7 billion. And then, [music] it got worse. The next auction, covering June 2026 onward, smashed [music] straight into the market price cap of almost $329 per megawatt day. And without that cap, the real price would have hit $389. [music] The cap had already saved consumers $13 billion. In December 2025, internal filings show PJM's next auction prices hit an even higher cap, but that wasn't the problem. PJM was short of its own safety target. Even after prices exploded, [music] the grid operator still couldn't buy enough power to guarantee its reserve margin. All across the country, prices are beginning to rise. The average family inside the PJM footprint faces [music] an estimated $70 more per month by 2028. The total bill across the next 8 years lands somewhere between $100 billion and $163 billion. The reason for all [music] of this is simple: demand. Data centers alone drove roughly 63% of [music] the entire price surge. The explosion in demand from AI server farms is happening [music] so fast that the grid is struggling to keep up. The cheapest, most reliable power in the country is being pulled out of the public pool and sent into corporate data halls. Ordinary families are left covering the rising costs of keeping [music] the rest of the system stable, and it's already on your bill. Chapter two, the great cannibalization. Most Americans still believe their country is leading a clean energy boom, but the real story is more complicated. In September 2024, Constellation Energy signed a 20-year power purchase agreement with Microsoft to restart the Three Mile Island Unit 1 in Pennsylvania. The plant sits just a few hundred feet from the cooling towers of Unit 2, [music] the reactor that suffered America's worst nuclear accident in 1979. Unit 1 ran for decades until it shut down in 2019, not for safety concerns, but because it stopped making enough money to justify staying online. The plant is set to come online in 2027, a full year ahead of the original 2028 target. The Department of Energy closed a $1 billion loan to lower borrowing costs, while Constellation itself put $1.6 billion into the restart. The plant's operating license was extended to at least 2054. [music] Microsoft gets the reactor's full output for the next two decades, enough electricity to power roughly 800,000 [music] average American homes for a year. The power itself still flows into the shared grid, but the [music] clean energy credit is claimed by Microsoft. So, while nothing physically changes in how the electrons move, the output is effectively reserved on paper for AI workloads all the way through 2047. The deal also delivers $3.6 billion in state and federal taxes. 3,400 jobs will be created across Pennsylvania, and $16 billion added to the state's economy. Washington has even called it the poster child of its energy and AI strategy. Politicians point to the investment and the jobs as proof. What they ignore is the community deprived of power. Shared national power is becoming private reserves and the households left behind end up paying more just to keep the grid running. Chapter 3, the five nines reality. People assume the answer to AI's power hunger is simply more renewables. Build enough solar in Nevada, add enough wind in Texas, throw in some big old batteries and the problem solves itself. Except it doesn't. AI data centers run on a reliable standard called five nines, 99.999% uptime or just about five minutes of downtime per year. That kind of precision exists for a reason. A single frontier model training run can cost tens of millions of dollars and run for weeks. One brief power flicker can corrupt the entire run forcing engineers to restart from the last saved to checkpoint. And that's where the grid problem becomes unavoidable. Solar produces nothing at night and wind can disappear for days at a time. Bridging those gaps to meet five nines reliability would require battery storage at a scale no country has ever built. The cooling problem is just as brutal. A single rack of Nvidia servers pulls about eight to 10 times more power compared to a standard server rack just a few years ago. All of this creates immense heat and that has got to go somewhere. New AI data centers now run on direct liquid cooling pumping water across the chips at flow rates that would supply small American towns. >> [music] >> And that's why hyperscalers are picking nuclear reactors. Nothing else delivers carbon-free around-the-clock power at the size they need. The patchy stuff gets left for the public grid to absorb along with every headache that comes with balancing it. A Reuters review of letters filed with PJM tells the rest of the story. While the cleanest power gets locked into corporate contracts, the dirtiest plants are getting a second life. About 23 coal, gas, and oil plants in the PJM region were scheduled to retire for 2025 or just after. Since January 2025, 13 have had their shutdowns delayed or canceled outright. Brandon Shores, a coal plant in Maryland that was the state's top climate polluter in 2022, will now run until May 2029. It's owned by the same Talen Energy that sells nuclear power to Amazon. The public is being told that green energy is the future, but the physical system is already splitting into two tiers. >> [music] >> Firm nuclear capacity secured by corporations and intermittent renewables carrying the rest of the load. [music] And even with firm power being locked up, the sheer scale of the new demand is now straining something more basic, the grid [music] itself. Chapter 4, Data Center Alley. Loudoun County, Virginia sits about 40 mi west of Washington D.C. A stretch of rolling farmland has given way to suburbs and endless warehouse-scale buildings. The region now handles roughly 25% of all North American internet traffic. The locals call it Data Center Alley, and at its center is zip code 20147, the closest thing the internet has to a physical brain. [music] The country now holds around 200 data center buildings with another 117 in development. Each new project adds more pressure on the grid. Most new campus proposals request enough electricity to run an entire country, roughly equal to 150,000 American households. One campus, [music] one company, sometimes even just one customer. To carry the load, Dominion Energy is building massive new transmission lines meant to move power into Loudoun County's data center belt. One of the biggest [music] proposals, the Golden to Mars line, would cut a 22-mi path through neighborhoods, schools, and farmland, all carried on steel towers rising hundreds of feet into the skyline. And the spending is scaling with it. Dominion's investment plan for 2025 through 2029 has climbed to $50.1 billion. But the cost doesn't stay local. It gets spread across the entire regional grid, 13 states in total. So, households far from Virginia will help pay for the infrastructure built to serve data centers they may never interact with. Virginia's own joint legislative audit and review commission ran the numbers. They said the state would need a brand new 1.5 gigawatt natural gas plant every 2 years for 15 straight years just to feed [music] the data center demand. These costs are buried in the utility bills and wholesale market fees that most customers never directly see, and the pressure isn't just coming from the wires. [music] It's also coming from what's plugged into them. Chapter 5, the efficiency trap. Nvidia's Blackwell platform is the current top-of-the-line AI chip rolling out across data centers through 2025 and 2026. >> [music] >> At the system level, it uses up to 25 times less energy per inference token than the previous generation. That should solve the whole problem, but it doesn't. Global data center electricity demand is still set to more than double by 2030. The International Energy Agency's base case puts global use at roughly 945 terawatt hours by then, up from 415 terawatt hours in 2024, slightly more than the entire annual electricity use of Japan. By 2035, the same forecast climbs to 1,200 terawatt hours. In the United [music] States alone, data centers could consume more electricity than all heavy industry combined, steel, cement, [music] chemicals, and aluminum put together. A 2024 Lawrence Berkeley National Laboratory report says US data centers made up 4.4% of all electricity use in 2023. [music] Within just a few years, that could nearly double or even triple, depending on how fast demand accelerates. >> [music] >> In raw terms, electricity use could jump from 176 terawatt hours to as much as 580 terawatt hours in 5 years. So, [music] how do both things happen at once? Efficiency improving while total use still explodes. Well, that question isn't new. A 19th century English economist named William Stanley Jevons figured it out long before anyone thought about AI. Watching steam engines get better at burning coal, he noticed something strange. England's total coal use kept going up, not down. Cheaper energy unlocked new uses and the explosion [music] of new applications buried whatever savings the efficiency had produced. That same trap has now closed on AI. When the cost [music] of a single query drops from cents to fractions of a cent, every product manager finds a new opportunity. Chatbots get embedded in customer service. AI overlays appear on every single search bar. Code helpers run quietly in the background of every developer's day. >> [music] >> Automated agents start handling the work that people used to do by hand. Each one is cheap on its own, but multiplied across the global economy, they [music] eat every watt of savings and demand orders of magnitude more on top. >> [music] >> Cars on American roads tell the same story. Engines got much more fuel efficient over the past 40 years. Yet, total US gasoline use barely budged. Americans simply drove farther and bought bigger trucks. [music] Cheaper miles meant more miles. The IEA expects about half of all new data center demand [music] through 2035 to be met by renewables with natural gas covering another large chunk of the gap. That just [music] shuffles how the energy is supplied. It doesn't solve the problem. And no chip breakthrough on the horizon changes it. If anything, the crisis accelerates because technology keeps getting better, not worse. America's trying to solve this by pushing more and more power through its already stressed wires to where the data centers are built. But, China is taking a different approach to the same problem. Chapter 6, China's East to West Master Plan. In February 2022, the Chinese state launched something it calls the East Data West Computing Initiative. It's a national restructuring of where computing actually happens. A coordinated effort to move data centers away from dense coastal cities and into inland regions where land is cheaper and power is more abundant. It's being executed at a speed and scale no Western government has attempted in our lifetime. It splits the country into eight national computing hubs and 10 data center clusters. The Eastern clusters sit near Beijing, Shanghai, and the greater Bay Area handling the fast response work for the cities where most of the economy lives. The Western clusters are placed on purpose in the empty interior across Inner Mongolia, Ningxia, Gansu, and the edges of the Gobi Desert. As of late 2024, China's National Data Bureau reported that roughly $6.1 billion had already flowed into the eight hubs. Why build out there? Well, that's because that's where China's huge hydro, wind, and solar power sits unused. The Western provinces produce far more electricity than their thinly spread populations can use. The surplus had historically gone [music] to waste because sending it was wasteful and expensive. The state's fix wasn't to build more wires. It was to move the demand. Computing tasks that don't need instant millisecond responses are now being pushed west along dedicated high-capacity fiber networks. And for most workloads, it works. Latency between Eastern cities and Western data hubs already meets China's 20-millisecond target for routine computing. The result is a physical reshaping of the internet itself. Data halls are built directly on top of the cheapest, [music] most abundant electricity in the country. In places like Inner Mongolia's Wulanchabu cluster, that strategy has already attracted around 10 gigawatts of planned data center projects [music] from companies including 21Vianet, Huawei, and ByteDance. New plants inside the national hubs must hit a power rating of 1.2 or lower, with some new builds hitting 1.04 compared to [music] a global average closer to 1 and 1/2. China's overall data center power use is forecast to grow 19% per year between 2025 [music] and 2030, faster than any other major economy. If you can't efficiently move the power to the computers, [music] you move the computers to the power. While the United States lets corporations privatize nuclear plants one reactor at a time, [music] China treats energy and computing as a single state asset. The geography itself wipes out the transmission choke points crippling the American system, [music] and there is no Loudoun County problem in China. America is breaking its power system into privately controlled zones built around individual corporate demand centers rather than a shared network. China is doing the opposite. It's reshaping geography itself so supply and demand sits closer together. American response to all of this is private, expensive, and almost invisible. Chapter 7, [music] the SMR Mirage. Through 2024 and 2025, the biggest names in American tech started pouring money into a new kind of nuclear plant called small modular reactors or SMRs. They're smaller than traditional reactors, built in factories, shipped in pieces, and assembled on site. Each unit typically produces between 50 and 300 MW. They can be scaled by adding modules instead of betting everything on a single multi-billion dollar mega project. Amazon led the push with a $500 million dollar investment into X Energy, a company developing high-temperature gas-cooled reactors. The plan is to unlock more than 5 GW of new nuclear capacity by 2039, enough to power roughly 3.8 million homes. The first project, the Cascade Advanced Energy Facility in Washington State, is designed to stack 12 reactor modules into three larger blocks totaling 960 MW on a footprint the size of a few city blocks. Google followed with a long-term deal with Kairos Power to bring a new generation of nuclear plants online in the coming decade. The first project, Hermes 2 in Tennessee, will feed electricity into the regional grid, but its purpose is effectively pre-committed to Google's data centers. It's also the first time a US utility has struck a deal tied directly to an advanced reactor design instead of a traditional plant. Meta followed that same path opening negotiations for multiple gigawatts of nuclear capacity. Equinix, which runs a large share of the world's data center infrastructure, signed on with Oklo. Prometheus Hyperscale came next, and every single one of these reactors is being designed to plug straight into or feed those campuses through behind-the-meter setups. That's when power moves from the generator straight to the customer through private lines, and the shared grid never gets touched. Most of the tariffs and rules that apply to grid-connected power don't apply either. And these reactors are private utilities for corporate use. The public is cut out at the blueprint stage. For most of the 20th century, American nuclear power came through national labs, federally funded research, and rate-regulated utilities that answer to state commissions. The benefits flowed, however imperfectly, to ordinary families. The biggest nuclear push in two generations is now being paid for by three corporations, >> [music] >> built for their exclusive use, and engineered to stay outside of the public system. A tech company isn't just a tech company anymore. Chapter 8, the sovereign tech utility. Microsoft, Amazon, and Google have stopped being software companies that happen to run cloud divisions on the side. Instead, they've become the largest unregulated energy buyers on the planet. Amazon's deal at the Talen Energy Susquehanna nuclear plant in Pennsylvania shows how messy this gets when private wires meet public rules. In March 2024, Amazon Web Services bought a data center campus next door to the Susquehanna plant for $650 million. They plan to increase the on-site load, and it'll be fed straight from the reactor through dedicated lines. FERC has already prepared a private power connection between Susquehanna and a nearby data center in 2015, allowing electricity to flow directly into a single customer. When that connection was later expanded, utilities pushed back hard. Their argument wasn't technical, it was about who ends up paying for the infrastructure when a single large user reshapes the system. The federal regulator ultimately blocked the expansion, siding with the concerns that the setup would shift hundreds of millions of dollars in grid costs [music] onto everyone else in the region. Even after that decision, the fight didn't [music] end. It moved into the courts, while the original private connection continues operating in the background. The fight cracked the rulebook wide open. The same arrangement is being copied across the country, regardless. Each deal looks reasonable alone, but stacked together, none of them answer to voters. The old utility system was built around everyone getting power. Utilities had to answer to regulators to keep the grid running for entire communities, not just their biggest customers. The new power players operate by a different rulebook. Their job is to win the AI race, and the energy [music] system is starting to reorganize around that mission. Almost nobody voted for this. There was no congressional debate or national plan, just a string of private [music] deals announced through company press releases. Pennsylvania Governor Josh Shapiro wrote letters [music] to PJM helping fast-track the interconnection. Every level of government has greased the [music] wheels. None of them has asked who the power's for, or who gets to use it, or who pays when the bill comes due. The shared system is being carved [music] into corporate territories, and most Americans don't yet realize that the shift is already happened. Chapter nine, the reckoning. China has grasped something America still hasn't accepted. Two systems are now competing for the future of AI, and they aren't playing the same game. On one side is China's system, a coordinated national build-out built around eight hubs and 10 massive computing clusters, where data centers are placed directly next to the country's cheapest, [music] most abundant power. The entire thing moves as one machine. Energy policy, transmission planning, latency t