The demand for carbon-based energy is yielding to the demand for silicon-based computing. Silicon not only requires GPUs, but its appetite for electricity has already surpassed that of carbon-based needs. Judging from the latest news over the past two weeks, "electricity" is flooding into the headlines of Silicon Valley's AI industry with unprecedented density. On May 11, Blackstone invested $1 billion in power generator VoltaGrid, while on the same day, Nvidia entered into a 5-year equity option agreement worth up to $2.1 billion with AI data center operator IREN. On May 6, Hut 8, a company transformed from a Bitcoin miner, signed a 15-year, 352MW take-or-pay lease in Beacon Point, Texas, valued at $9.8 billion, causing its stock price to surge nearly 30% that day. On April 20, Anthropic expanded its collaboration with Amazon: Anthropic committed to investing over $100 billion in AWS technology over the next decade and secured up to 5GW of new computing capacity for training and running Claude. 5GW is roughly equivalent to the installed capacity of several large nuclear power units. The most noteworthy aspect of this series of transactions is not the price, but the term used in Amazon's announcement—"binding constraint." This indicates that the hardest bottleneck for AI scaling is not chips, talent, or algorithms, but electricity. This aligns with the trend recently observed in California: the grid crisis faced by AI data centers has ignited a new wave of power plant construction. The competition has shifted from whose model is smarter to who controls the energy. On April 30, the North Dakota Public Service Commission unanimously approved a $110 million power project: 2 miles of transmission lines, a substation, and 280MW of capacity—all funded by AI data center developer Applied Digital, which then handed it over to the local power cooperative, Minnkota Power, for operation. An AI company funding its own power infrastructure and transferring it to a utility—this is not an investment, but an admission ticket. Construction began in April, with the line expected to be operational by September. The timeline for computing power is even tighter than that for electricity. This is not an isolated case. According to a TechCrunch report on April 27, data center demand has driven up the construction costs of natural gas power plants in the U.S. by 66% within two years, extending project timelines by 23%. The International Energy Agency's global energy assessment released in the same month shows that a full half of the new electricity demand in the U.S. by 2025 will come from data centers—a single industry consuming half of a superpower's new electricity demand. A month ago, Sam Altman admitted at the BlackRock Infrastructure Summit in New York: "Anything at this scale, it's just like so much stuff goes wrong." The example he cited was not an algorithm crash or a chip shortage, but his data center campus in Abilene, Texas, losing power after the state government cut off its electricity supply to prioritize residents during extreme weather. This marks the first time the world's most aggressive AI expander has publicly acknowledged the immense challenges of data centers. In the same week, according to Fortune, OpenAI abandoned an expansion plan in Abilene, while Crusoe and Microsoft took over an adjacent new campus with a total capacity of 2.1GW—one company exited, another took over, and what was transferred was not an algorithm, but electricity. For Chinese entrepreneurs and investors, there are at least two key takeaways. A prominent wealth management operator in California stated: "Based on our research, the most profitable wealth management product this year is likely to be AI data centers, offering returns of over 10% to ordinary investors." Meanwhile, a representative from Zhejiang Jinqun Technology, which is securing hundreds of millions in AI data center orders in Japan, noted: "We are also considering liquid cooling solutions in West Lake and coastal areas of Zhejiang. We'll observe how Silicon Valley approaches it first, then learn from them." Technology Hitting a Wall: Not a Lack of Electricity, but Electricity Failing to Reach the Racks It is reported that nearly half of the U.S. data centers scheduled to open in 2026 face delays or cancellations. These companies hold capital budgets of tens of billions of dollars; delays are not due to a lack of funds, but three bottlenecks simultaneously constraining progress. The first is the physical bottleneck. A single Nvidia GB200 rack has a power density of up to 120kW—five years ago, a rack required only 10kW, a twelvefold increase. Supporting high-voltage transformers, power distribution cabinets, and cooling systems are all struggling to keep pace. According to an industry chain analyst, delivery cycles for some high-voltage transformer models have extended to over two years—money can't buy immediate availability. "The global shortage of transformers is exceptionally severe," remarked a senior professional with long-term expertise in the power sector. The second is the institutional bottleneck. The U.S. has 66 power balancing authorities, 7 Independent System Operators, and over 3,000 independent power companies, with a median grid interconnection queue wait time of five years. Most transmission lines were built in the 1960s and 1970s and have long exceeded their design lifespan. In the AI industry, five years means three generations of chip upgrades—no company can afford to wait. The third is the scale bottleneck. The Electric Reliability Council of Texas (ERCOT) predicts that by 2032, electricity demand could quadruple from recent peaks. Data centers currently account for about 7% of total U.S. electricity consumption, with various institutions forecasting this proportion to reach between 12% and 17% by 2030. The direction is not in dispute; the debate centers on the speed and scale. The intersection of these three bottlenecks clarifies the structural issue: electricity is no longer just an operational cost but a condition for entry. No matter how powerful the model, without access to electricity, it remains theoretical. Capital Shifting Gears: Money Flowing from GPUs to Power Plants AI companies are not waiting for the grid to catch up; they are taking matters into their own hands. On April 29, the four tech giants released their earnings reports on the same day, with combined capital expenditure guidance for the full year 2026 totaling approximately $725 billion—Amazon around $200 billion, Microsoft about $190 billion, Alphabet $180 to $190 billion, and Meta $125 to $145 billion. Microsoft disclosed that roughly one-third of its capex flows to data center construction and power infrastructure, while Alphabet allocates about 40% to data centers and network equipment—power-related expenditures are shifting from marginal items in capital budgets to major components. A statement from Microsoft CFO Amy Hood that left Wall Street silent during the earnings call was: "Billions of dollars in cutting-edge Nvidia GPUs we have procured and received are sitting in our warehouses gathering dust—no data center has sufficient power to install them." Azure has an $80 billion backlog of unfulfilled orders, with the bottleneck not in chips but in power access; Google CEO Sundar Pichai acknowledged on the same day that "we are constrained by compute in the near term." Meta is taking the most capital-intensive path: according to TechCrunch, its supported data center expansion is advancing up to 10 dedicated natural gas power plants, with total generation capacity "enough to power the entire state of South Dakota." Meta's earnings call marked the first time it acknowledged that data centers "now need their own power plants," even exploring space-based solar power. The choice of natural gas over renewables is because AI data centers require 7×24 stable baseload power; the intermittency of solar and wind would halt training when darkness falls. The cost is tension with carbon neutrality commitments—according to Reuters, investors are already applying pressure, with the stock falling 6% post-earnings. Almost simultaneously, Bloomberg reported on May 6 that Microsoft is also internally discussing delaying or even abandoning its 2030 goal of 100% carbon-free energy on an hourly basis—carbon commitments are giving way to power expansion, evolving from a Meta-specific case to an industry consensus. Google is taking the lightest path: financial instruments. In March of this year, Google signed agreements with five power companies in the South Central and Midwest U.S., committing to proactively reduce data center electricity usage during grid stress, freeing up up to 1GW of demand response capacity—equivalent to the output of a nuclear power plant. Google isn't building power plants or touching heavy assets; instead, it uses Power Purchase Agreements to lock in long-term electricity prices and leverages demand response for better grid interconnection conditions—transforming from an "electricity buyer" into a participant in grid dispatch. The trade-off is remaining constrained by the grid during demand surges: Q1 saw Google Cloud's backlog exceed $460 billion, nearly doubling quarter-over-quarter. Microsoft has chosen a middle path: joint ventures. According to Reuters, Chevron, Microsoft, and Engine No. 1 have signed an exclusive agreement to explore building a 2.5 to 5GW natural gas-powered AI data center campus in West Texas—a collaboration between a traditional oil company and a tech giant to build power plants, unthinkable five years ago. On Amazon's side, regarding Anthropic's newly secured 5GW AWS computing capacity and $100 billion deal, Amazon's statement was: "Power, not just processors, is the hard constraint for AI scaling." All three paths share the same goal: bypassing the grid they cannot wait for and solving the power problem themselves. According to Morgan Stanley estimates, by the end of 2026, nearly 30% of new data center capacity will be built off-grid—a year ago, this proportion was close to zero. The shift from zero to thirty percent took only a year. Institutional Restructuring: FERC's June Deadline and Interstate Divisions On April 16, the U.S. Federal Energy Regulatory Commission (FERC) formally initiated a special rulemaking for the interconnection of large-load customers, setting a June action deadline. Politico's assessment is that FERC is entering "unprecedented legal territory." Institutions are being forced to respond, pressured from two directions. On one side, tech companies cannot wait—they have begun building their own power sources, bypassing the grid. On April 21, it was reported that a House Republican proposal seeks to exempt data centers with off-grid self-generation systems from federal energy regulations—if bringing your own power exempts you from regulation, the system itself encourages tech companies to become independent power systems. On the other side, the public is pushing back—the Maine legislature passed the nation's first state-level data center construction moratorium bill in the same week, which was vetoed by Governor Janet Mills. It is reported that Amazon, Microsoft, and Google have abandoned multiple multi-billion-dollar projects due to community opposition. The Democratic Governor of Pennsylvania and the Republican Governor of Virginia took the rare step of jointly submitting comments to FERC, urging that federal standards should be "a floor, not a ceiling." Local governments' investment promotion materials are shifting from "tax incentives" to "we have electricity, we have transformers, we have water." Who pays for all this is another battlefront. Harvard energy law scholar Ari Peskoe submitted comments to FERC, calling for an end to spreading the transmission upgrade costs triggered by data centers across all ratepayers—those who build data centers should pay. If the cost-causation principle is adopted, it will fundamentally alter the economics of data center siting. Wisconsin regulators have already ruled: data centers must bear their own energy costs at the full rate. China's Mirror Image: The Same Physical Limit, a Different Solution It is difficult to simply attribute all this to the U.S. grid being backward. In fact, similar anxieties exist across the ocean. Although Elon Musk has repeatedly expressed admiration for China's developed grid and cheap electricity: two state-owned operators manage unified dispatch, 45 ultra-high voltage lines span 40,000 kilometers, and industrial electricity prices are roughly half those in the U.S. However, the same Nvidia GB200 rack in China still has a power density of 120kW. China's "reservoir" is indeed much larger than America's, but at the final few meters to the rack, the "pipe diameter" is the same physical constraint—dedicated transformers, busbars, and liquid cooling systems must all keep up. From the "East Data, West Computing" bases in Guizhou and Inner Mongolia to the AI demand centers in Shanghai and Hangzhou, geographical and density mismatches are equally real. Some Chinese AI companies are offering technological responses different from those in the U.S. It is understood that companies like Zhejiang Jinqun Technology are advancing containerized liquid-cooled computing centers: servers, two-phase immersion liquid cooling systems, and high-voltage DC power supply modules are all pre-integrated into standard containers, ready for use upon delivery, with a 6-month deployment time (compared to three years for traditional methods), and a PUE that can drop below 1.05. "We also decided to build after seeing Jensen Huang's GTC interview, because in China, the power bottleneck also exists, and we are anxious too," said the aforementioned company representative. Two systems face the same physical limit. The U.S. solution is to turn tech companies into quasi-utilities—market mechanisms, capital competition, institutional restructuring. China's solution is national unified dispatch plus industrial technological innovation—containerized plug-and-play computing centers are another answer from Chinese engineers. Based on the difference between China's industrial electricity price (approximately $0.088 per kWh) and that of the U.S. (approximately $0.19 per kWh), a 100MW data center could save about $89 million annually in electricity costs—this is not a policy subsidy, but a systemic advantage. Overseas computing power deployments by Chinese tech companies are also accelerating. ByteDance is investing $2.1 billion to build a data center in Malaysia; GDS International's overseas business, DayOne, secured nearly $2 billion in financing, planning 1GW of capacity. Goldman Sachs estimates that total overseas data center investment by Chinese AI companies will reach $70 billion, targeting Southeast Asia and the Middle East. "We have also delivered containerized AI data centers in Malaysia based on client demand, taking just three months, saving years of construction time," added the Jinqun Technology representative. Second-Order Effects: Who Profits in Nvidia's Shadow? Where is the value in the AI industry chain actually migrating? Most investors focused on AI are watching Nvidia, but open the power supply unit of an AI rack—you'll find silicon, silicon carbide (SiC), and gallium nitride (GaN) all working simultaneously. They are not responsible for computation, but for efficiently delivering electricity to the GPUs. A 120kW rack power density means that every one percentage point improvement in power conversion efficiency can reduce cooling pressure by several kilowatts—this is not marginal optimization; it determines at the physical layer whether a data center can even power on. Germany's Infineon is currently the only company globally that masters all three of these power semiconductor materials. According to its financial reports, Infineon's AI data center revenue for fiscal 2025 is approximately €700 million, nearly triple that of the previous fiscal year; the target for fiscal 2026 is €1.5 billion; and for 2027, €2.5 billion—a tenfold increase in three years. Infineon operates at the layer that allows AI chips to receive power, and growth in this layer is approaching or even surpassing that of GPUs themselves. A New York secondary market investor commented: "It was by studying Nvidia's second-order supply chain that we discovered this stock and built a position." Nvidia itself is also using capital to bind downstream—on May 11, it entered into a 5-year equity option agreement worth up to $2.1 billion with AI data center operator IREN. Jensen Huang's explanation was: "When deploying AI factories at scale, you need deep, full-stack integration across five layers: compute, networking, software, power, and operations." These companies share a common characteristic: they do not need to bet on which AI model will win. Whether OpenAI, Google, or Meta ultimately prevails, data centers will need electricity, cooling, and wiring—this is the essence of second-order investment logic: not betting on the champion of the algorithm race, but on the infrastructure of the赛道 itself. The center of gravity of value in the AI industry chain is sinking from the algorithm layer to the infrastructure layer, and further to the power layer. Whoever controls power access controls the foundation of AI competition. The winner of the next decade may not be the company with the best model, but the one that can deliver electricity of sufficient density, in the right geographic location, at the right time.
Comments