E212｜The Trillion-Dollar AI Data Center Buildout: America's GDP Growth Rests Entirely on It
Summary
- AI infrastructure has gone from a tech-company capex item to a core driver of U.S. economic growth. 泓君 cites Jason Furman’s estimate that nearly all U.S. GDP growth in the first half of 2025 came from AI infrastructure; excluding information technology and software, growth was just 0.1%. As OpenAI restructures and lays the groundwork for a potential IPO, it has pledged $1.4T in future investment—creating a stark contrast between the tech giants’ breakneck infrastructure buildout and the 0.1% growth in traditional industries.
- The scale of this arms race has jumped from $500B to several trillion dollars. OpenAI has proposed 10 GW, 6 GW and 10 GW projects with NVIDIA, AMD and Broadcom, respectively, for a combined 26 GW; at the program’s $50B-per-GW estimate, that implies roughly $1.5T. 徐熠兴 believes the 10 GW Stargate target “may only be the very beginning,” with ambitions potentially reaching 10x that scale over the next 5-10 years; the industry view cited by the guests is that data center investment could reach $5T-$7T over the next five years.
- The real bottleneck is no longer an individual GPU, but the power and supporting supply chain needed to keep entire clusters running. The “Power First” flywheel is to secure power first, then buy GPUs, train better models, win users and revenue, and recycle profits into more land and power. That has produced a consensus that “Underinvestment is riskier than over investment”: excess land, power, data center capacity and GPUs can be absorbed internally, rented out or resold, while underinvestment could cost a company its right to survive, “because nobody wants to be Nokia.”
- Data center demand is shifting from pretraining toward inference and applications, but that will not reduce total compute demand. Two years ago, roughly 60%-70% of compute went to pretraining; earlier this year, inference accounted for about 60% and training 40%. 徐熠兴 sees inference potentially exceeding 80% in the future, and 王辰晟 broadly agrees. That shift means AI is beginning to generate revenue through subscriptions and applications. Large clusters also materially reduce operating costs and shorten model iteration cycles, reinforcing the Microsoft CEO’s view that “AI is valuable only when it is actually creating GDP.”
- The U.S. faces a three-layer shortfall in nameplate capacity, firm power and grid connections. The U.S. power system has grown by less than 1% a year over the past 20 years. The program estimates that the country needs roughly 80 GW of new capacity annually but is adding only 50-60 GW, implying a 20 GW annual gap that could accumulate to 100 GW over five years; solar’s capacity factor is about 25%, versus 85% for gas and 93% for nuclear. 泓君 relays 黄仁勋’s assessment: “You can have the GPUs, but without power, you still can’t run the data center.”
- Transmission lines, gas turbines and transformers may be harder to scale quickly than generation capital itself. New U.S. transmission lines typically take 7-12 years, while GE Vernova’s gas turbine backlog is already booked through 2028. According to public analysis, xAI has absorbed nearly 70% of U.S. gas turbine inventory, and Colossus 2 could deploy roughly 160 units. Transformer lead times have stretched from about 3 months to at least 18 months—and possibly 18-24 months—fulfilling Musk’s pun: “transformer lead transformer.”
- The investment beneficiaries extend from GE Vernova and nuclear power to electrical equipment, copper, storage and logistics, but valuations and fundamentals are not always moving together. 徐熠兴 believes Oklo’s surge reflects market expectations more than commercial operating progress; 王辰晟 calculates that some chemical fuel-cell alternatives are paying an additional $600M-$1B over five years to gain speed. The ultimate risk remains a domino effect through the funding chain: fraud at any company, weaker-than-expected AI adoption or a loss of trust could hit the capital cycle that is “disproportionately determining the U.S. economy.”
Deep dive
1. AI Infrastructure Is Absorbing U.S. Growth
泓君 cites Harvard economist Jason Furman’s estimate: nearly all U.S. GDP growth in the first half of 2025 came from AI infrastructure; strip out information technology and software, and growth falls to just 0.1%. Data centers are therefore no longer merely a tech-industry issue, but a major macroeconomic variable.
After completing its corporate restructuring and laying the groundwork for a potential IPO, OpenAI proposed investing $1.4T in AI infrastructure over the coming years. The day before the restructuring, it also wrote to the White House that China added 429 GW of power capacity in 2024 versus just 51 GW in the U.S., urging the government to prioritize AI projects in permitting.
The program’s overarching frame is a highly concentrated “trillion-dollar infrastructure buildout”: a handful of technology companies, working with financial giants, are beginning to exert a decisive influence on the U.S. economy. 泓君 closed by preserving the key question—if AI commercialization falls short, could a single trust shock in the funding transmission chain become “the first domino to fall”?
2. OpenAI Has Taken the Race to 26 GW and Trillion-Dollar Scale
徐熠兴 believes OpenAI’s public 10 GW Stargate target “may only be the very beginning,” with its 5-10-year ambition potentially reaching 10x that scale. He says this could correspond to an industry worth roughly $5T, and cites 黄仁勋 or consulting-firm estimates that data center infrastructure investment could reach $5T-$7T over the next five years.
泓君 believes data centers could account for 70% of U.S. GDP growth this year. She says she would not be surprised, then adds: “You may be underestimating it.”
王辰晟 breaks down the commitments: OpenAI is planning 10 GW with NVIDIA, 6 GW with AMD and 10 GW with Broadcom, for 26 GW in total. At $50B per GW, that represents roughly $1.5T over five years. OpenAI has also locked in monthly capacity of 900,000 wafers from Samsung and SK hynix, equivalent to roughly one-third of the DRAM market and 60% of the HBM market.
Competitors are each targeting a different bottleneck. xAI is sweeping up small gas turbines; Meta continues to buy low-cost power and land, planning a roughly 5 GW campus “large enough to cover most of Manhattan”; Google is buying up capacity across interconnects, fiber-optic cables and other supply chains. “Every giant is pushing hard. None of them wants to lose ground in this competition.”
Microsoft’s path has been more circuitous. It paused some construction and exited data center leases earlier this year, while its CEO publicly worried about overbuilding; it subsequently completed one of the world’s largest AI data centers. 徐熠兴 sees strategy fluctuating over the year, but Google and Microsoft’s existing data centers may already exceed 10 GW, so their slower growth should not automatically be read as a lack of aggression.
3. “Power First” Makes Overinvestment the Smaller Risk
王辰晟 explains the “Power First” chain in simple terms: whoever secures power first can buy more GPUs, train better models, win more users and revenue, then reinvest profits into more land and power. Energy is therefore the first link connecting models, market share and returns on capital.
The arms race is built around a shared judgment: “Underinvestment is riskier than over investment.” If the company that reaches the best model or AGI first captures most of the market, underinvestment could mean the rapid disappearance of its room to survive. Even as Wall Street questions returns and cash flow, the giants “don’t even blink.”
The losses from overinvestment are more bounded. Land, power, data center capacity and GPUs can all be used internally, rented out or resold. 泓君 cites an old Silicon Valley saying: “Bill always eats Andy.” As long as the hardware and infrastructure exist, software will eventually find a way to consume the resources.
泓君 frames the wager as whether a $4T market cap becomes $3T, or whether missing AI turns it into zero. Companies naturally prefer to tell the story of surviving the tide and going from $4T to $10T, but she explicitly adds the qualification: “That does not mean it will necessarily happen.”
4. Large Clusters Accelerate Training and Monetize Inference
王辰晟 cites Google’s economic calculation: compared with a distributed deployment of the same size, a 1 GW AI data center in Iowa can save $500M in annual operating costs, thanks to efficiency gains in transmission, cooling and operations. Large clusters are not pursuing scale for its own sake; they directly lower the unit cost of compute.
Training cadence further reinforces the need for scale. The program estimates that GPT-4 used roughly 16,000 H100s for 90 days to process 1.7T data points; GPT-4.5 may require 10^26 operations and potentially 2-3x as many cards. A GB200 cluster with 25,000 cards could also require 90-120 days of compute. The giants want to move from training once a quarter to iterating every 1-2 weeks.
The use of compute has reversed: two years ago, roughly 60%-70% went to pretraining; earlier this year, inference accounted for about 60% and training 40%. 徐熠兴 expects inference could exceed 80% in the future, emphasizing that training does not directly generate revenue—subscriptions and applications are what turn capex into cash flow. 王辰晟 broadly agrees.
Small, distributed compute is not useless. Companies such as Novita can use idle resources to offer compute at lower cost, but training requires high-speed interconnects between racks and a unified cluster. Third-party services may require 99.999% reliability, while internal training may need only 99.9%, which is why Stargate is locating much of its training in West Texas, where wind, solar, land and grid access are available.
5. The U.S. May Be Short 20 GW a Year, with a Larger Firm-Power Gap
徐熠兴 points out that the U.S. power system has grown by less than 1% annually over the past 20 years, becoming almost decoupled from GDP growth; even immediately doubling the rate would take it to only about 2%. Data centers may account for 40% of new load, with the rest coming from electric vehicles, reshoring manufacturing and other demand.
The program estimates that the U.S. needs roughly 80 GW of new capacity each year but is adding only 50-60 GW, creating an annual shortfall of about 20 GW and a potential five-year accumulation of 100 GW. The country’s total generation capacity is roughly 1,300 GW. New York uses about 6 GW on average and peaks at 11-12 GW, so 20 GW is equivalent to 2-3 New Yorks.
Looking only at nameplate capacity understates the problem. Of the roughly 50 GW added in a year, about 45 GW comes from solar, 5 GW from wind and less than 5 GW from thermal power, leaving effective output potentially below 20-25 GW. Solar’s capacity factor is about 25%, versus roughly 85% for gas and 93% for nuclear. The same “gigawatt” does not represent the same amount of electricity.
王辰晟 offers a reverse check: if 60 GW of data centers cost $50B per GW, they would require $3T, while the major companies’ total investment next year is broadly below $1T. On capex plans alone, the short-term requirement may therefore not be as high as 60 GW. Some institutions forecast that roughly 60% of new generation will come from gas and 40% from solar, wind and storage; new nuclear may arrive by 2028, while SMRs may not become a major source until 2030.
6. Having a Generator Does Not Mean the Data Center Has Power
徐熠兴 warns that data centers depend on the full generation-transmission-distribution system, not an isolated power plant. Generation accounts for roughly 50% of power-system investment, transmission 10%-20% and distribution 20%-30%; the U.S. grid may not even have enough capacity to absorb all new generation.
Stargate currently targets 10 GW, with roughly 7 GW already signed or announced, but an intention is not the same as successful grid connection. The project must find existing capacity while also creating new capacity on the grid; otherwise, even completed generation facilities will face resistance when they try to connect.
Technology companies are therefore beginning to build their own generators, power plants, substations, distribution networks and short-distance transmission lines. “The utilities have completely failed to keep up with their demand.” The most practical strategy is to place generation next to the data center and keep engineering constraints within the company’s economic and political reach.
7. xAI’s Buying Spree Exposes the Gas Turbine Supply Crisis
GE Vernova’s gas turbine business grew slowly over the past decade, reaching a 2019-2020 peak of only more than 70 units a year, at roughly 30-50 MW each. By comparison, about 4,000 aircraft engines roll off production lines globally every year. Gas turbines were previously viewed as a “sunset industry” that increased carbon emissions, so the supply chain did not reserve capacity for AI demand.
王辰晟 cites public information indicating that xAI has taken nearly 70% of U.S. gas turbine inventory. According to SemiAnalysis, if memory serves, Colossus 2 alone could be equipped with roughly 160 units. These turbines fill the 18-month gap between grid-connection approvals that take roughly two years and data centers that may need to go live in six months—a “short-term damage-control” measure, not a permanent optimal solution.
Large turbines offer higher efficiency and power output, but GE’s orders are booked through 2028. Aero-derivative gas turbines and smaller equipment from Caterpillar and others can help, but it may take 10 units to replace a single 300 MW plant, and expanding production still takes time. The supply-chain challenge is therefore spreading across more equipment and components.
Asked whether China could quickly fill the gap, 王辰晟 says turbine blades require specialized alloys and defense-grade safety standards, and cannot be solved through ordinary manufacturing scale alone. “Given enough time, it can definitely solve it.” The problem is that the energy race has not left the supply chain enough time.
8. Transformers Turn a 3 MW Project into an 18-Month Wait
王辰晟 recounts the example of Tesla’s Dojo. Roughly 1.5-2 years ago, the company wanted to build a small 3 MW cluster in Palo Alto with only a dozen-plus training units, but the city said there was no power and transformer lead times had stretched from 3 months to 18 months. Tesla ultimately bought 2 transformers itself, installed them for the government and handed them over in exchange for permission to use the facilities.
The scale gap captures the problem: “At the time it was only 3 MW. Now we routinely talk about 3 GW—a 1,000x difference.” The program also notes that transformer lead times could reach 18-24 months, making them an unavoidable bottleneck for power plants, distribution networks and data centers alike.
The upstream bottleneck is specialized material such as grain-oriented electrical steel. Based on 王辰晟’s figures, only one U.S. company produces it, with annual output of 250,000 tons; global output is roughly 5M tons, while Baosteel alone produces close to 2M tons. The U.S. restricted related imports through anti-dumping measures in 2016, 2020 and 2024, as well as the OBBBA law cited on the program, but domestic manufacturing cannot absorb demand in the short term.
9. 800-Volt DC Trades Standards for Power and Copper
Electricity is currently transmitted over long distances at roughly 350 kV, stepped down at local substations to 3.8-35 kV, then routed through UPS systems as 480 V or 415 V AC into the data center before being converted to 54 V DC for servers. NVIDIA’s proposed 800 V DC architecture is aimed first at the data center and rack, not the residential grid.
Rack power is rising rapidly: Hopper/H100 is around 30 kW, GB200 around 100 kW, and Vera Rubin and later products could reach 400 kW or even 1 MW. NVIDIA estimates that a 1 MW rack operating at 54 V would require roughly 200 kg of copper; a 1 GW data center using the old architecture could require as much as 500,000 tons of copper.
Higher voltage sharply reduces current and resistive losses. Transmitting 1 MW at 54 V could produce losses of 22%, while raising the voltage to 800 V cuts copper-transmission losses to 0.6%. When electricity is already the scarcest resource, this is not a marginal optimization; it helps avoid a situation where “the next shortage could be copper.”
The industry still runs mainly on 415 V AC, so it will first try 400 V or 415 V DC before moving to 800 V. The end state would use solid-state transformers at the entrance and eliminate some UPS equipment, raising end-to-end efficiency from roughly 92%-98% to 98.5% or even 99%. Data centers may account for about 5% of U.S. electricity consumption in 2025 and potentially double by 2030—enough to support a dedicated standard.
10. China Plans Power Centrally; the U.S. Is Stuck on Permits, Costs and Communities
OpenAI’s letter used figures of 429 GW of new capacity in China in 2024 versus 51 GW in the U.S.; later in the program, 泓君 cited a different set of figures for “this year”: 495 GW for China and 50 GW for the U.S. Whichever set is used, the guests’ point is that the two countries differ by an order of magnitude in construction speed.
徐熠兴 puts institutional differences first. China can centrally plan west-to-east power transmission, south-to-north power transmission and high-voltage DC networks; the U.S. relies mainly on regional grids. If a long-distance line crosses a farmer who refuses to cooperate, it may have to detour hundreds of miles and restart negotiations—“That is why the U.S. cannot build high-speed rail.”
New U.S. long-distance transmission lines generally take 7-12 years. Technology companies may move faster on financing and project management, but they cannot bypass land rights and permitting, which is why they prefer to build power plants near data centers and avoid large-scale interregional transmission.
Cost differences are equally important. China’s annual solar installations can equal the rest of the world’s combined total, and sometimes exceed it; large-scale storage equipment in the U.S. may cost roughly 2x as much as in China. Siting also requires balancing water and power: closed-loop liquid cooling may not consume large amounts of water, but water conservation often requires more cooling electricity, while saving power may depend on local water resources and provoke community opposition over shortages and pollution.
11. The Beneficiary Chain Is Long; the Fastest-Rising Assets May Not Monetize First
徐熠兴 sees gas as the most direct beneficiary because it can come online quickly and offers strong order visibility and pricing power, making GE Vernova the clearest winner. Nuclear could benefit from demand over the next five years, but is not a short-term solution. Electrical equipment, transformers, copper and other raw materials should also benefit as the entire power system expands.
Oklo has not yet generated electricity, and its executives’ most optimistic target for commercial operations is 2027. 徐熠兴 considers that date potentially too optimistic; the recent stock surge “reflects market expectations and sentiment more than fundamentals.” The U.S. has suffered a decades-long break in nuclear talent and industrial capacity, and the sector cannot be modeled at software speed.
The race for speed forces companies to accept materially worse economics. For the same 100 MWh, a gas turbine may require roughly $200M of capex, versus about $700M for certain alternatives, while annual fuel and operating costs may be an additional $20M-$50M. The extra five-year spending could reach $600M-$1B, but “being faster than you to AGI” is overriding cost discipline.
Hidden bottlenecks also include storage and operations. Excluding video models such as Sora and Veo, the storage shortfall could already reach 5%-8% by the end of next year. In the 100,000-card GPT-4 or Meta Llama 4 scenarios discussed on the program, an outage could occur every 32 minutes on average and take about 15 minutes to recover; a 1% weekly failure rate would mean 1,000 cards and roughly 100 tons of reverse logistics. “There are far too many bottlenecks across the entire supply chain,” and a failure at any one point could amplify the fragility of the capital cycle.