Standing inside an 800,000-square-foot warehouse in Mississippi on January 8, Governor Tate Reeves unveiled what he described as the single-biggest investment in state history: a $20 billion project by Elon Musk’s xAI that promised nearly two gigawatts of computing power in a sprawling complex that would transform the local economy. The audience was excited. The figures were astounding. And if there were any economists in the back of the room, they were most likely doing some quiet math and coming to a difficult conclusion.
Moments like the Mississippi announcement have occurred remarkably frequently as a result of the AI infrastructure boom; each one is bigger than the last and is delivered with the kind of assured finality that implies the funds are already obtained, the land is already graded, and the schedule is already set. According to Barclays analysts keeping track of the running tally, more than 50 gigawatts of such U.S. projects have been announced, with a similar number touted throughout Europe. Approximately 110 gigawatts of data center capacity are being planned globally. Next, perform the math that is often omitted from press releases.
| Topic | Details |
|---|---|
| Total Projected Cost | Global data center projects needed to power AI estimated at $6.6 to $7.5 trillion — based on Nvidia CEO Jensen Huang’s construction cost range of $60–$80 billion per gigawatt |
| Planned Global Capacity | Approximately 110 GW of data center projects already in planning stages globally — over 50 GW in the U.S. alone, with a similar figure announced across Europe |
| Cost Comparison | The U.S. Interstate Highway System — authorized by President Eisenhower in 1956 — cost roughly $500 billion in today’s dollars and took over three decades to complete; the AI build-out aims to spend 13 times that figure in approximately five years |
| Hyperscaler Cash Flow | Alphabet, Amazon, Meta, Microsoft, and Oracle combined are projected to generate $5.5 trillion in operating cash flow over the next five years — still well short of the infrastructure requirement |
| Amazon Bond Issuance | Amazon raised a record $37 billion in U.S. bond markets in a single month; Alphabet sold a rare 100-year bond tranche as part of a $32 billion debt package in February 2026 |
| BofA Projection | Bank of America analysts project up to $1 trillion in hyperscaler-related investment-grade bond issuance possible through 2030 — underscoring how much of this build-out will be debt-financed |
| Key 2025 Funding Rounds | OpenAI raised $40 billion at a $300 billion valuation; Anthropic closed $16.5 billion across two rounds at a $183 billion valuation; xAI raised over $10 billion |
| Infrastructure Funds Available | Nearly $700 billion committed to direct-lending and infrastructure funds managed by Brookfield, Blackstone, and others — significant, but a fraction of what full build-out would require |
| Cautionary Voice | Apollo Global Management’s chief economist Torsten Sløk warned that AI companies appear more overvalued than dot-com firms were at their peak |
| Notable Collapse | Blue Owl Capital withdrew from a planned $10 billion Oracle data-center deal linked to OpenAI — an early signal that capital commitments are not always as solid as press releases suggest |
A single gigawatt of computing infrastructure costs between $60 billion and $80 billion, according to Jensen Huang, whose company Nvidia provides the vast majority of the processing power needed to train and run AI models. The implied cost of everything currently announced, at the lower end of his range, is $6.6 trillion. It is close to $7.5 trillion at higher estimates. It’s not a rounding error. That exceeds the annual production of most nations.
The U.S. Interstate Highway System is a common comparison that merits greater consideration than it usually receives. In 1956, Eisenhower gave his approval. At about $500 billion in today’s dollars, it was the biggest public works project in American history. The budget was exceeded. It took over thirty years to complete. In about five years, the AI sector plans to spend thirteen times that amount. without the support of taxpayers. on the back of private capital markets, which are being asked to do something at a scale and speed they have never done before, according to any honest accounting. The market might be ready for it. Additionally, as it becomes more difficult to bridge the gap between ambition and available funding, that figure may be subtly revised downward.
The major tech firms are making an effort. The bond issuance figures make it very evident that they are making a great effort. After raising a record $37 billion in U.S. debt markets in a single month, Amazon went on to close a $17 billion deal denominated in euros. In February 2026, Alphabet sold a $32 billion package that included a 100-year bond tranche. These are sincere promises from sincere businesses with solid revenue streams.
Over the next five years, cloud services, advertising, software subscriptions, and related businesses are expected to generate $5.5 trillion in operating cash flow for Alphabet, Amazon, Meta, Microsoft, and Oracle combined. Analysts at Bank of America predict that bond issuance related to hyperscalers could reach $1 trillion by 2030. Nearly $700 billion is committed and available in direct-lending and infrastructure funds managed by Brookfield, Blackstone, and others. These figures are not insignificant. However, they still don’t meet the theoretical requirements of the announced build-out, leaving a gap that is known to all in the industry but that very few are publicly addressing.
When a sector’s goals exceed its funding by a significant enough margin that even optimistic scenarios fail to close the gap, a particular type of financial tension develops. Similar things happened during the dot-com era, but the financial numbers were lower and the physical infrastructure requirements were less stringent. For example, you could build a website with a server farm that could fit in a rented office, but you couldn’t build a gigawatt data center in a strip mall. Torsten Sløk, chief economist at Apollo Global Management, has publicly stated that AI valuations appear even more stretched than dot-com companies were at their height.
Although the industry tends to ignore this comparison, the math doesn’t fully allow you to ignore it. $300 billion for OpenAI. After closing $16.5 billion in funding, Anthropic is now valued at $183 billion. Before releasing a product to the public, Thinking Machine Labs, a startup started by former OpenAI technologist Mira Murati, secured a $2 billion seed round at a $12 billion valuation. These are the kinds of valuations that make the assumptions that demand will materialize at the rate currently projected by the models’ most enthusiastic users, scaling will continue indefinitely, and everything will work.
Whether all of those presumptions are true is still up for debate. The physical limitations are actual and getting worse. Large-scale data centers’ power demands cannot be met by the electrical grid capacity in many areas without major infrastructure upgrades, which themselves require years and money. In regions vulnerable to drought, the need for water for cooling has sparked political backlash. The supply of copper, which is required in massive amounts for the electrical systems that underpin all of this, is becoming more and more limited worldwide. These bottlenecks are not hypothetical. Project schedules, construction cost overruns, and Blue Owl Capital’s covert financing withdrawal from a $10 billion Oracle data center deal connected to OpenAI are examples of how they manifest. There is no withdrawal of press releases. Sometimes funding commitments do.
Observing this develop across debt market filings, infrastructure announcements, and quarterly earnings calls paints an honest picture of an industry that made massive commitments in 2025 and is now navigating the more difficult and drawn-out process of figuring out which of those commitments it can truly keep. The funds are genuine. The aspiration is sincere. However, $7 trillion has a way of making things clear that enthusiasm often obscures.