AI dreams crash into stark $7 trln reality

NEW YORK, April 7 (Reuters Breakingviews) - Artificial intelligence is squeezing the planet’s resources. Serious questions have yet to be answered about whether there’s enough labor, copper, water and other basics to build and operate all the data centers now under consideration. Even if AI titans manage to solve these problems, there’s an even bigger and stickier supply matter to work out: money.

Large-language hoopla ushered in the new year, ​picking up where it left off in 2025. Standing in an 800,000-square-foot warehouse, Mississippi Governor Tate Reeves on January 8 unveiled, opens new tab what he called the single-biggest investment in state history, a $20 billion project by ‌Elon Musk’s xAI for a sprawling complex, opens new tab with nearly 2 gigawatts of computing power. All told, more than 50 GW of such U.S. projects have been announced, according to a running tally being kept by Barclays analysts, with a similar figure touted across Europe.

The Reuters Iran Briefing newsletter keeps you informed with the latest developments and analysis of the Iran war. Sign up here.

It’s a staggering amount of computing power, and one that’s especially hard to comprehend given the similarly imposing financial commitments. The sheer size reflects exuberance over how AI models such as OpenAI’s ChatGPT and Google’s Gemini will upend productivity and economic output worldwide. A closer look, however, suggests it will be exceedingly difficult, maybe ​impossible, to fund all the envisioned expansion.

Data centers are measured by the amount of peak electricity it takes to operate and cool the servers. Roughly 110 GW of such projects are already in the planning stages, at ​least based on tallies from public statements. Nvidia (NVDA.O), opens new tab boss Jensen Huang, whose company supplies most of the processing power for training and running AI, reckons it costs, opens new tab $60 billion to $80 billion to ⁠construct a 1 GW compound housing thousands of server racks. His numbers are higher than most. Using a $36 billion per GW figure from Bernstein analysts would translate into some $4 trillion to pay for all the currently slated data centers. At the ​bottom of Huang’s range, however, the implied outlay is $6.6 trillion.

The ambition eclipses one of the largest public works projects in history. Authorized by President Dwight Eisenhower in 1956, the U.S. Interstate Highway System cost, opens new tab about $500 billion in today’s dollars. It ran considerably ​overbudget and took more than three decades to complete. AI proponents aim to spend 13 times as much over, roughly speaking, just five years.

Without taxpayers to foot the new technology bill, it will fall to private investors instead. Although AI hype is rampant and 13-digit sums are casually thrown around these days in reference to everything from corporate valuations to government interest payments, it’s an implausible sum to raise in such a short period of time.

Big technology companies pinning their futures on AI are tapping deep pools of money. Alphabet (GOOGL.O), opens new tab, Amazon.com (AMZN.O), opens new tab, Meta Platforms (META.O), opens new tab, Microsoft (MSFT.O), opens new tab and Oracle (ORCL.N), opens new tab ​have already committed significant amounts and, theoretically at least, have even more at their disposal. Combined, the fivesome is expected to generate $5.5 trillion of operating cash flow from selling software subscriptions, advertising, cloud storage and other goods and services over ​the next five years, according to estimates gathered by Visible Alpha.

They are also borrowing heavily to compete in the cutthroat data-center arms race. Amazon alone raised, opens new tab a record $37 billion in U.S. bond markets last month, followed by a $17 billion euro-denominated deal, opens new tab. Alphabet earlier sold a rare 100-year tranche ‌as part of ⁠its $32 billion debt package in February. All told, projections from BofA analysts suggest $1 trillion of such hyperscaler-related investment-grade issuance is possible through 2030.

Retirement funds and other large investors will be getting in on the AI action, too. There’s nearly $700 billion of capital committed to direct-lending and infrastructure funds managed by Brookfield Asset Management (BAM.N), opens new tab, Blackstone (BX.N), opens new tab and others, according to research firm Preqin.

Additional lending capacity lurks inside markets, as well. Morgan Stanley analysts anticipate about $50 billion annually on average from asset-backed and commercial mortgage-backed securities related to building AI-related facilities over the next few years. High-yield bonds and leveraged loans should provide another $150 billion through 2030, JPMorgan analysts forecast. Some of the data centers themselves might even be throwing off cash by then, although assuming a 9% yield ​on cost, the interest payments on associated debt will chew ​up a chunky amount.

Put it all together and, on ⁠paper at least, there should be about $7.5 trillion of funding available, comfortably ahead of the $6.6 trillion cost estimate, using Huang’s lower bound. Other factors bear closer scrutiny first.

For one thing, the surplus assumes that hyperscalers like Microsoft will devote every last drop of their operating cash flow to data centers for the next five years. Their shareholders might start squawking about missing ​stock buybacks and dividends, to say nothing of the lack of capital expenditure on anything else. Allocating every dollar sitting inside private credit and infrastructure funds would similarly ​create an intolerable amount of concentration ⁠risk and opportunity cost.

Moreover, the all-in economic costs may be higher still. The per-GW assessments exclude potentially vast expenses related to shoring up the reliability of old and overburdened utilities to safeguard electricity transmission and ensure that retail customers don’t suffer either. If the upper limit of Huang’s range turns out to be more accurate, because of inflation or other reasons, even the full amount of projected capital would fall short.

When added to some of the likely physical limitations of rolling out so much promised AI capacity, the ⁠money problem becomes ​even starker. Timelines could easily be extended, just as they were for U.S. highways. Engineers, or AI itself, may discover ways to significantly reduce the ​cost along the way. The technology’s anticipated productivity gains and revenue opportunities also could easily be exaggerated, which would greatly curb the enthusiasm of investors. The most likely outcome, however, is that a considerable slug of the announced data center plans will turn into so much vaporware. Even in today’s ​world of financial abundance, putting $7 trillion into one industry so quickly would be squandering a precious resource.

Follow Jeffrey Goldfarb on X, opens new tab and LinkedIn, opens new tab.

For more insights like these, click here, opens new tab to try Breakingviews for free.

Editing by Liam Proud; Production by Pranav Kiran

• Suggested Topics:
• Breakingviews

Breakingviews
Reuters Breakingviews is the world’s leading source of agenda-setting financial insight. As the Reuters brand for financial commentary, we dissect the big business and economic stories as they break around the world every day. A global team of about 30 correspondents in New York, London, Hong Kong and other major cities provides expert analysis in real time.

Sign up for a free trial of our full service at and follow us on X @Breakingviews and at www.breakingviews.com. All opinions expressed are those of the authors.

• X

• Facebook

• Linkedin

• Email

• Link

Purchase Licensing Rights

Jeffrey Goldfarb

Thomson Reuters

Jeffrey is Global Corporate Finance Editor, based in New York, coordinating deal-related coverage. Previously, he oversaw the Asia-Pacific region for Reuters Breakingviews from Melbourne and Hong Kong. He first joined Breakingviews in London as the global financial crisis began and later spent seven years in New York as U.S. Editor. Before becoming a columnist in 2007, Jeffrey covered banking, M&A, media, technology, international trade and healthcare for Reuters and BNA in New York, Washington, Phoenix and across Europe. He has a master’s in journalism from Columbia University and a bachelor’s in finance from the George Washington University.

• Email