AI firms are burning cash faster than markets can supply it — and the tab keeps rising
A $10 trillion buildout, negative free cash flow at the world's richest tech companies, and a race to borrow that is now competing with the U.S. government itself for investors' money.
On a single Monday in late October, three numbers landed that captured just how far the AI spending machine has outrun ordinary finance. Anthropic filed paperwork for what is expected to be the largest initial public offering in U.S. history, targeting roughly $100 billion in stock sales — and immediately warned investors that the proceeds would be only a down payment. After disclosing that it could not keep pace with rivals in the competition for capital, Meta said it planned to raise as much as $95 billion through new stock sales. And SpaceX finalized a $40 billion borrowing deal — described as the largest GPU financing ever — to pay for AI chips and equipment, less than five months after its own record-setting IPO.
Take it when you can. Everyone is going to need more.— Matt Orton, Chief Market Strategist, Raymond James Investment Management
That scramble reflects an arithmetic problem with no clean solution. In a widely cited draft research paper, Columbia University finance professor Stijn Van Nieuwerburgh put the cost of building out AI infrastructure — data centers, chips, and associated equipment — at more than $10 trillion through 2032. As a share of U.S. economic output, the average annual spending would surpass what the country committed, in combined terms, to electrifying American households in the early 20th century, constructing the interstate highway system, and laying the backbone of the 1990s dot-com internet boom, Van Nieuwerburgh's estimates show.
Investment analyst projections compiled by S&P Global Market Intelligence show that the five companies at the center of the buildout — Google, Amazon, Microsoft, Meta, and Oracle — are on track to post negative free cash flow next year. That is a reversal for corporations that were, until recently, among the world's most reliable cash generators. According to the Washington Post, Google spent $1.15 on AI chips, data center land, and other capital purchases for every dollar its business brought in during the most recent quarter. The company is covering the gap partly by borrowing and selling stock.
This AI thing better work out because if it doesn't ... we're going to have a problem.— Torsten Slok, Chief Economist, Apollo Global Management
Drawing on company releases and earnings calls, Wealthion calculated that Amazon, Alphabet, Microsoft, and Meta together have committed roughly $725 billion to $745 billion in capital spending for 2026 alone. Capital projects at Amazon alone are expected to cost about $220 billion this year, and chief executive Andy Jassy said in July that even that level of investment would leave capacity short of demand in both 2026 and 2027. Jassy attributed part of the increase from an earlier $200 billion estimate to higher memory costs.
We have clear line of sight to strong financial returns.— Andy Jassy, CEO, Amazon
According to Anthropic's IPO prospectus, formally released October 26, the company plans to spend no less than $518 billion over the next decade on AI infrastructure built alongside six partners, among them Google, Microsoft, and Amazon. When chip agreements with companies such as Broadcom are added in, Anthropic's total projected AI infrastructure costs exceed $600 billion, its S-1 filing shows. The company's prospectus stated: "We anticipate making significant and increased investments for the foreseeable future."
The capital competition is broad enough to reach the U.S. Treasury. Treasury Secretary Scott Bessent said in a CNBC interview in August that AI companies and the federal government are now going head-to-head for investors' cash. "It's causing a short-term competition for capital," Bessent said, according to the Washington Post. The U.S. government is paying more to borrow as a result.
At the infrastructure level, money is not the only constraint. Power is. Lawrence Berkeley National Laboratory estimated that U.S. data centers used 192 terawatt-hours of electricity in 2024 — 4.7% of total U.S. consumption — and projects that figure could reach 649 terawatt-hours, or 11.8% of U.S. electricity, by 2030. According to Reuters, lead times for generator step-up transformers topped 160 weeks in early 2026, while lead times for high-voltage circuit breakers hit 125 weeks in late 2025, up from 77 weeks in 2023. Micron chief executive Sanjay Mehrotra said the company had no "line of sight to when supply and demand will return to balance" for memory chips, and expected tight supply through 2027 and 2028.
Several major U.S. AI infrastructure projects faced delays or limits in 2026 due to power shortages, permitting obstacles, and financing gaps. Texas and New York halted construction of new AI data centers; Pennsylvania and Oregon imposed new limits. According to Heatmap, data center projects worth $260 billion were called off in 2026. In February, energy intelligence firm Currence put the share of large data center capacity scheduled for 2026 that could face delays at 30% to 50%, noting that roughly 16 gigawatts were on the schedule while only about 5 gigawatts were actually under construction at that point.
A class of smaller companies called neoclouds — independent operators that build AI-optimized data centers and rent computing power to leading AI firms — has become a flashpoint for investor anxiety. CoreWeave, one such provider with Microsoft, OpenAI, and Meta as leading customers, exemplifies the model: spend heavily upfront to outfit data centers, then collect rent checks later. Market research firm Gartner warned of a stark timing "imbalance" in this approach and said more than $3.5 trillion of AI infrastructure spending could be exposed to impairment risk by the end of 2030 if revenue does not materialize as projected. When neocloud companies have sought to go public, investors have either pushed back or required higher interest rates before agreeing to lend.
Some analysts see the pattern as a replay of earlier infrastructure bubbles. Techpolicy.press argued that more AI data centers will be built than customer demand can support — a dynamic it compared to the overexpansion of railroads and internet communications facilities in prior eras. In those cases, oversupply was followed by price collapse and contraction. The Bank for International Settlements, which advises government central banks, recently cautioned that a faltering AI boom could trigger "economywide recessions."
I'm not saying AI is going to go away, it's just not clear to me these guys are going to make money on it.— Christopher Wood, Global Head of Equity Strategy, Jefferies
The costs are already reaching consumers. Citing chip shortages driven by AI demand, executives at companies including Apple and Microsoft have announced higher prices for smartphones, laptops, video game consoles, and other products. Data center electricity consumption is putting pressure on power grids in parts of the country, including the mid-Atlantic region. The AI boom is also contributing to inflation, adding to affordability pressures for American households.
On the demand side, there are real signals of growth. Consumer AI spending reached an estimated $40 billion in 2026, more than triple the prior year's figure, according to one estimate cited by Techpolicy.press. In the six months ended June 30, Anthropic's AI service revenue climbed to $16.3 billion — a sharp rise from $1.1 billion in the same period a year earlier. Google, Microsoft, Meta, and Amazon all pointed to soaring AI-related sales in recent investor calls. According to the industry analysis firm Semianalysis, data center capacity in 2027 will be roughly twice what it was in 2026. Whether that supply expansion translates into sufficient revenue — and how quickly — is the central unresolved question hanging over the entire buildout.
Why it matters — The AI infrastructure buildout is now large enough to move inflation, stock markets, government borrowing costs, and consumer prices — meaning its success or failure will affect the financial lives of Americans well beyond the technology sector.
⚠ Not yet confirmed
- Consumer AI spending has reached $40 billion in 2026, more than triple what it was a year ago.
- Semianalysis estimates data center capacity will double in 2027 compared with 2026.
- Heatmap reported $260 billion worth of data center projects were cancelled in 2026.
- At least 47 major U.S. AI infrastructure projects stalled or were blocked in 2026 due to power shortages, permitting obstacles, and financing gaps, according to Dealroom.
Sources differ on Whether the AI buildout will generate sufficient returns to justify the investment: AI boosters including Amazon CEO Andy Jassy say there is 'clear line of sight to strong financial returns' and that demand from businesses is sustained and growing. (inquirer.com) vs Skeptics including Jefferies' Christopher Wood and the Bank for International Settlements warn returns may not materialize and risk of economywide recession exists if the boom falters. (inquirer.com)
Sources differ on Whether the AI data center buildout represents rational investment or a bubble: Techpolicy.press argues the industry is in a familiar bubble pattern — more investment than demand can support — comparable to railroad and dot-com overexpansion, likely to end in contraction. (techpolicy.press) vs Raymond James strategist Matt Orton says panic is out of control and AI has huge potential; investor concern is overblown. (inquirer.com)
Reported by finance-commerce.com, goldmansachs.com, dealroom.co, gartner.com, washingtonpost.com, dnyuz.com, techpolicy.press, wealthion.com, inquirer.com, axios.com
- Scott BessentUnited States Secretary of the Treasury
- Andy JassyAmerican businessman