$690 billion. That is how much the four biggest cloud companies plus Oracle plan to spend on AI infrastructure in 2026, roughly 70% more than last year.
The individual numbers are staggering. Amazon leads at $200 billion (up 52%). Alphabet plans $175-185 billion, nearly doubling its 2025 spend. Microsoft is in for $100-120 billion. Meta rounds it out at $60-65 billion. These are not projections from analysts - they are the companies' own guidance.
The Revenue Gap
Here is the problem: direct AI revenue in 2025 came in around $40-60 billion across the industry. AI-specific capital expenditure that same year was roughly $300 billion. That is a coverage ratio of 0.15x, meaning actual AI revenue covers about 15 cents of every dollar spent building AI infrastructure.
Sequoia Capital's David Cahn calculated that the AI ecosystem needs $600 billion in annual revenue to justify current spending levels. The industry generates somewhere between $50-100 billion. That gap is not closing fast enough to match the pace of investment.
Free Cash Flow Is Evaporating
The spending is already hitting balance sheets hard. Alphabet's free cash flow (the money left after a company pays for operations and capital investments) is projected to fall from $73 billion in 2025 to roughly $8 billion in 2026. That is a 90% decline. Bank of America estimates that AI capex will consume 94% of the Big Four's operating cash flow after dividends and buybacks.
Amazon's trailing twelve-month free cash flow has already compressed to $11.2 billion, with $200 billion in spending still ahead.
The Depreciation Time Bomb
Inference costs (what it costs to actually run AI models and get answers) are falling 50-200x annually, according to Epoch AI data. Processing a query that cost $20 per million tokens with GPT-3 in 2020 now costs about $0.07 for comparable quality. That 280-fold price drop is great for users but brutal for companies that just spent billions on hardware optimized for last year's workloads.
Amazon already took a $920 million write-down in Q4 2024 and shortened its estimated server lifespan from six years to five. Investor Michael Burry estimates hyperscalers will understate depreciation by $176 billion between 2026 and 2028, artificially inflating earnings by over 20%.
The Prisoner's Dilemma
So why keep spending? Alphabet CEO Sundar Pichai put it bluntly: "The risk of underinvesting is dramatically greater than the risk of overinvesting." Each company fears that pulling back will hand the market to whoever keeps building. Classic prisoner's dilemma.
MIT economist Dario Acemoglu offers a more skeptical view, projecting AI will deliver only a 1.1-1.6% GDP increase over the next decade, with just 5% of economic tasks cost-effectively automatable at current prices.
The historical parallel that should worry investors: the 1990s telecom bubble, where companies laid fiber optic cable based on the claim that internet traffic was doubling every 100 days (it was not). By 2001, only 5% of installed fiber capacity was in use. J.P. Morgan projects $300 billion in investment-grade bonds for AI data centers in 2026 alone, building the same kind of debt structure that made the telecom crash so painful.
Without AI spending, U.S. corporate capital expenditure would actually be negative right now. Pantheon Macroeconomics found that essentially all GDP growth in the first half of 2025 (1.4% annualized) came from AI investment. That makes this more than a tech story. If AI capex slows, the broader economy feels it.