The other I in AI: Cost of investment | Number Theory

A generational disruption to technology is also bringing a generational change, and perhaps a massive disruption to the global financial system

Nvidia today is the world’s most valuable company by market cap. A Financial Times story published yesterday had the following headline: Wall Street giants bet Nvidia’s AI chips will defy the laws of finance. A generational disruption to technology is also bringing a generational change, and perhaps a massive disruption to the global financial system. And that is the hyperscaler story playing out right now. Whether the math will add up is the biggest bet in the global economy.

Unsplash image

• 
  Big Tech is becoming heavy industry

  For much of the internet age, the world’s biggest technology companies were celebrated for building businesses that scaled with software. AI is rewriting that equation. Since ChatGPT’s arrival in late 2022, the five largest hyperscalers – Amazon, Microsoft, Alphabet, Meta and Oracle—have embarked on one of the biggest corporate investment cycles in history. Their combined capital expenditure nearly tripled from $162.3 billion in 2022 to $448.3 billion in 2025. That trajectory is only getting steeper, with analysts now expecting the figure to reach around $750-800 billion in 2026 and more than $1 trillion in 2027. Most of it is brick and mortar investment, data centres with chips, power, water and steel requirements.
• 
  The investment push is testing the limits of even the richest companies the world

  Big Tech can still afford the AI race, but even they are pushing the cash flow envelope. The five hyperscalers generated $180.9 billion in operating cash flow in the final quarter of 2025, up from $65.6 billion in early 2022. Cash capital expenditure, however, rose much faster, from $36.6 billion to $130.7 billion. By the first quarter of 2026, cash capital expenditure, which measures cash spent on property and equipment but excludes finance leases (they involve little cash upfront) had reached $148.4 billion, only marginally below the $157.9 billion produced by operations. Since GPT-4’s release, operating cash flow has grown at an annualised pace of about 23%, compared with 70% for cash capital expenditure. Back of the envelop extrapolation suggests capex could overtake operating cash flow by late 2026, leaving little or no free cash flow. To be sure, this is not a forecast, since today’s investment could itself lift future revenues.
• 
  The cloud is borrowing from the future

  Big Tech’s AI build-out is no longer being financed by cash alone. According to Moody’s Ratings, direct debt across the six hyperscalers it tracks, including CoreWeave, has reached roughly $460 billion. Their lease obligations are even larger. Moody’s estimates that the six have around $1.2 trillion in total lease commitments, including more than $820 billion tied to facilities that have yet to commence. This means tomorrow’s data centres are already claiming tomorrow’s cash flows. Leasing lets hyperscalers secure capacity without paying the full construction cost upfront, often leaving developers and special-purpose vehicles to raise the debt. It makes the build-out look lighter on today’s balance sheet, but not necessarily less binding. The borrowing has accelerated even faster than expected, with Amazon, Alphabet, Meta and Oracle having issued about $194 billion of bonds by early July.
• 
  The chip in the investment armour could be the proverbial chink

  Investments are long-term by nature. Not so much in cutting edge tech. Goldman Sachs estimates that compute will absorb $5.1 trillion, or two-thirds of projected AI infrastructure investment between 2026 and 2031. The real uncertainty is not whether those chips will keep working, but how long they will remain worth running. If new generation chips deliver sharp gains in performance and efficiency, older hardware can become economically obsolete long before it physically fails. In fact, with a three-year useful life, cumulative depreciation could approach $4 trillion, compared with roughly $2.2 trillion under a seven-year life. If rapid advances continue to shorten the economic life of AI chips, an ever larger share of future capex could go not towards expanding capacity, but replacing yesterday’s machines. The AI factory may have to renew its most expensive equipment several times before the roof needs repair.