Jessica Wachter, a finance professor at Wharton and former SEC chief economist, cuts through the AI hype with a stark accounting reality check. She estimates that hyperscalers will spend nearly $1.1 trillion on data center infrastructure by 2027. To justify this outlay and achieve a 15% return on capital, these companies must boost their own productivity by a factor of 2.7 by 2030. This isn’t just ambitious; it’s historically unprecedented in its compression. Wachter warns that if this productivity boom fails to materialize, the current buildout risks becoming the largest misallocation of capital in history, potentially leading to bankruptcy for those unable to service their debt.
The disconnect between spending and revenue is already glaring. Gary Gensler, former SEC chair and now an MIT professor, points out that while hyperscalers plan to drop $750 billion this year alone, total AI revenues sit between $150 billion and $200 billion. Alphabet, Microsoft, Amazon, Meta, and Oracle are betting the farm on future demand that remains unproven. Even cash-rich giants like Alphabet reported a $5.9 billion free cash flow deficit in the latest quarter, marking its first shortfall since going public in 2004. The era of endless free cash is over, replaced by heavy borrowing and negative free cash flow projections for the group.
This financial strain extends beyond balance sheets into the broader US economy, where AI investments could soon represent 3% of GDP. The risk isn’t merely that returns will be modest, but that they won’t cover the rising cost of capital as debt loads swell. If demand for compute power stagnates or models become more efficient, reducing the need for raw hardware, hyperscalers will still owe billions in interest payments. Investors are growing impatient, and the financial mechanisms passing these loans through the system mean the fallout could ripple well beyond Silicon Valley.
We are witnessing a bet on exponential growth that assumes AI will replicate the IT boom of the 1990s, but in a fraction of the time. The technology might deliver, but the timeline is unforgiving. If the productivity gains don’t arrive quickly enough, the physical monuments of this era—massive, energy-hungry data centers—could become stranded assets. The question isn’t whether AI is useful, but whether it can generate cash fast enough to pay for the pipes delivering it.
