NVIDIA CEO Jensen Huang promised 25x cost and energy reductions from Blackwell GPUs. Two years later, that number is 10x, based on case studies from private inference providers that do not disclose margins and are not profitable. Nobody in the press called this out. The overstatement was 150%. It did not make a single AI startup profitable, and it did not lower measurable costs in dollars.
Hyperscalers will have spent over $1.3 trillion on generative AI by end of 2026, with plans for $1 trillion more in 2027. Between March 2022 and July 2026, Meta, Google, Amazon, and Microsoft nearly tripled their property, plant, and equipment, from $498 billion to over $1.3 trillion, while none disclosed actual AI revenue. Capital expenditures now consume 24.4% of Amazon revenue, 33.7% of Meta, 37.3% of Microsoft, and 43.4% of Google. On-balance sheet debt doubled to $557 billion. Off-balance sheet exposure, per Nikkei, tops $1.35 trillion. Google alone carries $811 billion in contracted future spending commitments, up $661 billion in six months.
The core argument here is not that AI is bad technology. It is that the math does not work at any scale. Over $2 trillion in brand new revenue would be required to justify this buildout, and nothing LLMs currently do, have done, or are projected to do comes close to producing it. The original piece goes deeper into the bond markets, SPV structures, and how hyperscalers are engineering their balance sheets to obscure the true weight of what they are carrying. Read it for the mechanism, not just the conclusion.
[READ ORIGINAL →]