NVIDIA CEO Jensen Huang claimed Blackwell GPUs would reduce LLM inference costs by 25x. That number has since been quietly revised to 10x, verified only against private inference providers who are not profitable and do not disclose margins. Nobody in mainstream media called this a 150% overstatement. The new Vera Rubin chip now promises 10x more tokens per megawatt, benchmarked against an 18-month-old DeepSeek R-1 model. The pattern is worth noting.
Hyperscalers will have spent over $1.3 trillion on generative AI infrastructure by end of 2026, with another trillion planned for 2027. Meta, Google, Amazon, and Microsoft have nearly tripled their property, plant, and equipment since March 2022, from $498 billion to over $1.3 trillion, while adding $307 billion in on-balance sheet debt. Off-balance sheet obligations add another $1.35 trillion, per Nikkei. Google alone carries $811 billion in contracted future spending commitments, up $661 billion in six months. Capex now consumes 43.4% of Google revenue, 37.3% of Microsoft, 33.7% of Meta, and 24.4% of Amazon. None of these companies have disclosed actual AI revenue.
The bond markets are showing early stress. Amazon's $25 billion July bond sale attracted only 1.6x demand, against an industry average of 4x. Google's 100-year century bond has already lost 10% of its face value. Bloomberg Intelligence analyst Herman Chan estimates hyperscalers will need to raise $1.5 trillion in investment-grade debt over the next five years to sustain current spending. The full piece builds the math on what $2 trillion in brand-new revenue would actually require, and why no current LLM application gets close.
[READ ORIGINAL →]