Microsoft and Amazon have spent ~$590B combined on AI capex but show AI revenue run rates of $37B and $15B annually, with OpenAI and Anthropic accounting for 70-80% of each.
Key Takeaways
Microsoft’s $37B AI revenue run rate represents ~1.04% of its $293.8B AI capex spend; OpenAI alone accounts for ~71% of that run rate via Azure inference.
Amazon’s $15B AI revenue run rate is ~0.42% of $298B capex; Anthropic likely represents ~80% of that revenue and occupies 75%+ of Amazon’s AI GPU capacity via Project Rainier.
Big Tech (Microsoft, Google, Amazon, Meta) is projected to spend $800-900B on AI capex in 2026 and over $1T in 2027, totaling ~$2T by end of 2027.
Google and Meta refuse to disclose standalone AI revenues, citing vague metrics like “40% QoQ Gemini Enterprise growth” with no denominator and inconsistent Meta ad conversion figures.
CoreWeave’s 2025 revenue was 67% Microsoft-sourced (~$3.45B), used for OpenAI training; Microsoft’s “Fairwater” data centers with hundreds of thousands of GPUs are entirely reserved for OpenAI.
Hacker News Comment Review
Core methodological pushback: comparing current revenue to total cumulative capex ignores that recent capex has not yet produced revenue, making the ROI framing misleading without a time-adjusted model.
Commenters noted the author’s position is internally inconsistent: earlier claims that AI tools were not useful contradict the capacity constraint evidence he now cites as a problem.
General skepticism about the piece’s analytical rigor mixed with acknowledgment that specific structural points (circular revenue, hyperscaler dependency on two customers) are worth scrutinizing.
Notable Comments
@kev009: Argues the author conflates absurdity of current unit economics with long-term industry failure, drawing a parallel to .COM-era skepticism.
