Hyperscalers have committed $1.5T+ in AI capex through 2027 but current AI revenues are nowhere near the $3-6T needed to break even.
Key Takeaways
Microsoft spent ~$100B on OpenAI partnership alone, roughly 30% of its total capex since FY2023, yet estimated FY2025 AI revenue is only ~$17.9B.
Outside OpenAI and Anthropic, hyperscaler RPO growth is nearly flat; those two companies account for the majority of Microsoft, Google, and Amazon’s AI revenue backlogs.
Microsoft 365 Copilot has 20M subscribers, generating at most $7.2B annually at full price – against $88B in annual capex.
LLM costs scale linearly with usage: more revenue requires more GPUs, more data centers, more time, making margin improvement structurally difficult.
Anthropic at $45B annualized revenue (with zero opex) would not cover a single year of any one hyperscaler’s 2024 capex.
Hacker News Comment Review
Commenters split on the Uber analogy: some see deliberate subsidized land-grab with a plausible payoff if white-collar automation materializes; others note that Uber’s unit economics were at least directionally improvable, while LLM costs track revenue.
A recurring point: current cheap tokens are a subsidy window for builders, expected to end with either price hikes or advertising, consistent with rumors around Gemini Flash pricing changes.
Skeptics note developers willingly pay for AI tooling at prices they would never pay for equivalent human services, but whether that willingness survives subsidy removal is unresolved.
Notable Comments
@changoplatanero: frames spend as a rational attempt to dethrone Google given its accumulated capital moat, treating $500B+ as table stakes.
@bensyverson: “This is a golden era of subsidized tokens. It will not always be like this.”
