Is the AI rally a genuine productivity boom or a

The valuation gap between the hyperscalers and everyone else has widened to a point that feels uncomfortable even for longtime tech bulls. Nvidia's data-center revenue has roughly tripled in two years, yet a meaningful slice of that demand comes from the same handful of companies — Microsoft, Meta, Google, Amazon — that also happen to be Nvidia's largest shareholders or strategic partners. When those buyers are simultaneously investing in each other's model labs (OpenAI, Anthropic, xAI) and renting each other's cloud capacity, the cash starts to look like it's circulating inside a closed system rather than flowing from end-user adoption.

What would actual evidence of a bubble look like? A few concrete signals I'm watching:

• Revenue concentration: If more than 60 % of Nvidia's data-center sales still trace back to fewer than five accounts after the next two quarters, the "broad-based demand" narrative frays.
• CapEx payback horizons: Hyperscalers are guiding $200 B+ annual capex. At current inference pricing, the implied payback on a single H100 cluster stretches past 36 months — longer than the typical depreciation schedule.
• Model-api revenue vs. training spend: OpenAI and Anthropic together reportedly generate ~$4 B annualized API revenue while burning multiples of that on compute. The gap is being filled by equity rounds whose valuations assume near-monopoly pricing power forever.
• Secondary-market liquidity: Employee tender offers at Flatiron/Forge have slowed; when insiders can't exit at mark, the mark becomes aspirational.

None of these alone proves a bubble. Together they resemble the telecom build-out of 1999-2000 — massive infrastructure deployed ahead of verifiable demand, financed by circular equity swaps among the builders themselves.

The counterargument is straightforward: inference costs are dropping 4-5x per year, new modalities (video, agents, coding) unlock use cases that didn't exist six months ago, and enterprise contracts are shifting from pilots to multi-year commits. If that adoption curve holds, today's capex looks prescient, not excessive.

My read: the infrastructure layer (silicon, networking, power) is probably overbuilt for 2024-25 demand but correctly sized for 2027-28. The application layer is where the bubble risk lives — dozens of "AI wrapper" startups raising at 50x ARR with zero moat. That segment will compress hard; the hyperscalers and Nvidia will just grow into their multiples more slowly.

Curious what metrics others track to distinguish "expensive but justified" from "detached from fundamentals."