2026 Q3-Q4 — Stable Surface
OpenAI $40B raise closes.
Stock prices for public AI companies maintain. Revenue growth cited. No visible crack yet. Everyone points to Anthropic's profitability as validation of the whole sector.
Five structural walls. When any one fails, the bubble cracks. The cross-ownership network means failure propagates fast. Here's how much pressure each wall is under — right now.
Anthropic earned $559M operating profit in Q2 2026 (~$2.2B annualized) against a $965B valuation. To justify that valuation at a 10% discount rate, profit must grow to ~$96B annually within a decade — a 44x increase from today's run rate. This is possible given 130% QoQ revenue growth. It is not guaranteed.
Choose your depth. The data doesn't change — just the explanation.
The bubble cracks when ANY ONE of these fails. The cross-ownership network means failure propagates fast — Microsoft owns OpenAI, Amazon invested $8B in Anthropic, and they all rely on each other's revenue.
OpenAI: $25B ARR at 30x P/S. Anthropic: $44B ARR at 22x P/S but growing 130% QoQ. Dot-com peak was 30x P/S at DECELERATING growth. Anthropic's acceleration changes the math. OpenAI's losses don't. Status: YELLOW — elevated but not breaking.
CONFIRMED: agentic workloads consume 10-100x more tokens than chat. Enterprise budgets exploding. API prices falling 40%+ YoY. The math for replacing an $80K knowledge worker requires economics that don't close yet. This wall is under serious pressure.
Virginia already imports electricity for data centers. Texas ERCOT hits reserve margin wall by 2026-2027. FERC interconnection queue at record highs. Power permits being denied. The physical build-out cannot keep pace with announced demand. This wall shows visible cracks.
Microsoft invested $13B+ in OpenAI. Amazon invested $8B in Anthropic. OpenAI revenue includes $865.8M from Microsoft (circular). If OpenAI corrects 50%, Microsoft's balance sheet takes a hit, which affects Anthropic's fundraising, which affects Google. One failure cascades.
Physical infrastructure sabotage increasing. EU AI Act accelerating researcher exodus. 2028 election likely to feature aggressive AI regulation as major platform. Public sabotage and regulatory backlash hasn't hit investment calculus yet — but the pressure is building.
The open-source threat is arguably the most underpriced risk in AI valuations. Llama 4 and Mistral are within 15-20% of frontier performance at near-zero marginal cost. If open-source closes the gap entirely, frontier lab API revenue may compress to near-zero — the same dynamic that destroyed enterprise software margins in the 2000s.
Dot-com peak: 30x P/S with decelerating growth. OpenAI is there now. Anthropic's growth offsets its multiple. xAI is in a different category entirely.
GPT-4o, Claude Sonnet, Gemini 2.5 Pro all cutting prices. Confirmed and accelerating. Open-source substitution pressure intensifying.
From confidential projections reported in financial press. Revenue growing but losses accelerating. Cash flow positive projected 2030 — internally. Outside analysts say 2032-2033.
Stock prices for public AI companies maintain. Revenue growth cited. No visible crack yet. Everyone points to Anthropic's profitability as validation of the whole sector.
Revenue growth decelerated. Sophisticated investors model cash flow positive at 2032-2033, not 2030. Probability of down round: 50%. Probability of aggressive cost cuts: 70%. Narrative begins to shift.
xAI (133x), Cohere, AI21, Inflection — 2-4 of these either shut down, get acquired for pennies, or pivot radically. 80% probability by 2028. Stability AI: 90% probability of casualty.
Called "disciplined capital allocation." Everyone knows demand projections didn't materialize on schedule. When this happens: AI sector corrects 25-40%. Physical constraints (power, water, permitting) make the revision inevitable.
Not a bubble pop — every technology market consolidates this way (browsers, search, social, cloud). OpenAI's $750B valuation is 2-3x too high. $300-400B by 2030 is plausible if they hit numbers. The technology remains real and growing.
The dominant AI narrative fails its first major test in 2027-2028. Not a collapse. A failure of the specific claim that AI agents will replace most human knowledge workers within 3-5 years. The compute economics close for rote tasks — but the task complexity ceiling is real. Agents that autonomously write code still fail on judgment-intensive tasks requiring actual world understanding. The real displacement peak is 2030-2035, not 2026-2028. When this becomes undeniable in Q2-Q3 2027: narrative compression, valuation compression on the companies that priced in the aggressive timeline.