Where AI Value Lands

Source: AI Redistribution of Economic Surplus — The Skeptics and Bear Case

5 min read

AI Redistribution of Economic Surplus — The Skeptics and Bear Case

Citation

Internal synthesis document drafted in preparation for Where Value Lands. Surveys and braids together four priority skeptic works: acemoglu-simple-macroeconomics, goldman-sachs-too-much-spend, marcus-world-models-failure, zitron-circular-economics. Raw file: raw/AI Redistribution of Economic Surplus — The Skeptics and Bear Case.md.

Central claim

The dominant narrative — that generative AI and AGI will radically redistribute economic surplus, double GDP trajectories, and rewrite labor economics — is not supported by rigorous evidence across four interlocking pillars: cognitive/architectural limits, labor-economic ceilings on productivity, ROI shortfall versus a ~$1T capex cycle, and a structurally circular product-layer economy. The acute risks are not sentient AI or mass job destruction — they are capital misallocation, stalled aggregate productivity, and widening capital-labor inequality.

The four pillars

  1. Capability overestimation. LLMs are sophisticated next-token predictors, not reasoning agents. They lack causal world models and compositionality. The “scaling law” hypothesis underpinning trillion-dollar capex is contradicted by diminishing returns at the frontier. See scaling-wall.
  2. Productivity paradox. A task-based framework grounded in Hulten’s theorem caps 10-year US TFP gains at ~0.55–0.71% — orders of magnitude below McKinsey/GS optimistic forecasts. Only ~4.6% of US tasks are profitably automatable. See task-based-framework.
  3. ROI shortfall. ~$1T in hyperscaler+semi capex is not matched by enterprise-grade value creation. AI is a high-cost solution applied to low-value tasks, inverting the historical pattern of disruptive tech. Energy and grid constraints compound the problem. See ai-capex-cycle (folded into circular-ai-economy).
  4. Circular product-layer economics. 70–80% of hyperscaler “AI revenue” is consumed by OpenAI and Anthropic, themselves funded by VC and recycling that cash back to hyperscalers. Application startups operate at deeply inverted margins (Perplexity ~164% of revenue on compute). See circular-ai-economy.

Key quantitative anchors worth preserving

  • ~19.9% of US labor tasks theoretically exposed to gen AI; only ~23% of those are profitably automatable → ~4.6% of all US tasks.
  • Average task-level cost saving: ~15.4% → cumulative 10-year TFP ceiling 0.55–0.71%.
  • Hyperscaler historical capex ~1.7T.
  • Revenue required to break even: ~6T+.
  • Combined annual revenue of MSFT+META+AMZN+GOOG (all products): ~$1.599T.
  • OpenAI: ~50% of revenue on inference + ~75% on training; burned 5B in one year.
  • Perplexity: ~164% of revenue spent on compute in 2024.
  • 70–80% of MSFT/AMZN “AI compute capacity” consumed by OpenAI + Anthropic.

Stance — what this contributes to the inquiry

  • Provides the bear-case anchor the paper plan flagged in 05_open-questions.
  • Validates a specific shape of H2_u-curve-of-value: it strongly supports the middle dies claim (mid-stack startups bleed margin) but threatens the bottom: hyperscalers may be holding rapidly depreciating, mispriced assets. The bottom of the U may not be safe either.
  • Pressures H1_L0-L7-ladder: if the scaling wall is real, the autonomy gains imagined at L4/L5 may never materialize at production reliability, and L7 (intelligence as commodity) becomes a bet on a paradigm shift, not a glide path.

Contradictions and tensions surfaced

  • The U-curve hypothesis assumes the bottom of the stack is durably profitable. Zitron’s circular-economy thesis argues the bottom is currently propped up by VC pass-through, not enterprise revenue. If true, the U is upside-down: middle dies, but bottom dies too once the music stops.
  • Acemoglu’s productivity ceiling implies a smaller pie to fight over. If TFP gains over 10 years are ~0.7%, the “redistribution of economic surplus” framing of the paper needs to grapple with the possibility that there is no surplus to redistribute — only a reshuffling of capital from VCs to GPU vendors.
  • The “killer app has not emerged” critique directly threatens any application-layer winner-take-most thesis embedded in H3_orthogonal-axes-under-priced.

Sources surveyed inside this document

Suggested next moves

  • Read each of the four primary sources directly and ingest them as separate sources to verify the synthesis above.
  • Run /discover to surface a counter-bear-case (the strongest bull synthesis) so the paper does not become one-sided.
  • Once both sides are in the wiki, revisit H2_u-curve-of-value confidence and consider whether the “U” framing survives or needs to be reshaped (e.g., into an L, an inverted U, or a noisier landscape).

Graph View

Backlinks