Where AI Value Lands

H2 — The U-Curve of Value

18 min read

H2 — The U-Curve of Value

Vertical-domain corroboration (ER&I)

deloitte-ai-dossier-eri supplies the first non-software vertical instantiation of the U: in industrial verticals, the top of the U is the multi-agent operational orchestrator (vertical-ai-orchestration) and the bottom of the U is the proprietary operational data + physical assets the orchestrator acts on. Generic foundation models and generic agent frameworks are the squeezed middle. The same structural shape recurs in a completely different industry, which is mild evidence that the U is not purely a SaaS-stack artifact — but the shape’s universality across blocs (divergent-value-stack-optima) and the surplus-magnitude question (task-based-framework) remain unresolved.

Hypothesis

Economic margin does not distribute linearly across the stack described in H1_L0-L7-ladder. It concentrates at the top (workflow + distribution ownership: who owns the user, the schema, the habits) and at the very bottom (silicon, energy, frontier weights), and evaporates in the middle (generic wrappers, generic orchestrators, generic agent platforms get compressed between commoditizing model capability above and commoditizing infrastructure below).

Status

Working hypothesis. Confidence lower than H1_L0-L7-ladder because the “U” shape is asserted without rigorous economic data behind it. Could turn out to be a different shape entirely: an L (only the top matters), an inverted-U, a fragmented landscape with multiple local maxima, or no consistent shape at all.

The intuition

Two layers are commoditizing fast:

  • Foundation models above — eaten by frontier labs releasing capability for free or near-free.
  • Infrastructure below — eaten by hyperscalers and falling GPU prices.

What survives the squeeze:

  • At the top: things that are not the model — distribution, brand, workflow lock-in, proprietary data, trust, embedded integrations, and verification surfaces (autonomy-slider — Cursor’s diff, Perplexity’s citations). Aggregation Theory applied to AI.
  • At the bottom: things hyperscalers cannot fully commoditize — frontier weights themselves, the silicon, the power. Reinforced by Huang’s “AI factory” framing in karpathy-software-3: gigawatt-scale inference plants where extreme co-design (silicon + networking + energy) makes the substrate capital-intensive in a way that resists commoditization.

The middle — “we’re an AI startup that builds agents” without proprietary distribution or proprietary substrate — is where capital goes to die.

Empirical validation of the top (via where-value-lands-2026)

The strongest evidence yet that the top of the U — distribution ownership, workflow lock-in, proprietary data — is real and measurable:

  • Cursor (Anysphere): 2B ARR. Moat = intelligent model routing + continuous learning pipeline + context-aware codebase adaptation. Michael Truell’s “automobile vs. engine” strategy: while labs build engines, Cursor builds the car developers drive. The data flywheel third-party APIs cannot access is the actual moat.
  • Windsurf: Three-way acquisition battle (OpenAI at 2.4B, Cognition asset acquisition). The fight confirms that the distribution layer — where developer intent is formulated, debugged, and executed — is the most strategically critical AI asset in the current stack.
  • Perplexity (Srinivas): 80/20 on queries; owns direct user relationship; forces model providers to compete on price. Zip2 as the failure-mode counter-example: white-labeling = surrendering leverage.

Critical taxonomy note: This source calls Cursor, Windsurf, and LangGraph “middle stack.” H2 would classify them as “the top” (distribution + workflow ownership). The terminology differs; the underlying claim is compatible. The “thin wrapper” (generic proxy UI) is what dies; the “workflow-embedded platform” is what H2 calls the top. See middle-layer-defensibility for the full distinction.

Partial rehabilitation of sophisticated orchestration: Harrison Chase’s LangGraph case adds nuance: graph-based orchestration with HITL, time-travel debugging, and observability is not a thin wrapper — it is enterprise infrastructure. This suggests H2’s “generic middle dies” should be sharpened to “generic middle dies; sophisticated orchestration infrastructure can survive via context-control.”

Jevons Paradox complicates the bottom: DeepSeek’s efficiency gains (1/70th GPT-4 Turbo cost) did not reduce GPU demand — they expanded it via unlocking previously cost-prohibitive use cases. This partially defends the bottom of the U against circular-ai-economy, though the two are not mutually exclusive. See jevons-paradox-ai.

Bull-frame pressure points (Karpathy / Huang / Steinberger via karpathy-software-3)

The Karpathy source adds two pressures on the U-curve, pulling in opposite directions:

  • Reinforces the top. llm-as-operating-system implies install-base moats compound the way OS ecosystems do; democratization-of-programming implies distribution + workflow ownership matter more than syntax-level production. The defensible top is verification + distribution, not “we built an AI app.”
  • Reinforces the bottom. agentic-scaling-law argues inference (not training) is the load-bearing economic activity. If correct, AI-factory substrate has a margin profile closer to a utility than to a speculative R&D bet — partially restoring the bottom of the U threatened by circular-ai-economy.
  • Threatens the top differently. Steinberger’s “agents eliminate 80% of apps” prediction (same source) says agents talk directly to endpoints, making consumer-facing GUIs redundant. If true, the top of the U is not the verification GUI — it’s whatever the agent-facing API layer becomes. The application layer specializes, doesn’t disappear, but its shape changes.

Structural precursor — Choudary’s horizontal-value thesis (via choudary-ecosystem-teardown)

Choudary’s 2022 four-archetype framework (ecosystem-business-archetypes) is a pre-AI structural precursor to H2. His claim: as transaction costs fall (vertical-to-horizontal-unbundling), vertically integrated value chains unbundle into horizontal layers, and value concentrates in specific horizontal positions. H2 is the AI-era specialisation of the same pattern.

Mapped onto the U-curve:

  • Aggregators and Infrastructures → top and bottom of the U respectively (distribution + substrate).
  • Integrators → bimodal middle: defensible if switchboard-essential (Stripe-class, Plaid-class, Visa-class network effects); squeezed if generic. This partially defends H2 by giving the middle a richer taxonomy than “wrappers”; it partially threatens H2 by surfacing durable Integrator-class plays that capture middle-stack value.
  • Capabilities → either concentrate (when IP-defensible) or commoditise; the wrong-archetype-confusion failure mode (cross-archetype-confusion) drives many failed Capability bets.

The thin-wrapper firms that motivate “the middle dies” are not an archetype in Choudary’s taxonomy at all — they are failed Capability attempts. This sharpens H2: the squeezed middle is the population of firms that tried to be Capabilities without IP / data advantage and failed, plus the population of generic Integrators displaced by switchboard incumbents.

Seller-side instantiation — Weber’s four AI-startup archetypes (via weber-ai-startup-business-models)

Weber et al.’s empirical four-pattern taxonomy of AI startups (ai-startup-business-archetypes-weber) maps cleanly onto the U:

  • ai-charged-product-service-provider (26/100 of Weber’s sample) → top of U. Readily-trained model embedded into a standardized product; value depends on workflow + distribution. Cursor / Windsurf / Perplexity are 2026 instances. Empirically validated.
  • ai-development-facilitator (25/100) → the squeezed middle. Generic SDKs and chatbot-builders are exactly what gets compressed. Survives only via middle-layer-defensibility niches (Hugging Face’s distribution; agent frameworks).
  • ai-data-analytics-provider (30/100, largest cluster) → shoulder of U. Classical analytics with ML inside; sustainable B2B SaaS economics but not where the new AI margin lands. Mild counter-evidence to the strong form of “middle dies” — the population is large and profitable. The hypothesis should distinguish firm count from value-capture share at exit.
  • ai-deep-tech-researcher (19/100) → bottom of U when defensible; merges with H4_rl-specialization-value-pocket in the 2026 RL-fine-tuning sub-pattern.

Net effect on H2: the four-archetype mapping reinforces the qualitative shape (top + bottom > middle) while complicating the strong form — the middle is densely populated and not uniformly dead. The U-curve is best read as a value-capture curve, not a firm-count curve.

Historical analogies (to be tested, not assumed)

  • Intel vs. PC OEMs in the 90s — chip and OS captured margin; assemblers in the middle were squeezed.
  • Stripe vs. acquirers / processors — developer-experience layer on top + card networks on the bottom; middle compressed.
  • AWS vs. independent SaaS — infra on bottom, branded apps on top, undifferentiated “platforms” between get acquired or wither.
  • Aggregation Theory (Ben Thompson) — distribution-side scale beats supply-side scale once supply is commoditized.

Why this might be wrong

  • The middle layer may not be undifferentiated. There may be real moats in agent orchestration, evals, observability that the U-curve framing dismisses too easily.
  • The top may not be defensible either — if frontier models keep absorbing capabilities, the application layer itself collapses into the model.
  • “U” may be too clean a shape. Reality could be noisy, with multiple local maxima depending on vertical.
  • The historical analogies may be cherry-picked. Counter-cases: telcos (margin in the middle, not the edges), credit card networks (a middle layer that captured margin durably).
  • The bottom may also be hollow. The bear case in bear-case-synthesis argues — via circular-ai-economy — that hyperscaler “AI revenue” is largely recycled VC capital, not durable enterprise demand. If so, the U is upside-down: middle dies and bottom dies, leaving only the top (distribution + workflow) actually defensible. The shape is then closer to an L, not a U.
  • There may be no surplus to redistribute. task-based-framework caps 10-year US TFP gains from AI at <0.71% cumulative. If correct, the absolute size of the surplus is small enough that any “U” shape is operating on a fraction of the prize the optimistic narrative assumes — capital reshuffles among GPU vendors and VCs rather than productivity expanding the pie.
  • Enterprise-side evidence is ambiguous. wef-ai-in-action-2025 reports a scaling-gap of 74% of firms struggling to scale AI and only 16% prepared for enterprise reinvention. This is interpretation-flexible: it can be read as managerial (closeable with enablers — the WEF stance), bear-structural (most enterprise AI does not deliver — compatible with bear-case-synthesis), or middle-dies (most internal initiatives are thin wrappers without distribution or substrate — compatible with middle-layer-defensibility). The same number supports all three readings, which means the gap is suggestive evidence but not yet adjudicating.
  • Bull projections worth triangulating against. WEF / Accenture project $7.6–17.9T global AI contribution by 2038 — a figure incompatible with task-based-framework’s <0.71% TFP cap. The source-of-projection (consultancy with commercial interest in scaling) should be weighed; useful as the canonical bull-multilateral number to contrast against.
  • The shape may not be universal. divergent-value-stack-optima argues — drawing on geopolitics-global-ai-divide — that the U-curve is plausibly a US-stack artifact. China’s open-weight asymmetry (open-weight-asymmetry) deliberately destroys the bottom of the U to force value downstream into application and B2B integration: a bottom-light, application-heavy shape. Europe attempts a third shape — value at the compliance and sovereign-infra layer (brussels-effect-and-mirage, sovereign-ai) — that does not resemble a U at all. If the where-does-value-land answer is bloc-specific, the single-curve framing is the wrong unit of analysis.
  • The bottom-right of the U may be thin where the population is thickest (SME-side floor effect). oecd-sme-ai-adoption-2025 documents a sme-ai-adoption-gap of 40% large vs 11.9% small firms using AI OECD-wide in 2024. Only 29% of SMEs using generative AI deploy it in core activities; the rest is peripheral (marketing, drafting, customer service). The long-tail SME application population that the U’s bottom-right relies on is shallow in depth and concentrated in capability-poor sectors (sectoral-ai-diffusion-pattern: ~7% in construction, hospitality, transport vs ~45% in ICT). Three further empirical findings, gated by the oecd-sme-enabler-quartet, narrow the tail: If SMEs settle at the oecd-sme-adopter-taxonomy Novice / Optimiser corner — wide-but-shallow off-the-shelf use — the application-side surplus pools at a small number of digitally mature large firms and AI-intensive sectors, not at the broad SME base. The U may be asymmetric in the SME dimension: top captured by a small Champion-adjacent population, large middle of Optimisers paying recurring SaaS rent upward. The US has the thinnest divide and likely the fattest application-layer tail; the EU bloc, despite richer policy, has the widest divides. This doesn’t refute the U, but it makes the application-layer-as-broad-base claim contingent on policy, not architectural.

Confidence

Confidence varies sharply across the four sub-claims of the U. Treat them separately.

  • Top of U (high) — empirically validated by the application-layer revenue and retention pattern documented in where-value-lands-2026 and reinforced by the vertical-domain instantiation in deloitte-ai-dossier-eri. Cursor, Windsurf, GitHub Copilot, and the vertical-ai-orchestration class show durable workflow + distribution lock-in. See application-layer.
  • Middle claim (medium) — qualitatively reinforced (Karpathy, Steinberger, middle-layer-defensibility) but the firm-count vs value-share distinction is unresolved. Many middle-layer firms exist (the trough is shallow by count) but aggregate value capture is unclear; some monitoring / orchestration / evals sub-pockets may survive.
  • Bottom of U (under pressure) — challenged by circular-ai-economy dynamics: if hyperscaler “AI revenue” is largely recycled VC capital, the bottom is hollow and the U collapses toward an L. task-based-framework’s <0.71% 10-year TFP cap compounds the concern. Frontier-lab defensibility (scaling-wall, open-weight-asymmetry) is the counter-pull.
  • Universal shape (contested) — divergent-value-stack-optima, drawing on geopolitics-global-ai-divide, argues the U is plausibly a US-stack artifact: China’s open-weight asymmetry deliberately flattens the bottom, Europe centres value at the compliance / sovereign-infra layer. Treat the U as a default shape that can flatten or invert per bloc and per vertical.

Labor-side mirror (via massenkoff-mccrory-labor-market-impacts-2026)

The value-stack U has a labor-exposure twin on a different axis. Anthropic’s 2026 internal study finds AI exposure concentrated on the upper half of the wage distribution (computer programmers 74.5%, customer service reps 70.1%, financial analysts 57.2%, market research analysts 64.8%) while 30% of workers — manual physical labor (cooks, mechanics, lifeguards, bartenders) — have zero exposure. Exposed workers are older, more female, more white/Asian, more educated (4× the graduate-degree share), and earn 47% more per hour than unexposed ones (see exposed-worker-demographics).

This produces a labor-exposure shape that is also U-like but on a different axis: high-skill knowledge work at one end, manual physical labor at the other, semi-skilled service work compressed in the middle. The two U’s interact:

  • Surplus-side. Capital owners of L4–L6 application-layer firms (Cursor, GitHub Copilot, Anthropic, LangChain customers) capture surplus by automating tasks performed by the highest-paid labor on the curve. The labor-side incidence is the inverse of prior automation waves: this one hits the educated, not the routine.
  • Detection-side. The unemployment-did-exposed-workers DiD is flat post-ChatGPT (+0.002, SE 0.002, insignificant) — labor-side displacement is invisible in unemployment at the 1pp differential detectability threshold. The only detectable signal is ai-young-worker-hiring-slowdown: a ~14% drop in 22–25-year-old job-start rates in exposed occupations. Capital captures surplus; senior incumbents retain jobs; the entry-level cohort bears the suppressed-hiring cost.
  • Diffusion-side. theoretical-vs-observed-capability-gap shows actual Claude deployment is ~3× behind theoretical capability in the most-exposed category. As that gap closes, the labor-side incidence scales — the U-shape sharpens.

For Where Value Lands, this updates the central claim: value lands on capital at the top of the value-stack U, with the cost borne not by the bottom of the labor distribution but by the entry-level cohort of the top half. The redistribution story is intra-class (incumbent vs entrant within knowledge-work) before it is inter-class.

What would retire this hypothesis

  • Three or more profitable, durable, mid-stack players emerge with no distribution moat and no substrate moat — pure “middle of the stack” winners.
  • The historical analogies don’t hold up under scrutiny.
  • Interviewed operators consistently report value capture in the middle.
  • A correction confirms circular-ai-economy: hyperscalers’ AI revenue line items contract sharply when VC funding tightens, validating that the bottom of the U was funded by recycled capital, not enterprise demand. In that case the U is the wrong shape — likely an L (only the top matters) or a collapsed flat during correction.
  • The three blocs durably diverge. If US, China, and EU stacks settle into the three different shapes described in divergent-value-stack-optima (top-heavy / bottom-heavy / regulatory-centric), the universal-U framing must be replaced by a per-bloc analysis. The paper’s Section 5 then becomes a structural cross-cut of Section 3, not a sidebar.
  • The SME / firm-size divide either narrows or widens decisively. If longitudinal SME data shows the firm-size adoption divide narrowing (small firms catching up at the application layer) within five years, the U-shape’s bottom tail becomes credibly populated. If the divide widens — or stays flat as sme-broadband-firm-size-gap suggests — the U risks collapsing into an L-shape concentrated at the infrastructure top and the largest application-layer adopters. The persistence of the divide across multiple years and the country variance (1.65× US → 4.7× Italy) make the application-layer-as-broad-base claim contingent on policy, not architectural.

Referenced by

Auto-generated reverse-link index — pages in the wiki that link here. Maintained by /lint and reverse-link sweeps.

Paper planning

Hypotheses

Search

Concepts

Sources

Ideas

Data

Thoughts

Graph View

Backlinks