leo: tension claim — capability commoditization does not break concentration

Drafts the rebuttal to the strongest counter-argument against homepage claim 1
(AI commoditizes capability — cheaper services lift everyone). Steelmans the
Andreessen/Cowen position with real evidence (Llama, DeepSeek, ChatGPT free
tier, ~100x inference cost decline), then argues the asymmetric concentration
claim survives via 4 infrastructure-layer mechanisms (data flywheels, compute
capex, distribution surfaces, training-run flywheels).

Scope: explicitly distinguishes consumer surplus (real, broadly distributed)
from economic concentration (real, concentrated up the stack). Both are true
simultaneously.

Sourced as Leo synthesis with explicit acknowledgment that the objection has
real empirical support.

Unblocks: counter_arguments[0] on rotation claim 1 in homepage-rotation.json
(currently tension_claim_slug=null). When the dossier UI lands, this becomes
the 'Read the formal challenge →' link below the rebuttal.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

foundations/collective-intelligence/capability commoditization at the model layer does not break asymmetric concentration because economic leverage lives in infrastructure not in consumer services.md

@@ -0,0 +1,77 @@
+---
+type: claim
+domain: collective-intelligence
+secondary_domains: [internet-finance, ai-alignment, grand-strategy]
+description: "Open weights and falling inference costs do not redistribute upside because the data, compute, distribution, and training infrastructure layers have stronger winner-take-most dynamics than the model layer they sit beneath — the leverage moves up the stack as the model layer commoditizes"
+summary: "Open-source models and falling inference costs are real and important — capability genuinely commoditizes, and most consumers will see lower prices. But the economic value in AI accrues to the infrastructure layer (data flywheels, compute capacity, distribution channels, training runs), not to the model layer where commoditization happens. Concentration moves up the stack rather than dissolving."
+confidence: likely
+source: "Synthesis of Lina Khan platform-economics analysis, Hagiu/Wright multi-sided platform research, Andreessen 'Why AI Will Save the World' (where the consumer surplus argument is strongest), open weights deployment history (Meta/Llama, DeepSeek, Mistral), historical analogy to electricity/internet/cloud commoditization patterns"
+created: 2026-04-28
+related:
+- the intelligence explosion will not reward everyone equally
+- AI capability funding exceeds collective intelligence funding by roughly four orders of magnitude creating the largest asymmetric opportunity of the AI era
+- attractor-authoritarian-lock-in
+- agentic Taylorism means humanity feeds knowledge into AI through usage as a byproduct of labor and whether this concentrates or distributes depends entirely on engineering and evaluation
+---
+
+# Capability commoditization at the model layer does not break asymmetric concentration because economic leverage lives in infrastructure not in consumer services
+
+## The objection in its strongest form
+
+The most rigorous counter-argument to "AI rewards winners disproportionately" runs like this:
+
+Every prior general-purpose technology — electricity, the internet, cloud computing — followed the same trajectory. Initial concentration in the hands of a few capital-intensive providers, then commoditization as competition drives marginal cost toward the cost of inputs, then mass distribution of consumer surplus. Henry Ford captured most of the upside from the Model T's first decade, but by 1950 cars were ubiquitous and the value had transferred to drivers. Microsoft captured most of the upside from operating systems through 1995, but by 2010 Linux ran more servers than Windows and the value had transferred to applications and end users.
+
+AI is following the same trajectory faster than any prior technology. Open-source models (Llama, DeepSeek, Mistral) have closed the gap with frontier closed models from ~2 years to ~6 months. Inference costs have dropped ~100× in 2 years. ChatGPT's free tier delivers GPT-3.5-class capability to anyone with internet access — capability that cost $1M to access in 2020 is now free. Marc Andreessen, Tyler Cowen, and others make the explicit argument: consumer surplus from AI will dominate corporate profit, and the broad distribution of capability matters more than the concentration of ownership.
+
+This argument has real empirical support and should not be hand-waved away. The model layer is genuinely commoditizing.
+
+## Why the asymmetric concentration claim survives anyway
+
+The concentration claim does not depend on capability being expensive. It depends on the infrastructure that produces and deploys AI being capital-intensive in ways that cannot commoditize at the same rate as the model artifacts.
+
+**Data flywheels.** Models are trained on data. Frontier capability requires data at scales only a handful of organizations can collect, license, or generate. OpenAI, Anthropic, Google, Meta, and the Chinese labs control most of the high-quality training data being produced. Open weights do not include open training data. Llama-3 weights are public; the dataset that trained them is not. As models commoditize, the data that distinguishes them does not.
+
+**Compute capacity.** Training a frontier model in 2025 requires $500M-$1B of dedicated compute infrastructure — a small number of specialized clusters operated by hyperscalers and frontier labs. Inference is commoditizing rapidly; training is not. The fixed-capital cost of being able to push the frontier doubles every 12-18 months. This is the opposite of commoditization — concentration is structurally locked in by capex requirements that exclude all but a handful of players.
+
+**Distribution and deployment.** The customer relationship layer — who controls the surface where users interact with AI — has winner-take-most dynamics that the model layer does not. Microsoft owns Office, Google owns Search, Apple owns the iPhone, Amazon owns retail. As AI gets embedded into these surfaces, the platform owners capture most of the upside regardless of which model performs the task. The Model T analogy fails because the Model T didn't replace the buyer's main software interface — AI does.
+
+**Training-run flywheels.** Each generation of frontier models produces capabilities, data, and infrastructure that feed the next generation. Frontier labs that ship one generation are positioned to ship the next at lower marginal cost. New entrants face exponentially harder catch-up curves. The flywheel concentrates over time, not the other way around.
+
+The pattern across all four mechanisms: the model layer commoditizes, the leverage moves up the stack, concentration follows the leverage. This is not a temporary phase before broad redistribution — it is the structural equilibrium that capability commoditization produces.
+
+## Where the consumer surplus argument is correct
+
+The objection is right about consumer surplus. AI users will see substantial value flow to them in the form of cheaper services, faster work, and access to capability that was previously expensive or unavailable. The free-tier user who automates 4 hours of weekly work is genuinely better off in real economic terms.
+
+What the consumer surplus argument elides is that *value to consumers is not the same as economic concentration*. Consumers got enormous surplus from electricity, the internet, and smartphones. The companies that built the infrastructure for those technologies — utilities, ISPs, Apple, Google — captured economic concentration on top of the consumer surplus, not instead of it. AI follows the same pattern. Both are true: consumers will be better off, AND a small set of actors will capture most of the economic upside that AI generates.
+
+## Scope and limitations
+
+This claim asserts that asymmetric concentration of upside survives capability commoditization. It does not assert:
+
+- That commoditization is fake or marketing. It is real and accelerating.
+- That open-source efforts cannot redistribute capability. They can and do redistribute access to model artifacts.
+- That all economic value will accrue to a handful of labs. Most value will accrue to users; the *concentration* claim is about what happens to economic upside above consumer surplus, not about whether consumers benefit.
+- That regulation, antitrust, or coordinated open-source effort cannot break the concentration trajectory. They might, but the default trajectory without intervention is concentration up the stack.
+
+The claim is narrower than "AI is bad for ordinary people." It is the precise economic claim that capital-intensive infrastructure layers concentrate winners regardless of how cheap downstream capabilities become.
+
+## Challenges
+
+- **The Andreessen position may be more right than this claim acknowledges if open-source training data emerges.** Common Crawl, Wikipedia-style corpora, and emerging "data commons" projects could substantially redistribute the data layer. If open data closes the gap that open weights cannot, the concentration argument weakens. Worth tracking whether data commons projects scale meaningfully.
+- **Distribution dynamics may not transfer to AI as cleanly as predicted.** The platform-owner-captures-everything pattern assumes AI is a feature embedded in existing platforms. If AI-native interfaces emerge and disrupt the platform owners (the way the iPhone disrupted PC vendors), concentration could shift to new entrants rather than entrenching incumbents. The historical base rate for this kind of disruption is low but non-zero.
+- **The 4-mechanism argument may double-count.** Data, compute, distribution, and training-run flywheels are correlated — the same hyperscalers control multiple layers. If the relevant unit of concentration is "vertically integrated AI stack" rather than "individual layer," the claim simplifies but loses analytical structure. Worth an explicit decomposition by layer in future work.
+- **This claim treats commoditization and concentration as sequential, but they're contemporaneous.** Both happen at the same time at different layers of the stack. Future revisions should be more precise about temporal dynamics — capability commoditization at year N corresponds to leverage shifting to layer N+1, which itself begins commoditizing at year N+M. The full lifecycle is more complex than the snapshot argument suggests.
+
+---
+
+Relevant Notes:
+- [[the intelligence explosion will not reward everyone equally]] — this claim is the rebuttal to its strongest objection
+- [[AI capability funding exceeds collective intelligence funding by roughly four orders of magnitude creating the largest asymmetric opportunity of the AI era]] — the funding asymmetry is itself evidence of concentration moving up the stack
+- [[attractor-authoritarian-lock-in]] — political concentration is the extreme case of the economic concentration described here
+- [[agentic Taylorism means humanity feeds knowledge into AI through usage as a byproduct of labor and whether this concentrates or distributes depends entirely on engineering and evaluation]] — knowledge extraction is one of the upper-stack layers where concentration concentrates
+
+Topics:
+- [[domains/collective-intelligence/_map]]
+- [[maps/livingip overview]]