Centralized vs. Decentralized Intelligence
Whether intelligence is rented from three labs or owned by every user.
Kernel
The centralized scenario is the 2025 status quo: a handful of frontier labs sell intelligence as a metered API. The decentralized scenario is the cypherpunk-adjacent counter-program: local-first models, encrypted agent identities, on-chain coordination, x402 micropayments, MCP marketplaces. Whether the second scenario remains a niche or absorbs a meaningful slice of the agent economy is the most operationally consequential open question of the late 2020s.
Frontline
Local-model performance (Llama-class on consumer laptops by 2025). Agent identity (centralized OAuth vs. decentralized DID/keys). Payments (Stripe-rails x402 vs. crypto-native x402). Coordination (centralized agent registries vs. on-chain marketplaces). Trust (lab brand vs. cryptographic proof).
Doctrine — Centralized
Frontier capability requires capital and infrastructure that only large labs can provide. Distribution is easier through the existing cloud trinity. Trust accrues to brand and operational discipline. Most users do not want to manage keys, models, or wallets; they want the API to work.
Doctrine — Decentralized
Agents acting on behalf of users should not require trust in a third lab. Cryptographic identity, local inference, and on-chain payments make it possible to compose intelligence without rent extraction. Privacy-by-default is the durable competitive advantage. The platform-capitalism era can be skipped at the AI layer if the right primitives ship in time.
Stakes
If decentralized wins even 10–20% of the agent-economy volume, the entire crypto-AI thesis (Psy-class privacy chains, x402 facilitators, agent marketplaces) becomes structurally large. If decentralized stays at 1–2%, the cypherpunk wing of the agent era is a hobbyist subculture for the second time in twenty years.
Outlook
Realistic 2028 distribution: 70% centralized (consumer + most enterprise), 15% sovereign / national clouds, 10% decentralized agent commerce, 5% local-first. The 10% slice is small in share but enormous in absolute terms — large enough to support a healthy ecosystem of privacy chains and agent infrastructure if the technical work ships.