Market thesis
What the model market actually needs
Open weights are abundant. Proprietary data is not. The next wave of value is domain fine-tunes sold as API access, with governance, gating, and per-token settlement that enterprises can audit.
The shift
Teams no longer only consume frontier closed APIs. They take open bases (Llama, DeepSeek, Mistral-class, and peers), continue train on private corpora, and productize the result. Hugging Face is distribution for weights. Inference hosts sell raw capacity. What is missing is a marketplace where the specialist capability is the SKU, the host keeps compute and weights, and buyers get one key with clear access policy.
Problems buyers and hosts hit
No clean way to resell a fine-tune without leaking it
Shipping weights or a naked endpoint invites cloning. Hosts need proxied access, approval queues, and competitor domain blocks.
Provenance is becoming procurement hygiene
Enterprises increasingly ask where a model came from: base weights, fine-tune lineage, license, usage restrictions, and security signals. Catalogs like Cisco's AI Supply Chain Provenance Explorer make this explicit for open models.
Metering is a stream of micro-payments
Agent workloads fire thousands of tiny bills. Monthly invoices lose the real-time audit trail. Chains not built for microtx make per-call settlement uneconomic.
Aggregation is not the same as specialist markets
OpenRouter-style routers unify frontier APIs. Bittensor-style networks price decentralized commodities. Neither is a full host-owned specialist marketplace with enterprise access policy + transparent fee split.
Where blockchain helps
- Usage + settlement receipts. Append-only events for every billable completion and the three-way platform / model / compute split. Litheum targets this microtransaction regime rather than large-transfer L1 patterns.
- Listing and credit integrity. Marketplace listing events and credit grants become auditable state transitions, not opaque admin rows.
- Provenance anchors. Hashes of model cards, dataset licenses, and evaluation bundles can be pinned on-chain while heavy artifacts stay off-chain (object storage / IPFS-style). Industry work on model provenance and training-data lineage is converging here.
- Agent-native payments. DeFAI agents that buy inference continuously need a rail where fee overhead does not dominate unit economics.
What should stay off-chain
Approval queues, allowlists, competitor email domain blocks
Heartbeats, regions, degraded status, private VPC endpoints
OpenAI-compatible /v1 paths, mg_ keys, playground, SDKs
Credits, pricing pages, host USP and data-moat storytelling
Human review of access requests, policy notes, org identity
Weights, eval sets, long logs (hash on-chain, store off-chain)
Cautus Intelligence product mapping
- Hosts publish OSS or fine-tuned listings on their own compute.
- Buyers discover models with origin, USP, gates, and live status.
- Access policy is enforced before inference (open / approval / allowlist + blocks).
- Gateway meters tokens and debits credits.
- Litheum records usage and settlements for continuous auditability.
Research context informing this thesis includes enterprise demand for AI model provenance and supply-chain visibility, open-model ecosystem growth, blockchain applications for training-data and model lineage, and decentralized AI market experiments. Cautus Intelligence is a product demo, not a claim of exclusive category ownership.