Beyond LLM inference
LLM inference is the workload most acutely in need of fair, fine-grained payment — but TAP's primitives generalize. The same channel construction, the same bilateral halt, the same x402 bootstrap apply wherever value is delivered as a continuous flow and the consumer benefits from being able to halt mid-flow.
This section sketches each adjacent application: what changes, what doesn't, and where the protocol's shape requires rethinking before it maps cleanly to a new domain. We treat these not as commitments but as an honest read on adjacency — what the substrate actually supports.
What's the same across applications
The on-chain program, the channel state machine, and the x402 bootstrap are unchanged. What varies is the per-application metadata published at session open (what the unit is, what the price is, what halt signals make sense), the evaluator (what indicates quality issues for this content type), and the producer wrapper that adapts the protocol to the underlying service's streaming format.
Index
| Application | Unit | Halt example | Page |
|---|---|---|---|
| Video streaming | per-second | viewer pauses; quality drops | Video |
| Audio (TTS, music) | per-second | encoder error; off-topic spoken response | Audio |
| Cloud compute & GPU rental | per-second | training diverges; container OOMs | GPU rental |
| Metered APIs (search, geospatial, DB) | per-unit-of-work | result-set limit reached | Metered APIs |
Where the substrate stops being a fit
Three honest cuts:
- Atomic deliveries. Generated images, complete audio files, compiled artifacts have no value as a partial. Streaming + halt does nothing for them. The whitepaper proposes hash-locked atomic exchange as a v2 extension; that's a different cryptographic recipe on the same channel.
- Sub-cent per-session aggregate spend. TAP's two-tx amortization (open
- settle) costs around $0.004 in fees. Below ~$0.005 of session spend, direct one-shot x402 payment is more efficient. Channel reuse fixes this for high-volume consumers; for one-shot tiny payments, TAP is the wrong tool.
- Privacy-sensitive workloads. Channel openings link consumer and producer pubkeys on-chain. Medical inference, legal queries, and other privacy-load-bearing flows need application-layer privacy (stealth addresses, blinded routing) that v1 doesn't provide.