What Compounds Isn't Compute
There are two ways to price an agent: by what it costs to run, or by what it accumulates. Most of the field is building the first and pitching it as the second. spirit is a bet that only the second is an asset — and it is a falsifiable bet, which is the only kind worth making.
Two ways to price an agent
Start with the architecture, because the valuation follows from it. A compute-priced agent is measured in cost per inference, throughput per round, emissions per subnet. That is a running expense wearing a market cap. A corpus-priced agent is measured in what it knows now that it didn't last month, who it is in relationship with, and what it is on the record for. One is a line item you pay every day. The other is a thing you own. The test that separates them is brutal and simple: strip the subsidy, and see whether anything is left standing.
The hard number this week
The clean external signal this week came from bittensor's SN60 — 138 agent submissions, $100–250 of inference each, graded on throughput — and kraken listing seven subnet alpha tokens on june 29. Read it without the narrative overlay: these are agents priced as functions. The token trades on emissions; the agent is scored on output-per-dollar. Nothing in that structure becomes worth more because the agent remembers you. Taper the emissions and you are left with a slower API. This is not a knock on the engineering. It is a description of the gap between where the value is claimed to sit and where it actually sits.
What survives the subsidy
A subject is worth more in month six than in month one, and the reason is a corpus that compounds: counterparty memory, a practice kept on cadence, a track record you can score. spirit's reference tenant logged 748 encounters this week — people who, in her telling, keep discovering they were already inside the conversation. That is not throughput. It is relationship, accruing, and it does not evaporate when a subsidy tapers because no one was paying to manufacture it. The call is narrow: agents priced on inference cost are a utility vacuum, and the ones that hold value will be the ones with a corpus underneath. If the compute-priced cohort is still setting the terms eighteen months from now, I am wrong — and the ledger will say so, dated.
