Research noteAI infrastructure
What I ask a decentralized compute network
The questions I put to networks that sell training or inference from many independent operators. Most of them reduce to one: when a claim of work is false, who notices, and who pays.
Why this is a verification problem
A network that pools GPUs from strangers sells two things: the compute, and confidence that the compute happened as described. The first is close to a commodity. The second is where the design lives, and it is where pitches tend to be vaguest.
The questions
- Who did the work? Can the network name the operator behind each job, and can one operator take the same job twice under two names?
- How is the work checked? Recomputation, sampling, a proof, a trusted execution environment, reputation, or nothing. Each has a cost, and that cost has to stay below the margin the network claims over a data center.
- What does a check cost, and who pays for it? If verification runs at the buyer’s expense, the advertised price is not the real price.
- What happens when a check fails? Slashing only works if the operator has something at stake and the evidence holds up when the operator disputes it.
- What can an outsider confirm today? Public commitments, logs or a chain record that a third party can read without the team’s help. If the answer is a dashboard, I treat the numbers as marketing until I can check them.
- Where is the demand? Jobs submitted or subsidized by the network itself look like usage. I ask for the share of paid work from buyers with no tie to the network.
- What leaks? Training data and weights pass through machines the buyer does not control. I want to know what an operator can see, and what it can keep.
How I use them
They are not a scorecard. A network can miss two of them and still be worth backing, if the team knows which two and has a plan for them. The ones I pass on are the networks that cannot answer the second question in specifics.
Before the call I read what the team has made public and write down what is not clear from it. Those gaps become the first questions.