The thing I keep getting stuck on with OpenGradient is that “verifiable AI” sounds much cleaner than it actually is.
At first, I liked the idea for the obvious reason. Developers get to plug AI into apps without dealing with GPUs, model hosting, APIs, or a bunch of messy offchain setup.
But then the uncomfortable part clicked.
The AI workloads people actually care about may not be the ones that can be verified in the strongest way.
The small, simple stuff fits the clean cryptographic story better. But once you get into agents, DeFi, lending, trading, risk models, and larger inference jobs, people start caring a lot more about speed and cost.
And that is where trust sneaks back in.
It does not vanish. It just moves.
Instead of trusting an API provider, you may be trusting a secure hardware environment. That can still be better, but it is not the same as having everything mathematically proven.
What makes OpenGradient interesting is that it sits right at this intersection. The network is trying to make AI computation more accessible and verifiable at the same time, but those goals do not always pull in the same direction.
The more useful and demanding the workload becomes, the more likely it is that practical tradeoffs start mattering.
That is why I find myself less interested in total inference numbers and more interested in the trust assumptions underneath them.
How much of the activity is actually zk-proven?
How much relies on TEE attestations?
That distinction may end up being more important than most people realize.
The more I think about it, the tradeoff itself is not the problem.
The problem would be pretending it is not there.
