
An analysis published on October 26, 2026 put the position plainly: for large language models, hardware attestation is the realistic option today, while optimistic verification is cheap and adequate for open models where privacy is not needed, and zero-knowledge machine learning may stay impractical for years. That is a clear statement of what works. It is also, read carefully, a statement about which models each technique works on, and that distinction carries the whole thesis.
The cost spread is not a detail. Zero-knowledge machine learning has super-linear proving overhead, making it infeasible for billion-parameter models: the leading protocol generates a proof for one forward pass of a 13-billion-parameter model in about 15 minutes, and a 100-token answer needs 100 such proofs, putting a short reply near 25 hours. A March 2026 benchmark of five zkML libraries ran on a 3.4-million-parameter network, four orders of magnitude below the models in question.
Optimistic verification inverts the trade. Inference runs at near-native cost and is assumed correct unless challenged, which is cheap, and the dispute window introduces latency that breaks composability for anything downstream. It also requires reproducing the full inference trace during a challenge, which rules it out wherever the input or the weights are confidential.
Trusted execution is the production answer at roughly 5 percent overhead, hardware-attested, available on current datacenter accelerators. Per-query cost comparisons put the hybrid attested approach near $0.07 against about $0.06 for optimistic and $45 or more for zero-knowledge. On cost there is no contest, and that is why attestation is what is deployed.
Note where the settlement costs land, because the thesis name points at the wrong half. Verifying a hardware attestation signature on-chain costs on the order of 40,000 gas. An optimistic claim costs around 21,000. A zero-knowledge verifier costs roughly 350,000.
All three are affordable. Settling a proof of inference on-chain is a solved engineering problem at any plausible volume, and nothing about the blockchain leg is the constraint. A contract can check a signature chain and record the result for a fraction of a cent.
So the interesting question was never whether the verification could be settled on-chain. It is what the verified statement actually says, and a signature check tells a contract only that some attested hardware vouched for some measurement. Everything that matters sits in what that measurement is bound to.
Here is the strongest argument against this being meaningful adoption, and it survives the engineering working perfectly.
A hardware attestation proves that a specific binary ran inside genuine attested silicon, and it reports a hash of what was loaded. A verifier can check the vendor's signature chain and confirm the measurement. What a verifier cannot do is establish that the measurement corresponds to the model it believes it is buying, because that binding requires independent knowledge of the model's weights. For a closed frontier model, only the provider has them. The provider publishes the commitment, attests to having run the thing it committed to, and no outside party can check the commitment against anything.
The proof is therefore sound and self-referential. It makes substitution within a provider detectable, a real gain worth having, since a provider cannot silently serve a cheaper model once committed. It does not make the provider's description of its own model checkable by anyone: verification is relative to a claim, and the party that authored the claim is the party being verified.
The boundary condition is testable. Where weights are published and hashed independently the binding is external and the attestation becomes a proof about a known artifact, which is exactly the case the cheapest technique already serves and not the case institutions are asking for, because the models they want are the ones whose weights nobody outside the lab has seen.
Three conditions, ordered by how far away they are. The weight commitment has to be externally verifiable, through publication, third-party escrow or a disclosed attested build process, because without it the strongest available proof is a proof about the provider's own assertion. The attestation chain has to be trusted at the hardware vendor level, which moves the root of trust to a chip manufacturer rather than removing it. And the output has to matter for a reason that model identity addresses.
That last condition is the one most deployments fail. The institutional risk in using a model is usually that the answer is wrong, not that the server lied about which model produced it. Attestation addresses the second and is silent on the first, so a firm buying verifiable inference to manage model risk has bought the wrong product.
The honest non-hype case is narrower and real. Where a counterparty is paid per inference and could profit by serving something cheaper, attestation is a direct commercial control, and splitting verification into an enclave while computation runs on an untrusted accelerator has shown throughput gains of six to twenty times over naive enclave execution. That is a working audit mechanism for a metered service, a smaller claim than settling AI on-chain and the part that is true.
The thesis fails on any of three findings. If no frontier provider ever publishes an externally checkable weight commitment, attestation stays a within-provider control and never becomes third-party verification. If zero-knowledge proving costs do not fall by several orders of magnitude, the only technique that verifies computation without trusting hardware remains confined to small models. And if buyers keep selecting providers on latency and price without requiring attestation, then the demand the thesis assumes does not exist.
Four things are trackable and all are public. Whether any major provider publishes a model hash that a third party can reproduce. Whether procurement terms in regulated sectors begin requiring attested inference as a condition rather than a feature. Published zkML proving times for models above a billion parameters. And whether enclave memory limits, long the binding constraint on in-enclave inference, move with new silicon generations.
The constructive reading is that a real control arrived and is being oversold. Attestation at five percent overhead beats nothing, on-chain settlement is trivially cheap, and the mechanism genuinely solves provider substitution. What it does not do is make a closed model's behavior checkable from outside the lab, and the gap between those statements is where most of the enthusiasm sits.
For informational purposes only. Not an offer to buy or sell any security. Available only to accredited investors who meet regulatory requirements.