Last week's research call with a small team in Zurich left me circling a thought I'd been resisting for about six months. The decentralized-AI conversation in 2026 is still mostly about the supply side, who has GPUs, whose orchestration is cleaner, whose tokenomics aligns operators best. That's the conversation I wrote about a few weeks ago when I argued that AI x Web3 is three markets, not one. The supply-side market, decentralized compute, is the most real of the three. I still believe that.
What I'd been underweighting was the layer above it. The trust layer. The part that proves the inference your protocol delivered is actually the inference your customer paid for. Verifiable inference, and the zkML and adjacent work that makes it possible, is the bridge most dAI conversations skip past, and I think it's the category that decides which dAI bets are real in 2027 and which ones quietly retire.
The Trust Problem, Concretely
Imagine a decentralized inference network running a Mistral or Llama variant. A customer submits a query. That query gets routed to one of thousands of operators running the model. An operator returns a result. And the customer pays.
The simple question: how does the customer know they got the inference they paid for? How do they know the operator ran the actual model on the actual prompt, rather than a smaller model? Rather than a cached lookup table? Rather than a deliberately compromised variant that subtly shifts outputs? How does the network know the operator didn't take shortcuts that the customer can't detect?
In 2026, the answer for almost every decentralized-inference protocol is some combination of operator stake at risk, reputational scoring, statistical sampling, and trust-but-verify after the fact. Those are real mechanisms. They work most of the time for most of the customer base. They do not work for the customer base that actually pays the highest dollar value per query, regulated industries, agentic AI deployments where the output drives downstream automated actions, defence and government work, anything where the cost of a single bad inference is meaningfully higher than the cost of the inference itself.
The trust gap is the reason the highest-value AI customers in 2026 still default to a known, centralised provider, not because the centralised provider is more accurate, but because the audit and accountability surface is settled. You can sue OpenAI. You cannot easily sue an anonymous Bittensor subnet operator who returned a degraded inference last Thursday.
What "Verifiable" Actually Means
The cleanest version of solving this is verifiable inference, a cryptographic proof, generated alongside the inference itself, that the specific model was run on the specific input and produced the specific output. The customer doesn't have to trust the operator. The customer verifies the proof and the answer is settled mathematically.
The reason this hasn't been the default architecture is straightforward: until recently, generating cryptographic proofs of model execution at any meaningful model size was either impossible or prohibitively expensive. A zk-proof of a small classifier inference was a research curiosity in 2022. A zk-proof of a small transformer inference cost roughly 1000x the actual inference. The category was real in academia and economically irrelevant in production.
What's changed in 2026, and the reason the trust-layer conversation has moved from "interesting in theory" to "starting to be operationally tractable," is a handful of breakthroughs across the zkML stack. Proof generation costs for small-to-medium transformer inferences have dropped roughly two orders of magnitude in two years. The proof-system constants are improving faster than the model-size constants are growing. The cost overhead is now somewhere in the range of 10–50x rather than 1000x. That's still expensive, but it's expensive in the way that paying for an enterprise SOC 2 audit is expensive. Some customers will pay for it. The customers who pay for it are the high-dollar-value ones.
Who's Building It
The credible teams in this category in 2026 fall into roughly three groups.
Pure verification protocols — teams whose whole product is the proof layer rather than the inference itself. Modulus Labs, Giza, EZKL, and the newer wave coming behind them. These teams are building the zkML primitives that other inference networks plug into. The bet is that verification becomes an embedded service across the inference stack, the way audit firms became embedded across enterprise software.
Integrated dAI networks with verification built-in — teams running both the inference network and the verification layer as a single product. Ritual is the clearest example. The bet is that verification is too important to be an optional add-on, so the network with credible verification by default has a structural moat over the network that bolts it on later.
Hybrid trust models — teams using a mix of cryptographic proof, trusted execution environments (TEE), and economic mechanisms to deliver a trust property without requiring full zk-proof costs. This is the category I think is most likely to actually power production workloads in 2026–2027, because the cost economics work today and the trust properties are good enough for most regulated use cases. The work happening at Phala, the TEE-plus-token designs coming out of several teams, and the Eigen-secured verification primitives I touched on in Bridge Notes #3 all sit in this group.
The bet I'd make in 2026: in the next 18 months, the production deployments of verifiable inference happen mostly in the hybrid-trust category, with the pure-zkML category compounding underneath them as the proof-system costs continue to drop. The two are complements, not substitutes.
Why This Composes With The Rest Of The Stack
The verifiable-inference thesis matters most where it composes with other on-chain primitives. Three composition points are worth flagging.
Composition one: agentic AI on-chain. The on-chain agent thesis I was cautious on in Bridge Notes #2 becomes meaningfully more credible if the agent's inference is provably correct. The reason agentic AI on-chain in 2026 is still small is partly trust — users don't authorise agents to spend their balances without confidence that the agent's reasoning won't be compromised. Verifiable inference moves that trust question from social to mathematical.
Composition two: regulated industries. The regulatory bridge I talked about in Bridge Notes #1 extends here. Regulated industries, healthcare, financial services, government, need not just AI that works but AI that they can prove worked correctly to an external auditor. Verifiable inference is the natural primitive for that.
Composition three: model authenticity. As the model layer fragments, open-source models, proprietary fine-tunes, region-specific deployments, knowing that the model you queried is the model you think you queried becomes its own problem. Verifiable inference solves it. This is a smaller wedge than the first two but it's a category that's going to grow as the model layer continues to fragment.
What I'd Underwrite
For the portfolio at PRIM3 in 2026, the verifiable-inference layer is one of the few places I'd actively look to add new positions even though we already have dAI exposure. Specifically: hybrid-trust deployments with credible early enterprise pilots, zkML proof-system teams with research-grade roadmap and pragmatic shipping discipline, and inference networks that have made verification a default architectural choice rather than an optional plugin.
I'd be cautious about a few things: pure-research zkML plays with no commercial roadmap, "verifiable AI" teams whose actual product is mostly TEE-as-a-service without a credible cost-cutting differentiation against the existing cloud-TEE providers, and integrated networks whose tokenomics depend more on verifier-set incentives than on real customer-paid demand for verification.
The Forbes-Quotable Line
If I had to compress this to one sentence: the next wave of decentralized AI doesn't compound on compute supply — it compounds on whether the inference can be proven, and the protocols building that proof layer are the ones I'd back into 2027.
The next Bridge Notes will turn back to the meta question I've been circling all year — what two decades of operating across Web3 cycles actually taught me, and which of the patterns I keep seeing are repeating in 2026 and which one isn't. If you're building in the verifiable-inference category and the framings here land, or push back, I'd genuinely like to talk: LinkedIn or Telegram.