EZKL

EZKL is led by Jason Morton, an academic-turned-founder whose zkonduit team has shipped a continuously maintained, broadly adopted open-source library rather than a perpetual research preview. The execution signal is the artifact itself: ezkl is the zkML tool most other teams reach for, and the cadence of releases on GitHub reflects a working engineering operation. Against Modulus Labs, which produced strong research but ultimately folded into Tools for Humanity, EZKL has kept an independent, shipping product alive. Lookout would upgrade further on evidence of a sustained commercial deployment, and downgrade if maintenance velocity slowed or key contributors departed.

The differentiation is real and specific: EZKL takes models in the ONNX standard and compiles them to Halo2 circuits, giving a general path from mainstream ML tooling to zk proofs that few rivals match in developer accessibility. That ONNX-in, proof-out ergonomics is the reason it became the reference zkML library rather than one of many. Versus general-purpose zkVMs like RISC Zero, EZKL is more specialized and currently more efficient for the ML-circuit case, but that very specialization is the long-term risk if zkVMs close the gap. The view strengthens on benchmarks showing EZKL proving larger models at acceptable cost, and weakens if a general zkVM matches it on ML workloads.

EZKL has no token, so there is no emission schedule or value-accrual mechanism to evaluate, and Lookout treats that as neutral rather than negative for an infrastructure library. The open question is whether verifiable inference ever monetizes as a protocol with a token, or remains a developer tool monetized — if at all — as a service or hosted prover. Compared with token-live peers like Giza, EZKL retains full optionality to design economics around proven demand. A credible token tied to real proving volume would move this positive; continued tool-only status leaves it flat.

Adoption is real but developer-stage: EZKL is the most-used zkML library and a common dependency in research and prototypes, which is meaningful mindshare for a primitive. The gap is that library downloads and citations are not the same as production applications generating sustained proving demand, and zkML's killer app has not yet arrived. Against RISC Zero, whose Bonsai and Boundless have live networks and paying-adjacent usage, EZKL's traction is narrower and less monetized. The axis moves up on a flagship production deployment processing real volume, and stays neutral while usage is largely experimental.

EZKL has raised roughly $3.81M with Bloomberg Beta, Lemniscap and Symbolic Capital among its backers — a respectable seed for an open-source tool but modest against the capital in the broader proving sector. The cap table is credible and crypto-aware, yet small relative to RISC Zero's $50M-plus or even Giza's $8M-plus. That keeps EZKL efficient rather than dominant, dependent on lean execution rather than spend. The view would improve with a larger follow-on signalling institutional conviction, and weaken if runway constraints throttled development.

EZKL sits squarely in the verifiable-AI narrative — proving that a model ran honestly is exactly the trust primitive an on-chain AI era needs — but it is tooling beneath that story rather than a name that owns it. The narrative is genuine and durable, yet zkML specifically has cooled relative to the broader 'verifiable compute' and zkVM framing that captured more mindshare. Against RISC Zero's universal-ZK pitch, EZKL's zkML-library framing is more niche and less investable. The narrative strengthens if verifiable inference becomes a regulatory or trust requirement, and stays neutral while it remains a builder's primitive.

The defining risks are technical and structural rather than existential: zkML proving overhead is still prohibitive for large models, which caps the near-term addressable market, and general-purpose zkVMs could commoditize purpose-built ML circuits. With no token, there is no market-structural or unlock risk, but also no on-chain upside vehicle. Relative to Modulus Labs, which exited via acquisition, EZKL's risk is slow demand rather than abandonment. Lookout would de-risk on falling proving costs and a production use case, and flag it upward if a zkVM rendered bespoke zkML redundant.