The Future of L2s: zk-Rollups for Machine Learning

Centralized AI models are black boxes. Users must trust providers on outputs, data usage, and fairness. This is antithetical to verifiable computation.

• No Audit Trail: Cannot prove a model wasn't trained on proprietary data.
• Result Opacity: Impossible to verify inference was run correctly without re-execution.
• Market Fragmentation: Proprietary silos prevent composable AI services.

ZK-rollups like StarkNet and zkSync provide the settlement layer for verifiable ML. The L2 becomes a global, trustless coprocessor.

• Stateful Proofs: Model weights and inference traces become part of the chain's provable state.
• Native Composability: Verified AI outputs trigger DeFi actions, gaming logic, or DAO votes atomically.
• Cost Scaling: Batching proofs for millions of inferences drives cost toward ~$0.01 per proof.

General-purpose zkEVMs are inefficient for linear algebra. Projects like Modulus, EZKL, and RISC Zero are building zkVMs optimized for tensor operations.

• Hardware Acceleration: Leveraging GPUs and FPGAs for 100-1000x faster proof generation.
• Framework Integration: Direct compilers from PyTorch and TensorFlow graphs to zk-circuits.
• Prover Market: Decentralized proving networks emerge, creating a $1B+ compute marketplace.

On a zk-rollup, data provenance and usage rights are programmable. This enables new economic models impossible in Web2.

• Tokenized Datasets: Ownership and licensing fees are enforced on-chain via smart contracts.
• Proof-of-Training: Models can cryptographically attest their training data lineage, enabling ethical AI markets.
• Federated Learning at Scale: Contributors to a model can be compensated without exposing raw data, powered by zk-SNARKs.

Today's ML compute is a pure cost. On an L2, it becomes a transaction-generating, fee-earning service layer.

• Micro-Monetization: Charge per verified inference, enabling pay-per-use AI.
• L2 Sequencer Revenue: Proving fees and transaction ordering fees create sustainable validator economics.
• Cross-Chain AI: Verified outputs are portable assets, bridging to Ethereum, Solana, and Cosmos via protocols like LayerZero.

Hardware improvements (Moore's Law) are slowing. ZK-proof efficiency (ZK Law) is improving exponentially. The crossover point is near.

• Proving Overhead: Falling from 1,000,000x to ~100x vs. native execution within 24 months.
• Regulatory Tailwind: GDPR, AI Acts demand auditable AI, making zk-proofs a compliance tool.
• First-Mover Protocols: Worldcoin, Gensyn, Ritual are already building the stack; L2s are the logical settlement home.

Proving the execution of a large neural network on Ethereum would cost > $1M in gas and take days. This kills any practical application.

• Solution: Dedicated zk-Rollups like Modulus Labs' zkOracle and Giza's zkML Rollup.
• Mechanism: They run the ML model off-chain and generate a succinct ZK proof of the inference result.
• Outcome: The L1 only verifies the proof, reducing cost to ~$1-10 and time to ~1 minute.

ML models are large state objects (GBs). Storing them on-chain is prohibitive, but off-chain storage breaks verifiability.

• Architecture: Rollups like Risc Zero's zkVM and EigenLayer AVSs use off-chain DA layers (e.g., Celestia, EigenDA) for model data.
• Trade-off: This creates a security/cost continuum—higher security with Ethereum DA, lower cost with external DA.
• Key Entity: EigenLayer restakers can secure these DA layers, creating a cryptoeconomic security pool for ML states.

Raw verifiable inference is a commodity. The value is in permissionless, composable AI agents with provable outputs.

• Example: A trading agent on Aevo that proves it followed its strategy, or a gaming NPC with verifiable behavior.
• Protocols: Modulus' 'Based AI' and Giza's Actions framework let developers deploy these agents as smart contracts.
• Revenue: Fees are extracted via model inference calls, creating a new Machine Learning Economy atop the rollup.

GPU-optimized ZK provers (like Ulvetanna's Binius) are still in development. Today's proofs are slow and expensive for large models.

• Current State: Proving a ResNet-50 inference can take 10+ minutes and cost ~$50 on a CPU.
• Future Path: Custom ASICs (like those from Ingonyama) and parallel GPU proving are required to reach sub-$1, sub-minute proofs.
• Risk: This creates a centralization pressure around high-performance proving farms in the short term.

Generating ZKPs for large ML models is computationally explosive. The specialized hardware (e.g., FPGAs, ASICs) required creates a centralization vector and could make inference costs prohibitive versus centralized clouds.

• Prover time for a single inference could reach minutes to hours, destroying UX.
• Capital costs for prover hardware could exceed $1M+, limiting node operators.
• Risk of a winner-take-all hardware market dominated by entities like Ingonyama, Ulvetanna.

ML models are stateful. A zk-Rollup must provably track model weights and training data provenance. Storing this on-chain is impossible; storing it off-chain recreates data availability problems akin to Celestia's core thesis.

• Model checkpoint sizes can be terabytes, far exceeding Ethereum calldata limits.
• Reliance on off-chain data committees or EigenDA introduces new trust assumptions.
• Corrupted or unavailable data halts the entire inference network.

ML models often process real-world data (sensors, APIs). A zk-Rollup can prove correct execution, but cannot guarantee the integrity of the input data. This recreates and amplifies the blockchain oracle problem.

• Requires trusted oracles like Chainlink to feed data, creating a central point of failure.
• Adversarial inputs can force a model to produce valid but malicious outputs.
• The system's security collapses to the weakest oracle in the stack.

ZK-ML frameworks (e.g., EZKL, RISC Zero) force developers into constrained environments. They must write circuits, not code, dealing with finite field arithmetic and circuit compiler bugs.

• Tooling immaturity leads to longer dev cycles and audit complexity.
• Circuit bugs are cryptographic and irreversible, unlike smart contract bugs which can be forked.
• Risk of vendor lock-in to specific proof systems (Halo2, Plonky2) and hardware.

A decentralized, privacy-preserving network for AI inference is a regulator's nightmare. It enables uncensorable models, potentially trained on copyrighted or restricted data. This guarantees aggressive scrutiny.

• OFAC sanctions risk for operators, similar to Tornado Cash.
• Model provenance becomes a legal requirement, conflicting with ZK's privacy.
• Global fragmentation as jurisdictions (EU AI Act, US EO) take hostile stances.

The value of verifiable inference is unclear for most applications. End-users may not pay a 10-100x premium for a ZK proof when a centralized API is "good enough." The utility must justify the cost.

• Fee market dynamics could price out all but the highest-value use cases (e.g., on-chain trading bots).
• Prover subsidies may be required indefinitely, creating unsustainable tokenomics.
• Without a killer app, the network becomes a solution in search of a problem.