Lilypad

The problem

Cloud compute is a duopoly. Distributed compute markets have existed conceptually for a decade but have never cleared economically at scale — matching demand to permissionless supply requires protocol-level scheduling, verification, and settlement that legacy stacks cannot provide, and every prior attempt has hit the same problems: unverifiable results, unreliable providers, unstable pricing, and workloads that don't fit generic compute markets. The class of workload that actually creates demand for distributed compute is inference at scale — a workload the current cloud duopoly prices aggressively but that has real economic room for a decentralized alternative if the reliability and verification problems can be solved.

Key challenges

Permissionless compute networks have to solve four hard problems in one system: (1) job scheduling that respects heterogeneous provider capability, (2) result verification without trusting the provider, (3) payment settlement that clears at market-clearing prices, and (4) an operational surface that doesn't require workload authors to understand any of the above. Each problem is a research area; combining them into a live network is where the engineering actually happens.

What we built

AR Data contributed engineering on the Lilypad network — compute job scheduling, provider verification, and result attestation. The focus was on making distributed compute usable for AI inference workloads, where cost sensitivity and latency budgets create real demand for an alternative to the cloud duopoly. Scheduling matches workloads to providers based on capability declarations (GPU class, VRAM, network) and reputation signals. Result verification uses attestation patterns appropriate to the workload — deterministic replay for reproducible workloads, sampling for stochastic ones. Settlement clears through the network's native payment layer with hold-and-release semantics on verification.

Our approach

1. 1

   Workload-specific verification strategy

   Verification isn't one-size-fits-all. Deterministic workloads get replay verification; stochastic ones get statistical sampling. Picking the right strategy per workload is what makes verification tractable.

2. 2

   Capability declarations plus reputation

   Scheduling can't rely on provider self-report alone. Combining capability declarations with observed reputation signals gives the matcher a defensible view of what each provider can actually deliver.

3. 3

   Hold-and-release payment semantics

   Payment on submission is exploitable. Payment on verified completion aligns incentives — providers only get paid for work that verifies.

4. 4

   AI inference as the initial workload class

   General-purpose compute markets have failed. Focusing on AI inference — a class with real demand and clear verification patterns — is a more tractable starting position.

Key architectural decisions

Results

• Compute job scheduling with capability + reputation matching
• Provider verification workflows tuned per workload class
• Result attestation with hold-and-release payment semantics
• AI inference as the primary target workload
• Contributions to the Lilypad network's operational surface
• Patterns transferable to other decentralized-compute contexts

Impact

Lilypad represents the class of infrastructure that decides whether decentralized compute stays theoretical or becomes real. The engagement produced operational patterns for scheduling, verification, and settlement that inform how we approach any permissionless-market problem — and it's why our decentralized-tech consulting continues to include this space rather than retreating to storage alone.

Tech stack