Sparo

Relay unlocks idle GPU capacity through multi-party exchanges, making AI compute faster, cheaper, and more accessible.

Comment

Inspiration

AI compute is becoming one of the world's most valuable resources, yet much of it remains inefficiently allocated. Companies reserve expensive GPUs they don't fully utilize, while others face high prices, limited availability, and long wait times.

We realized the problem isn't always a shortage of compute. Sometimes, the compute already exists, but the agreements governing it prevent that capacity from reaching the people who need it.

Traditional GPU marketplaces connect individual buyers and sellers. But what happens when a deal requires three or four parties to cooperate before anyone benefits?

That question inspired Relay.

What it does

Relay is a clearinghouse for AI compute. Instead of simply listing GPUs for rent, Relay discovers and coordinates multi-party exchanges that make existing computing commitments more efficient.

For example, imagine a company paying for six GPUs it isn't using. Another company urgently needs those GPUs, while a third provider has cheaper replacement capacity available.

Relay identifies an arrangement where the first company can release its idle capacity, the second gets compute faster and at a lower cost, and the third earns revenue from otherwise unused infrastructure.

The key innovation is conditional multi-party clearing: each participant's commitments are coordinated as part of one agreement, rather than relying on independent transactions that might leave someone worse off.

Relay supports the full lifecycle:

  1. Discover: Track GPU market prices and available compute opportunities.
  2. Clear: Find economically beneficial direct and multi-party exchanges using a deterministic optimization engine.
  3. Approve: Coordinate agreement terms and approvals across participants.
  4. Execute & Verify: Run compute workloads and verify execution before permitting payment.
  5. Settle: Coordinate invoices, payments, provider payouts, and financial reconciliation.

The result is a more flexible market where compute can move economically between organizations without requiring physical GPU migration.

How we built it

We built Relay as an integrated compute marketplace, optimization engine, execution platform, and financial coordination system.

Our backend combines market intelligence, deterministic deal clearing, multi-party agreement coordination, workload execution, and financial operations.

Core technologies and integrations:

  • Apify: Collecting GPU market listings and reference pricing data.
  • Tiger Data: Time-series market data, pricing history, and analytics integration.
  • CREAO: Coordinating agent workflows, investigations, and operational tasks.
  • Elastic: Indexing financial evidence and supporting AI-assisted exception investigations.
  • Coinbase x402: Testnet payment workflows and machine-to-machine payment infrastructure.
  • Solana: Testnet provider settlement and payout tracking.
  • Docker: Executing containerized workloads and gathering verification evidence.
  • Python / TypeScript: Backend services, financial logic, optimization, and APIs.
  • Figma: Designing Relay's interface and visual identity.

At its core, Relay uses a deterministic clearing engine to evaluate whether proposed exchanges satisfy resource requirements, deadlines, budget limits, provider permissions, switching costs, and economic feasibility.

We also built financial controls around the transaction lifecycle, including version-specific approvals, separation of duties, invoice verification, exception management, double-entry accounting, and reconciliation.

Rather than allowing AI agents to independently authorize financial decisions, Relay separates automated coordination from human-controlled approvals.

Challenges we ran into

One of our biggest challenges was making multi-party transactions genuinely feasible rather than simply matching buyers with available GPUs.

An exchange might look profitable on paper but fail because of timing constraints, incompatible hardware, missing provider authorization, or the cost of switching capacity. We had to model these dependencies and ensure the entire arrangement was viable before proposing it.

Another challenge was integrating execution with financial settlement. A signed agreement doesn't necessarily mean useful compute was delivered. We designed the system to verify workload execution before allowing settlement to proceed.

Coordinating multiple external services also introduced challenges around API contracts, asynchronous workflows, data consistency, and degraded service states.

Finally, building reliable financial operations required careful handling of approval authority, billing discrepancies, blocked payments, and ledger reconciliation.

Accomplishments we're proud of

We're particularly proud of building beyond a conventional GPU rental marketplace.

Relay can evaluate conditional, multi-party exchanges rather than treating each compute transaction independently.

We implemented an end-to-end financial workflow connecting verified usage, invoicing, exception detection, human-approved corrections, payment instructions, settlement tracking, and balanced ledger entries.

We also integrated AI-assisted financial investigations while keeping economic decisions deterministic and sensitive financial actions subject to explicit authorization.

Most importantly, Relay demonstrates how existing compute commitments can become more flexible without requiring companies to abandon the financial and operational controls they depend on.

What we learned

Building Relay taught us that infrastructure markets are not just matching problems. They're coordination problems.

Optimizing compute allocation requires understanding technical compatibility, economic incentives, contractual obligations, and operational risk simultaneously.

We also learned how important it is to design AI-assisted systems with clear boundaries between recommendations, deterministic decisions, and human authority.

What's next for Relay

Our long-term vision is to make AI compute commitments more liquid, accessible, and efficient.

We want to expand Relay with additional compute providers, richer utilization forecasting, improved market intelligence, and more sophisticated multi-party clearing strategies.

We also plan to develop stronger provider integrations so authorized capacity transfers and workload execution can be coordinated across real infrastructure.

Ultimately, we envision Relay as a financial and operational coordination layer for AI compute: a system that helps organizations unlock the value of resources they've already committed to while giving others faster, more affordable access to the capacity they need.

Built With

Share this project:

Updates