Boundless taps idle crypto GPUs to cut AI inference costs

Boundless Network Inc. is expanding its distributed GPU network of about 4,000 units to run AI model inference as managed infrastructure, with early benchmarks claiming up to 50% cost savings compared to hyperscalers for asynchronous tasks. The company, which built the network over four years for zero-knowledge proof generation in crypto, is now tuning the same coordination layer for AI workloads. A full product launch is planned for later this summer, with signups available via a waitlist on its website.

What happened

Boundless assembled its GPU network to meet transaction-proving demand in crypto, pooling supply and scheduling compute-heavy jobs across scattered machines. The company is now tuning that coordination layer for AI inference, treating it as managed infrastructure rather than a spot market. The network includes consumer-grade cards and GPUs originally bought for crypto mining, which Boundless argues can handle many inference workloads without requiring scarce data-center chips.

The company also plans to give its native token, ZKC, a role in the AI network. Operators would stake ZKC to join, with the stake size tied to how much they can earn. The existing zero-knowledge proving network will continue running in parallel.

Why AI builders should care

Inference costs are becoming the dominant factor in production AI spending. Gartner expects inference to account for 55% of AI-optimized infrastructure-as-a-service spending this year and over 65% by 2029. For teams running multiple models or high-volume agent workflows, every percentage point of cost reduction directly affects how much they can deploy.

Boundless challenges the assumption that expensive data-center GPUs are always necessary. By tapping idle crypto hardware, the network could offer a cheaper alternative for workloads that don't need instant responses. The token-based staking model also introduces a new incentive structure for distributed compute, potentially attracting operators who already own GPUs for crypto purposes.

Practical implications

For developers, the immediate takeaway is a potential new option for asynchronous inference tasks such as batch processing, background model evaluation, or non-real-time agent reasoning. The 50% cost savings claim applies to these workloads, not latency-sensitive ones.

The staking mechanism means operators can earn by contributing compute, which could increase supply and further lower costs. However, the token model also introduces volatility and regulatory considerations that traditional cloud services don't have.

Boundless will keep its ZK proving network running alongside the AI expansion, which may allow hardware to be dynamically allocated between the two workloads based on demand. This dual-use approach could improve overall utilization compared to dedicated inference clusters.

Caveats

The cost savings and performance claims are based on early benchmarking provided by Boundless, not independent third-party testing. Actual results may vary once the service launches and scales.

The product is not yet generally available. A full launch is planned for later this summer, and current access is limited to a waitlist. Pricing, model support, and reliability SLAs have not been disclosed.

Running AI inference on consumer-grade GPUs may introduce trade-offs in throughput, memory capacity, and energy efficiency compared to data-center hardware. Teams should benchmark their own workloads before committing to the network.

Finally, the integration of a token-based access model adds complexity around token price fluctuations, staking lockups, and potential regulatory scrutiny. Builders should evaluate whether the cost savings outweigh these operational risks.