Boundless is a zero-knowledge (ZK) proving infrastructure provider making blockchain networks faster and more interoperable. In the company’s own words, “Boundless is a protocol for accessing verifiable compute on any chain and rewarding nodes for contributing useful work.”

Running on a RISC-V-based zkVM, the company’s operations rely on sustained, predictable GPU compute at scale: workloads that are uniquely sensitive to memory bandwidth, latency, and hardware consistency. As the ZK ecosystem matures, reliable, high-performance infrastructure has become a competitive differentiator for Boundless.

The challenge: When cloud flexibility becomes a liability

Boundless was running GPU-intensive ZK proving workloads on AWS (primarily L40-class instances), evaluating both on-demand and spot market capacity. But what began as a practical cloud-first strategy quickly revealed structural limitations incompatible with production ZK workloads.

The problem: AWS GPU spot instances operate on two-minute termination notices, which are fundamentally incompatible with long-running ZK proof generation jobs. AWS spot placement scores for 100-500 GPUs of specific instance types were consistently low across regions, confirming to the team that capacity wasn’t available when they needed it most.

Shared cloud infrastructure also introduced “noisy neighbor” performance variances that made it impossible to tune CUDA workloads for the consistent, high-memory-bandwidth behavior ZK proving requires. On-demand GPU pricing also became prohibitively expensive at scale, with AWS spend reaching approximately $2,300 per day.

Finding Latitude.sh

During Hetzner’s publicized withdrawal from crypto-related workloads amid the Ethereum boom, Latitude.sh emerged as a recommended provider in community discussions on alternatives.

On discovering Latitude.sh, the value propositions were immediately clear to Boundless:

• Bare-metal, single-tenant hardware with no noisy neighbors

• Cloud-like, API-driven provisioning for operational agility

• Genuine GPU availability at scale

• A significantly better cost structure than AWS on-demand pricing.

The migration

During the migration from AWS to Latitude.sh, the team moved GPU-dependent ZK-proving workloads off AWS infrastructure and onto Latitude.sh bare metal, maintaining residual AWS services where they were still needed 

Immediately, the Boundless team noticed that infrastructure was delivered exactly as advertised, and they felt supported by the Latitude.sh team throughout the transition.

The results

For Boundless, the financial impact of switching to Latitude.sh was immediate and significant: a $860-per-day cut in AWS spend, a reduction of over 60%.

Beyond cost savings, the move to Latitude.sh unlocked hardware capabilities that are unavailable in cloud environments. The RTX 6000 Pro with 96 GB VRAM enables MIG (Multi-Instance GPU) partitioning into 4x 24 GB vGPUs – a configuration that pushed ZK proving throughput from approximately 10-11 MHz to ~20 MHz on RISC Zero’s v2 circuit.

The benefits

• Performance: Single-tenant bare metal eliminates the shared infrastructure overhead that hurts ZK workloads on AWS. Now, every GPU cycle belongs to Boundless – and as a result, they’ve doubled their proving throughput and removed noisy neighbors.

• Hardware quality: RTX 6000 Pro-class cards with 96 GB VRAM enable MIG partitioning, a workload configuration that public cloud providers simply don’t support.

• Pricing predictability: Spot interruptions and unreliable capacity have been replaced with reliable, budgetable compute, resulting in a cost reduction of over 60% compared to AWS for similar GPU capacity.

• Support: The team at Latitude.sh are experts in infrastructure-critical workloads, and they provide timely, knowledgeable support as needed.

By using Latitude.sh to get bare-metal GPU performance at a fraction of cloud cost, the Boundless team can now focus on features, not abstractions.

Want to try that for yourself and boost your ZK-validation performance? Get started on Latitude.sh for free.