[ Prover farm with GPU scheduler ] →
H100 / A100 / 4090 pool, deadline-aware job prioritization, NUMA-aware placement, 75%+ per-card utilization in typical setups.
> solutions / zk
Managed operations for ZK prover farms and rollups
Proof generation is GPU-bound, deadline-bound, parallel. Dropped jobs = missed blocks. We build the farm, the queue, the retries, and per-circuit benchmarks on SP1, RISC Zero, Boundless, Brevis.
We arrive with a working stack: SP1 and RISC Zero as the RISC-V zkVM baseline, Boundless and Brevis as decentralized prover marketplaces, Jolt and Halo2 for custom circuits. For each farm we wire up a GPU scheduler with deadline-aware prioritization, kv-cache for circuit segments, retry queues with exponential backoff, and missed-proof-window alerts.
The whitespace we sit on: the "operator for operators" on prover marketplaces. Succinct, Boundless and Brevis launched as marketplaces in late 2025; no one occupies the "we run your GPU farm on these marketplaces with a signed SLA" position. We take it.
> stack we operate
ZK subset. The platform layer is identical across ICPs.
ZK:
SP1
RISC Zero
Boundless
Brevis
Jolt
Halo2
Platform:
Kubernetes
Terraform
Ansible
Prometheus
Grafana
Loki
OpenTelemetry
PagerDuty
> what we deploy
Concrete deliverables for ZK teams. Each ships end-to-end with repo, IaC and runbooks.
[ Proof-job retry queue ] →
Idempotent submit, exponential backoff, dead-letter queue with per-circuit analysis, alert on pre-deadline ETA miss.
[ Per-circuit benchmark catalogue ] →
Measurements on SP1 / RISC Zero / Jolt per circuit: time-to-prove, peak memory, optimal GPU type. Artifact cache for precompiles.
[ Operator on Boundless / Brevis / Succinct ] →
Marketplace registration, bidding strategy per circuit type, reputation tracking, payout reconciliation, marketplace switching by marginal cost.
[ ZK rollup ops: prover + sequencer + DA ] →
Proving stack for your own rollup: coordination across sequencer, batcher and provers, L1 finality monitoring, stall playbook.
> what we operate 24/7
After handoff the pager lives with us. Coverage tuned for ZK farms:
- Prover liveness against network deadlines: alert if proof ETA exceeds 70% of the proof window.
- GPU health watchdog: ECC errors, thermal throttling, xid signals trigger preempt and job migration.
- Queue depth / lag metrics per circuit: alert on >20% drift off baseline.
- Auto-retry on failed proof generation with deterministic seed for reproducibility.
- Versioned runbooks: prover crash, OOM on witness generation, stuck marketplace bid, L1 fork-choice inside the proof window.
- Monthly perf review: new precompiles, refreshed benchmarks, load rebalancing across marketplaces.
> migration scenarios
What we move without losing payouts or skipping proofs.
single marketplace to mix
Move from a single prover marketplace to a portfolio of 2-3: differentiate by circuit type, hedge against single-venue downtime.
cloud GPU to bare-metal
Farm moved off AWS / GCP onto dedicated GPU at Latitude.sh / DataPacket: typical 50% cost-per-proof reduction, no deadline misses.
SP1 to RISC Zero (or back)
Parallel shadow proving for correctness, gradual cutover by circuit type, latency baseline pinned along the way.
scale-up for incentive season
Prover capacity 5x in 7-10 days: H100 burst sourcing, auto-onboarding into marketplaces, post-season wind-down playbook.
FP32 to FP16 / mixed precision
Quantization-aware proof generation where the circuit allows: 30% reduction in time-to-prove, re-verify check per batch.
onto in-house proving stack
From an external prover service onto your own farm: breakeven calc, gradual cutover, audit of key material and precompiles.
> cases
Anonymized. NDAs cover names; the numbers are real.
ZK rollup · 6 mo · validator ops + prover farm · slashing: 0 · missed proof windows: 0
SP1 marketplace operator · 4 mo · 32 H100 · top-3 by win-rate · cost-per-proof -45%
RISC Zero proving service · 8 mo · 16 H100 + 24 4090 · ETA accuracy ±8%
Brevis incentive season · 8 weeks · 5x burst capacity · 99.4% on-time delivery
> SLA tiers
Three coverage levels. For production proving with a signed deadline SLA we recommend Silver or higher.
| Tier | Response p95 (Sev-1) | Coverage | Incident report | Engineer hours / mo |
|---|---|---|---|---|
| Bronze | 30 min | Business hours, 5×8 | Within 48h | 40 |
| Silver | 15 min | 24/7 on-call rotation | Within 24h | 80 |
| Gold | 5 min | 24/7 with dedicated engineer | Within 12h | 160+ |
> FAQ
Across all of them. SP1 and RISC Zero for the RISC-V zkVM approach, Jolt and Halo2 for custom circuits, Boundless / Brevis / Succinct as the marketplace layer. If you have a collab with a specific provider, we plug into their toolchain.
Sev-1. Escalation in 5-15 min (by tier), root cause in 12-24h, postmortem with concrete action items. Architecturally we try to catch misses early: an alert on ETA >70% of the proof window triggers a pre-emptive migration to a free GPU.
Depends on the circuit and target throughput. Baseline for one production circuit: 4-8 H100 or 16-24 RTX 4090. For a marketplace bid: start with 8-16 H100 and scale by win-rate. Send the circuit type and target proofs/hour; we'll send a sized estimate in 24h.
Yes. It's one of our whitespace bets. We register as your operator on Boundless / Brevis / Succinct, run the bidding strategy, watch reputation, reconcile payouts. Staking material and payout wallets stay with you.
Yes. That includes sequencer + batcher + prover as a coordinated pipeline, L1 finality observability, and a stall playbook. We can build the stack from scratch or take an existing setup into operations.