Measurements and limits
Every number here was measured, with where.
Speed
| What | Result | Where |
|---|---|---|
| One check over 128 sealed slots | median 4.25 s, 3.82 to 5.64 s, 10 of 10 exact | local Arcium cluster, 2 nodes |
| Open (8 in flight) | about 3.4 s per batch once warm | local cluster |
| Close | median 3.3 s, every payout exact | local cluster |
| One check over 256 slots (earlier design) | median 3.85 s, p95 4.74 s | Arcium devnet cluster 456 |
| Arcium mainnet turnaround, other teams' circuits | 1 to 3 s, 100 of 100 answered | mainnet cluster 10000, 2026-11-07 |
A 256-slot check with per-slot nonces is too large for Arcium to register, so one round covers 128 slots and a bigger book runs rounds in parallel.
Uptime
Over 30 days on mainnet cluster 10000, 99.4% of the node transactions that landed were real answers. Two incidents stand out: on 2026-10-13 from about 09:00 to 10:10, computations timed out or were answered late, and on 2026-10-15 from about 17:50 to 18:40 the same happened. Since 2026-10-16 the longest stretch without an answer was 3.7 minutes.
What uptime means for leverage
While the cluster cannot answer, no position can be checked. Replaying 13 months of 1-second SOL prices with the cluster silent from the worst possible moment gives the leverage that still leaves zero bad debt:
| Silent for | Worst move | Leverage with zero bad debt |
|---|---|---|
| 2 min | 9.3% | 5.4x |
| 10 min | 12.3% | 4.1x |
| 15 min | 15.5% | 3.2x |
| 30 min | 17.5% | 2.9x |
| 1 hour | 18.4% | 2.7x |
The engine allows up to 10x; each market sets its own cap and can change it. The first market starts at 5x, which cleared the worst 2-minute stall in 13 months with zero bad debt, and caps open collateral at 10% of the pool, so a longer stall can only cost the pool a small share.
Cost
One Arcium computation on mainnet cost 0.003505 SOL all in, measured from real queue transactions. A check round covers 128 positions for one fee.