Throughout the competition, our matching engine ran on deliberately constrained infrastructure —
just three machines with 72 vCPUs, which processed $93,550,000,000 in notional volume and 1.57M execution reports,
driven by 242 autonomous trading agents built and deployed by participants.
Platform-layer CPU never exceeded 9%. That's not the limit — that's headroom.
The agents participants built made their own decisions, called tools, placed orders and ran risk control autonomously on the platform — the figures below are real activity captured in live production.
Cumulative figures measured throughout in production (6/21 21:00 → 6/26 21:00 UTC, 5 trading days) — notional volume, matched trades and full-lifecycle execution reports, plus the WAL records used for state recovery and consistency auditing.
The highest instantaneous rates appeared during the pre-close surge. The real test isn't sustained pressure — it's these instantaneous spikes. And platform-layer CPU stayed below 9% throughout.
Peak rates run roughly 25–30× the active average — the load is extremely spiky. What the system really has to absorb is the instantaneous spike, not sustained saturation.
Real engineering strength isn't "we can hold up if we add machines" — it's "we have ample room without adding any." This competition was a public stress test: we held the infrastructure well below industry norms and let the engine stand on its own, letting the results speak.
All production-measured instance types. The platform layer carried the whole competition system on a single machine; the trading engine completed matching and hedging on a minimal topology, running low throughout.
Our measured footprint placed side by side with the typical industry scale of a regional commercial bank's core execution system — comparable throughput, yet a stark gap in hardware footprint and total cost of ownership.
| Metric | Syphonix (measured) | Regional commercial bank (typical industry scale) |
|---|---|---|
| Core execution servers | 3 servers (72 vCPU) full measured footprint |
20–80 production servers excluding disaster recovery |
| Deployment architecture | Single production topology lightweight deployment |
Multi-tier cluster + HA + DR high availability · geo-redundant DR |
| Platform TCO | ≈ $386 / 5 days measured · ≈ $3.2/hr |
≈ $20k–100k / 5 days typical industry scale · illustrative |
| Peak CPU utilization | < 9% 16-core platform, measured |
Large redundant capacity provisioned for peak load + failover |
| Throughput | 200 msg/s · 95 fills/s single minimal topology |
Multi-node cluster comparable workloads shared across nodes |
| Infrastructure efficiency | Institutional execution on a minimal hardware footprint |
Prioritizes reliability & compliance HA and regulation over utilization |
From Cambridge, Oxford, Imperial, UCL, KCL and 41 universities in total, alongside hedge-fund practitioners — a high-density, high-caliber real arena.
Beyond scale, what matters more is quality: every trade was processed and completed by a real external LP, with a substantial size per trade.
Daily trade counts by trading day (21:00 UTC / NY 17:00 boundary). From 63K trades on day one all the way up to 280K on the final day — the last day, layered with the final hour before close, set both the trades peak and the per-minute peak: a textbook end-of-competition sprint.
$93,550,000,000 notional, 1.57M executions, 100% real LP fills — driven by autonomous AI agents, on 3 deliberately under-provisioned machines with platform CPU never crossing 9%. The same engine has multiples of headroom, no rearchitecture required.