See the architecture. Then see it live.
Effiniti makes the physical layer self-adaptable. It creates dedicated optical paths per-flow in 150 microseconds, over standard Ethernet, with zero packet loss, so GPUs stop waiting and one network does the work of three. This page is the proof. Every claim carries a number, and every number is measured or on the roadmap.
Reshapes Layer 1 in 150µs, without disruption
Full Layer 1–4 integration in a standard 2U form factor. Creates optimal optical bypasses in 150 microseconds without disrupting active traffic. Drop-in compatible with existing datacenter infrastructure, with hot-swappable electro-optical modules for continuous operation and incremental upgrades.
Distributed intelligence, decided per flow
Each NIC independently detects, classifies, and optimizes flows, classifying traffic within 2–3 packets. Zero-wait operation: traffic flows immediately while optimization happens in parallel, with seamless path migration and zero packet loss. Open platform: a standalone card, or integrated into a partner SmartNIC or DPU.
Dedicated physical paths, created on demand.
Dynamic Optical Bypass creates dedicated physical paths on demand, per-flow, delivering latency comparable to a direct fiber connection while cutting hop count from 3–5 to 0–2. Whether a link is formed electronically or optically inside the switch is an implementation choice. The point is that the physical layer continuously reconfigures, without disruption, at microsecond speed.
The switch reshapes Layer 1. The card decides, per flow, what Layer 1 should be. Neither waits on a central controller. Reconfiguration is normal operation, not a disruptive event, because the switch and the card are co-designed for it.
Detect
The SNIC sees a new flow and classifies it within 2–3 packets. Traffic starts moving immediately.
Decide
The card decides, per flow, what the physical path should be. No round-trip to a central controller.
Reshape
The switch forms a dedicated optical bypass in 150µs, without disrupting active traffic.
Migrate
The flow moves to its new path with zero packet loss. One fabric, every workload.
Why a faster controller runs out of room.
SDN with optical circuit switching puts a millisecond controller on top of a static physical layer and steers a few circuits network-wide. It is the closest existing approach, and it is not the same thing. Effiniti makes the physical layer itself adapt, per-flow, at microsecond speed.
| Dimension | SDN + optical circuit switching | Effiniti · self-adaptable |
|---|---|---|
| What adapts | A controller on top of a static physical layer | The physical layer itself |
| Granularity | A few circuits, network-wide | Per-flow |
| Reconfiguration time | Milliseconds | 150 microseconds |
| Decision path | Central controller round-trip | Distributed, on the card |
| Disruption on change | Reconfiguration is an event | Normal operation, zero packet loss |
| Networks required | Still three or more | One adaptive fabric |
| Adoption | New layer to operate | Standard Ethernet, top-of-rack swap |
Owning the comparison is owning the category. The account is technical and fair, never a smear.
Utilization from below 45% to above 85%.
A static Layer 1 leaves GPUs waiting: over 32% of xPU time is spent waiting for data, and utilization sits below 45%. A faster controller helps a little. Effiniti roughly doubles xPU utilization by making one adaptive network out of three.
Figures are Astrape's technology-page numbers, shown live at the Demo Center. Every published result carries the condition it was measured under.
| Form factor | 2U |
| Layers integrated | L1–4 |
| Per-port speed | 100G – 1.6T |
| Port count | 48 – 256 |
| Bypass latency | 150µs |
| Packet loss on migration | 0 |
| Modules | Hot-swap electro-optical |
| Stack | Standard Ethernet · SONiC |
| Classification | 2–3 packets |
| Operation | Zero-wait, parallel |
| Path migration | Seamless · 0 loss |
| Decision model | Distributed, no controller |
| Deployment | Standalone card |
| Integration | Partner SmartNIC / DPU |
| Lock-in | None · open platform |
| Hops | 0–2 (from 3–5) |
Product renders of the switch and the SNIC card go here as they are produced. No stock imagery: the actual switch, the actual architecture, the actual numbers.
A scaling line, not a one-off.
EFFINITI-100 is in pilot now, scaling to 200, 600, 2000, and 8000 as port speeds arrive. An early deployment grows rather than gets replaced. For a cautious infrastructure buyer, the roadmap itself is a de-risking argument, and dates met are the deepest form of trust.
| EFFINITI-100 | 24 × 100G | Pilot now |
| EFFINITI-200 | 48 × 800G | Roadmap |
| EFFINITI-600 | Scaling | Roadmap |
| EFFINITI-2000 | 256 × 1.6T | Roadmap |
| Scale target | 16M+ servers | Multi-site |
The savings compound, because they come from one change.
You stop buying, cabling, powering, and staffing the same capacity three times over. One fabric, one network to manage, any port serving any workload.
Argue with the architecture. Then watch it run.
Deep-tech networking is believed when it is seen. Come to Belgium and watch Effiniti reconfigure the physical layer per-flow, with your own scenarios and our engineers.