SCOI Atlas · L0 Infrastructure
L0C Interconnect Fabric
NVLink, InfiniBand, Ethernet-based scale-out — the pipes between accelerators.
TL;DR · Direct answer
L0C Interconnect Fabric is a sublayer of L0 Infrastructure in the Supply Chain of Intelligence™ (SCOI) by Anand Arivukkarasu. It is a required capability but not on its own a durable moat. For inference, network is usually a solved problem; for training, it is the problem.
What actually matters at L0C
- Training run efficiency is dominated by all-reduce bandwidth, not FLOPs.
- The Ethernet-vs-InfiniBand war is real; both will exist and both will lock you in differently.
- Cross-DC training is not yet routine; assume single-facility for frontier work.
The startup lens
For inference, network is usually a solved problem; for training, it is the problem.
Vertical lens — how this plays across categories
Foundation model labs
Interconnect is a Board topic, not an infra topic.
Fine-tune-heavy SaaS
Prefer providers that expose topology, so LoRA jobs land on the right islands.
Edge AI
Local interconnect (PCIe gen5, CXL) will define next-gen edge boxes.
How to defend L0C
- Provider selection based on topology transparency.
- Job schedulers that pack workloads to fabric islands.
Other sublayers in L0 Infrastructure
L0A
Silicon & Memory
Accelerators (Nvidia, AMD, TPU, Trainium, custom), HBM, and the memory hierarchy that dominates training cost.
L0B
Data Centers
Powered shells, colo, hyperscale — the physical containers of intelligence.
L0D
Compute & State Infrastructure
Vector DBs, object stores, feature stores, KV caches — persistent state around ephemeral compute.
L0E
Edge & On-Device Compute
NPUs in phones, laptops, cars, wearables and industrial devices.