Supply Chain of Intelligence™(SCoI)
    Full framework
    pickaxe

    Layer L0

    Infrastructure

    The shovels. Chips, data centers, networking, cloud, edge, what is needed to process intelligence.

    Why it matters

    The compute substrate. NVIDIA doesn't care which model wins, they sell to all of them. When L2 commoditizes, value accrues to L0.

    The Shovels & Mining Equipment

    Before anyone finds gold, someone has to build the pickaxes, drill rigs, and mine shafts. In AI: NVIDIA builds the GPUs, CoreWeave builds the data centers, hyperscalers run the clouds. No shovels → no gold rush. Shovel sellers outlast most miners.

    The 5 sublayers

    L0a

    Silicon & Memory

    GPUs, TPUs, custom AI accelerators, plus HBM and high-bandwidth memory (SK Hynix, Micron, Samsung), the invisible bottleneck behind every chip cycle

    L0b

    Data Centers

    Physical facilities housing compute at scale

    L0c

    Interconnect Fabric

    Networking between chips, racks, regions, clouds

    L0d

    Compute & State Infrastructure

    On-demand compute, scheduling, and durable agent state (checkpointing, workflow state, runtime memory stores)

    L0e

    Edge & On-Device Compute

    Local inference on phones, vehicles, sensors, endpoints

    , Layer diagnostic card · SCOI v1

    Is a company really at L0?

    The compute substrate: chips, interconnect, data centers, cloud, edge, the shovels of the AI gold rush.

    Inclusion tests · include if ALL

    • Sells or operates physical compute capacity (silicon, racks, regions, edge devices).
    • Revenue scales with tokens/inferences processed, not seats or end-users.
    • Model-agnostic, wins whether OpenAI or Anthropic wins.

    Exclusion tests · exclude if ANY

    • Rents compute and resells access wrapped as 'AI infrastructure'.
    • Sells dev-tools that run on someone else's compute.
    • Calls itself 'AI infra' but the durable asset is a model (L2) or a workflow (L5).

    The L0 removal test

    Remove L0 and L2 cannot train, L7 cannot serve. There is no software substitute for compute capacity.

    Economic work this layer does

    Turns L-1 power and silicon into addressable, schedulable compute that L2 and L5 can rent.

    Canonical examples

    • NVIDIA

      Owns the GPU + CUDA stack. Sells to every layer above without picking sides.

    • CoreWeave

      Owns the data-center build-out and GPU fleet others lease against.

    • AWS / Azure / GCP

      Hyperscalers, they own L0 capacity plus the L4 distribution rails on top.

    Anti-examples · look-alikes that fail

    • GPU-arbitrage startups

      Reselling rented capacity. No structural cost advantage, no moat.

    • Most "AI cloud" wrappers

      Marketing label over a thin scheduler on someone else's GPUs.

    • Inference-API startups w/o silicon

      L4/L6 dressed as L0. First margin compression kills them.

    Disagree with a classification?Open the classification table →

    Who's playing here

    NVIDIAAMDTSMCCoreWeaveEquinix

    Verdict: Shovel sellers win every gold rush.

    L-1 ResourcesL1 Data