company · explainer · draft

OpenAI (training fabric fragment)

OpenAI has published resilient AI training-cluster networking (MRC multipath RDMA + static SRv6 source routing) and capacity partnership notes (e.g. Stargate / OCI). ChatGPT/API public edge (CDN, anycast, PoPs, ASN) is not publicly documented in the curated primaries — keep it in Known unknowns.

In one sentence. Teaching the published MRC/SRv6 training-fabric fragment without inventing an inference-edge anycast map or freezing paper figures as live cluster inventories.

Why it exists

Large GPU training fabrics need high throughput under failures. MRC spreads a transfer across many paths and stops using failed paths quickly; static SRv6 embeds path (uSID) so switches need not run adaptive routing that fights MRC (OpenAI MRC post + PDF). Cross-link: `/architectures/spine-leaf-clos/` for Clos teaching contrast — MRC paper compares multi-plane themes, not an OpenAI site map.

Visual walkthrough

Walkthrough

MRC multipath RDMA + static SRv6

Published training-cluster networking themes: spray across paths, stop using failed paths, embed path with static SRv6. Edge/CDN for ChatGPT remains a known unknown.

Step 1 / 3
RDMA transferSpray pathsGPU trainerRDMA NICMultipath fabricmany pathsPeer GPU

A large training transfer spreads across many fabric paths (MRC multipath RDMA). Single-path RDMA collapses when a link fails; MRC is published as resilience for AI training fabrics.

Control vs data plane

Control plane

Static SRv6 source routing with dynamic routing disabled in switches (as published). This is cluster fabric control, not Internet BGP anycast. Stargate posts name compute partnerships without disclosing edge/CDN control planes.

Data plane

Multipath RDMA (MRC) across high-speed training fabrics; NICs/switches called out in the paper are product examples in that publication window — era-label them. Inference/API packet paths are unpublished.

Request / packet path

North–south

Training north–south is partner datacenter networking (Microsoft, OCI, etc.) named in capacity posts — fabric internals beyond MRC remain private. Public ChatGPT/API edge path is a known unknown.

East–west

East–west inside the training cluster follows MRC/SRv6 multipath themes. Non-training WAN between sites is unpublished.

Scaling & math

Do not invent cluster sizes, oversubscription ratios, or site lists beyond illustrative paper figures / marketing capacity posts. Reject third-party “OpenAI vs Anthropic network” rumor articles as non-primary.

When it breaks

  • Training transfer throughput collapses when paths fail.

    Cause. Without multipath resilience, single-path RDMA suffers from failed links.

    Mitigation. MRC stops using failed paths quickly and sprays across many paths as published.

  • Adaptive switch routing fights multipath spraying.

    Cause. Dynamic routing in switches conflicts with MRC path selection.

    Mitigation. Paper/post: static SRv6 source routing embeds path; disable adaptive routing that fights MRC.

  • Cannot debug ChatGPT/API edge with primary topology docs.

    Cause. Public edge/CDN/anycast/ASN architecture not published in curated sources.

    Mitigation. Keep edge in Known unknowns; do not invent Cloudflare-like diagrams.

Misconceptions

  • “MRC/SRv6 posts disclose ChatGPT’s public CDN/anycast edge.” — They document training-cluster networking.
  • “Partnership capacity posts are a full site inventory.” — They name partners/regions at marketing depth; fabric beyond MRC stays private.

Reference expression

Primary: MRC networking post; MRC/SRv6 PDF; Stargate compute infrastructure post. Optional adjacent: Habitat storage / PostgreSQL scaling posts are product storage narratives, not edge topology. review: draft.

Standards & sources

Known unknowns

  • ChatGPT / API public edge architecture (CDN, anycast, PoPs, ASN) — not publicly documented.
  • Non-training WAN between sites, peering policy, and DDoS posture.
  • Exact cluster sizes, oversubscription ratios, and site lists beyond illustrative paper figures / marketing capacity posts.
  • Partner datacenter fabric internals beyond the MRC publication.

Check yourself

  1. MRC as published primarily addresses…

    • Internet anycast PoP maps for ChatGPT
    • Multipath RDMA resilience in AI training fabrics
    • Airbnb SmartStack discovery
    • Starlink Ka gateway licensing

    Answer: Multipath RDMA resilience in AI training fabrics. OpenAI MRC post/PDF: training-cluster multipath RDMA.

  2. Static SRv6 in the MRC design is used to…

    • Replace all GPU NICs with Wi-Fi
    • Embed path (uSID) so switches need not run adaptive routing that fights MRC
    • Announce AS13335
    • Configure Front Door VIP rings

    Answer: Embed path (uSID) so switches need not run adaptive routing that fights MRC. MRC/SRv6 PDF: static source routing; dynamic routing disabled.

  3. OpenAI public edge (CDN/anycast/ASN) for ChatGPT/API is…

    • Fully documented like Unimog
    • A known unknown in curated primaries
    • Identical to Netflix Open Connect AS2906
    • Proven by PeeringDB alone as primary

    Answer: A known unknown in curated primaries. Phase 7 ship rule: edge = unknowns.

  4. Stargate compute posts should be read as…

    • Full packet-level fabric dumps
    • Capacity / partnership narratives — not edge topology disclosures
    • FCC Gen2 satellite authorizations
    • Service Fabric 2018 live maps

    Answer: Capacity / partnership narratives — not edge topology disclosures. Capacity partnerships ≠ published edge architecture.