Skip to main content
BSS/OSS Academy

The AI infrastructure standards landscape

No single body governs AI infrastructure. Each layer of the stack draws on a different mix of formal standards, industry consortia, open specifications, open-source conformance and vendor reference architectures; the table maps them layer by layer, including where proprietary technology still dominates.

  • DMTFFormal standard

    Infrastructure management

    Redfish (2026.1), with a 2026 work register extending it to UALink fabrics

  • IEEEFormal standard

    Benchmarking · Storage & data · Scale-out network

    IEEE 3404-2025, sharing AI data and models across computing centres; IEEE 2937-2022, performance benchmarking for AI server systems; 800 Gb/s Ethernet (802.3df-2024); 1.6 Tb/s (P802.3dj) still in ballot

  • IETFFormal standard

    Scale-out network

    The IP foundations (routing, transport, ECN); no AI-specific working group

  • ISO/IECFormal standard

    AI governance & risk · Facility: power, cooling, building

    ISO/IEC 22237 data-centre facilities series; ISO/IEC 30134-2 defines PUE; ISO/IEC 42001:2023, AI management systems: governance of AI, not hardware

  • JEDECFormal standard

    Chips, memory & I/O

    HBM memory standards, HBM4 published April 2025

  • TIAFormal standard

    Facility: power, cooling, building

    ANSI/TIA-942-C (2024), data-centre infrastructure

  • CXL ConsortiumConsortium specification

    Chips, memory & I/O

    Compute Express Link 4.0 (November 2025)

  • InfiniBand Trade AssociationConsortium specification

    Scale-out network

    InfiniBand architecture, which also defines RoCE (RDMA over Ethernet)

  • MLCommonsConsortium specification

    Benchmarking

    MLPerf Training, Inference and Storage benchmarks: results, not certifications

  • OIFConsortium specification

    Chips, memory & I/O

    Implementation agreements for electrical and optical I/O, incl. co-packaged optics

  • Open Compute ProjectConsortium specification

    Scale-up interconnect · Servers & racks

    Open specifications for servers, racks, power, cooling and networking; describes itself as not a standards body; Ethernet for Scale-Up Networking, a 2025 workstream with early operator requirements

  • PCI-SIGConsortium specification

    Chips, memory & I/O

    PCI Express 7.0 (June 2025)

  • SNIAConsortium specification

    Storage & data

    Storage.AI (a 2025 project, not yet a standard), and an alliance with OCP on AI storage and memory

  • UALink ConsortiumConsortium specification

    Scale-up interconnect

    UALink 200G 1.0 (April 2025), then four further specifications (April 2026)

  • Ultra Ethernet ConsortiumConsortium specification

    Scale-out network

    UEC Specification 1.0 (June 2025), now 1.0.3: Ethernet transport for AI and HPC

  • CNCFOpen source & conformance

    Platform software

    Kubernetes, and the Certified Kubernetes AI Conformance programme (Nov 2025)

  • Open Container InitiativeOpen source & conformance

    Platform software

    Container runtime, image and distribution specifications

  • OpenTelemetryOpen source & conformance

    Platform software

    Vendor-neutral traces, metrics and logs (CNCF graduated)

  • NVIDIAVendor / proprietary

    Scale-up interconnect · Servers & racks

    Reference architectures; contributed GB200 NVL72 rack and tray designs to OCP (2024); Proprietary GPU-to-GPU interconnect; NVLink Fusion (2025) licenses it to partner silicon

  • ASHRAE TC 9.9Guideline, certification or framework

    Facility: power, cooling, building

    Thermal and liquid-cooling guidelines for data processing environments

  • NISTGuideline, certification or framework

    AI governance & risk

    AI Risk Management Framework 1.0 (2023): voluntary risk guidance

  • Uptime InstituteGuideline, certification or framework

    Facility: power, cooling, building

    Tier certification: a private certification scheme, not a formal standard