idr M. Han, Ed. Internet-Draft Zheng Ruan Intended status: Standards Track China Unicom Expires: 6 May 2027 2 November 2026 BGP Extension for AI Compute Service Metadata draft-ruan-idr-ai-compute-service-metadata-01 Abstract This document defines four AI inference-specific Sub-TLVs for the BGP Edge Metadata Path Attribute specified in [I-D.ietf-idr-5g-edge-service-metadata]. They advertise model capabilities, aggregated inference latency, optional cost preference, and selected long-lived KV cache prefix identifiers. Authorized CATS Path Selectors (C-PSes) and other policy decision functions may use this information to select suitable service contact instances and forwarding paths. This document reuses the existing optional non- transitive attribute. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 6 May 2027. Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components Han & Zheng Ruan Expires 6 May 2027 [Page 1] Internet-Draft BGP Extension for AI Compute Service Met November 2026 extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License. Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 2. Terminology and Definitions . . . . . . . . . . . . . . . . . 2 3. AI Inference Sub-TLVs . . . . . . . . . . . . . . . . . . . . 3 3.1. AI Model Capability Sub-TLV (TBD1) . . . . . . . . . . . 3 3.2. Inference-SLA Sub-TLV (TBD2) . . . . . . . . . . . . . . 3 3.3. Inference-Billing Sub-TLV (TBD3) . . . . . . . . . . . . 4 3.4. KV-Cache-Prefix Sub-TLV (TBD4) . . . . . . . . . . . . . 4 4. 4. Processing and Selection . . . . . . . . . . . . . . . . 5 4.1. Model Correlation . . . . . . . . . . . . . . . . . . . . 5 4.2. Advertisement Control . . . . . . . . . . . . . . . . . . 5 4.3. Use by CATS . . . . . . . . . . . . . . . . . . . . . . . 5 5. Operational and Security Considerations . . . . . . . . . . . 6 6. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 6 7. Normative References . . . . . . . . . . . . . . . . . . . . 7 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 7 1. Introduction Distributed AI inference services may be deployed at multiple sites, each supporting different models and service conditions. One service contact instance [RFC10053] may provide access to multiple models. A CATS Path Selector should be able to determine which sites support a requested model before evaluating model-specific performance, available resources, and network paths. The Edge Metadata Path Attribute [I-D.ietf-idr-5g-edge-service-metadata] already carries generic service capability, delay, and resource indicators. This document adds AI-specific model eligibility, generation latency, optional relative cost, and a limited set of persistent prefix identifiers. Generic metrics SHOULD continue to use the existing Sub-TLVs. 2. Terminology and Definitions * Model-ID: ASCII-encoded identifier of the AI inference model, with a maximum length of 64 octets. * AI Model Capability: A declarative indication that a service contact instance can reach a specified model service . A declaration is not a guarantee that a new request can currently be admitted. Han & Zheng Ruan Expires 6 May 2027 [Page 2] Internet-Draft BGP Extension for AI Compute Service Met November 2026 * TTFT: Time to First Token, measured in milliseconds according to an explicitly configured workload profile. * TPOT: Time per Output Token, measured in milliseconds according to an explicitly configured workload profile. * Fixed-Prefix-Key: An opaque identifier for a long-lived, stable KV prefix, including a fixed system prompt prefix or a specially designated persistent customer-specific prefix. It MUST NOT contain raw prompts or serialized KV tensors. 3. AI Inference Sub-TLVs The four Sub-TLVs below reuse the header, encoding rules, and processing constraints of [I-D.ietf-idr-5g-edge-service-metadata]. The 8-bit Length field counts Value octets only; multi-octet fields use network byte order. The Sub-Type codepoints are TBD. 3.1. AI Model Capability Sub-TLV (TBD1) This Sub-TLV advertises the AI models available through a service contact instance, including model identity, revision, and capability characteristics, to support model-aware service selection. The fields appear in the following order: 1. Model-ID (variable-length string): ASCII-encoded identifier of the AI inference model, with a maximum length of 64 octets. 2. Model-Characteristics (uint32): Bitmask indicating the model's characteristics, including general-purpose capabilities, industry-specific specialization (e.g., healthcare, finance, or industrial applications), and specialized capabilities (e.g., reasoning, coding, or multimodal understanding). Multiple characteristics MAY be indicated simultaneously. 3.2. Inference-SLA Sub-TLV (TBD2) This Sub-TLV carries aggregated inference latency measurements for one model service. 1. Model-ID (variable-length string): ASCII-encoded identifier of the AI inference model, with a maximum length of 64 octets. 2. Avg-TTFT (uint16): Average Time-To-First-Token latency measured in milliseconds. 3. Avg-TPOT (uint16): Average Time-Per-Output-Token latency measured in milliseconds. Han & Zheng Ruan Expires 6 May 2027 [Page 3] Internet-Draft BGP Extension for AI Compute Service Met November 2026 4. Current-TPS (uint32): Real-time token generation throughput expressed as tokens per second. 5. Pending-Queue-Num (uint32): Count of queued unprocessed inference tasks waiting for GPU accelerator resources. 3.3. Inference-Billing Sub-TLV (TBD3) This Sub-TLV advertises relative inference cost indicators for individual models to support cost-aware service selection. It is updated only when the corresponding cost policies change. The fields appear in the following order: 1. Model-ID (variable-length string): ASCII-encoded identifier of the AI inference model, with a maximum length of 64 octets. 2. Cost-Index (uint16): Relative inference cost indicator derived from cache hit/miss characteristics and associated computational costs. A lower value indicates a lower estimated inference cost. Values are comparable only within a common administrative cost policy. 3.4. KV-Cache-Prefix Sub-TLV (TBD4) This Sub-TLV advertises a limited set of stable, long-lived KV cache prefix identifiers associated with individual models to support cache-aware service selection. The fields appear in the following order: 1. Model-ID (variable-length string): ASCII-encoded identifier of the AI inference model. 2. Cache-Compatibility-ID (uint32): Administratively assigned identifier representing the cache compatibility profile, including the model weights, tokenizer configuration, and relevant inference settings. 3. Prefix-Class (uint8): Indicates the prefix category: 1 for a fixed system prompt prefix, and 2 for an explicitly authorized, long-lived customer-specific prefix. 4.Fixed-Prefix-Key (128 bits): Opaque identifier of a reusable KV cache prefix, used to identify candidate inference sites with matching persistent prefixes. Han & Zheng Ruan Expires 6 May 2027 [Page 4] Internet-Draft BGP Extension for AI Compute Service Met November 2026 Only a bounded number of long-lived prefixes associated with fixed system prompts or explicitly retained customer-specific contexts MAY be advertised. Short-lived KV cache entries derived from dynamic user contexts, conversation histories, or individual requests MUST NOT be advertised. 4. 4. Processing and Selection 4.1. Model Correlation The receiver correlates the four AI inference Sub-TLVs using the Model-ID and the associated service route. The AI Model Capability Sub-TLV identifies the models available through a service contact instance. The Inference-SLA, Inference-Billing, and KV-Cache-Prefix Sub-TLVs provide additional information for those models. A receiver MUST NOT consider a model available solely based on an Inference-SLA, Inference-Billing, or KV-Cache-Prefix advertisement without a corresponding AI Model Capability declaration. 4.2. Advertisement Control AI Model Capability and KV-Cache-Prefix information SHOULD be updated only when the corresponding model availability or long-lived prefix information changes. Inference-SLA measurements MUST be aggregated, and their advertisement SHOULD use configurable thresholds and hysteresis to limit unnecessary BGP updates. Inference-Billing information SHOULD be updated only when the associated billing parameters change. BGP updates MUST NOT be triggered by individual inference requests, cache hits, cache misses, or cache eviction events. The number of advertised Fixed-Prefix-Key entries MUST be bounded by administrative policy. All advertisements MUST follow the capability negotiation, distribution scope, and update control procedures defined in [I-D.ietf-idr-5g-edge-service-metadata]. 4.3. Use by CATS An authorized CATS Path Selector (C-PS) MAY use the advertised AI inference metadata to identify suitable service contact instances and select forwarding paths. The C-PS first identifies candidate service contact instances supporting the requested Model-ID through the AI Model Capability Sub-TLV. It may then consider the corresponding Inference-SLA, Han & Zheng Ruan Expires 6 May 2027 [Page 5] Internet-Draft BGP Extension for AI Compute Service Met November 2026 Inference-Billing, and KV-Cache-Prefix information, together with network metrics and local policies, when selecting a suitable service contact instance and forwarding path. For cache-aware selection, the Fixed-Prefix-Key and Cache- Compatibility-ID MAY be used to identify candidate service contact instances with potentially reusable KV cache prefixes. A matching identifier does not guarantee actual cache availability, which remains subject to verification by the inference service layer. The selected forwarding path is executed by the ingress CATS- Forwarder. Fine-grained inference instance scheduling and request- level KV cache management remain outside the scope of this specification. 5. Operational and Security Considerations Model inventories, latency, billing preferences, and customer- specific cache identifiers may be sensitive. Originators MUST limit metadata dissemination to authorized domains according to the non- transitive Edge Metadata attribute and its export policies. Non- transitivity alone is not a confidentiality mechanism. Identifiers MUST be consistently interpreted within their scope. Opaque keys SHOULD be difficult to guess; hashing alone does not guarantee privacy. Implementations SHOULD validate input values, stale metadata, and model-version compatibility, and MUST bound advertisement size. BGP is not a complete model registry or dynamic KV-cache directory. 6. IANA Considerations This document requests four assignments from the Edge Metadata Sub- TLV registry established by [I-D.ietf-idr-5g-edge-service-metadata]: +==========+=====================+ | Sub-Type | Description | +==========+=====================+ | TBD1 | AI Model Capability | +----------+---------------------+ | TBD2 | Inference-SLA | +----------+---------------------+ | TBD3 | Inference-Billing | +----------+---------------------+ | TBD4 | KV-Cache-Prefix | +----------+---------------------+ Table 1 Han & Zheng Ruan Expires 6 May 2027 [Page 6] Internet-Draft BGP Extension for AI Compute Service Met November 2026 7. Normative References [RFC10053] Li, C., Ed., Du, Z., Boucadair, M., Ed., Contreras, L. M., and J. Drake, "A Framework for Computing-Aware Traffic Steering (CATS)", RFC 10053, DOI 10.17487/RFC10053, October 2026, . [I-D.ietf-idr-5g-edge-service-metadata] Dunbar, L., Majumdar, K., Li, C., Mishra, G. S., and Z. Du, "BGP Extension for 5G Edge Service Metadata", Work in Progress, Internet-Draft, draft-ietf-idr-5g-edge-service- metadata-34, 25 September 2026, . Authors' Addresses MengYao Han (editor) China Unicom Beijing China Email: hanmy12@chinaunicom.cn Zheng Ruan China Unicom Beijing China Email: ruanz6@chinaunicom.cn Han & Zheng Ruan Expires 6 May 2027 [Page 7]