8.4 Supply Chain Mapping
Original Problem in the Paper
Paper motivation/open problem: “Mapping the AI supply chains can allow policymakers to better understand the complex ecosystem involved in the development and deployment of AI systems”; identifying key actors/processes lets policymakers target interventions. Existing export controls on chip exports to Russia/China face enforcement difficulties, AI chips may be smuggled, and understanding resource flows helps combat chip/hardware smuggling. Open problem question: technical methods for an auditable log of all actors and contributions across AI development from data collection to deployment.
July 2026 Update & Trajectory
Supply-chain mapping is operational in narrow digital layers: SPDX AI-SBOM profiles, C2PA AI/ML content credentials, model/data cards, OECD HAIP reports, NIST value-chain actions, and export-control due-diligence guidance. Hardware supply-chain mapping and enforcement remain only partially tractable: BIS guidance and licensing rules exist, but smuggling/diversion, cloud access, shell customers, and multi-jurisdiction semiconductor chokepoints keep the governance problem open. No verified July 2026 end-to-end auditable log covers all actors from raw data/minerals/chips through training, fine-tuning, deployment, and outputs.
Deployed / Operationalized
- SPDX 3.0 and AI-SBOM concepts represent AI systems as machine-readable knowledge graphs of software dependencies, models, data assets, prompt templates, agents, licenses, compliance, ethics/security attributes, and audit trails.
- C2PA AI/ML guidance uses Content Credentials for datasets, software, models, training-data partitions, fine-tuning datasets, model outputs, hashes, signatures, ingredients, and attestations.
- NIST GenAI Profile includes value-chain/component integration risk, third-party resource monitoring, training-data/model documentation, transparency artifacts, and incident-response value-chain details.
- OECD HAIP Reporting Framework invites developers/deployers/providers to report practices across advanced-AI value chain and publishes reports.
- BIS 2026 advanced-computing guidance clarifies license requirements for entities headquartered/parented in Country Group D:5 or Macau even when located elsewhere; BIS extended authorized IC designer timeline to Dec 31 2026.
- CSET AI-chip work maps chokepoints: U.S./allies dominate AI chip design, advanced fabs, EDA, and semiconductor manufacturing equipment.
New Tractable Vectors
- Build AI-SBOM graphs linking deployed agent outputs to prompt, model version, fine-tune data, base model, libraries, APIs/tools, licenses, vulnerabilities, and supplier attestations.
- Cryptographically bind model, dataset, and output provenance with C2PA manifests/hashes/signatures and sidecar manifests for non-media files.
- Use export-control due-diligence records, customer location/parentage, cloud-account telemetry, and chip identifiers to prioritize diversion investigations.
- Standardize procurement questionnaires and audit trails around NIST value-chain risks and HAIP/OECD reporting fields.
- Trace harmful output root causes through prompt/model/data/agent graphs during incident response.
Key Open Questions
- Physical chip traceability: tamper-resistant identifiers, resale tracking, and privacy-preserving verification across distributors, cloud providers, and end users.
- End-to-end integration of hardware, energy, data, labor, model, software, prompt, agent, deployment, and downstream-output supply chains.
- Confidentiality versus auditability for proprietary training data, model weights, security controls, and customer identities.
- Detecting and preventing circumvention via shell entities, cloud compute resale, chip smuggling, and geographically distributed inference/training.
- Interoperability among C2PA, SPDX, model cards, data sheets, EU training-data summaries, cloud logs, and regulatory reporting.