Back to Dashboard
Assessment / Compute/ 3.2.1

3.2.1 Definition of Chip and Cluster Specifications for Model Training

2026 Governance Status: Narrowly operationalized; decentralized-training gap open

Original Problem in the Paper

Paper motivation: compute governance can target advanced AI because training/deployment requires large compute, but blunt chip restrictions risk overbreadth. Dedicated open problems: derive metrics/specs capturing chips/clusters relevant to AI while excluding ordinary compute; understand how throughput, memory bandwidth/capacity, and interconnect bandwidth affect suitability; close loopholes like downgraded chips still useful for AI; understand decentralized training and efficiency/cost of many small clusters/chips versus fewer powerful ones.

July 2026 Update & Trajectory

Policy has operationalized chip definitions using TPP/performance density, HBM controls, data-center authorizations, and 2026 entity-headquarters guidance. Technical work has made decentralized/low-communication training more feasible. But the scientific problem remains only narrowly operationalized: export-control metrics are proxies, still need continual revision, do not fully encode cluster/network/software stack suitability, and decentralized training progress creates new governance gaps.

Deployed / Operationalized

  • BIS advanced-computing ECCNs classify AI chips by total processing performance and performance density, with notes covering GPUs, TPUs, neural processors, accelerators, FPGAs, and ASICs.
  • BIS added/refined HBM and data-center/validated-end-user controls, acknowledging memory bandwidth/capacity and cluster/data-center context matter for advanced AI.
  • BIS May 31 2026 guidance clarifies licenses remain required for advanced computing items to entities headquartered in Country Group D:5 or Macau even when located elsewhere.
  • DiLoCo/OpenDiLoCo operationalize distributed low-communication training across poorly connected or geographically distributed clusters, indicating cluster-size/geography assumptions are less stable.

New Tractable Vectors

  • Map export-control thresholds to commercially available accelerators and HBM stacks using TPP, performance density, memory bandwidth density, and .z-item containment.
  • Study global/distributed training efficiency using open DiLoCo/Hivemind implementations across continents and countries.
  • Assess compliance and diversion risk at data-center/customer-entity level, not only chip SKU level.
  • Compare many-smaller-chip clusters versus fewer high-end GPUs under communication-efficient algorithms.

Key Open Questions

  • Specification that robustly captures frontier-training suitability across compute, HBM, interconnect, topology, sparsity/quantization, compiler/runtime, and energy/cooling constraints.
  • Governance thresholds that remain effective as vendors optimize just below TPP/performance-density cutoffs.
  • Detecting and attributing decentralized training runs split across regions, accounts, providers, and chip classes.
  • Quantifying when inference clusters or post-training/fine-tuning clusters become strategically equivalent to pretraining compute.
  • International harmonization and auditability of data-center controls without blocking benign scientific/business workloads.

Evidence & Primary Sources