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 and legal-status gaps open

Original Problem in the Paper

The paper presents compute governance as a possible lever for advanced AI governance because training and deployment can require large computing resources. It also warns that compute governance can be blunt: restrictions should ideally target chips and compute clusters relevant to policy-relevant AI development while excluding ordinary scientific, business, or casual high-end computing. The paper's open problems include deriving chip and cluster metrics for AI suitability; understanding how throughput, memory bandwidth, memory capacity, and interconnect bandwidth affect AI workloads; closing downgraded-chip loopholes such as A800-style compliant-but-useful variants; understanding decentralized training across geographically disparate clusters; and comparing many weaker chips with fewer stronger chips at similar theoretical throughput.

July 2026 Update & Trajectory

Policy has narrowly operationalized chip definitions through export-control metrics such as total processing performance (TPP), performance density, and HBM-related controls. BIS's May 31, 2026 headquarters guidance also clarifies that license requirements continue for advanced computing items going to entities headquartered in Country Group D:5 or Macau, or whose ultimate parent companies are headquartered there, even when the entities themselves are located elsewhere. But the regulatory picture is legally delicate: the January 2025 AI Diffusion Rule revised advanced-computing IC controls and updated Data Center VEU authorization, while BIS announced in May 2025 that it would not enforce that rule and planned formal rescission; GAO later stated that the AI Diffusion Framework remains in the CFR until rulemaking is complete and that Commerce's blanket non-enforcement announcement is itself subject to Congressional Review Act requirements.

Technical work has made decentralized and low-communication training more feasible in research demonstrations, including DiLoCo's 8-worker low-communication result and OpenDiLoCo's two-continent/three-country billion-parameter-scale experiments. The sources reviewed here do not establish that decentralized training is equivalent to centralized high-end clusters at frontier-training scale. The scientific problem therefore remains only narrowly operationalized: export-control metrics are proxies, they require continuing revision, they do not fully encode cluster/network/software-stack suitability, and low-communication training progress widens the gap between chip-SKU controls and geographically distributed training capability.

Deployed / Operationalized

  • BIS advanced-computing ECCNs classify covered AI-relevant chips using TPP and performance-density thresholds; the January 2025 AI Diffusion Rule text for ECCN 3A090.a used TPP of 4800 or more, or TPP of 1600 or more plus performance density of 5.92 or more, and Note 2 covered GPUs, TPUs, neural processors, in-memory processors, vision processors, text processors, co-processors/accelerators, adaptive processors, FPLDs, and ASICs.
  • BIS added controls for certain high-bandwidth memory and expressly described HBM as providing memory capacity and bandwidth needed for advanced AI models and supercomputing applications.
  • The January 2025 AI Diffusion Rule updated Data Center VEU authorization for advanced computing IC exports, but that data-center framework should be treated as caveated historical/legal context because BIS announced non-enforcement/planned rescission in May 2025 and GAO stated in 2026 that the framework remains in the CFR until rulemaking is complete.
  • BIS May 31, 2026 guidance confirms that license requirements continue for advanced computing items, including ECCNs 3A090/4A090 and related .z paragraph items, for entities headquartered in Country Group D:5 or Macau, or with ultimate parents headquartered there, even when located elsewhere. A separate BIS FAQ updated June 17, 2026 clarifies application to 3A090.a, 4A090.a, and related .z items.
  • DiLoCo and OpenDiLoCo demonstrate research-grade, low-communication training across poorly connected device islands and, for OpenDiLoCo, geographically distributed clusters across two continents and three countries; this makes cluster-size and geography assumptions less stable at least for research-scale and billion-parameter training.

New Tractable Vectors

  • Map export-control thresholds to commercially available accelerators and HBM stacks using TPP, performance density, memory bandwidth density, and related .z-item containment.
  • Reproduce and stress-test distributed-training efficiency using open DiLoCo/OpenDiLoCo/Hivemind implementations across regions, providers, continents, and countries, while keeping claims bounded to demonstrated model scale and utilization.
  • Assess compliance and diversion risk at the data-center, customer-entity, and ultimate-parent level, not only at the chip-SKU level, using the May 31, 2026 BIS headquarters guidance and FAQ as concrete legal test cases.
  • Compare many-smaller-chip clusters with fewer high-end GPUs under communication-efficient algorithms, including sensitivity to interconnect bandwidth, failure/unavailability, and geographically distributed workers.
  • Track how legal status changes—non-enforcement, rescission rulemaking, replacement rules, and CRA review—affect which technical thresholds are practically enforceable at a given time.

Key Open Questions

  • A specification that robustly captures frontier-training suitability across compute, HBM, interconnect, topology, sparsity/quantization, compiler/runtime, and energy/cooling constraints, rather than relying only on export-control proxy metrics such as TPP and performance density.
  • Governance thresholds robust to downgraded or cutoff-adjacent chips that remain useful for AI workloads, as illustrated by the source paper's A800-style loophole concern.
  • Detecting and attributing decentralized training attempts split across regions, accounts, providers, chip classes, and legal entities without assuming all important training occurs inside one tightly connected cluster.
  • Quantifying when inference clusters, post-training clusters, or fine-tuning clusters become strategically relevant in ways similar to pretraining compute for particular governance objectives.
  • International harmonization and auditability of chip, HBM, and data-center controls without blocking benign scientific, business, or other non-frontier workloads.
  • Maintaining source-defensible public metrics as legal instruments change faster than hardware, networking, and distributed-training methods.

Evidence & Primary Sources

  • The source paper, *Open Problems in Technical AI Governance*, frames compute governance as motivated by large compute requirements for AI training/deployment and identifies the relevant open problems: hardware specifications for AI suitability; distinctions from ordinary high-end compute uses; throughput, memory bandwidth/capacity, and interconnect bandwidth; A800-style loopholes; decentralized training; and many smaller clusters/chips versus fewer larger ones. (2024-07-20; v2 2025-04-16): https://arxiv.org/abs/2407.14981
  • The Framework for Artificial Intelligence Diffusion rule revised advanced-computing IC controls, added controls on model weights for certain advanced closed-weight dual-use AI models, and updated Data Center VEU authorization. In ECCN 3A090.a, it used TPP/performance-density thresholds including TPP of 4800 or more, or TPP of 1600 or more plus performance density of 5.92 or more; Note 2 listed GPUs, TPUs, neural processors, in-memory processors, vision/text processors, co-processors/accelerators, adaptive processors, FPLDs, and ASICs. (2025-01-15): https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion
  • The December 2024 BIS rule added controls for certain high-bandwidth memory; the rule states that HBM provides necessary memory capacity and bandwidth for advanced AI models and supercomputing applications, and added HBM-related controls including ECCN 3A090.c using memory bandwidth density. (2024-12-05): https://www.federalregister.gov/documents/2024/12/05/2024-28270/foreign-produced-direct-product-rule-additions-and-refinements-to-controls-for-advanced-computing
  • The January 2025 BIS due-diligence rule revised the EAR to provide additional due-diligence procedures for advanced computing ICs and recapped earlier controls that adjusted parameters for advanced computing ICs critical for advanced computing and AI applications and imposed measures addressing circumvention risk. (2025-01-16): https://www.federalregister.gov/documents/2025/01/16/2025-00711/implementation-of-additional-due-diligence-measures-for-advanced-computing-integrated-circuits
  • BIS announced on May 13, 2025 that it planned to publish a regulation formalizing rescission of the AI Diffusion Rule and that enforcement officials had been instructed not to enforce that rule. (2025-05-13): https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens
  • GAO decision B-337935 states that Commerce announced non-enforcement and planned rescission of the AI Diffusion Rule; that Commerce explained the AI Diffusion Framework remains in the CFR until rulemaking is complete; and that Commerce's non-enforcement policy announcement is a rule subject to Congressional Review Act submission requirements. (2026-05-12): https://www.gao.gov/products/b-337935
  • BIS May 31, 2026 guidance clarifies that licenses are required to export advanced computing items to entities headquartered in Country Group D:5 or Macau, or with ultimate parent companies headquartered there, even if the entities themselves are located outside Country Group D:5 or Macau. It names advanced computing items including ECCNs 3A090.a/.b, 4A090.a/.b, and related .z paragraph items, and explains the interaction with the January 2025 AI Diffusion Rule and BIS's May 2025 non-enforcement policy. (2026-05-31): https://www.bis.gov/media/documents/bis-guidance-may-31-2026.pdf
  • A BIS FAQ updated June 17, 2026 clarifies that the May 31 guidance applies to 3A090.a ICs and other advanced computing items containing such .a ICs, including 4A090.a and related .z items such as 5A002.z.1.a. (2026-06-17): https://www.bis.gov/media/documents/May-31-FAQ.pdf
  • DiLoCo proposes Distributed Low-Communication training for language models on poorly connected islands of devices; on C4, DiLoCo with 8 workers performed as well as fully synchronous optimization while communicating 500 times less, and the arXiv record shows a September 23, 2024 revision. (2024-09-23): https://arxiv.org/abs/2311.08105
  • OpenDiLoCo is an open-source implementation/replication using Hivemind; its abstract reports training across two continents and three countries with 90-95% compute utilization and scaling to billion-parameter models. (2024-07-10): https://arxiv.org/abs/2407.07852