Compute Governance: Tracking and Limiting Access to Frontier Training Runs

Why Compute Became the Governance Choke Point

As frontier artificial intelligence systems have grown more capable, policymakers have searched for a practical point of control. Regulating algorithms directly is difficult because model architectures and training techniques evolve constantly and are hard to inspect from the outside. Regulating model weights after training is also difficult, since weights can be copied, leaked, or distributed instantly once they exist. Compute, meaning the physical chips and data center capacity used to train large models, is different: it is manufactured by a small number of companies, requires enormous capital investment, consumes vast amounts of electricity, and physically exists in identifiable locations. That combination of scarcity and visibility is why compute governance, the practice of using access to advanced computing hardware as a policy lever, has become the centerpiece of frontier AI regulation over the past several years.

The logic is straightforward even if the implementation is not: if a small number of chip designers and a small number of cloud providers control the vast majority of the hardware capable of training frontier-scale models, then governments can exert meaningful influence over who trains what by controlling who gets access to that hardware, and by requiring disclosure when very large training runs happen at all.

The US Executive Order and Training-Run Reporting

A central piece of the US approach requires AI developers to report training runs that exceed a defined compute threshold, measured in floating-point operations. The same executive order extends the reporting obligation to cloud computing providers, requiring them to report large training runs conducted by non-US customers once those runs cross the same threshold. The effect is to create visibility into frontier-scale training activity regardless of whether it happens inside a US company’s own data centers or on rented cloud infrastructure serving a foreign customer.

This reporting-based approach is deliberately lighter-touch than an outright licensing regime for training itself. It does not require government pre-approval to train a large model; it requires disclosure after the fact, or as the run is underway, so that regulators have visibility into where the frontier of capability is moving. That distinction, between reporting and licensing, has been one of the more durable design choices in compute governance, since it lets policymakers keep track of frontier developments without directly halting research.

Chip Export Controls Tighten Further in 2026

While reporting obligations govern what happens after a chip is already in a data center, export controls govern who can obtain the chip in the first place. A final rule published on January 15, 2026 tightened the technical thresholds used to define which chips are restricted for export to certain destinations. Under that rule, chips exceeding a Total Processing Performance threshold of 21,000, or a bandwidth threshold of 6,500 gigabytes per second, remain fully restricted for export to specified destinations. Alongside the tightened thresholds, the same rule package introduced a 25 percent tariff on advanced computing semiconductors that meet those same performance thresholds, adding a financial lever on top of the licensing restrictions.

By late 2025, license applications for advanced computing exports to China and Macau were governed by a broad presumption-of-denial policy, meaning that, absent a specific policy exception, applications were expected to be rejected rather than approved. In early March 2026, the Department of Commerce went further and circulated a draft rule that would have required licenses for virtually all AI chip exports globally, a dramatic expansion beyond destination-specific restrictions. That draft was ultimately withdrawn on March 13, 2026, after industry and allied-government pushback, illustrating how contested and fast-moving this area of policy remains even within a single year.

The Reported Escalation to Models Themselves

Compliance-focused publication ComplianceHub.Wiki, in an analysis titled “From Chips to Models,” described the export control apparatus as having climbed steadily up the AI stack since 2022: starting with chips, extending to semiconductor manufacturing equipment, then to raw compute and the cloud access that delivers it, and, as of mid-2026, reportedly reaching deployed models themselves. According to that reporting and related coverage, a June 12, 2026 letter issued under Commerce Secretary Howard Lutnick instructed Anthropic to disable its Fable 5 and Mythos 5 models worldwide for foreign nationals, including its own foreign-national employees.

This claim should be treated with appropriate care since it traces to a relatively small number of sources reporting on a single, unusually specific episode. What multiple outlets appear to agree on is that the restriction was subsequently lifted, roughly three weeks later, after Anthropic agreed to additional safeguards, and that the episode is widely viewed as the first time export-control authority was pointed directly at a deployed commercial model version rather than at chips, manufacturing equipment, or raw compute access. Legal commentators have since debated whether the underlying statutory authority actually extends this far, making the episode as much a live legal question as a settled precedent.

Whether or not the details of the Lutnick letter are eventually confirmed in full by additional sources, the broader trajectory it illustrates, export control authority extending further up the AI stack over time, from raw silicon toward the software and models running on top of it, is consistent with the pattern described across multiple compute governance analyses of 2025 and 2026.

International Responses Beyond the United States

Compute governance is not solely a US phenomenon. South Korea’s AI Basic Act entered into force on January 22, 2026, establishing risk management obligations for high-performance AI systems trained using 10 to the 26th power floating-point operations or more, with phased enforcement scheduled to begin in 2027. That threshold sits in a similar order of magnitude to the compute levels discussed in comparable frontier-model provisions elsewhere, reflecting a degree of informal convergence among regulators internationally on what counts as a frontier-scale training run, even without a formal shared international standard.

The European Union has likewise built compute-based thresholds into its AI regulatory framework, using training compute as one of the signals for identifying general-purpose AI models with systemic risk. Taken together, the emerging picture is one of parallel national and regional compute-reporting regimes rather than a single unified global compute governance treaty, with the United States, the European Union, and South Korea each defining thresholds independently, occasionally converging in magnitude but not through formal coordination.

Why Compute Thresholds Are an Imperfect Proxy

Despite the growing regulatory infrastructure built around compute thresholds, researchers at organizations such as the Institute for Law and AI have raised a structural concern: compute is a proxy for capability and risk, not a direct measurement of either. Algorithmic efficiency improvements steadily reduce the amount of training compute needed to reach a given level of model capability, which means a threshold calibrated to capture today’s most capable models can gradually capture fewer and fewer of tomorrow’s most capable models, even as those newer models pose comparable or greater risk.

This creates what researchers describe as a genuine regulatory dilemma. If thresholds are lowered over time to keep pace with algorithmic efficiency gains, then a growing number of smaller developers, including startups and academic labs that lack the compliance resources of large frontier labs, get pulled into reporting and licensing obligations originally designed for a handful of well-resourced companies. If thresholds are left unchanged, they gradually stop capturing the most capable and most risky systems at all, since those systems can increasingly be trained below the threshold. Neither option is clearly superior, and the tension between them has no fully settled resolution in current policy design.

The Institute for Law and AI’s recommended framing treats compute thresholds as an initial screening filter rather than a complete risk assessment, useful for identifying which general-purpose AI models warrant closer regulatory scrutiny, while acknowledging that other signals, such as demonstrated capability benchmarks or estimated effective compute after accounting for algorithmic efficiency, are conceptually better proxies for actual risk even though they are considerably harder to measure consistently across developers.

What This Means for AI Developers and Compliance Teams

For organizations training large models or providing cloud infrastructure to those that do, three practical implications follow from the current state of compute governance. First, reporting thresholds should be tracked continuously rather than treated as a one-time compliance check, since both the US executive order threshold and international equivalents like South Korea’s AI Basic Act are subject to periodic revision. Second, export control exposure is no longer limited to hardware; the reported Anthropic episode suggests that deployed models themselves may increasingly fall within the scope of export authority, a much harder category to plan around than a chip specification sheet. Third, algorithmic efficiency gains that reduce training compute requirements, generally a positive development for cost and accessibility, can inadvertently move a given project across or beneath regulatory thresholds, so compliance teams need to reassess reporting obligations whenever a significant efficiency improvement is adopted, not just when compute budgets increase.

More broadly, compute governance is likely to keep expanding in scope rather than settling into a fixed set of rules, given the pace of change seen already in 2026 alone, from tightened chip thresholds in January, to a withdrawn global licensing draft in March, to a reported model-level restriction in June. Any organization operating at the frontier of AI capability should expect the regulatory perimeter to keep moving rather than assume today’s rules are the final version.

• Reporting versus licensingThe US executive order requires disclosure of large training runs rather than pre-approval, a lighter-touch design than the licensing regime used for chip exports.
• Cloud providers as reporting intermediariesUS cloud providers must report large training runs conducted by non-US customers, extending visibility beyond domestic developers alone.
• Presumption of denialBy late 2025, advanced computing export license applications to China and Macau were governed by a presumption that applications would be denied absent a specific exception.
• Model-level export control is contestedThe reported Anthropic episode extended export authority to a deployed model rather than hardware, a move legal commentators are still debating on statutory grounds.
• International convergence without coordinationSouth Korea’s AI Basic Act and the European Union’s systemic-risk provisions use compute thresholds in a similar order of magnitude to US provisions, despite no formal joint standard-setting process.

Glossary

Compute governance

The use of access to advanced computing hardware, such as chips and data center capacity, as a policy lever for overseeing frontier artificial intelligence development.

Total Processing Performance (TPP)

A technical metric used in export control rules to classify chip performance and determine whether a given chip is restricted for export to certain destinations.

Training-run reporting threshold

A defined level of computation, typically measured in floating-point operations, above which a training run must be disclosed to a government agency.

Presumption of denial

An export licensing policy under which applications are expected to be rejected by default unless a specific exception applies, rather than evaluated case by case with a neutral starting point.

Algorithmic efficiency

Improvements in model architecture or training methods that reduce the amount of compute needed to reach a given level of capability, gradually eroding the reliability of static compute thresholds.

Effective compute

An estimate of a model’s capability-adjusted compute usage that attempts to account for algorithmic efficiency gains, proposed as a more accurate but harder-to-measure alternative to raw training compute.

FAQs

What is compute governance?

Compute governance is the practice of using control over advanced computing hardware, such as chips and data center capacity, as a policy lever for overseeing frontier AI development. It is considered more trackable than regulating algorithms or model weights directly, since chips are manufactured by a small number of companies and physically exist in identifiable locations.

What does the US executive order require for AI training runs?

It requires AI developers to report any training run exceeding a defined compute threshold, measured in floating-point operations, and requires US cloud computing providers to report large training runs conducted by non-US customers once those runs cross the same threshold.

What changed in chip export controls in January 2026?

A final rule published January 15, 2026 tightened restrictions so that chips exceeding a Total Processing Performance of 21,000, or bandwidth of 6,500 gigabytes per second, remain fully restricted for export to certain destinations, alongside a new 25 percent tariff on qualifying advanced computing semiconductors.

What happened with the reported global AI chip licensing rule in March 2026?

The Department of Commerce circulated a draft rule in early March 2026 that would have required licenses for nearly all AI chip exports worldwide, a major expansion beyond destination-specific restrictions. That draft rule was withdrawn on March 13, 2026 after industry and allied-government pushback.

Did the US government really force Anthropic to disable its models?

As reported by ComplianceHub.Wiki and corroborated by several other outlets, a June 12, 2026 letter under Commerce Secretary Howard Lutnick reportedly led Anthropic to disable its Fable 5 and Mythos 5 models worldwide for foreign nationals. The restriction was reportedly lifted roughly three weeks later after Anthropic agreed to additional safeguards, and the underlying legal authority remains debated.

Why do critics say compute thresholds are an imperfect proxy for AI risk?

Algorithmic efficiency improvements steadily reduce the amount of compute needed to reach a given capability level, so a fixed threshold calibrated to today’s most capable models gradually captures fewer of tomorrow’s most capable ones, even as those models pose comparable or greater risk, according to researchers at the Institute for Law and AI.

How does South Korea’s AI Basic Act use compute thresholds?

South Korea’s AI Basic Act, which entered into force on January 22, 2026, establishes risk management obligations for high-performance AI systems trained using 10 to the 26th power floating-point operations or more, with phased enforcement beginning in 2027, a threshold in a similar order of magnitude to comparable US and European provisions.

Is compute governance the same everywhere in the world?

No. The United States, European Union, and South Korea each define their own compute-related thresholds and reporting obligations independently, without a single unified international treaty. Their thresholds sometimes converge in order of magnitude, but enforcement mechanisms and scope differ considerably between jurisdictions.

For related coverage of how governments are restricting access to advanced hardware, see our explainer on chip export controls and our analysis of sovereign AI compute strategies. Readers may also want our comparison of superintelligence governance proposals, our roundup of AI risk management frameworks compared, and our overview of global AI ethics standards. For a look at how similar tracking concerns apply to brain data rather than chips, see our piece on neurorights and legal protection for brain data.

• From Chips to Models: How AI Became a Controlled Commodity, and What Compliance Teams Should Expect Next, ComplianceHub.Wiki
• When Washington Switched Off an AI Model: Fable 5, Mythos 5, and the Export-Control Precedent That Should Worry Every Compliance Team, ComplianceHub.Wiki
• US lifts export controls on Anthropic’s frontier cybersecurity AI models, The Record from Recorded Future News
• White House lifts export control on Anthropic that froze its most advanced models, CNN Business
• Legal Considerations Related to the Anthropic Export Controls Directive, Just Security
• The Role of Compute Thresholds for AI Governance, Institute for Law and AI
• AI Chip Export Controls in 2026: What Changed and What It Means, ECCN Finder