AI development is, at present, extremely hardware-intensive. Training the models that sit at the frontier of capability requires clusters of thousands of specialized chips running for weeks or months, consuming power in the range of tens of megawatts. The chips are designed by a small number of companies, fabricated using equipment made by an even smaller number, and sold through supply chains that pass through specific national jurisdictions. This hardware concentration is one of the few features of frontier AI development that makes near-term regulation practical.
Compute governance is the approach to AI regulation that tries to exploit this concentration as a lever. Rather than directly regulating what AI systems can or cannot be built (a question that requires technical judgments most regulators cannot make), compute governance regulates the hardware required to build them. Control the compute, and you have some degree of control over who can develop frontier AI, at what scale, and with what degree of reporting and transparency.
The hardware chokepoints
These chokepoints are not permanent. China is actively investing in domestic semiconductor production with the goal of reducing dependence on the Western supply chain. Current domestic chips remain significantly behind the frontier — a gap that imposes real costs on Chinese AI development but does not make advanced AI development impossible with sufficient time and resources. The governance window created by existing hardware concentration is real but finite.
What compute governance measures already exist
The US government has implemented export controls on advanced AI chips as the most significant compute governance measure to date. Controls announced in October 2022 restricted the sale of Nvidia's A100 and H100 chips to China, Russia, and other specified countries. Subsequent rounds of controls in 2023 and 2024 extended restrictions to chips designed specifically to fall below the original thresholds, closing the most obvious circumvention route. The controls have imposed real costs on Chinese AI labs: documented chip smuggling efforts, black market pricing, and reduced capability of Chinese training clusters all indicate that the restrictions have had effect.
The Biden administration's October 2023 Executive Order on AI included a requirement that developers report to the government for training runs above approximately 10^26 floating point operations — roughly the compute required to train a GPT-4-scale model. This reporting requirement was a first step toward compute-based oversight: giving the government visibility into who is running large training runs before systems are deployed, rather than after.
Several researchers have proposed that AI chips could be manufactured with hardware mechanisms allowing their location and use to be tracked by the chip manufacturer or a designated international body. Because all chips already contain unique identifiers, extending this to active tracking is technically feasible. A chip-tracking regime could allow an international oversight body to verify that compute is not being used in unauthorized training runs — creating a verification mechanism analogous to nuclear materials accounting, but for AI compute.
What compute governance cannot do
Three significant limitations constrain the effectiveness of compute governance as a safety tool. First, circumvention: chips can be purchased through third countries before restrictions take effect, stockpiled, or substituted with slightly less capable chips that fall below regulatory thresholds. The US controls have been circumvented in some documented cases, and the smuggling and third-country routing patterns have been extensive enough to prompt multiple rounds of tightening.
Second, algorithmic efficiency: as AI training becomes more efficient, the compute required to train a given level of capability decreases. A compute threshold set at the right level today may be far too permissive in three years as efficiency improves. Governance frameworks built around fixed compute thresholds need mechanisms to adjust those thresholds as the relationship between compute and capability evolves.
Third, and most importantly, compute governance does not address the alignment problem. A system trained using properly licensed hardware, reported to the appropriate authorities, and within a sanctioned compute threshold can still be deeply misaligned. Compute governance can control the pace and scale of frontier AI development and provide visibility into who is training what. It cannot ensure that what they are training is aligned.
This is why compute governance is best understood as one component of a broader framework alongside alignment verification requirements and international treaty obligations — not as a standalone solution. The Foundation's three policy proposals treat compute governance as the near-term lever that buys time and visibility while the harder technical and diplomatic work of alignment verification and international treaty negotiation advances.
Common questions.
An approach to AI regulation that uses control over the specialized hardware required to train large AI models as a regulatory lever. Because frontier AI training requires specific chips produced by a handful of companies in specific national jurisdictions, controlling the production, sale, and use of this hardware provides a degree of control over who can develop frontier AI and at what scale — without requiring regulators to make technical judgments about AI systems directly.
US export controls restricting the sale of advanced AI chips (initially the Nvidia A100 and H100, later extended) to China and other specified countries are the most significant existing measure. The Biden administration's October 2023 AI Executive Order also introduced reporting requirements for training runs above approximately 10^26 FLOP. Various proposals for chip-level tracking mechanisms exist but have not yet been implemented.
Partially, and decreasingly over time. Current hardware concentration creates real leverage: the supply chain passes through specific chokepoints (TSMC, ASML, Nvidia) that are subject to US and allied jurisdiction. China is actively working to reduce this dependence through domestic semiconductor investment, and the governance window created by existing concentration is finite. Export controls have imposed real costs on Chinese AI development but have not halted it. International compute governance, including a chip-tracking regime, would be more robust but requires multilateral agreement.
No. Compute governance can slow frontier AI development and provide visibility into large training runs. It cannot ensure that systems trained within the governance framework are aligned, corrigible, or safe to deploy. It is best understood as a near-term lever that buys time for alignment research and international treaty negotiation to advance — not as a solution to the underlying safety problem.