Compute Governance

What is Compute Governance, and why do policymakers choose to regulate compute rather than AI models directly?

Compute governance refers to a set of policy tools built on a core logic: rather than trying to directly review or restrict an AI model's capabilities, algorithms, or training methods (which are fundamentally information — hard to physically track, and easy to circumvent controls on via the internet), it's more effective to regulate the physical resources required to train frontier AI models — high-end AI accelerator chips (like Nvidia H100- or H200-class chips).

The logic behind this policy choice is straightforward: chips are physical objects, requiring fabrication plants with an extremely high technical barrier to entry, with global production capacity highly concentrated among a small number of companies, making shipping and transaction records comparatively easy to track. This means "who bought how many chips, and where were they shipped" is information regulators can grasp far more easily than "who trained what model." And since Compute Scaling itself is the core engine driving frontier AI capability gains, controlling where compute flows is, to some degree, equivalent to controlling who has the capacity to train the next generation of the most powerful models.

What problem does Compute Governance aim to solve, and how did this policy tool evolve?

Compute governance policy goals typically involve two intertwined but logically distinct considerations: national security (preventing specific countries from acquiring frontier AI capability that could be used for military or surveillance purposes) and AI Safety (preventing capability development from outpacing Alignment research, reducing loss-of-control risk). In practical policy, these two considerations often get discussed together, but the underlying policy logic and stakeholders don't fully overlap.

Taking the US as an example, compute control policy evolved rapidly starting in 2022: it initially used a "Total Processing Performance" (TPP) threshold to define which chips would be controlled, with the threshold and calculation methods revised multiple times since. Between late 2025 and early 2026, policy underwent a major shift — the US government at one point rescinded the "AI Diffusion Framework," which had divided the global market into three access tiers, shifting instead to "case-by-case review" rather than "presumption of denial" for exports of specific chips, paired with a 25% tariff on those same chips. This rapid policy reversal itself reflects the ongoing political tug-of-war compute governance faces between "national security control" and "industry competitiveness."

What specific mechanisms does Compute Governance operate through, and what practical challenges does it currently face?

Compute governance's enforcement mechanisms broadly include several layers: export controls (restricting which countries or entities specific chip specs can be sold to), import tariffs (such as the 25% tariff on H200-class chips), and the newer concept of "location verification" — building technical mechanisms into chips that confirm their actual deployment location, making it easier for regulators to verify whether chip flow complies with export license conditions. New US rules that took effect in January 2026 shifted export review for certain chips from "presumption of denial" to "case-by-case review," but attached multiple conditions including end-user identity verification, security measures, and data center operator requirements.

In practice, the biggest challenge lies at the enforcement level: chip smuggling, routing through third countries to evade controls, and friction between "sovereign AI" programs (nations or regions wanting to build independent compute infrastructure under their own jurisdiction) and existing export control frameworks all undercut compute governance's actual effectiveness. Research institute reports also note that export controls alone aren't sufficient to fully prevent specific countries from developing advanced AI capability — which is why some policy researchers argue compute governance needs to be paired with other tools like compute thresholds, forming a multi-layered governance architecture rather than betting everything on export controls alone.

How does Compute Governance help readers make sense of AI industry news?

Whenever you see headlines like "country announces new chip export rules" or "a chip maker's stock swings on policy news," understanding the logic of compute governance helps readers grasp the actual policy goals and stakeholders behind such news — the parties directly affected by these policies go far beyond chip manufacturers alone, extending to cloud service providers, data center operators, server integrators, and virtually any enterprise deploying AI training or inference workloads across borders, all of whom may need to adjust their compliance strategy in response to new rules.

Compute governance also connects directly to the technical phenomenon of Compute Scaling: if a country or company is cut off from access to top-tier chips, this doesn't just affect short-term model performance — over the long run, it more likely affects whether that country or company can keep pace with the capability gains delivered by the compute scaling curve. This is also why compute — along with the chips, power, and capital behind it — is increasingly seen as a strategic resource comparable to oil. Understanding compute governance's policy logic is effectively understanding one of the most central front lines in today's AI geopolitical competition.