Compute Governance

Compute governance refers to policy and control over the physical inputs to AI capability — semiconductors, data centers, and the electricity to run them. Computing power is the physical foundation of AI capability, and control over it determines who can train and deploy frontier systems. It is one of the few AI inputs that is both essential and subject to policy leverage.

Components

The compute stack has three principal inputs. Chips are advanced semiconductors (GPUs, TPUs) required for large-scale training; their supply chain is concentrated in a few countries, with US design, TSMC fabrication, and ASML lithography. Data centers are the facilities for training and inference, where location decisions involve energy costs, regulatory environment, and security. Energy, the electricity supply, is a potential bottleneck: if AI firms lack sufficient power, it could limit overall progress.

Snapshot

Power and demand figures (newest on top):

Supply-chain chokepoints

Compute governance operates through chokepoints, not through the industry as a whole. The binding chokepoints on AI accelerator supply through 2026–2027 are not wafer fabrication but advanced packaging, memory, lithography, and leading-edge fabrication:

• TSMC CoWoS advanced packaging — oversubscribed through 2026 (~110k wafers/month target capacity).
• HBM (SK Hynix, Samsung, Micron) — all three vendors sold out through 2026 under long-term contracts; shortages persist into late 2027.
• ASML EUV lithography — single-supplier monopoly; export-controlled to China.
• Leading-edge fabrication — TSMC dominant; SMIC roughly one generation behind at poor yields.

Policy leverage attaches to whichever chokepoint is currently binding. The December 2024 BIS HBM rule tracked the migration of the binding constraint from logic to memory. See Semiconductor Supply Chain for the full map and (Source: Raw Sources/SemiAnalysis - CoWoS and HBM Supply Chain.md) for the empirical basis. The Chinese alternative stack, and HBM as a shared bottleneck across it, is documented in (Source: Raw Sources/SemiAnalysis - Huawei Ascend Production Ramp.md).

Policy levers

The policy instruments available for compute governance include:

• Export Controls (AI): restricting China's access to advanced chips is the primary US compute governance tool.
• Data center siting and permitting.
• Energy infrastructure investment — grid interconnects, transmission, generation.
• Power as a chokepoint (see below).
• Federal procurement as a demand signal.
• Investment restrictions on adversary access to the US compute ecosystem.
• Disclosure regimes for large training runs (site, power draw, counterparty) — modeled on BIS chip-export reporting (RAND — AI's Power Requirements Under Exponential Growth (2025)).
• Compute thresholds as the trigger for regulatory coverage, paired with a spend threshold. ARI's August 2026 federal blueprint would cover only developers meeting both a 10^26 FLOP training-compute test and a $100 million aggregate annual training-spend test, the two conditions conjunctive. Its authors state the limitation of the compute leg directly — that algorithmic efficiency gains mean models trained with less compute may eventually match the performance of current frontier models — and build in a fixed-cycle regulatory review to adjust the threshold as the proxy degrades, with the stated aim of eventually replacing it with a capabilities-based definition developed with technical assistance from CAISI. The design treats the compute threshold as a temporary administrative convenience rather than a durable definition of the frontier.
• Compute allocation as the object of governance rather than compute access. The pacing proposals reframe the lever: IFP's August 2026 report proposes that if a threshold of automated AI R&D risk is exceeded, policy should incentivize reallocating compute and talent away from those activities and toward diffusion and safety research, and its disclosure recommendations anchor on the share of compute devoted to autonomous work. ARI's transparency programme makes the same measure the primary auditable metric in its quarterly filings. See Recursive Self-Improvement (RSI).

Power as a compute chokepoint

Work by Lennart Heim and co-authors argues that power, not chips, is becoming the binding constraint on frontier AI (RAND — AI's Power Requirements Under Exponential Growth (2025); Epoch AI — Can AI Scaling Continue Through 2030?). The argument holds that power has governance-relevant properties that chips partly lack. Power is slow to build, taking 5–15 years for new nuclear and 3–7 years for gas with interconnect, while accelerators ship in 12–18 months. It is geographically fixed: unlike chips, power cannot be rerouted via shell companies or third countries. It is visible, in that interconnect queues, substation upgrades, and power-purchase agreements leave public paper trails. And it is already politically contested, with utilities, ratepayer advocates, and local governments holding standing.

The quantitative basis for treating this as a near-term issue is captured in the Snapshot table above: the largest single training run is projected to move from above 100 MW today to 2–16 GW by 2030 (Epoch and RAND brackets); US AI data-center aggregate power is projected to grow from roughly 5–8 GW in 2024 to 40–80 GW in 2030, or 8–15% of US electricity in the central RAND scenario; and accommodating AI plus baseline growth would require US grid capacity additions at three times the historical rate.

On this account, chip export controls addressed the 2020–2024 chokepoint, while power siting, transmission, and disclosure are emerging as the 2025–2030 chokepoint. The America's AI Action Plan's NEPA streamlining push is described as a direct response to the power constraint.

Strategic role

Sullivan and Feldman identify compute as the first of six sources of US AI power (Geopolitics in the Age of Artificial Intelligence). It is foundational across nearly all eight worlds in the Eight Worlds Framework, though its relative importance varies. In worlds where catching up is hard, compute advantages compound and controlling the stack matters most; in worlds where catching up is easy, compute controls buy time but may not be decisive; and if China develops an alternative compute stack, chip controls become less effective. Nation-state compute independence is the commercial expression of this logic, tracked under Sovereign AI (product).

Several forecasts have been advanced about the near-term direction of US compute governance. RAND-aligned analysis based on a pattern of escalating enforcement need projects (issued May 2026) that BIS will receive a material enforcement-headcount expansion in the FY2027 appropriation, with a resolution criterion of a public appropriation showing a BIS enforcement-staffing increase greater than 50% over FY2025 by 2027-09-30. Based on BIS proposed rulemaking, one forecast (issued 2025) holds that the cloud-compute KYC rule will finalize and bind, with final Federal Register publication of KYC requirements on cloud rental of restricted compute expected by 2026-12-31. Drawing on Huawei Ascend 950PR and DeepSeek V4 optimization, a forecast (issued May 2026) projects that US export controls will be extended to cover Chinese-domestic compute, resolved by a BIS rule, EO, or DOC guidance explicitly addressing Chinese-domestic-compute deployment for AI training by 2027-12-31. Citing the Anthropic-Google $200B commitment and the Colossus 1 pattern, a forecast (issued May 2026) projects that a frontier-AI compute commitment above $10B will undergo a formal national-security review, resolved by public disclosure of a CFIUS, Team Telecom, or equivalent review of an AI compute deal by 2027-05. Finally, on a Procurement-Driven AI Governance reframe trajectory, a forecast (issued May 2026) projects that compute governance will become the load-bearing governance mode relative to pre-release vetting or capability-threshold approaches, resolved if the compute-governance regime (export controls plus KYC plus sovereign-AI subsidies) becomes the most-cited US AI policy gate in mainstream coverage by 2027-12-31.

See also

• Lennart Heim — compute-governance scholar (RAND, previously GovAI).
• Epoch AI — quantitative baselines for training-compute and power trends.
• Sovereign AI (product) — nation-state compute independence as the commercial expression of compute governance logic.
• Semiconductor Supply Chain — chokepoint map.

• related: Responsible Innovation at the Frontier (ARI, August 2026) — the conjunctive 10^26 FLOP / $100M coverage test and its stated proxy degradation
• related: How Should the US Prepare for Increasingly Automated AI R&D? (IFP, August 2026) — compute allocation, rather than access, as the governed quantity

Sources

• Geopolitics in the Age of Artificial Intelligence (Foreign Affairs, 2026-01-27)
• RAND — AI's Power Requirements Under Exponential Growth (2025) — Pilz, Mahmood, Heim (2025); policy-facing power forecast and chokepoint framing.
• Epoch AI — How Much Power Will Frontier AI Training Demand in 2030? — per-run power trajectory.
• Epoch AI — Can AI Scaling Continue Through 2030? — four-bottleneck physical-input analysis.
• (Source: Raw Sources/SemiAnalysis - CoWoS and HBM Supply Chain.md) — CoWoS and HBM as the binding 2026 chokepoints.
• (Source: Raw Sources/SemiAnalysis - Huawei Ascend Production Ramp.md) — Chinese alternative stack; HBM as shared bottleneck.
• Semiconductor Supply Chain — chokepoint map.