AI Safety Power Map

Jurisdiction over most of the frontier developer and accelerator base, and over the export of the accelerators everyone else depends on. Expand it and the authority turns out to sit in specific agencies rather than in "the US".

Power levers: Export controls on accelerators and semiconductor equipment; Procurement as a very large customer; Permitting and grid authority over frontier-scale sites.

No curated dependencies yet — relationship coverage in this snapshot is deliberately incomplete rather than inferred.

Public sources: whitehouse.gov

Administers the export controls that decide which accelerators may be sold where. One of the clearest places where a single administrative body conditions frontier compute access.

Power levers: Entity list and licence decisions; Performance thresholds on exportable accelerators.

Public sources: bis.gov

Sits on the grid side of the compute stack, where transmission and interconnection decide how quickly a frontier-scale site can actually draw power.

Power levers: Transmission and interconnection policy; Grid reliability authorities.

No curated dependencies yet — relationship coverage in this snapshot is deliberately incomplete rather than inferred.

Public sources: energy.gov/oe

Sets the terms for one of the two frontier developer ecosystems, and is the counterparty that most shapes how the other one is governed.

Power levers: Approval and licensing of model release; Industrial policy and domestic accelerator programmes.

No curated dependencies yet — relationship coverage in this snapshot is deliberately incomplete rather than inferred.

Public sources: cac.gov.cn

The most developed binding regime aimed specifically at general-purpose models, and therefore a live constraint on how frontier systems are deployed into a large market.

Power levers: General-purpose model obligations; Codes of practice and enforcement.

Public sources: EU AI Office

Evaluation capacity rather than authority. Its power is epistemic: what it measures becomes what the debate is about.

Power levers: Pre-deployment evaluations; Public technical reporting.

No curated dependencies yet — relationship coverage in this snapshot is deliberately incomplete rather than inferred.

Public sources: aisi.gov.uk

Appointment and removal authority over OpenAI leadership. Expanding OpenAI is what shows that some decisions are the board’s and not the executive’s.

Power levers: Appoints and removes leadership; Structural and mission constraints.

Public sources: openai.com

Holds board-appointment rights designed to survive commercial pressure. A governance instrument, not an advisory body.

Power levers: Elects a portion of the board; Mission-protective structure.

Public sources: The Long-Term Benefit Trust

Executive authority over OpenAI as of the snapshot date, and one of the loudest voices setting the public agenda about what is coming.

Power levers: Deployment and release timing; Capital and compute deals.

Public sources: openai.com

Executive authority over Anthropic as of the snapshot date, and a primary author of the public case for treating scaling as a safety problem.

Power levers: Deployment and release timing; Public risk framing.

Public sources: anthropic.com

Executive authority over Google DeepMind as of the snapshot date, inside a parent company with its own compute and distribution.

Power levers: Research direction; Release posture within Google.

Public sources: deepmind.google

Controlling authority over Meta as of the snapshot date, including the decision to release frontier-adjacent weights openly.

Power levers: Open-weight release decisions; Capital allocation at scale.

Public sources: ai.meta.com

Controlling authority as of the snapshot date, with unusual ability to move capital and physical build-out quickly.

Power levers: Capital and site build-out speed; Public agenda setting.

Public sources: x.ai

Executive authority over the company that allocates the scarcest input in the stack. Who gets accelerators, and when, is partly a commercial decision.

Power levers: Allocation and roadmap decisions; Pricing and supply commitments.

Public sources: nvidia.com

A frontier developer whose release decisions move the whole field’s expectations. Expand it and the authority splits between the executive and the foundation board.

Power levers: Frontier training runs; Release and access policy.

Public sources: openai.com

A frontier developer that also supplies much of the public safety framing. Expand it and a trust holds part of the board authority.

Power levers: Frontier training runs; Responsible-scaling commitments.

Public sources: anthropic.com

A frontier developer inside a company that owns its own compute and one of the largest distribution surfaces in existence.

Power levers: Frontier training runs; Own accelerators and cloud; Default distribution.

Public sources: deepmind.google

Sets the open-weight frontier, which changes what every other actor’s deployment decisions are competing against.

Power levers: Open-weight releases; Very large owned build-out.

Public sources: ai.meta.com

A frontier developer distinguished less by model lead than by how fast it can put compute and power on the ground.

Power levers: Rapid site build-out; Capital concentration.

Public sources: x.ai

Both a developer and a cloud, which makes it one of the few actors outside the US stack that controls several layers at once.

Power levers: Open-weight model releases; Regional cloud capacity.

No curated dependencies yet — relationship coverage in this snapshot is deliberately incomplete rather than inferred.

Public sources: alibabacloud.com

Very large distribution and a well-resourced research effort, developing under a different regulatory regime.

Power levers: Consumer-scale deployment; Talent concentration.

No curated dependencies yet — relationship coverage in this snapshot is deliberately incomplete rather than inferred.

Public sources: bytedance.com

Demonstrated that frontier-adjacent capability can arrive from outside the assumed set of actors, which is itself a fact about how controllable the frontier is.

Power levers: Efficient training results; Open-weight releases.

No curated dependencies yet — relationship coverage in this snapshot is deliberately incomplete rather than inferred.

Public sources: deepseek.com

Supplies the compute a frontier developer trains on, which makes a commercial relationship into a dependency on someone else’s capacity plan.

Power levers: Capacity allocation and scheduling; Enterprise distribution.

Public sources: azure.microsoft.com

Supplies frontier training and serving capacity to more than one developer, which makes it a shared dependency rather than a private arrangement.

Power levers: Capacity allocation; Own accelerator programme.

Public sources: aws.amazon.com

The single most concentrated point in the stack: nearly every frontier training run depends on its accelerators, and allocation is discretionary.

Power levers: Who receives accelerators, and when; Roadmap and interconnect.

Public sources: nvidia.com

Manufactures the accelerators. A chokepoint that is geographically concentrated as well as technically concentrated.

Power levers: Leading-node capacity; Advanced packaging.

Public sources: tsmc.com

Sole supplier of the lithography needed for leading-node manufacturing. Its power is almost entirely structural rather than agenda-setting.

Power levers: EUV tool supply; Servicing and upgrade dependence.

Public sources: asml.com

High-bandwidth memory is a real constraint on accelerator output, and the supplier base for it is small.

Power levers: HBM capacity and qualification.

Public sources: skhynix.com

Site-specific generation and storage for frontier-scale compute. Little global agenda power, considerable leverage over one build.

Power levers: Generation and storage at a specific site.

Public sources: sbenergy.com

The utility on the other side of a specific frontier-scale interconnection. Contextual power that is decisive locally and invisible globally.

Power levers: Interconnection timing; Generation commitments.

Public sources: entergy-louisiana.com

A federal power producer whose supply decisions condition specific frontier-compute sites.

Power levers: Bulk power supply; Transmission commitments.

Public sources: tva.com

A municipal utility with real veto and delay leverage over one particular site — the clearest case that power here is contextual rather than global.

Power levers: Local interconnection and permitting; Delay.

Public sources: mlgw.com

Appointment and removal authority over leadership.

Evidence: OpenAI — Our structure — OpenAI's own account of its structure, in which the Foundation governs the Group.

Holds rights to elect part of the board.

Evidence: Anthropic — The Long-Term Benefit Trust — Anthropic's announcement of the trust and the board seats it elects.

Executive authority as of the snapshot date.

Evidence: OpenAI — Our structure — OpenAI's own account of where executive authority sits and what the Foundation retains over it.

Executive authority as of the snapshot date.

Evidence: Anthropic — Company — Anthropic's own leadership page.

Executive authority as of the snapshot date.

Evidence: Google DeepMind — About — Google DeepMind's own account of its leadership and remit.

Controlling authority as of the snapshot date, held through a dual-class structure rather than through the chief-executive role alone.

Evidence: Meta Platforms — Annual report on Form 10-K (filed 2026-01-29) — Meta's own filing: Zuckerberg "is able to exercise voting rights with respect to a majority of the voting power of our outstanding capital stock and therefore has the ability to control the outcome of all matters submitted to our stockholders for approval". Establishes control, not merely the job title.

Founder and chief executive as of the snapshot date. The company is private, so the ownership and voting structure behind that authority is not publicly inspectable and is not claimed here.

Evidence: SpaceXAI — Company — The organisation's own company page, naming its founder and chief executive. It establishes the executive role; unlike a listed company there is no filing establishing voting control, which is why this edge is stated at the narrower level.

Executive authority as of the snapshot date.

Evidence: NVIDIA — Jensen Huang, Founder, President and CEO — NVIDIA's own executive biography.

The primary cloud partner for frontier training and serving, which makes a commercial arrangement into a dependency on someone else's capacity plan.

Evidence: Microsoft — The next phase of the Microsoft/OpenAI partnership — Microsoft's own statement that it remains OpenAI's primary cloud partner and that OpenAI products ship first on Azure.

Training and serving capacity for frontier runs.

Evidence: AWS — Trainium customers — AWS's own customer page, carrying Anthropic on training and serving Claude on Trainium.

A second cloud counterparty on a multi-year infrastructure agreement — and the reason this supplier sits on more than one frontier path at once.

Evidence: OpenAI — Amazon partnership — OpenAI's own announcement of a multi-year Amazon infrastructure agreement, including Trainium capacity.

Lithography without which leading-node manufacturing does not happen.

Evidence: ASML — EUV technology training centre in Taiwan — ASML's own release, naming the counterparty rather than the product line: "In 2010, we shipped the first prototype EUV lithography system to TSMC… In 2017, we shipped the first production-ready system, the TWINSCAN NXE:3400, to TSMC." Dated 2020, so it establishes the supply relationship rather than current volumes.

Manufactures the accelerators, including advanced packaging.

Evidence: NVIDIA — Blackwell platform arrives — NVIDIA's own launch release, stating Blackwell GPUs are manufactured on a custom TSMC 4NP process.

High-bandwidth memory constrains accelerator output.

Evidence: NVIDIA — SK hynix AI factory — NVIDIA and SK hynix's own multiyear memory partnership announcement.

Accelerator supply for very large owned build-out.

Evidence: NVIDIA — Meta builds AI infrastructure with NVIDIA — NVIDIA's own release on the Meta partnership, covering deployment of millions of Blackwell and Rubin GPUs.

Accelerator supply for rapid site build-out.

Evidence: NVIDIA — Spectrum-X networking for Colossus — NVIDIA's own release describing Colossus as a 100,000-GPU system used to train Grok.

Selected to build and operate generation for a specific 1.2 GW frontier-scale site.

Evidence: OpenAI — Stargate / SB Energy partnership — OpenAI's own announcement selecting SB Energy to build and operate its 1.2 GW Milam County site.

The utility counterparty for a specific named frontier-scale site.

Evidence: Entergy — Entergy Louisiana to power Meta's Richland Parish data centre — Entergy's own news release naming Meta and the specific site it will power.

Bulk power supply to a specific site, under a board-approved firm power arrangement.

Evidence: TVA — Board approved resolutions — TVA's own record of the board resolution approving a firm power arrangement above 100 MW for the xAI site.

Local interconnection and permitting for the same site — little global power, decisive locally.

Evidence: MLGW — xAI — The utility's own published page on service and grid capacity for the xAI site.

Export controls decide which accelerators may be sold where.

Evidence: BIS — Revision to License Review Policy for Advanced Computing Commodities (2026) — The rule itself, which sets licence policy by naming NVIDIA H200-class accelerators and equivalents., Commerce Control List — 15 CFR Part 774 — The standing list the licence requirement is drawn from.

General-purpose model obligations condition deployment into the EU.

Evidence: Regulation (EU) 2024/1689 — the AI Act — The binding text. It establishes obligations on providers of general-purpose models as a class rather than naming this provider, so the edge is stated at that level.

General-purpose model obligations condition deployment into the EU.

Evidence: Regulation (EU) 2024/1689 — the AI Act — The binding text. It establishes obligations on providers of general-purpose models as a class rather than naming this provider, so the edge is stated at that level.