Skip to content

NVIDIA

Note on inclusion: NVIDIA is not an AI model company, but it is the core supplier for AI infrastructure — every frontier AI lab depends on its GPUs. NVIDIA’s role in AI governance runs through supply-chain controls, not through a model-level safety framework. Its policy behaviour (response to export controls, government lobbying, open-source strategy for its in-house models) has far-reaching consequences for global AI governance.

  • Founded: 1993
  • Headquarters: Santa Clara, California, USA
  • CEO: Jensen Huang 黄仁勋
  • Market capitalisation: over $4 trillion at end-2025 (top global tier)
  • Core business:
    • Data Center GPUs (AI training / inference): H100 / H200 / B100 / B200 / GB200 (Blackwell)
    • CUDA ecosystem: the de facto standard for AI software development
    • Nemotron series: NVIDIA’s in-house open-source LLMs (smaller scale, ecosystem-focused)
    • Omniverse / Isaac: industrial simulation and robotics AI
    • DGX / Project Digits: integrated AI systems
  • Business model: hardware sales + cloud services (DGX Cloud) + software subscriptions + verticals (automotive, robotics, healthcare)

Strategic positioning: “selling shovels” in the AI revolution

Section titled “Strategic positioning: “selling shovels” in the AI revolution”

Institutional consequences of GPU near-monopoly

Section titled “Institutional consequences of GPU near-monopoly”

NVIDIA holds over 90% share of the data-centre AI GPU market (2025 data). This near-monopoly position produces three institutional consequences:

  1. Export-control focal point: U.S. export controls on China are almost synonymous with controls on NVIDIA exports
  2. CUDA lock-in: the AI software ecosystem depends deeply on CUDA; open-source alternatives (AMD ROCm, Intel oneAPI) have progressed slowly
  3. Geopolitical pressure: any U.S.–China AI policy shift directly affects NVIDIA revenues
  • Anthropic / OpenAI / Google DeepMind / xAI / Meta / Mistral all run primarily on NVIDIA GPUs
  • The Stargate Project (OpenAI + Microsoft + Oracle): multi-hundred-billion-dollar NVIDIA purchase commitments
  • Colossus (xAI, Memphis): 100k+ H100s in a single cluster
  • Meta Llama training: 2024–2025 single training runs used tens of thousands of H100s
  • Chinese companies are export-restricted but remain important NVIDIA customers (via the downgraded H20, now the tariffed H200 variant)

Deep dive: the evolution of export controls and NVIDIA’s two-sided game

Section titled “Deep dive: the evolution of export controls and NVIDIA’s two-sided game”
DateEvent
2022-10BIS first restricts A100 / H100 exports to China
2023H800 / A800 (downgraded variants) launched to work around the controls
2023-10BIS tightens further; H800 / A800 also controlled
2024NVIDIA launches H20 (further downgraded), China-specific
2024-05H20 controls under discussion; NVIDIA holds large H20 inventory
2025-01Trump EO 14179 triggers a reassessment of export-control policy
2025-10B100 / B200 / GB200 (Blackwell) placed under China export ban
2025-12-08Trump policy reversal: H200 exports to China permitted, with 25% tariff
2026-01-12House passes the Remote Access Security Act (369–22), extending export controls to cloud services
2026-01China pauses new H200 orders from Chinese tech companies pending policy review
  • Toward the U.S. government: Jensen Huang is close to the Trump administration (multiple White House visits in 2025) and supports the AI Action Plan’s “export American AI” vision
  • Toward the Chinese market: continues launching compliant downgraded variants (H800 → H20 → tariffed H200), with Chinese 2026 orders exceeding 2 million units against NVIDIA inventory of 700k
  • Public stance: Huang has repeatedly argued that export controls “are accelerating China’s indigenisation” — simultaneously voicing industry concern and preparing the ground for later loosening

Significance of the Remote Access Security Act (2026-01)

Section titled “Significance of the Remote Access Security Act (2026-01)”
  • First extension of export controls to cloud services: Chinese companies had been accessing 2,300+ Blackwell GPUs by renting offshore (Lambda Labs, AWS international, etc.)
  • House passed 369–22 on 2026-01-12 (rare bipartisan unanimity)
  • If the Senate passes and the president signs → cloud services become subject to export controls, implicating NVIDIA DGX Cloud and other cloud providers
  • Deep consequences for AI governance: compute as a governance lever is further strengthened

NVIDIA has no RSP / Preparedness / FSF-equivalent safety framework because its products are hardware + general-purpose software, not end-use AI systems.

But NVIDIA has distinctive AI-governance tooling:

  • CUDA License: terms can include export-control compliance requirements
  • NIM / AI Enterprise: enterprise AI deployment tools including content filtering and model monitoring
  • Nemotron open-source models: NVIDIA’s in-house open-source LLMs (Llama derivatives + small-scale in-house) — a “showcase tool” rather than a primary product
  • 2023 White House Voluntary Commitments: not signed (NVIDIA is not an AI model provider)
  • Frontier Model Forum: not a member
  • GPAI Code of Practice: not applicable (NVIDIA does not provide GPAI-model-class primary products)
  • California SB 53: no public position
  • EO 14179 / 14365: publicly supportive (NVIDIA is a key player in the AI Action Plan)

Project Digits and the “consumer-grade AI workstation”

Section titled “Project Digits and the “consumer-grade AI workstation””

In 2025 NVIDIA launched Project Digits (a personal AI computer at a $3,000 price point) aimed at individual developers:

  • Can run 200B-parameter models locally
  • Bypasses cloud-service export controls
  • Proliferation of compute at the consumer hardware layer may undermine the efficacy of binding compute thresholds (EU 10^25 FLOP / California 10^26 FLOP)
TypeDocumentLinkSubpage
Usage policyNVIDIA End User License Agreement (incl. export-control clauses)nvidia.com/en-us/about-nvidia/legal-info
Model / toolsNemotron series + NeMo frameworkdeveloper.nvidia.com/nemo
Export-control complianceExport Compliance pagenvidia.com/en-us/about-nvidia/export-compliance
  • United States:
    • BIS export-control focal point: NVIDIA is the principal object of U.S. AI export enforcement
    • CFIUS: not applicable (NVIDIA is a U.S. company)
    • EO 14179 / AI Action Plan: supportive and a beneficiary
    • DoD contracts: yes, though not the main business line
  • China:
    • Downgraded-GPU supply (H20 → tariffed H200) continuously adjusted
    • Anchor-customer relationships with Chinese cloud providers (Alibaba Cloud, Tencent Cloud, ByteDance Volcano Engine)
    • Chinese 2026 procurement exceeds 2 million units (H200 leading)
  • European Union:
    • Delivered indirectly via AWS / Azure / GCP / Mistral Compute and other channels
    • Not directly subject to the AI Act (GPUs are not AI systems)
    • General regulation still applies (e.g. CE marking for electronic products)
  • Other export-restricted regions: full embargoes in Russia, Iran, North Korea, Cuba, etc.

NVIDIA is not an AI model company, but its weight in AI governance is comparable to major AI labs:

  1. Compute is the underlying lever of AI governance: EO 14110’s 10^26 FLOP threshold, California SB 53, and the EU AI Act systemic-risk GPAI provisions are all compute-indexed — compute allocation, pricing, and access controls determine how effective AI-governance policy can be
  2. Primary object of export-control policy: BIS AI-chip controls targeting China and the Remote Access Security Act treat NVIDIA as the principal regulated entity
  3. Key intermediary in geopolitical dynamics: the material foundation of U.S.–China AI competition
  4. Ecosystem lock-in: CUDA’s de facto standard status means any AI-governance measure must reckon with CUDA compatibility

Conclusion: without understanding NVIDIA’s role in the AI supply chain, one cannot understand how contemporary AI governance actually operates.

  • Single-vendor dominance in AI infrastructure; the foundation of the global AI industry
  • A beneficiary of Trump deregulation and the acceleration narrative
  • Navigates both sides of export controls: publicly supportive of policy (Huang’s lobbying) while continuously launching compliant downgraded products
  • Does not take the safety-framework path: as a hardware supplier, not a direct participant in model-layer self-regulation