AI Chip Governance Supply Chain

AI Chip Governance Supply Chain refers to the policies, frameworks, hardware mechanisms, and institutional arrangements that govern the production, distribution, and use of semiconductors critical to advanced AI systems. The field blends supply chain security, export regulation, hardware engineering, and governance design to address risks ranging from smuggling and unauthorized proliferation to longer-term concerns about ungoverned AI scaling. It is a subcategory of Compute GovernanceConceptCompute GovernanceThis is a comprehensive overview of U.S. AI chip export controls policy, documenting the evolution from blanket restrictions to case-by-case licensing while highlighting significant enforcement cha...Quality: 58/100.

Quick Assessment

Key Links

Supply Chain Topology

The AI chip supply chain flows through a small number of chokepoints, from equipment manufacturing through fabrication to chip design and ultimately to AI labs and data centers. The following diagram illustrates the key dependencies.

flowchart LR
  subgraph EQUIPMENT["Equipment Suppliers"]
      ASML["ASML
(EUV Lithography)"]
      AMAT["Applied Materials
(Etch/Deposition)"]
  end
  subgraph MEMORY["Memory Producers"]
      SKH["SK Hynix
(HBM Memory)"]
  end
  subgraph FOUNDRIES["Foundries"]
      TSMC["TSMC
(90%+ advanced chips)"]
      SMIC["SMIC
(China; 5nm via DUV)"]
  end
  subgraph DESIGNERS["Chip Designers"]
      NVDA["Nvidia
(90-95% AI accelerators)"]
      AMD["AMD
(Instinct MI300)"]
  end
  subgraph CONSUMERS["AI Infrastructure"]
      LABS["AI Labs & Hyperscalers
(Google, Microsoft, Meta, etc.)"]
  end

  ASML -->|EUV machines| TSMC
  ASML -.->|Restricted| SMIC
  AMAT -->|Etch/deposition tools| TSMC
  AMAT -->|Etch/deposition tools| SMIC
  SKH -->|HBM chips| NVDA
  SKH -->|HBM chips| AMD
  TSMC -->|Fabrication| NVDA
  TSMC -->|Fabrication| AMD
  SMIC -->|Domestic fabrication| LABS
  NVDA -->|"GPUs (H100, B200)"| LABS
  AMD -->|"GPUs (MI300)"| LABS

  style ASML fill:#dbeafe,stroke:#2563eb
  style AMAT fill:#dbeafe,stroke:#2563eb
  style SKH fill:#fef3c7,stroke:#d97706
  style TSMC fill:#d1fae5,stroke:#059669
  style SMIC fill:#fce7f3,stroke:#db2777
  style NVDA fill:#ede9fe,stroke:#7c3aed
  style AMD fill:#ede9fe,stroke:#7c3aed
  style LABS fill:#f3f4f6,stroke:#6b7280

Key takeaway: The dashed line from ASML to SMIC represents the export control barrier. Without EUV, SMIC uses multi-patterning DUV techniques, which are slower and less efficient but have achieved 5nm-class results.

Key Players Compared

The following tables draw from the FactBase to compare key metrics across the major semiconductor supply chain companies.

Revenue

Headcount

Market Capitalization

Overview

The global supply chain for advanced AI chips is extraordinarily concentrated. Taiwan's TSMCOrganizationTSMCComprehensive overview of TSMC's role as the critical chokepoint for global AI hardware, covering business model, geopolitical risk, AI safety implications, and controversies including Arizona expa... manufactures over 60% of all semiconductors globally and more than 90% of the most advanced chips.¹ The Netherlands' ASMLOrganizationASMLASML is the world's sole supplier of EUV lithography machines, making it a critical chokepoint in AI compute governance and U.S.-China technology competition; this article provides a thorough, well... supplies the extreme ultraviolet (EUV) lithography machines without which cutting-edge chips cannot be fabricated. U.S. firms dominate chip design—Nvidia in particular has become synonymous with AI training hardware—while Japan and South Korea contribute critical materials, packaging tools, and memory. This concentration is not merely an economic fact; it is the structural foundation on which AI chip governance is built.

Because no country can manufacture frontier AI chips without inputs from this small cluster of allied nations, that cluster possesses significant collective leverage. Export controls, hardware-embedded mechanisms, and multilateral coordination are all strategies for converting that physical leverage into enforceable governance. The core idea animating the field is that, unlike software—which can be copied instantly and monitored only with difficulty—chips are physical objects that move through trackable supply chains, can be embedded with hardware controls, and must pass through a small number of chokepoints before reaching end users.²

The field is evolving rapidly. U.S. policy has oscillated between administrations, with the Biden administration's 2025 AI Diffusion Interim Final Rule expanding controls on model weights and chip performance thresholds, while the second Trump administration introduced its own modifications and tariffs. Meanwhile, proposals for Hardware-Enabled GovernanceApproachHardware-Enabled GovernanceRAND analysis identifies attestation-based licensing as most feasible hardware-enabled governance mechanism with 5-10 year timeline, while 100,000+ export-controlled GPUs were smuggled to China in ...Quality: 70/100 mechanisms—features built directly into chips to enforce export compliance, verify location, and limit certain uses—have moved from academic papers to active policy discussion. Whether the governance infrastructure can keep pace with the technology it aims to regulate remains deeply uncertain.

Supply Chain Chokepoints and Key Actors

Chip Designers

NvidiaOrganizationNVIDIANVIDIA holds 90-95% of the AI accelerator market, with its H100, H200, and B200 GPUs powering virtually all frontier AI training. Revenue reached $130.5B in FY2025 (114% YoY growth), driven almost ... is the dominant supplier of GPUs used in AI training and inference. Its H100, H200, and Blackwell series chips are the primary targets of U.S. export controls aimed at China. Nvidia has responded to these controls by developing China-specific downgraded chips (the A800, H800, and H20) that attempt to remain below regulatory performance thresholds. Nvidia CEO Jensen HuangPersonJensen HuangJensen Huang co-founded NVIDIA in 1993 and has served as its CEO since. Under his leadership, NVIDIA's GPUs became the dominant hardware platform for training large AI models. NVIDIA's market capit... has publicly evaluated the H20 chip's status under export controls. The company's position on hardware governance has been cautious: Nvidia has reportedly piloted customer tracking software but has resisted proposals for hardware kill-switches or hard functionality throttles, with company representatives arguing that restricting access would stifle global innovation and weaken U.S. competitive advantage.

AMDOrganizationAMDAdvanced Micro Devices (AMD) is the principal competitor to Nvidia in AI accelerator chips, with its Instinct MI300 series GPUs targeting AI training and inference workloads. Founded in 1969 and he... is the principal competitor in AI accelerators and, alongside Nvidia, Intel, and Apple, already incorporates chip features that researchers argue could be repurposed for governance functions—such as location verification pings and usage monitoring—without revealing sensitive operational details.

Intel occupies a dual role as both a chip designer and, through Intel Foundry Services, a fabricator. It is also a major supplier of FPGAs (via its Altera acquisition), which are used in AI inference and edge applications.

Google designs its own Tensor Processing Units (TPUs) for internal AI workloads. Google has been noted for TPU-level tracking capabilities, and its vertically integrated model—designing, deploying, and controlling access to its own AI chips through cloud services—provides a degree of de facto governance that externally sold chips lack.

Foundries

TSMC (Taiwan Semiconductor Manufacturing Company) is the critical chokepoint in global AI chip fabrication. Its geographic concentration in Taiwan creates both geopolitical fragility and governance leverage: any chip requiring TSMC's most advanced nodes must pass through Taiwan. TSMC has cooperated with U.S. export control enforcement by declining orders from sanctioned entities and is expanding fabrication capacity in Arizona under CHIPS Act incentives.³

Samsung operates both advanced logic fabs and dominates high-bandwidth memory (HBM) production, a key bottleneck in AI chip supply chains. Samsung supplies Chinese AI firms (Alibaba, Tencent, ByteDance) with HBM under ongoing supply agreements.

SMICOrganizationSMICSMIC is China's largest semiconductor foundry and a central actor in AI hardware geopolitics, with detailed coverage of its technology trajectory, state backing, U.S. export controls, and controver... (Semiconductor Manufacturing International Corporation), China's leading foundry, has reportedly manufactured 5nm electronic circuits, challenging earlier assumptions about the effectiveness of export controls on manufacturing equipment.⁴ This development is cited as evidence that controls must be continuously updated as Chinese domestic capability grows.

Intel Foundry is positioned as an alternative Western foundry, with the U.S. government and partners such as Amazon Web Services investing in Intel's 18A node as a supply security hedge.

Equipment Suppliers

ASML (Netherlands) is the sole global supplier of EUV lithography machines, which are required to manufacture chips at advanced nodes (sub-7nm). Dutch export restrictions on ASML equipment—aligned with U.S. policy—are one of the most effective chokepoints in the current governance architecture. Without EUV machines, no country can independently manufacture the most advanced AI chips.

Applied MaterialsOrganizationApplied MaterialsApplied Materials is the world's largest supplier of semiconductor manufacturing equipment, providing deposition, etching, and inspection tools used across the chip fabrication process. Founded in ..., Lam Research, and Tokyo Electron supply the etching, deposition, and other semiconductor manufacturing equipment that, together with ASML's lithography tools, constitute the full toolkit of advanced chip fabrication. U.S. export controls have progressively expanded to cover these tools, and the December 2024 expansion of the Foreign Direct Product Rule (FDPR) extended U.S. jurisdiction to any chip manufactured anywhere using U.S.-origin equipment or software.

Deloitte has predicted that at least $30 billion will be spent on critical technologies including EUV lithography and HBM tools by 2026 amid escalating trade barriers.

Governance-Specific Hardware Startups

No major dedicated startups focused solely on governance-specific hardware have been publicly identified in available research. The primary proposals for hardware-enabled governance involve modifying firmware and secure enclave features in chips from existing major manufacturers, rather than purpose-built governance chips. Research from CNAS has documented that AMD, Apple, Intel, and Nvidia chips already contain features—such as location verification mechanisms and change-control hooks—that could be adapted for governance purposes with appropriate firmware updates and policy mandates.⁵

Policy Architecture

Export Controls

The primary instrument of AI chip governance is the U.S. Export Administration Regulations (EAR), administered by the Bureau of Industry and SecurityOrganizationBureau of Industry and SecurityBIS is a high-relevance regulatory actor for AI governance due to its semiconductor export controls and emerging AI diffusion rules; this article provides an unusually thorough, well-sourced, and b... (BIS) within the Department of Commerce. The core tools include:⁶

The Commerce Control List (CCL) classifies chips and semiconductor equipment under Export Control Classification Numbers (ECCNs). October 2022 rules added new ECCNs for high-performance AI chips; January 2025 rules added ECCN 4E091 covering AI model weights.

The Foreign Direct Product Rule (FDPR) extends U.S. jurisdiction extraterritorially: any chip manufactured anywhere in the world using U.S.-origin design software, manufacturing equipment, or components is subject to U.S. export controls. The December 2024 expansion of the FDPR brought essentially all advanced chips—regardless of where they are fabricated—under U.S. regulatory reach.

The Biden Administration's AI Diffusion Rule (January 2025 Interim Final Rule) introduced a tiered country framework, imposing worldwide licensing requirements and a "rebuttable presumption" of control on third-party assembly and test firms. This extended compliance obligations to server manufacturers and data center operators. However, the rule was rescinded by BIS on May 13, 2025—before its May 15, 2025 compliance date—and never took effect.

The Trump Administration's Chip Export Modifications (2025) included eliminating certain Validated End-User (VEU) authorizations for semiconductor manufacturing equipment producers and introducing a 25% tariff on advanced AI chips not intended for the U.S. supply chain.

The H200 Export Policy (January 13, 2026) established a separate case-by-case review rule for exports of chips like the Nvidia H200 to China, requiring that Third-Party Partner (TPP) exports to China remain below 50% of U.S. shipments and that no U.S. supply be delayed.⁶

Proposed Legislation

Several pieces of legislation have been introduced in the U.S. Congress to strengthen chip governance:

• The Chip Security Act proposes a two-phase framework: an initial 180-day mandate for location verification on exported chips, followed by a feasibility assessment within two years for anti-tamper features or functionality throttling. It grants the Commerce Secretary post-export monitoring authority.⁷
• The AI Overwatch Act (introduced December 2025) requires congressional review of advanced AI chip export licenses to China.⁸
• The GAIN AI Act mandates that U.S. businesses receive priority access to advanced AI chips before exports to China or countries of concern.
• The STRIDE Act directs the State Department to coordinate semiconductor export controls with partner nations.

Hardware-Enabled Mechanisms (HEMs)

Hardware-Enabled GovernanceApproachHardware-Enabled GovernanceRAND analysis identifies attestation-based licensing as most feasible hardware-enabled governance mechanism with 5-10 year timeline, while 100,000+ export-controlled GPUs were smuggled to China in ...Quality: 70/100 mechanisms represent a proposed evolution beyond list-based export controls toward enforcement built directly into chips. The CNAS working paper "Secure, Governable Chips" documents existing chip features from AMD, Apple, Intel, and Nvidia that already support governance functions—including location verification pings, usage monitoring, and change-control hooks.⁵

Proponents argue that HEMs could enable more surgical governance: instead of blanket restrictions that burden all users above a performance threshold, chips could verify their own location, flag unauthorized transfers, limit certain use cases (such as training runs above specified compute thresholds), or require periodic license renewal. A staged rollout has been proposed: short-term firmware updates linked to export licenses for data center AI chips, followed by deeper hardware integration in future chip generations.

The CNAS research recommends coordinating HEM development through a NIST-led working group, with input from chip designers (AMD, Apple, Intel, Nvidia), the CHIPS Program Office, and allied governments (Netherlands, Taiwan, South Korea, Japan, UK, EU).⁵ The National Advanced Packaging Manufacturing Program has been identified as a venue for R&D on tamper-proof encasing, though no specific funding amounts have been allocated.

Industry Stances on Hardware Governance

Industry positions on hardware-enabled governance range from cautious engagement to active resistance.

Nvidia is the only major chipmaker actively building governance technology. In December 2025, Nvidia piloted location-verification software for Blackwell AI GPUs, using GPU telemetry and communication latency to landmark servers to estimate chip country of operation.⁹ The system uses confidential computing capabilities already in the chips, is customer-installed (not hidden), and Nvidia plans to make it open-source. However, the company has rejected kill switches: Chief Security Officer David Reber wrote in August 2025 that kill switches are "permanent flaws" and "a gift to hackers," referencing the failed 1993 NSA Clipper Chip as precedent.¹⁰

AMD and Intel have similarly cooperated with export controls while publicly emphasizing innovation and competitiveness concerns. Both companies have existing chip features that researchers argue are HEM-compatible, but neither has committed to deploying them for governance purposes absent government mandate.

Google presents a different model: its TPU chips are used internally and accessed by external customers only through cloud APIs, providing inherent usage controls. Google has been noted for TPU-level tracking capabilities, and its integrated model means governance is partially built into its business architecture rather than being an add-on.

TSMC has been cooperative with U.S. export control enforcement, declining orders from sanctioned entities. Its geographic situation—concentrated in Taiwan—makes it both a geopolitical vulnerability and a natural enforcement chokepoint, and TSMC's cooperation is essential to the current governance regime.

The Semiconductor Industry AssociationOrganizationSemiconductor Industry AssociationThe Semiconductor Industry Association (SIA) is the U.S. trade association representing the semiconductor industry. SIA advocates for policies that promote chip innovation and competitiveness, incl... has called for a more measured approach that balances security with economic and innovation priorities, reflecting the industry's general concern that overly stringent rules push foreign customers toward South Korean and Taiwanese alternatives, eroding U.S. competitive advantage over time.

Investment Levels and Funding Gaps

Global chip revenue reached approximately $630 billion in 2024 and is projected to reach $910 billion by 2026.¹ U.S. private investment catalyzed by the CHIPS and Science Act totals over $630 billion.³ The four major hyperscalers (Google, Amazon, Microsoft, Meta) have projected combined capital expenditure of $650 billion in the AI infrastructure build-out. OpenAI has signed agreements for 900,000 wafers per month by 2029 for its Stargate project.

Against this backdrop, investment specifically in governance-oriented hardware R&D appears marginal. No dedicated public fund for HEM development has been identified in available research. The primary sources of governance-relevant research are policy think tanks—CNAS, the Georgetown Center for Security and Emerging TechnologyOrganizationGeorgetown CSETCSET is a $100M+ Georgetown center with 50+ staff conducting data-driven AI policy research, particularly on U.S.-China competition and export controls. The center conducts hundreds of annual gover...Quality: 43/100, Open PhilanthropyOrganizationOpen PhilanthropyOpen Philanthropy rebranded to Coefficient Giving in November 2025. See the Coefficient Giving page for current information.Quality: 52/100-funded organizations, and academic labs—whose budgets are small relative to chip industry capex.

The research organizations most actively working on AI chip governance policy include:

• CNAS (Center for a New American Security)OrganizationCenter for a New American SecurityCNAS is a moderately important national security think tank with substantive but secondary relevance to AI safety, approaching AI risks through a geopolitical/national security lens rather than exi...: Authored the "Secure, Governable Chips" working paper, the most detailed public treatment of HEM design. Published work on allied coordination for export controls.
• IAPS (Institute for AI Policy and Strategy): Active on compute governance, export controls, and hardware security.
• GovAI (Centre for the Governance of AI): Publishes research on AI governance frameworks, including compute governance dimensions. See also AI Governance and PolicyCruxAI Governance and PolicyComprehensive analysis of AI governance mechanisms estimating 30-50% probability of meaningful regulation by 2027 and 5-25% x-risk reduction potential through coordinated international approaches. ...Quality: 66/100.
• Georgetown CSETOrganizationGeorgetown CSETCSET is a $100M+ Georgetown center with 50+ staff conducting data-driven AI policy research, particularly on U.S.-China competition and export controls. The center conducts hundreds of annual gover...Quality: 43/100: Analyzes semiconductor supply chains, export control effectiveness, and allied coordination.
• Academic labs: Computer science and security researchers at MIT, Stanford, and Carnegie Mellon have contributed to technical understanding of HEM feasibility.

Identifiable dedicated funding for hardware governance mechanisms includes:

This amounts to roughly $10M in philanthropic funding versus $400B+ in annual AI infrastructure spending—a ratio of approximately 1:40,000. No dedicated Focused Research Organization for hardware governance exists, and no VC-funded startup is focused specifically on governance-specific hardware.

Proposed Funding Priorities

Coefficient GivingOrganizationCoefficient GivingCoefficient Giving (formerly Open Philanthropy) has directed $4B+ in grants since 2014, including $336M to AI safety (~60% of external funding). The organization spent ~$50M on AI safety in 2024, w...Quality: 55/100 has described the field as "still very funding-constrained." Several organizations have identified areas where additional funding could advance hardware governance research:

The window may be closing. Lennart HeimPersonLennart HeimCompute governance researcher. Affiliated with RAND, Epoch AI, and GovAI. Leading expert on AI compute supply chains and governance. Co-authored RAND working papers on hardware-enabled mechanisms (... (RAND/GovAI) has argued that compute governance may only be viable for "a decade or two" before algorithmic efficiency improvements make chip controls less effective. Investments made now in hardware governance mechanisms could have outsized impact compared to the same investments made later.

Criticisms and Concerns

Enforcement Limitations

Current export controls face fundamental enforcement challenges. BIS relies primarily on checking buyers against a blacklist, but shell companies can be established for a few thousand dollars in hours or days, while investigations take years. Because officials have no reliable means of tracking who possesses AI chips after export or how they are being used, controls are applied as blanket restrictions on all shipments above certain performance thresholds.

The scale of smuggling is significant. CNAS estimated a median of ~140,000 chips diverted to China in 2024, representing 6–10% of China's AI compute capacity.¹¹ The Megaspeed case illustrates how intermediaries exploit regulatory gaps: the Singapore-based firm allegedly diverted $2 billion in restricted Nvidia chips through a Malaysian subsidiary. In December 2025, Operation Gatekeeper exposed a $160 million smuggling ring exporting H100 and H200 GPUs via shell companies.¹² In March 2026, Supermicro's co-founder was arrested in connection with a $2.5 billion smuggling operation diverting servers to China via Southeast Asian front companies—$510 million in a single six-week window.¹³ Estimated profits from just three smuggling cases in 2024 exceed BIS's entire annual enforcement budget.

De-Americanization Risk

A concern raised by multiple analysts is that blanket controls incentivize foreign firms to design U.S. components out of their supply chains entirely. If Chinese or European chip designers can source manufacturing equipment from non-U.S. suppliers, the FDPR's extraterritorial reach becomes ineffective. This "de-Americanization" dynamic could reduce U.S. leverage over time, producing the opposite of the intended governance effect.

Innovation and Competition Concerns

Critics argue that hardware control proposals would allow chip designers, fabricators, or the U.S. government to determine the purposes and manner of advanced chip use, creating restrictions on the ability of commercial or academic entities to train AI models. In a concentrated AI market, such controls would disadvantage new competitors and harm innovation. The concern is not merely economic: if governance mechanisms entrench the position of incumbent firms and governments, they may serve industrial policy goals more than safety goals.

Geopolitical Complications

Export controls have geopolitical costs. Restricting advanced chips exclusively to U.S. companies risks alienating allies—Nvidia has secured a deal to supply 260,000 advanced AI chips to South Korea's government and industrial partners, a deal that stricter controls could jeopardize. China's SMIC has reportedly manufactured 5nm electronic circuits despite restrictions, suggesting that controls delay rather than prevent adversary capability development.⁴ A cycle of misperceptions about rivals' motives is identified in the research literature as a risk to strategic stability, and trust-building measures are argued to be prerequisites for effective international AI governance agreements.

Policy Volatility

Frequent policy reversals—Biden's AI Diffusion Rule, Trump's modifications, scrapped frameworks—create procurement uncertainty for global semiconductor firms. The second Trump administration reversed an April 2025 export freeze on Nvidia chips in December 2025, subject to conditions. This volatility strains business relationships, delays deployments, and undermines the credibility of U.S. governance commitments to allied partners.¹⁴

Key Uncertainties

• HEM feasibility and adoption: Whether chip manufacturers will voluntarily implement hardware governance features, or whether government mandates will be required—and whether mandates can survive legal challenge—remains open.
• Allied coordination durability: The effectiveness of U.S. export controls depends on Dutch (ASML) and Japanese equipment makers, Taiwanese fabricators, and South Korean memory producers maintaining aligned policies. Defection by any major allied supplier would substantially reduce governance leverage.
• China's domestic capability trajectory: SMIC's 5nm progress challenges assumptions underlying current control thresholds. If China develops meaningful independent capability in advanced fabrication, the leverage model breaks down.
• Threshold obsolescence: As AI algorithms and chip architectures improve, the compute required for dangerous capabilities may fall below current performance thresholds, requiring continuous updates to control parameters.
• Governance vs. industrial policy: The extent to which U.S. chip governance serves genuine safety goals versus protectionist industrial policy goals is contested and shapes whether allies and international institutions will view it as legitimate.

Sources