Computational Neutrality and Computational Sovereignty for Global Compute Access

Hong Yanqing's article in Law Science Magazine, 2026 issue 5, analyzes how U.S. export controls can constrain Chinese firms' remote use of third-country data center compute even without chip imports. It proposes computational neutrality for general commercial workloads and computational sovereignty for critical workloads as complementary drafting principles.

Publication and thesis

Hong Yanqing's article "Building a Global Compute Access Order: Centered on Computational Neutrality and Computational Sovereignty" appeared in Law Science Magazine, 2026, issue 5. The public version is the longer draft; the journal text is condensed. The analysis is dated 17 May 2026.

The paper's starting point is that cross-border access to advanced compute now covers remote invocation of capacity in third-country data centers, not only physical delivery of chips. Chinese firms that do not buy chips or import them can still face export license conditions, customer identification, use review, and entity-list risk when they rent capacity abroad. The stated purpose of an international access order is to reduce that shock. The paper also treats China as a future compute supplier and argues that computational neutrality and computational sovereignty should apply to all dominant suppliers, including future Chinese ones.

"Compute access" is defined as continuous, metered, schedulable remote use of chips, clusters, and orchestration. The paper separates that question from data access and from model access. Ownership of a chip, and whether the chip enters China, is not treated as the whole legal test.

Three linked concepts are introduced. Overseas data center quasi-regulation describes how item, license, and authorization conditions shape facility operations without the exporting state acting as the host-state regulator. A compliance impact zone is the factual range of customer policy created when platforms convert one jurisdiction's requirements, plus their own risk judgment, into onboarding, allocation, and termination rules. An access nexus is the proposed statutory hook that would treat remote invocation itself as a controlled event.

January 2026 U.S. measures

In January 2026, the U.S. Department of Commerce Bureau of Industry and Security (BIS) issued a revision to advanced computing item license review policy. For specified case-by-case reviews of advanced computing exports to China and Macau, applicants may have to submit customer identification procedures and remote end-user lists. The paper states that this path can extend chip licensing into post-deployment service relationships, within the listed items, destinations, and license conditions. It is not described as a universal duty on all third-country data centers.

The same month, the U.S. House of Representatives passed H.R.2683, the Remote Access Security Act, which would add an authorization in the Export Control Reform Act (ECRA) for remote access to items. The paper uses the House-passed text as the proposed legislative direction. Existing license conditions and the proposed access authorization sit at different legal levels, but both move the focus from where a chip is shipped to who keeps calling the capacity it hosts. Platform contract clauses on review, suspension, and termination are identified as the channel into customer relationships.

Quasi-regulation of overseas data centers

Third-country data centers are not automatically outside U.S. Export Administration Regulations (EAR) effects if a transaction involves EAR-controlled advanced computing items. Final-use, end-user, and anti-circumvention conditions can pass along the supply chain into operations. The paper calls resulting limits "factual blocking" when they cut practical access without a total service ban: refusal, review delay, capacity limits, or instability. It does not claim that all Chinese customers have been fully excluded.

EAR section 734.3 sets the item-jurisdiction starting point, including certain U.S. items and U.S.-origin items. Some foreign-made items may be pulled in by de minimis or foreign-direct-product rules. Jurisdiction over an item does not by itself answer which cloud services need a license. After deployment, a chip can serve many customers, so purchaser, facility operator, and caller may be different parties.

The 2024 Data Center Validated End User (VEU) authorization is cited as a historical sample of operationalized review: customer relationships, physical and logical security, identity and access management, records, reporting, and follow-up review. The January 2026 rule, on a case-by-case path, asks for customer identification, IaaS remote end users, and measures against unauthorized access and model-weight transfer. Destination, party, item, and time still have to be assessed separately. The paper warns against collapsing those programs into one global cloud-service license.

May 2025 anti-circumvention industry guidance lists, as a red flag for further inquiry, an IaaS provider that cannot confirm its users are not China-headquartered entities. The flag does not disappear because the provider sits outside China and Macau. Red flags are not standalone bans. Related statements point to the EAR "knowledge" standard and possible license requirements where model training and WMD or military-intelligence end uses or end users are involved. Entity-list risk can affect later access to EAR items even without a completed EAR violation.

Quasi-regulation is summarized as conditional (supply and authorization terms, not a substitute for host-state regulation), continuous (compliance does not end at chip delivery if later use and users remain in the license), and public-private (platforms implement the effect through contracts and account policy).

Compliance impact zone

Quasi-regulation explains how state rules enter a facility. It does not, by itself, explain why a platform would exclude a customer the law does not clearly bar. The compliance impact zone is that extra range: legal duties plus preventive risk policy. Four incentives are listed: ongoing verification cost as control and use change; limited order revenue versus license or listing risk; interpretive space between red flags and prohibitions; and unified global policies.

"Default deny" is defined as presuming non-supply from an identity category before a workload-specific risk judgment, or setting conditions that cannot be completed in practice. License rules show remote end-user and related controls in some applications. Anti-circumvention guidance shows some customer relationships treated as risk clues. Platform contracts show review, pause, and termination tools. Together they support a transmission mechanism and the possibility of preventive exclusion. They do not, on their own, prove systematic default denial of all Chinese firms in every third-country market. The paper keeps rule analysis separate from market frequency, which would need refusal notices, review times, instance outcomes, peer comparison, or interviews.

Harm types and two principles

Four harm types are listed. Access harm: the customer cannot obtain a resource or permission. Time harm: identity, control, and use checks overrun the task window. Capacity harm: scale, performance, or priority is too low to finish the job even though some service remains. Continuity harm: instance reclaim, account limits, or contract termination after work has started, including interrupted training and blocked checkpoint reads.

General commercial workloads, absent a specific restricted use, raise a consistent, explainable market-entry question. Some public-service, critical-infrastructure, or strategic-research tasks need higher continuity because interruption costs and replacement time are high. Public-body status or a "foundation model" label does not make every task critical. Task function, tolerable downtime, and substitute capacity are the proposed tests.

Computational neutrality would constrain nationality-based default denial of general commercial workloads. Computational sovereignty would require minimum continuity, judicial controllability, auditability, and exit capability for critical workloads. The two principles are complementary drafting tools, not an existing unconditional international right to obtain compute.

Six modules are proposed: scope and workload layering; computational neutrality; due diligence discipline; refusal and cutoff procedure; computational sovereignty; and trust safeguards. Security controls would remain available, but access limits would need an explainable basis, a reviewable process, and a workable exit, applied equally to all dominant compute suppliers, including the rule-makers.

The paper's claimed contribution is to turn scattered access risk into identifiable duties so security governance, platform responsibility, and user continuity can be tested in one frame.