The Trump administration runs a protection racket on Chinese firms’ access to American AI compute.
Jensen Huang lobbied President Trump to keep Chinese AI firms on American compute, and won it for a year. A year ago, he called the term “China hawk” a “badge of shame” and “unpatriotic.” Before his departure, Huang denied that Nvidia had a diversion problem and took his case directly to Trump for fewer regulations and greater access to China, despite the national-security warnings his own government had on file. The man selling the access spent the year personally vouching for the supply chain.
Megaspeed, a Singapore-based Nvidia business partner, is accused of smuggling chips to China. Lawmakers cited alleged funneling worth billions of dollars and blamed Commerce for moving too slowly. The Wall Street Journal reported that Alice Huang, a former top Megaspeed executive whom American officials suspect is a Chinese intelligence asset, had an affair with Jensen Huang. No court verdict is required to describe the year. The alleged diversion route had a trapdoor. The scandal is not a verdict. The year is the indictment.
The Commerce Department’s new export rule is the instrument of seizure. It took effect immediately, even as companies have about 60 days to offer comments, and it restricts Chinese access to computing power powered by U.S. chips in Malaysia, Thailand, Indonesia, and Singapore. It also tightens restrictions on Chinese chip designers. MSI’s earlier report on foreign access to AI chips described an administration moving toward tighter controls; Huang’s ouster supplied the accelerant. The consultation is retrospective: contracts and compliance systems are already changing while the official record is still being assembled. The rule asks the market to absorb the decision before the government has finished explaining it.
It is true that chip diversion to Chinese AI companies is a serious concern and that lawmakers had pressed Commerce to close overseas loopholes for years. Huang’s commercial interests conflict with those concerns, so his lobbying cannot be treated as disinterested security advice. Corporate self-interest does not entitle an executive to set national-security policy. The trouble is that Washington’s remedy is not limited to conduct that can be proved against a particular end user. It makes Chinese access itself the condition to be policed.
Oracle supplies the cleanest measure of force. According to people familiar with the matter, it cancelled a $7 billion, five-year agreement to lease AI computing capacity to Tencent in Southeast Asia after U.S. officials raised concerns. The company did not respond to a request for comment. The cancellation followed official concern rather than a judicial finding of illegality; that is the mechanism worth noticing.
That channel is not a smuggler’s tunnel. It is commerce—signed, renewable, and consistent with the stated policy right up until officials raised concerns. Regulated firms do not need a final prohibition to understand that a contract involving a Chinese customer may become a federal investigation. Official concern was enough to make the commercial risk unacceptable.
At the infrastructure layer, exporting a chip and exporting compute are not synonyms. A cloud customer generally does not receive a boxed processor; it rents slices of computing time—virtual machines, containers, or model endpoints—running on hardware in a facility it may never see. The new rule reaches that service layer: Chinese access is restricted even where the servers stand in third countries. An API key—the token that lets software call a service—can carry geopolitical nationality.
Where a customer takes trained weights out of a provider’s environment, files cross the border and the compute capability travels with them. Where the workload remains on the provider’s cluster and is served back over an interface, the answer travels and the capability stays. A rule reaching both the chips and access to them therefore changes the contractual allocation of power: who may rent the cycles, who may resell them, and who audits the workload. Kate Crawford opens Atlas of AI with the observation that “AI is made from vast amounts of natural resources, fuel, and human labor.” Compute is infrastructure, physical and locatable, not a spell.
The enforcement costs have already migrated into private balance sheets and paid compliance labour. Since the Nvidia affair broke, H.K. Park, who leads the national-security practice at Crumpton Global, has received triple the number of requests from companies seeking checks on suppliers, business partners, and customers in Southeast Asia. Corporate boards have hired the firm to determine whether they face China-linked exposure; some companies worry they could become the next target of a U.S. government investigation.
I spent three years in Ottawa on cleared federal contract work, and the lesson that outlasted the briefings was modest: an export-control regime is a stack of binders, and the binders perform exactly as well as the last person in authority wanted them to. The binders did not change this month. The person did. Every company downstream is now paying—compliance retainers, cancelled contracts, disclosure paperwork—for the privilege of learning that the constraint it was promised was a brochure. That is not voluntary friction. It is the state assigning multinational firms the job of proving which customers fall outside a Chinese-access prohibition.
To be fair, the administration’s rule has a real strategic setting. Beijing had imposed export controls on rare earths, and the preceding year’s China policy had mostly suspended, abandoned, or softened antagonistic steps. There was also a plausible enforcement case: advanced chips sold through third countries become difficult to trace once servers, subcontractors, and resellers enter the chain. The trouble is that the chosen rule does not stay with that conduct. It makes Chinese access the routing condition for U.S.-chip compute in four foreign markets, then makes private firms police the condition.
The rule’s supporters supply a true half and suppress the variable that should control the remedy. Advanced chips can be diverted; export controls can impose costs; firms must obey the rules. Fine. But a restriction broad enough to exclude lawful Chinese users from four foreign markets imposes those costs whether or not any particular shipment was diverted. A measure aimed at diversion should make diversion easier to identify and punish. This one makes the customer’s passport—or, more often, the nationality of the entity controlling it—part of the compute quote. It is a security strategy with a customs department made of other companies’ compliance teams.
In fairness to the detente—in the narrow sense in which stability between nuclear powers is a virtue—great powers have always managed dangerous trade this way, and a cold peace outperforms a hot one on any ledger that counts people. The trouble is that this detente was administered by the parties with inventory to move. The policy of record had American companies balancing commercial interests against national security when selling sensitive technology. Read it slowly: the seller was asked to conduct the buyer’s background check on himself.
Scott Bessent and Marco Rubio advanced proposed rules that had been languishing. Commerce Secretary Howard Lutnick had resisted restrictions that might anger China, and Trump defended him before boarding Air Force One for the Asia-Pacific Economic Cooperation meetings in Shenzhen. Some Commerce Department staff were forced out after advocating tighter restrictions. Cory Doctorow’s account of why platforms misbehave is a theory of removed constraints—competition, regulation, interoperability, labour—and the one missing here was regulation. “When a company is bigger than the government,” Doctorow wrote, “it gets damned hard to credibly threaten to punish that company.” Washington spent a year demonstrating the corollary: when the government had decided not to threaten the company, the constraint was not weakened. It was absent.
Not an argument. A scandal with a name on it. Policy that moves only on scandal is not policy; it is compensation. Washington did not develop a settled position on compute diversion. It developed a schedule.
Cory Doctorow and Rebecca Giblin’s Chokepoint Capitalism describes a basic mechanism: when a powerful intermediary stands between creators and audiences, it can squeeze both sides because neither can route around it. Washington has not abolished the chokepoint; it has changed its toll collector. Instead of one dominant cloud platform deciding which customer receives capacity, a state system now decides which China-linked customer may use American hardware at all. That can produce a legitimate security benefit. It can also make U.S. compute a protected market rather than an export, with the protection sold as compliance software, advisory retainers, and legal exposure.
My father spent thirty years on the bar mill at Manitoba Rolling Mills, where a change in gauge altered what the line could make and what the customer had ordered. The trade was simple enough to explain at the shop floor: change the setting and the machine follows, whether or not the order was sensible. The new rule is a change in gauge at national scale. The machines—American chips in Malaysian, Thai, Indonesian, and Singaporean facilities—follow immediately. The shop floor now has to discover what the setting was supposed to accomplish.
That makes Harold Innis’s staples question newly practical, although the staple here is compute rather than grain, fish, or newsprint. Control of a communications-and-compute staple lets its owner set the terms on which everyone else trades. Singapore is this cycle’s entrepôt, the city that exists to transship what other people produce, and the legal authority over a customer can now be decided in Washington even when the service sits in Singapore. The old geography of extraction is not disappearing; it is acquiring a cloud-control layer and a compliance checklist.
The proposed disclosure rule for Chinese AI models is described as gentler than a ban. Airbnb and Coinbase are among the prominent companies using Chinese tools, often because they are less expensive and easier to customize than models from OpenAI and Anthropic. It is true that the Chinese models are cheaper and that their customization options matter to procurement departments. A cost line is a cost line. The trouble is that the procurement logic priced the model and never priced the dependency.
Disclosure locates the tenancy after it has been built. A questionnaire with national-security language attached is not a light-touch policy when every answer becomes a branch in a board’s risk model. Firms bear the cost of proving their use of a permitted tool is permitted.
Joseph Hoefer, chief AI officer at Monument Advocacy, which represents companies including Netflix and Waymo, said, “Companies that assumed the light-touch approach would hold through 2028 are going to have to recalibrate.” His second sentence is more revealing: “National security gives all sides cover to deal in a way that just wasn’t there before.” Businesses with a relationship to China, or an intermediary, are paying more attention to the incoming Congress than to the president’s words. That is the market’s own way of confirming that the detente is over.
Sen. Pete Ricketts, one of the lawmakers pressing for more restrictions, put the new momentum plainly: “There are those of us who wanted to have this regulation and now it’s becoming obvious why we need to have it.” He added, “This gives us more momentum to go back and do more restrictions.” Elizabeth Warren and Alexandria Ocasio-Cortez, who had demanded AI regulation, now have a scandal doing their scheduling. The House Select Committee on China, created in 2023 and sidelined during the friendly years, has better prospects after the Nvidia affair. Huang and Nvidia have been called to testify before congressional committees in the coming weeks.
Earlier MSI reporting on Trump’s closeness to Huang made the personal dimension visible. The sequence does not prove the rule unnecessary, but it shows how the argument moves: personal scandal supplies political demand; political demand makes broader commercial exclusion easier to present as administrative housekeeping. The scandal is not the policy’s evidentiary foundation. It is its accelerant and its shield.
Tim Wu’s warning in The Curse of Bigness is that concentrated private scale becomes a rival to public accountability—that bigness itself, regardless of price, is a threat to liberty. The present rule performs the inverse operation: public authority is used to police corporate access at a scale private firms cannot reach. There is no competition on either side of that arrangement. One side uses law to turn national allegiance into a condition of market access; the other side must hire counsel to discover which customers remain legally visible. The rule’s proponents assert that the security benefit exceeds the rent and friction the restriction creates. The enforcement record must demonstrate it.
Lina Khan’s analysis of Amazon in Amazon’s Antitrust Paradox shows how a formally plural market can conceal a gatekeeping relationship that smaller suppliers cannot escape. The same principle applies to capacity. A firm can have several cloud vendors on paper and still have no practical substitute for a class of American-chip workloads once nationality and end-user rules are applied. The policy narrows access without owning the data centres or the chips. Regulatory permission can do what physical concentration would otherwise have to accomplish.
The strongest objection on the other side is not that every diversion allegation is false or that Chinese AI development is harmless. It is that a narrow enforcement rule aimed at false end-user declarations, prohibited end uses, and customs laundering would be harder to evade and would require less sacrifice from lawful commerce. The public evidentiary baseline remains unsettled: the account of alleged customer diversion ran into a nationwide ground stop for airlines, bank-account glitches, and money-transfer delays. Those breakdowns do not prove that diversion is absent, but neither do they supply a clean audit for a rule broad enough to govern lawful access in four foreign markets. The defensible rule is available, though less convenient: verified end-user and end-use declarations, penalties and audits for false declarations, responsibility on the party that misstates the customer, notice and comment before effect, public enforcement results, and expiry unless Congress renews. None of this requires a new doctrine; it requires the government to show that broad exclusion prevents diversion better than targeted enforcement.
A policy this broad should state what it has achieved: fewer diverted chips, fewer Chinese access routes, or a measurable change in the computation available to Chinese AI firms. It has produced a known private cost—Oracle’s cancelled $7 billion agreement—and a known compliance boom. It has not yet produced the public metric by which its reach is supposed to be judged. Export controls are supposed to be instruments, not statements of intent. The instrument should at least be auditable.
A protection racket here is a structure, not a guess about secret payments: privileged market access for sellers inside the perimeter, compliance burdens for firms outside it, and a security rationale broad enough to make the boundary itself part of the commodity. The rule does not merely restrict a product. It makes permission to compute part of the product’s terms.
The Commerce Department has about 60 days to learn what its rule does to a market it has already changed. The comments may change the text; the contracts, compliance systems, and board votes caused by the rule are already on paper. Speed is useful. It is not a substitute for the record. A rule you can lobby away is not a rule. It is a price.