What If China Needs Nvidia to Lose?

In late October 2026, the New York Times reported that American officials had concluded that Alice Huang, a businesswoman tied to the Singapore data-center company Megaspeed International, was a Chinese intelligence asset, and that she had a close relationship with Jensen Huang, chief executive of Nvidia, the leading maker of artificial intelligence (AI) chips.1 Within two weeks, Huang published a letter apologizing for his personal conduct,2 Nvidia's board opened an independent investigation,3 and on November 11 the board accepted his resignation.4 The Information then reported that her network had helped move about 240,000 restricted chips, worth at least $7.5 billion, into China, a figure Reuters could not independently verify.5

Start with what is not known. The reporting on Alice Huang (also known as Huang Le) and Megaspeed rests largely on anonymous sources, and it is contested. Attorney General Todd Blanche says the Justice Department has reached no conclusion about whether she was a Chinese spy,6 and Singapore, which has investigated Megaspeed for false declarations since February 2025, says she "was not investigated for espionage."7 Neither she nor Megaspeed has been charged.8 Nothing public shows that Jensen Huang knew of any diversion or that anyone used his personal life as leverage; President Trump says Huang told him he had "absolutely no idea."9 Nvidia says the diversion allegations against Megaspeed "were thoroughly investigated and debunked last year,"10 and Megaspeed denies wrongdoing.11 Beijing calls the story a fabrication.12 The title of this Article names a logic a state could follow. It is not a finding that any state, service, or person followed it.

So the question here is narrower, and I think more useful. Assume the hypothetical: if a state did run a chip channel as a campaign, would the strategy make sense? Two readings of the scandal are available. The first holds that Beijing benefited from any such access and is therefore injured by its exposure and by the turmoil at Nvidia, which lost about $430 billion in market value over two sessions.13 Reuters Breakingviews columnist Robyn Mak put it this way: the fallout "may end up hurting China's AI prospects more than it helps," because Washington will now face pressure to close the legal workarounds Chinese firms use.14 Emily Kilcrease of the Center for a New American Security (CNAS) told The Verge that without the latest Nvidia chips, China "cannot close the gap with US frontier capabilities."15 The second reading holds that destabilizing the dominant American supplier can advance China's position. The China analyst Bill Bishop called the affair "a win for" China's Ministry of State Security (MSS), which would now "get to watch as politicians and shorts attack him and Nvidia," although he takes the contested intelligence link as given.16 This Article develops the second reading without that assumption. Both can be true in sequence rather than in contradiction.

My answer is this: the strategy would be rational. Disruption at Nvidia still costs China's frontier labs today, but it serves Beijing's industrial policy once domestic chips can carry the workloads the state cares most about, and the public record shows that point approaching, not reached. The logic holds whether or not the Megaspeed reporting turns out to be true. It does not need the reporting to be true.

The sequence works like this. While domestic chips lagged by a wide margin, any channel that delivered higher-performing silicon was useful.17 Once Huawei and other suppliers showed they could absorb a substantial share of Chinese demand, even while still behind on a per-chip basis, the calculus began to shift.18 Continued dependence on the foreign channel became a liability. Evidence that the channel had been used could be turned into pressure for tighter controls, reduced reliability of the foreign option, and faster migration onto the domestic stack. Washington's reaction shows the first step: Senate Minority Leader Chuck Schumer called for new chip restrictions within days, and House chairmen pressed for four chip bills.19 At that point, the relevant metric is no longer whether the latest foreign chip is 30 or 50 percent faster. It is whether the adversary retains a chokepoint over the compute that trains and runs the models that matter for national power.

Henry Farrell and Abraham Newman call the use of such a chokepoint weaponized interdependence. A state that controls a hub every other network must pass through can watch what flows through it or cut off the flow, and the state denied access pays the price. In their words, "[f]ocal points of cooperation have become sites of control."20 Nvidia's chips and software are that kind of hub for AI. Farrell and Newman describe the state that holds the hub. This Article asks what the state on the other end does, and its answer extends their framework rather than restating it: the target builds its own hub, then treats the old one's troubles as an opportunity. A sequence is incentive-compatible when each step serves the actor's interest on its own terms, so no one has to plan the whole path for it to unfold.

This Article proceeds in six parts. Part I explains why access mattered while China lagged. Part II describes the threshold at which the channel's value flips. Part III explains why, past that threshold, the contest turns on control of the stack. Part IV takes the limits seriously, including the gaps in the public record. Part V recommends how to test the argument, and Part VI answers the strongest objection.

I. ACCESS WHILE LAGGING

A. The Gap Is Measured in Years

The performance gap between Nvidia's leading parts and China's best domestic accelerators has been measured in years, not months. On Epoch AI's estimate, Huawei's newest Ascend chip delivers roughly seven times less compute than Nvidia's current flagship, which puts Huawei about four years behind.21 Three constraints keep the gap open: manufacturing yields, memory, and software.

Yields on the processes available without extreme ultraviolet (EUV) tools remain lower. EUV lithography prints the smallest chip features with very short-wavelength light, and shorter wavelengths print smaller features.22 China's leading foundry, Semiconductor Manufacturing International Corporation (SMIC), has no EUV tools, so it reaches advanced nodes by exposing each layer several times with older deep ultraviolet machines. Every extra pass adds cost and error, and fewer chips on each wafer come out working; that share is the yield. Gregory Allen of the Center for Strategic and International Studies puts SMIC's yield on Huawei's Ascend 910B at about 20 percent.23 A failed chip is a chip China paid for and cannot use.

High-bandwidth memory (HBM) is constrained too. HBM is a stack of memory chips placed right beside the processor so it can be fed data fast enough to stay busy, and a fast chip with too little of it sits idle.24 Since December 2024, Washington has required a license for advanced HBM bound for China,25 and Huawei is moving its newest chips onto its own "proprietary, low-cost HBM."26 A chip China can make but cannot feed is not yet a substitute.

The software stack, though improving, does not yet match the depth of Nvidia's Compute Unified Device Architecture (CUDA) ecosystem. CUDA is the platform of drivers, compilers, and libraries that most machine-learning software is written for.27 Huawei's counterpart, the Compute Architecture for Neural Networks (CANN), is still maturing. Huawei pledged to open-source it, but developers find it "less user-friendly than CUDA," and Epoch AI says it "remains years behind."28 Leaving Nvidia is a software problem as much as a silicon problem. A lab that moves to Huawei has to rewrite and retune its code, and an Nvidia chip lets it skip that cost while the domestic stack catches up.

B. Why the Channel Mattered

Under those conditions, chips that reach Chinese entities through intermediaries, third-country routing, or other workarounds are not marginal. Prosecutors in New York allege that one network diverted more than $510 million in servers to China in about three weeks of 2025,29 and in Texas, defendants admitted exporting at least $160 million in Nvidia H100 and H200 chips.30 Researchers at CNAS estimate that smuggled chips supplied between 1 and 40 percent of China's 2024 training compute, with a median estimate of 10 percent.31 At that scale, the channel likely made training and deployment faster or cheaper than they would otherwise have been, and it bought time for the domestic effort.

Export controls made the channel more valuable, not less, because they confirmed that open commercial access could not be assumed. Export controls here are the Commerce Department's license requirements for selling advanced chips into China; the October 2022 rule set a presumption that most such licenses would be denied.32 A license rule does not erase Chinese demand for the chip. It pushes that demand off the open market, and whoever can still deliver the part captures the price gap the rule creates. The company at the center of one California smuggling case was "founded shortly after the Commerce Department began requiring licenses,"33 and in a Georgia case, defendant Matthew Kelly called selling restricted servers into China "a lucrative business right now," with "millions of dollars in profits per order."34 The tighter the rule, the more the remaining channel is worth, and the more its exposure later costs the supplier at the other end of it.

C. What an Intelligence Service Would Want

An intelligence service tasked with supporting national technological goals would have a clear interest in keeping such channels open for as long as the gap persists. China's National Intelligence Law gives it the legal footing to ask: "All organizations and citizens shall support, assist, and cooperate with national intelligence efforts."35 The interest would be less in the chips as objects for reverse engineering, a possibility United States officials have raised though no public case documents it with Nvidia parts,36 than in the operational compute the chips provide while the domestic alternative matures. Facilitating or protecting a flow of restricted silicon would reduce the immediate penalty of export controls and lower the political cost, inside China, of the time required to close the manufacturing and design gap.

The same service would want to understand the supplier: its decision-making, its tolerance for risk, and the points at which political or personal vulnerability could be used to sustain the flow. Reporting on the Megaspeed matter describes an alleged effort to collect intelligence and influence Nvidia's chip policy. It does not describe blackmail or coercion of Jensen Huang, and nothing public indicates that any leverage was used.37 None of this requires the channel to be the sole or even the largest source of chips. It requires only that the channel be material while the alternative is still scaling.

II. THE THRESHOLD AND THE REVERSAL

A. The Threshold Is Not Parity

The threshold is not full parity. It is the point at which domestic silicon can cover a large enough fraction of priority workloads that losing the foreign channel no longer threatens the overall effort. Priority workloads are not one market. Running finished models for users (inference), training frontier models, and stocking state-funded data centers are three different tests.38 Domestic chips pass them unevenly. They are most usable for inference, weakest for frontier training, and in state-funded data centers Beijing can simply require them; Reuters reports that new state-funded projects must use domestic chips.39

The market data show the threshold approaching. International Data Corporation (IDC) counts Huawei at 28 percent of China's data-center AI accelerator shipments in the first half of 2026, measured in units.40 TrendForce put Huawei's full-year 2025 share at 20.3 percent.41 Labs are moving too. iFlytek says it trained its models entirely on Huawei hardware,42 DeepSeek has used Huawei's Ascend chips for part of its training,43 and Zhipu supports running its GLM-5 model on Ascend chips.44 Together, those facts indicate that the threshold has been approached even though it has not been crossed at the frontier.

A 28 percent share of units is not 28 percent of the computing power, though. Epoch AI estimates that Huawei will produce "roughly 25× less compute than Nvidia in 2026."45 The threshold concerns coverage of priority workloads, not raw power, which is why it can be approached while the compute gap stays wide.

B. Why the Channel's Value Flips

Once that condition is met, the channel's strategic value flips. What had been a source of needed compute becomes a lever Washington can tighten,46 a political liability for the American supplier, and one more reason for Beijing's own push away from foreign chips. Beijing has framed that push around backdoor risk and antitrust law; no official Chinese statement ties it to smuggling or to the Megaspeed exposure.47 Disruption, whether from scandal, tighter enforcement, or the supplier's lost ability to argue for continued sales, imposes short-term costs on labs that still prefer the higher-performing parts. Chinese labs "still reach for Nvidia hardware when they can get it."48 The same disruption reduces the reliability of the foreign option, strengthens the internal case for more domestic capacity, and removes a prominent advocate for continued sales.

Jensen Huang was that advocate. He called the export controls "a failure,"49 and he resigned on November 11.50 For a state whose newest five-year plan pledges to "adopt unconventional measures to drive decisive breakthroughs" in integrated circuits, backed by a national chip fund with 344 billion renminbi in registered capital, these second-order effects are not side issues.51 In the later stage of the sequence, they are the point.

C. The Intelligence Dimension

The same logic applies to the intelligence dimension. In the hypothetical, a channel that was protected or facilitated while it was useful could, once domestic scale exists and someone else exposes it, be exploited as evidence in a broader campaign. The Megaspeed record does not show even that much. Reporting credits United States surveillance and investigators with the discovery, and nothing public indicates that Beijing timed or caused the disclosures.52 Even the Taiwanese newspaper that reported, citing unnamed United States sources, the most elaborate version of the spy story says the United States government chose the moment to act.53

In the hypothetical, the goal is no longer to keep a specific flow alive at all costs. It is to ensure that the foreign supplier's position is weakened when the domestic alternative can absorb the demand that shifts. That requires no perfect timing and no single master plan. It requires only that the service keep the option to exploit the channel's political vulnerability once the industrial condition has changed. That is a claim about options, not proof of a plan, and Part V names the facts that would count against it.

III. FROM FRACTIONAL GAINS TO CONTROL OF THE STACK

While the gap is large, small differences in chip performance compound into large differences in training time, cost, and model capability. DeepSeek's founder has said Chinese firms "must use two to four times the computing power to achieve the same results."54 Once both sides can train and run highly capable models on domestically controlled compute, those fractional advantages matter less for the strategic balance than three other things: the ability to deny the other side reliable access to improvements, control of the software and systems layer, and assurance that one's own models do not depend on an adversary's hardware. China has not reached that condition; Epoch AI expects Huawei to "likely remain around four years behind" through 2030.55 This Part describes where the contest is heading, not where it stands.

Commoditization of the base model does not end competition. It changes the terrain. A model commoditizes when capable weights become cheap or free to copy, so the weights stop being the scarce thing. Compute does not commoditize the same way. As Gregory Allen puts it, "There is no ceiling on the demand for intelligence."56 The contest moves toward who can field models at scale without external permission, who controls the data and the fine-tuning for national-security uses, and who can impose costs on the other's compute supply.57

In that environment, a Cold War framing is not rhetorical. It describes a condition in which the leading edge of commercial AI is also the substrate for intelligence analysis, military planning, industrial control, and information operations. Beijing's military-civil fusion plan lists "high-end general-purpose chips" among the technologies it wants moving between civilian and military use.58 A state that approaches usable domestic scale in accelerators would have a stronger interest in measures that slow or destabilize its rival's unchallenged lead than in preserving a channel that mainly serves the catch-up period. The channel that once supplied the missing compute becomes, at the later stage, an instrument for highlighting dependence, justifying controls, and accelerating the shift onto sovereign silicon.

IV. LIMITS AND ALTERNATIVE READINGS

A. What the Record Does Not Show

The sequence is incentive-compatible. It is not proven as a completed, centrally directed operation in any specific case. The strongest contrary report deserves a direct answer. The New York Times, citing anonymous officials, reports that the MSS recruited Alice Huang to collect intelligence on Jensen Huang.59 Other officials told CNN (Cable News Network) that the intelligence gathered so far "does not prove that she was acting as a spy for Beijing."60 I treat the MSS account as an allegation, and nothing in this Article depends on it being true.

The timeline also cuts against reading the episode as a play by Beijing. United States officials were examining Megaspeed by 2024; a Commerce Department export-control official visited its Malaysian data center that September.61 Singapore opened its own investigation in February 2025 into "potential false declarations in relation to the purchase of servers."62 The New York Times first reported on Megaspeed's chip purchases on October 9, 2025,63 and the personal scandal came a year later, in late October and November 2026.64 Public attention to the channel built in stages, and it began well before the scandal sharpened it. On the current record, United States counterintelligence work and press reporting explain that attention. A later-stage exploitation reading is possible in principle, but nothing public supports it.

B. Profit Explains the Flows

Commercial actors have independent reasons to move chips across borders for profit. Export controls create arbitrage, and renting chips housed in Singapore or Malaysia data centers to Chinese firms is, under current rules, legal.65 Personal and corporate risk-taking can produce the same flows without intelligence tasking. As a former Commerce official told CNN, "An intelligence operation wasn't necessary to achieve that result."66 Profit is the explanation the record supports best. That is why this Article argues from incentives rather than from the Megaspeed reports: the incentive exists whether or not anyone acted on it.

C. Domestic Scale Is Partial

Domestic scale, moreover, remains partial. Frontier training still gains from the best available foreign chips, so the short-term costs of disruption are real. When DeepSeek tried to train on Ascend chips, according to Epoch AI, "Huawei sent its own engineers in and still couldn't complete a training run."67 Huawei's founder, Ren Zhengfei, told People's Daily there is "no need to worry about the chip problem," because stacking and clustering chips yields results comparable to the most advanced level, though he conceded in the same interview that Huawei's single chip still lags the United States by a generation.68 Claims like that overstate the present position in the same way that claims of total dependence understate it.69

What the sequence does show is that the simpler argument (disruption at Nvidia mostly harms China because China still needs the chips) is incomplete once domestic volume is material. Beijing already treats Nvidia access as a liability in some segments. The Cyberspace Administration of China summoned Nvidia over alleged security risks in its H20 chip, and Nvidia reports that Beijing restricted its licensed H200 sales.70 The simpler argument treats a channel that was valuable during catch-up as if it kept the same value after a usable alternative existed. The strategic interest in weakening the foreign supplier's dominance follows from the arrival of that alternative, and it does not depend on any service having caused a disclosure.

V. JUDGE THE CHANNEL BY WHAT CAN BE TESTED

An argument that fits every outcome explains nothing. If a channel stays hidden, that looks like facilitation; if it is exposed, that looks like exploitation. Richards Heuer's advice to intelligence analysts applies: state in advance what evidence would count against the claim.71 The useful question for Washington, then, is not whether Beijing planned the Megaspeed story. It is whether the industrial conditions that would make a later-stage strategy rational are arriving. Three tests answer it.

Coverage. Do domestic chips cover priority workloads? Watch IDC's share figures, which count units, alongside compute estimates like Epoch AI's, which weight each chip by performance.72 Rising unit share with a flat compute gap means inference and procurement are moving while the frontier is not.

Conduct. Does Beijing treat foreign chips as a shrinking necessity? The security summons, the antitrust finding, the reported domestic-chip rule for state-funded data centers, and the restriction of licensed H200 sales all point that way.73 A Beijing that buys every Nvidia chip it is allowed to buy would point the other way.

The gap. Does disruption at Nvidia narrow or widen the aggregate compute gap? The Institute for Progress estimates that Huawei's 2025 output was only 1 to 4 percent of United States production.74 While numbers like that hold, disruption at Nvidia costs Beijing's frontier labs more than it gains them, and the later stage has not arrived.

The policy recommendations follow from those tests. Keep export controls aimed at aggregate compute, where they still bite. Fight diversion through formal, public tools, such as the location-verification requirement in the pending Chip Security Act, which would let authorities confirm where a covered chip actually is.75 And treat the intelligence allegations the way Singapore and the Justice Department have: as matters for investigation, not for headlines.76 A scandal is a poor instrument of export policy. It moves fast, it rests on anonymous sourcing, and, if the argument here is right, it can serve the rival's industrial policy as easily as Washington's.

VI. THE STRONGEST OBJECTION

The best argument against this Article comes from Chinese labs themselves: compute is still their binding constraint. Lin Junyang, technical lead of Alibaba's Qwen team, says United States compute "may overall exceed ours by one to two orders of magnitude."77 DeepSeek's founder says "bans on shipments of advanced chips are the problem."78 CNAS analysts add, citing Goldman Sachs, that China can domestically produce only about 8 percent of the advanced AI chips it needs.79 On that view, disruption at Nvidia still threatens the overall effort, and the later stage is a fantasy.

This is a good argument, and much of it is right. For frontier training today, China still needs the chips, and disruption costs it. But the objection measures the wrong market. In inference and in state-funded data centers, Beijing can already require domestic parts, and it is doing so.80 Political substitution can run ahead of physical substitution. A state can treat a foreign supplier's troubles as useful in the segments it controls while its best labs still want the best chips. The objection describes the frontier. The sequence describes the market as a whole.

What would change my mind? Two things. First, evidence that Chinese firms still cannot run priority workloads, inference included, without Nvidia parts even as Huawei's share rises. Second, evidence that Beijing keeps buying every Nvidia chip it can get instead of restricting them. The record so far cuts my way on the second: Beijing has restricted licensed H200 sales even as its labs want them.81 The first is open, and it is the test that matters most.

CONCLUSION

Facilitated access to restricted chips while China lagged would be rational for any actor tasked with supporting national technological goals. The same access becomes a potential instrument of pressure once a domestic supplier can cover a substantial share of demand. At that threshold, the competition is no longer mainly about which chip is marginally faster on a public benchmark. It is about who controls the compute that trains and runs the models, and who can impose costs on the other's supply.

The recent reporting on one alleged channel does not show that the later stage has been deliberately triggered, and nothing public suggests it was. It does show why the overall sequence (secure access during the gap, build the domestic option under the pressure of controls, then treat disruption of the foreign leader as politically useful even while it still costs China's frontier labs) fits both the industrial record and the logic of a sustained strategic rivalry. The reading that China, dependent on chips it cannot make, loses whenever Nvidia stumbles describes an earlier moment better than the present one.