After Jensen Huang: China’s AI Breakthroughs and the Battle for Control of Computing Power

Beijing is building, not merely buying

The contradiction at the centre of Jensen Huang’s downfall is that Washington has been trying to prevent China from obtaining advanced American processors just as Beijing has been discouraging Chinese companies from relying on them.

For the United States, export controls are intended to preserve an advantage in computing power. For China, dependence on American chips creates a vulnerability that Washington can exploit. Both governments, from opposing positions, have come to regard the same technology as a strategic resource.

China’s response is no longer simply to acquire Nvidia products through whatever channels remain open. It is developing a competing system of processors, software, AI models and industrial applications.

The policy is explicit. The State Council’s August 2025 AI Plus programme called for the integration of artificial intelligence into manufacturing, agriculture, scientific research and public administration. An industrial plan issued by eight departments in January 2026 sought reliable supplies of essential AI technologies by 2027, alongside 100 industrial datasets and 500 representative application scenarios.

This is not merely a race to produce the most impressive chatbot. Beijing wants artificial intelligence to improve factory productivity, industrial design, transport and research, with domestic suppliers providing the infrastructure. The programme forms part of a wider effort to make technological progress less vulnerable to foreign restrictions.

The results are substantial but uneven. China has a huge industrial market, engineering talent and state support; it still faces obstacles in advanced chip manufacturing, high-bandwidth memory and equipment. Political pressure to buy domestic products may also impose costs on companies that could perform better with imported hardware.

Huang understood those trade-offs. His argument for continued American sales to China was that excluding Nvidia would encourage customers to adopt Chinese alternatives. Beijing’s calculation was the reverse: using Nvidia might deliver better performance today but weaken China’s independence tomorrow.

Huawei’s chips and China’s open-weight models

Huawei is trying to compete with Nvidia not just chip by chip but across entire computing systems. In September 2026 it announced the Atlas 960E SuperPoD, designed to connect up to 4,096 neural processing units, and outlined a roadmap for Ascend 960 processors in 2027.

The significance lies in the links between processors. Training advanced AI models depends on moving information rapidly across enormous computing clusters. Nvidia’s strength includes its networking technology and the CUDA software environment, which is deeply embedded in research and commercial development. Huawei is building alternatives, including its CANN software ecosystem.

These announcements are not proof that Huawei has matched Nvidia at commercial scale. Manufacturing yields, memory supplies, software compatibility and energy efficiency remain demanding challenges. But China is increasingly attempting to replace the system around the chip, rather than the chip alone.

Its model developers are making a parallel challenge. DeepSeek, Alibaba’s Qwen team, Moonshot AI, Zhipu and MiniMax have released powerful models, including systems with publicly available weights. Moonshot’s Kimi K3 and Alibaba’s Qwen releases during 2026 illustrated the scale of that competition.

Open-weight does not necessarily mean fully open-source: training data and methods may remain undisclosed. Nor does a large parameter count guarantee a superior model. Nevertheless, downloadable models let universities, businesses and governments operate and adapt AI without purchasing every interaction from a proprietary provider.

The commercial contrast is significant. OpenAI and Anthropic rely heavily on selling access to controlled systems. Chinese developers can distribute model weights while earning money from hosting, customisation and services. Free weights do not eliminate the expense of running large models, but they can reduce dependence on any single supplier.

For countries outside the two superpowers, this may matter as much as benchmark rankings. A model that can be operated locally offers greater control over costs, sensitive data and national requirements.

The relationship between chips and models is important. An open-weight system can, in principle, be adapted for different hardware, including Chinese processors. That does not make migration automatic: software libraries, memory requirements and optimisation still matter. But it weakens the lock-in created when both the model and the computing platform must be purchased from one commercial ecosystem. China’s approach may thus compete with Silicon Valley on two fronts at once: the machinery that performs computation and the software that gives the machinery value.

The chips waiting in Hong Kong

Early-November reporting described approximately 600,000 Nvidia H200 processors held in Hong Kong while import approval into mainland China remained uncertain. Chinese AI companies wanted the processors, yet Beijing was reluctant to authorise purchases on the desired scale. Limited batches had reportedly entered the mainland.

Nvidia disclosed in its July-quarter regulatory filing that Chinese restrictions had frustrated sales permitted under American export licences. It also recognised that policies encouraging domestic alternatives threatened its Chinese market.

The dilemma is practical. Chinese developers competing in rapidly advancing AI cannot always wait for local hardware to mature. Their government, however, does not want a new generation of companies locked permanently into Nvidia’s products.

Some companies have therefore sought computing power abroad. Reuters reported in November that Chinese developers were seeking additional access to Nvidia-powered servers in overseas data centres, including in Southeast Asia. Renting computing capacity is different from importing processors: the hardware may never enter China, even if a Chinese customer uses it remotely.

That distinction is reshaping American policy. The Remote Access Security Act, passed by the US House in January 2026 but not enacted by 14 November, would extend export-control authority over certain remote computing services. Washington fears that access to foreign-based servers can defeat restrictions on physical chip shipments.

For Singapore and Malaysia, the issue is more complicated. Both have legitimate ambitions to host international data centres. Neither wants illegal diversion, but neither automatically treats every American restriction as a prohibition under its own laws. Their sovereignty, commercial interests and role as technology hubs are now caught between the two powers.

This distinction also complicates enforcement. An investigator may establish where a server is installed without knowing who ultimately uses it. Cloud services can be resold, accessed through intermediaries or shared among customers. Regulators therefore need evidence about contracts, access rights and actual end uses, not only customs records. Expanding controls to remote access may answer one legal question while creating difficult new questions about jurisdiction and compliance.

The Megaspeed money trail

The controversy involving Huang cannot be understood without the financial network around Megaspeed International, a Singapore-based buyer of Nvidia-powered computing equipment, and its Malaysian affiliate Speedmatrix.

On 13 May, Culper Research published *NVIDIA: The China Problem*, alleging that much more Nvidia computing revenue depended on Chinese demand than direct sales figures indicated. Culper held a short position in Nvidia shares and stood to profit from a fall in the price. Its estimates therefore require independent scrutiny.

Yet the records it highlighted raised specific questions. Culper reported that Megaspeed’s liabilities and equity had expanded from roughly $33 million at the end of 2023 to about $3 billion a year later, driven largely by approximately $2.9 billion in refundable deposits. It also alleged that Speedmatrix imported around $4.6 billion in equipment between December 2024 and January 2026.

Of particular interest was financing involving Apex Enterprise Solutions, a Singapore company linked to Alibaba. Subsequent Reuters reporting confirmed an Alibaba-linked financing arrangement involving Speedmatrix. That corroborates the existence of the connection; it does not prove that Alibaba financed smuggling or that Nvidia knowingly assisted unlawful transfers.

The questions are where equipment was installed, who financed it and who controlled the computing services. Ownership, financing, physical shipment and remote use are different matters. A Malaysian data centre can legitimately host American servers and sell their computing capacity to customers abroad. Illegality depends on the actual transactions, applicable restrictions and any false declarations—not simply on a Chinese connection.

Separately, American prosecutors charged individuals associated with Super Micro Computer in March over an alleged $2.5 billion server-diversion scheme. That prosecution shows the seriousness of suspected export-control violations, but charges in one case cannot establish wrongdoing by every company in the regional computing trade.

What Singapore and Malaysia established

Singapore’s Ministry of Home Affairs clarified the position on 12 November. Its police and customs authorities had independently begun investigating Megaspeed and related parties in February 2025 over possible false declarations about servers. The inquiry was not launched at Washington’s request.

Alice Huang, also known as Huang Le, was investigated in connection with her management responsibilities at Megaspeed—not for espionage. Police questioned her in June 2025 and held her passport during their inquiry. In February 2026 they concluded there was insufficient evidence to establish offences relating to the suspected commercial declarations. They nevertheless assessed that she had knowingly provided inaccurate information to investigators. She received a stern warning, was barred from returning to Singapore and left the country.

Investigations into Megaspeed and other individuals remained open. Singapore also said it had received limited assistance despite requests for American information relevant to server movements, including matters involving Supermicro.

Those findings neither clear all the companies nor substantiate American intelligence claims about Alice Huang. They concern different questions and different legal standards.

Malaysia was equally careful. On 13 November its trade ministry said it had never cleared Megaspeed or Speedmatrix of possible wrongdoing. Speedmatrix operated as a tenant at Malaysian data centres and brought in its own equipment. Importing servers did not itself require a permit, but movements of certain advanced American chips out of or through Malaysia have required authorisation since July 2025.

Prime Minister Anwar Ibrahim said investigators had not then found indications that the relevant chips had been exported from Malaysia. That was a description of evidence available at the time, not an assurance that no diversion could ever be established.

The difference between these official findings and broader claims is not a technicality. A false customs declaration, an unauthorised re-export, prohibited remote access and espionage are distinct acts requiring distinct evidence. Even proof of one would not automatically establish the others. Collapsing them into a single story of Chinese wrongdoing would obscure what investigators have actually established and expose a newspaper to legitimate criticism.

Bloomberg meanwhile reported documents linking restricted Nvidia-powered servers to a data centre in Shanghai through a company associated with Alice Huang. The precise shipping route, licences and ownership arrangements still require verification. A Shanghai connection does not, without more, establish that equipment was illegally re-exported from Malaysia or that Huang acted for Chinese intelligence.

Intelligence allegations and Huang’s resignation

The personal crisis broke into public view on 29 October when the New York Times reported suspicions that Alice Huang had links to China’s Ministry of State Security and had developed a relationship with Jensen Huang. Subsequent reporting described intelligence assessments and possible access to sensitive information. Beijing rejected the allegations; the underlying classified material has not been made public.

There is a clear distinction between an intelligence suspicion, a documented relationship and proof of espionage. Reuters has said it could not independently verify whether Alice Huang worked for the Chinese government. No public finding has established that Jensen Huang knowingly assisted Chinese intelligence.

On 8 November, Huang acknowledged marital infidelity. Separate allegations about relationships inside Nvidia raised questions about his conduct and disclosures as chief executive. The board commissioned an independent investigation and said it had no reason to believe company secrets or national-security information had been compromised.

Huang resigned as chief executive, president and director on 11 November; the change was announced the next day. Chief financial officer Colette Kress took over as interim chief executive and Harvey Jones became chair.

Reuters subsequently reported that Huang had misrepresented his relationship with Alice Huang to senior Nvidia colleagues. If that account is borne out, it presents a corporate-governance issue regardless of whether any espionage occurred. Directors must be told about relationships that may create material conflicts of interest.

Huang’s departure is therefore not proof of an intelligence offence. It arose amid overlapping concerns about personal judgement, disclosure, a major customer under investigation and the political sensitivity of Nvidia’s business with China.

A struggle inside Washington

The timing of the revelations increased suspicion about political motives. Huang had been an influential advocate of controlled semiconductor trade with China and had developed access to President Donald Trump. His fall strengthened those favouring tighter restrictions.

That does not prove the disclosures were engineered. The first major report appeared before the 3 November midterm elections, and no public evidence identifies an intelligence agency, rival company or political faction as the organiser of a campaign to remove him. The distinction between who benefited and who caused an event is indispensable.

What can be demonstrated is a fierce policy disagreement. On 13 November, Treasury Secretary Scott Bessent criticised the Commerce Department for not placing Megaspeed on a restrictive export list. He said Treasury, State, the Pentagon and the US Trade Representative had supported action for months. Commerce Secretary Howard Lutnick controlled the relevant decision.

Questions also arose about Nvidia’s contacts with Commerce over Megaspeed. Such contact is not inherently improper; companies routinely discuss investigations affecting customers. Improper intervention would require evidence about the communications and the decisions they influenced.

Senators Elizabeth Warren and Chris Coons sought preservation of records concerning the intelligence warnings, export licences and dealings with Nvidia. Their questions are the right ones: when were officials warned, what protective steps were taken, why were proposed restrictions delayed and did commercial interests affect national-security decisions?

The answers may expose weaknesses in American export enforcement. They may also show that some intelligence accusations have been stated more confidently than the public evidence warrants. Neither possibility should be ruled out in advance.

The financial system behind the AI boom

Nvidia’s power rests not only on technology but on its position within the enormous financing of AI infrastructure. The company sells processors, invests in customers and sometimes commits to buying computing services from them.

In January 2026, Nvidia announced a $2 billion investment in CoreWeave, alongside plans for extensive new AI infrastructure. A separate arrangement, initially valued at approximately $6.3 billion, committed Nvidia to purchase certain unsold CoreWeave capacity under specified conditions.

These agreements can help finance expensive data centres and give investors confidence that computing capacity will be used. They can also spread risk across interconnected companies. If demand disappoints, Nvidia may face weaker chip sales while holding investments in customers and obligations linked to capacity purchases.

That does not establish accounting fraud or fictitious demand. Nor does every dollar Nvidia invests return automatically as revenue. But investors should examine how much infrastructure expansion is supported by independent end-user demand and how much by financing and purchasing commitments within the industry itself.

CoreWeave’s November forecast placed its 2026 revenue at about $13 billion to $13.3 billion against capital expenditure of $37 billion to $39 billion. Such spending may be justified by future returns, but it leaves the business exposed to slower demand, financing costs and rapidly changing technology.

Huang’s departure leaves these questions to Nvidia’s new leadership. They are central to assessing whether the AI boom is building durable productive capacity or taking on commitments that growth may struggle to support.

The comparison with earlier infrastructure booms is useful only up to a point. Heavy spending ahead of revenue can finance genuinely transformative industries. It can also leave overcapacity when forecasts prove too optimistic. What matters is the eventual utilisation of facilities, pricing power, operating costs and the quality of customers’ demand. Those measurable questions are more illuminating than declaring the entire AI industry either a bubble or an inevitable triumph.

Who will control computing power?

The contest extends beyond Nvidia. Its advanced processors rely on Taiwan’s TSMC, memory suppliers in South Korea and elsewhere, and specialised equipment including ASML’s lithography machines. America leads in important parts of design and software but does not control every part of production.

Washington is attempting to coordinate allied supply chains through initiatives such as Pax Silica, while China is building domestic substitutes and making competitive models available internationally. Neither country can instantly sever itself from the other’s technological and commercial networks.

The stakes are rising because AI increasingly assists the research needed to improve future AI. Anthropic reported that Claude played a leading role in roughly 26 per cent of measured AI research and development tasks in August 2026, while acknowledging methodological limits and no fully autonomous activity in the categories measured. This does not establish that machines can independently create unlimited successors. It does show why access to processors, electricity and research infrastructure may become even more consequential.

America has leading private laboratories, advanced chip design and formidable financial resources. China has a vast industrial base, coordinated investment and a growing open-weight model ecosystem. Both face weaknesses: the American system risks concentrated corporate power and enormous capital exposure; China risks inefficiency, manufacturing bottlenecks and excessive political direction.

Huang’s resignation does not decide which approach will prevail. Nor do unproven espionage allegations resolve whether chips were unlawfully diverted or American export controls were properly administered.

The enduring issue is broader. Governments are trying to restrict access to computing power while corporations finance and control much of the infrastructure that produces it. Chinese competitors are building alternative chips, software and models, potentially giving other countries more choices.

For China, the strategic answer increasingly lies in building rather than buying. For Washington, the question is whether denying access can preserve an advantage when restrictions also encourage substitutes. For everyone else, the question is whether artificial intelligence will remain concentrated in a handful of corporations and states or become accessible enough for others to shape its future.

The fall of Jensen Huang has exposed that struggle. It has not settled it.