← Back to Analysis
·Jon Kelly

The Nvidia crisis is a test of who controls AI

Downloadable models can make AI cheaper and more widely used. They also give companies more choice over whose chips and clouds they pay for. A spy scandal has now removed Nvidia’s chief executive, erased hundreds of billions of dollars of value in days, and turned the company’s succession into a Washington power struggle.

Downloadable models can make AI cheaper and more widely used. They also give companies more choice over whose chips and clouds they pay for. A spy scandal has now removed Nvidia’s chief executive, erased hundreds of billions of dollars of value in days, and turned the company’s succession into a Washington power struggle.

EuroOilWatch · 18 November 2026 · Analysis after Nvidia’s earnings

In July, when the Chinese AI model Kimi rattled chip stocks, Jensen Huang argued that Wall Street had the economics backwards. Give more people access to capable AI, he told Axios, and they would buy more computing power. “Free AI should be great for hardware,” he said. Four months later, Huang has resigned as Nvidia’s chief executive over a scandal involving an alleged Chinese spy, the shares are in their longest weekly losing streak since 2012, and Washington is tightening access to American AI chips while openly questioning whether Nvidia followed its own export rules. Huang’s Axios interview · Nvidia’s leadership announcement · NBC News on the resignation

The question Huang raised now has a second half. If cheaper AI drives up demand for computing, how much of that demand will Nvidia get — and who will be running the company that sells it?

The scandal that removed Jensen Huang

The crisis unfolded over two years, then over one weekend. A short timeline:

  • June 2024 — Huang is photographed with Alice Huang (Chinese name Huang Le), a Chinese-born executive at Megaspeed, a Singapore-based data-centre firm. The two are not related; they are seen together again in May 2025, travelling in the same car.
  • 2024 — the Commerce Department’s Bureau of Industry and Security begins investigating Megaspeed and Alice Huang over the alleged diversion of Nvidia chips; in September an official visits Megaspeed’s Malaysian site and finds Nvidia equipment still boxed. The New York Times later reported that in August 2025 Nvidia’s general counsel sought a meeting on the case; the inquiry was halted and moved to the Justice Department, which opened a criminal investigation into Megaspeed.
  • October 2025 — President Trump is briefed on the matter and chooses not to act, according to the Times.
  • Spring 2026 — Secretary of State Marco Rubio oversees the drafting of a highly classified memo warning about Huang’s relationship. It circulates among officials with top clearances before Trump’s visit to China in May — a visit Huang joins.
  • 29 October 2026 — the Times reports that Huang was seen as the target of a Chinese intelligence plot, and that Alice Huang had allegedly been recruited by the Ministry of State Security to pass information about him.
  • 5–8 November — House committee chairs request a closed-door briefing with Huang. Nvidia’s board announces an independent investigation into his conduct, and Huang admits marital infidelity in an apology to employees.
  • 9 November — the shares fall as much as 9% before closing down 4.3%, and investors file a class action accusing Nvidia of concealing the relationship. The CIA director and the director of national intelligence brief congressional leaders; Trump says the FBI is examining the affair; China’s foreign ministry says it is “nothing but smears against China”.
  • 12 November — Huang resigns, after reporting also linked him to a second affair with a subordinate executive. The board names chief financial officer Colette Kress as interim chief executive.

The Times on Rubio’s memo · Reuters on Nvidia’s probe · Nvidia’s 8 November statement · A Wikipedia timeline of the case

The company at the centre of it

Megaspeed is where the espionage allegations and the chip-export story meet. The firm began life in 2015 as 7Road International, a Singapore-domiciled arm of a Chinese video-game investment group, was renamed Megaspeed International in 2023, and was sold to a Singaporean owner in late 2023 — becoming an Nvidia Cloud Partner four days after the sale closed. It grew into one of Nvidia’s largest Southeast Asian partners just as Chinese buyers used Singapore to route around American chip controls. According to The Information, Alice Huang helped move at least $7.5 billion of Nvidia chips to Chinese buyers through Megaspeed, its Malaysian subsidiary Speedmatrix and other channels; the New York Times had earlier reported that Megaspeed bought more than $4.8 billion of Nvidia chips in two years. The company insists it is a “legitimate and lawful business”. Singapore says it investigated Alice Huang over possible false declarations, not espionage, found insufficient evidence that she had committed an offence, then warned her and barred her from re-entry. It says its probe into Megaspeed and six others continues, and that it does not enforce other countries’ export controls — while acting firmly against fraud. It has agreed to keep discussing the case with Washington. China’s government calls the spying story “smears against China”. BBC on Megaspeed and Alice Huang · Singapore’s Home Affairs Ministry, via Mothership · Guardian live coverage

Rubio’s memo and Bessent’s ultimatum

The political fight runs through two cabinet officers. Rubio’s spring 2026 memo put the administration on notice months before the storm broke; Trump had been briefed as early as October 2025 and declined to act. On 11 November, senators Elizabeth Warren and Chris Coons asked Rubio — who holds both the State Department and national security adviser portfolios — along with Commerce Secretary Howard Lutnick and chief of staff Stephen Miller to preserve all records concerning Nvidia chip exports and the reported intelligence operation, and to brief Congress. Nvidia’s board, for its part, has established a committee to oversee national security and export controls. Senate Banking Committee · Nvidia’s board announcement

No one is applying more pressure than Treasury Secretary Scott Bessent. On Sunday he told Fox News that Nvidia needs a “dramatic shakeup”, that “these shell companies were obvious”, and asked whether the company had pressed “these seemingly illicit sales to goose their revenues and their earnings”. “If this were a bank and they hadn’t complied like this,” he said, “the OCC would be in there right now … a lot of the executives would be removed.” He called Huang “an incredible innovator” but said that did not excuse the company from its export-rule obligations, urged the board to hire the next chief executive from outside the company, and wants answers from the board and its outside counsel within ten days on how Nvidia handled US export restrictions. Two days earlier, on CNBC, he had disclosed that Treasury, War, USTR and State had favoured putting Megaspeed on the entity list “for months … But Commerce has the pen.” Fortune on the fight over Huang’s return · Yahoo Finance on Bessent’s “dramatic shakeup” · Yahoo Finance on the ten-day deadline

The huge drop in stock value, in numbers

The market reaction is the largest in Nvidia’s history. When the board’s investigation became public on 9 November, the shares fell as much as 9% before closing down 4.3%. On 12 November, resignation day, they dropped 4.3% after falling as much as 14% at the open. In the week ending Friday 13 November the stock lost 13% — its sixth consecutive weekly decline and its longest losing streak since 2012 — marking, by The Wall Street Journal’s count, the largest weekly loss of market value on record, and taking Nvidia’s market capitalisation below $4 trillion. CNBC · WSJ live coverage · Reuters

The damage has two layers. Roughly $1.7 trillion has gone since the broader AI sell-off began in early October; the scandal itself accounts for about $600 billion of that. On Monday the shares closed at $156, about 16% below their 6 November close (OilWatch calculation from Yahoo Finance data). Fortune reported that a short bet of about $8 million had netted nearly $200 million; retail investors, by contrast, have been buying the dip at record scale. Guardian live coverage · Fortune · Yahoo Finance on retail buying

D.A. Davidson’s Gil Luria cut his price target from $300 to $200 while keeping a buy rating, saying “the market has appropriately adjusted for this transition”; Oppenheimer said “we expect fundamentals to take a backseat near term”. Nvidia reported on Tuesday evening, and the numbers did not look like a company in crisis: revenue of $115.8 billion, up 103% and about $7 billion above forecasts, with data-centre revenue of $107.6 billion. It guided to $130 billion next quarter, a period that includes an extra week, still assuming no data-centre revenue from China. Kress presented them as interim chief executive — “Our momentum is real,” she said — and the shares rose about 4% after hours. Anthropic is expected to begin trading on Thursday. CNBC · Nvidia’s third-quarter results · CBS News on the board

Can Huang come back?

Huang’s allies have formed a “war room” to restore him, and an open letter had gathered more than 165,000 signatures by Monday — including Elon Musk, Epic Games’ Tim Sweeney and Uber co-founder Travis Kalanick, plus about 7,500 people identifying themselves as current Nvidia staff — roughly 18% of the workforce, by the campaign’s own unverified count. Trump indicated on Monday that he would not mind Huang being reinstated. Against that stand Bessent’s opposition, a request from three House committee chairs for a closed-door briefing, and an independent investigation that has barely begun. UBS estimates that if Kress scales back Nvidia’s investing, more than $50 billion could be removed from the market. On Tuesday evening the board gave its answer: its chair, Harvey Jones, said it “does not intend to reinstate Jensen as CEO”, and the search covers internal and external candidates. Fortune · Guardian live coverage · CBS News on the board

What changes when you can download the model?

With a closed model, you generally use it through a provider’s app or API. The provider supplies the computing power and runs the service. An open-weight model makes the trained parameters available to download. You can choose a hosting provider, run it on your own infrastructure, or adapt it within the terms of its licence. Running it still costs money and requires engineering; the Open Source Initiative’s definition goes further and also requires code and information about the training data. Open Source Initiative definition

The distinction cuts across national lines. OpenAI released its downloadable gpt-oss models, and Nvidia publishes its own Nemotron models. This week, Alibaba unveiled Qwen4-Max as a hosted flagship, keeping its weights closed, and said smaller Qwen4 models would get downloadable weights, with no date given. One company can sell access to its most capable model through a hosted service while releasing other models for wider adoption. OpenAI’s gpt-oss release · Nvidia’s Nemotron releases · Alibaba Cloud’s release record · Reuters on Alibaba’s launch · TNW on the open-weight plans

Why cheaper AI can still be good for Nvidia

Nvidia’s shares fell about 17% on 27 January 2025, a week after DeepSeek released its downloadable R1 model; the company recovered by arguing that the spread of capable models would increase demand for inference — the computing required each time a trained model does a job. The distinction matters. Downloading a model lets an organisation skip the bill for training it from scratch. It still pays every time that model processes an invoice, reviews a contract or answers a customer. If falling costs bring enough new work into production, chip demand can rise even as the computing required for a given task falls. This autumn’s crash is a different kind of shock — a governance and export-control crisis, not a demand scare — but the underlying question is the same: who captures the work. DeepSeek’s R1 release · Reuters on the sell-off and Nvidia’s response

So far, open models have spread alongside booming Nvidia sales. Data-centre revenue reached $89 billion in the quarter reported in August, up 117% from a year earlier, and guidance for the following quarter assumed zero data-centre compute revenue from China — some perspective on the immediate sales exposure. The quarter reported on Tuesday took data-centre revenue to $107.6 billion, up 110%, on the same zero-China assumption. The longer-term issue is which hardware new AI services are built around. Nvidia’s August results and guidance · Nvidia’s November results

Washington can restrict the chips. Developers can choose where models run.

The Commerce Department issued a rule this week aimed at restricting remote access to advanced US chips through data centres in third countries, and said the rule was “written by and at the direction of the Department of Treasury and the National Security Council”. Such restrictions make it harder for Chinese developers to obtain computing power for their next model. They also give Chinese operators a reason to invest in making existing models work on domestic chips. Huawei said in April that its Ascend 950-based systems would support DeepSeek V4 — a statement of intended compatibility that leaves performance, cost and deployment scale open. US Bureau of Industry and Security · Reuters on Huawei and DeepSeek

In the US, hardware competition is coming from the biggest cloud providers as well. Amazon already uses its own Trainium chips for Claude with Anthropic, and Google’s Ironwood is its first TPU designed specifically for inference. The scandal gives both efforts an unexpected tailwind: every question about Nvidia’s governance is an argument for its customers to keep a second supplier. AWS on Trainium and Anthropic · Google on its Ironwood TPU

The power demand can outlast the stock-market story

The IEA’s 2026 central projection puts global data-centre electricity use at 950 TWh in 2030, up from 485 TWh in 2025. The location of that demand matters to utilities and industrial power buyers. If more AI runs on hardware owned by cloud providers, enterprises or regional operators, the next wave of demand can land on different grids and under different power contracts. A more competitive chip market can coexist with a scramble for grid connections. IEA energy and AI outlook

Huang’s July argument may still prove right. Cheaper AI can create more work for chips. His departure has put Nvidia in the spotlight just as companies and governments are making choices about where that work will run — and whether the company collecting that demand is led by Kress, an insider promoted by the board, or an outside successor of the kind Bessent is pressing for is now an open question. For Nvidia, a growing AI market and a growing share of it are two separate bets.


Weekly Fuel Security Briefing

Every Thursday: reserve status changes, price movements, and supply-risk signals across all 27 EU countries — in one concise email.

Read by logistics operators, procurement teams, and energy analysts.