A report alleging Jensen Huang's departure as Nvidia chief executive has placed U.S. restrictions on China-bound AI computing at the center of a potential leadership crisis, claiming that more than 1.5 gigawatts of controlled Nvidia GPUs ultimately reached China despite export controls.
The document says Nvidia's board investigated Huang's conduct, including relationships involving a China-based customer and several subordinates, before he stepped down. It also says Colette Kress, Nvidia's chief financial officer, was named interim chief executive while the company searches for a permanent replacement.
Neither the reported board findings nor the circumstances surrounding Huang's departure were independently detailed in the material provided. The claims, if borne out, would put Nvidia's management, sales practices and role in the global AI infrastructure build-out under unusually intense scrutiny.
Alleged China shipments frame leadership claims
The report's central allegation concerns the scale of restricted AI-chip shipments reaching China. Its tracking model estimated that over 1.5GW of Nvidia GPUs subject to U.S. controls flowed into the country, excluding chips specifically designed to comply with American export rules.
A gigawatt measure refers to the electricity capacity associated with operating large volumes of computing equipment, rather than a chip count. At that scale, the alleged shipments would represent substantial AI data-center capacity, where thousands of high-performance graphics processing units can be linked to train or run advanced models.
The document links the alleged diversion of controlled chips to pressure from U.S. intelligence agencies and lawmakers seeking tighter oversight of China's access to computing power. It presents Huang's exit as occurring amid an increasingly difficult conflict for Nvidia: demand for its hardware remains deeply connected to global cloud infrastructure, while Washington has sought to limit the most capable AI systems available to Chinese entities.
RASA could extend compliance to remote compute
The report identifies the Remote Access Security Act, known as RASA, as a legislative focus. According to the document, the House of Representatives passed the measure on Jan. 12 by a 369-to-22 vote, following unanimous approval by the House Foreign Affairs Committee in April 2025. The bill has since been referred to the Senate Banking Committee, where it has not yet received a vote, the report says.
RASA would seek to address a route around hardware export restrictions: remote access to high-end computing located outside China. Instead of purchasing a controlled chip, a customer can rent processing time from a cloud provider that operates the hardware in another jurisdiction.
That approach has helped create a market for neocloud companies, specialist infrastructure operators that lease AI computing capacity rather than offering the broader range of services associated with large technology platforms. The report says providers in Asia, Australia, the Middle East, Europe, the Nordics, Canada and the United States could face expanded compliance obligations if RASA becomes law.
Such rules could force cloud and data-center operators to determine more precisely who is using their computing capacity, where that customer is based and whether a transaction provides access to controlled AI capabilities. The effects would reach beyond chipmakers, extending to hosting companies, system integrators and businesses that resell GPU capacity.
Succession questions reach engineering and supply contracts
The report describes Nvidia as a company with a highly centralized decision-making structure under Huang. It says Huang had previously stated that roughly 60 executives reported directly to him, despite Nvidia having about 42,000 employees.
That concentration raises practical questions about the company's succession process, particularly in its data-center division, which has become Nvidia's largest source of revenue. The document names Ian Buck, associated with Nvidia's data-center organization, and Jonah Alben, a senior GPU engineering executive, as internal figures discussed for a more technically focused succession.
Kress's reported appointment as interim chief executive would provide continuity in finance and customer relations, but it would not immediately settle longer-term questions about leadership over chip architecture, manufacturing allocation and large infrastructure agreements.
The report points to Nvidia's disclosed commitments linked to AI infrastructure as one reason the succession process could carry broader consequences. It says Nvidia reported $530 billion in off-balance-sheet total guarantees in its fiscal second quarter, up from $184 billion in the previous quarter. That figure included $36 billion in take-or-pay commitments, which require payment for contracted capacity whether or not it is fully used, according to the document. Nvidia also reported a $99 billion equity investment portfolio, it says.
These arrangements can support rapid deployment of AI systems by giving suppliers and infrastructure partners greater certainty around demand. They can also leave a company more exposed if capacity spending slows, counterparties struggle financially or chip supply terms change.
HBM supply could become less favorable
The document also forecasts pressure on Nvidia's privileged position in high-bandwidth memory, or HBM, a specialized memory technology essential for training and serving large AI models. It estimates SK Hynix could supply HBM4 to Nvidia in 2027 for about $3.00 to $3.30 per gigabit, compared with an estimated $3.70 to $4.10 per gigabit for customers such as AMD, Broadcom, Amazon and Google.
The report argues that these price advantages may have relied partly on Huang's direct involvement with suppliers. It projects that more than 70% of relevant supply capacity could be allocated away from Nvidia by 2028 as contracts move toward standard commercial negotiations.
That projection remains an estimate rather than a disclosed supplier allocation plan. Yet it captures a genuine competitive pressure in AI hardware: AMD, hyperscale cloud groups and custom-chip designers are all competing for advanced packaging and HBM supply that has remained constrained.
China's software ecosystem adapts around restrictions
Export restrictions have also encouraged Chinese developers to optimize AI software for domestic accelerators. The report cites open-source projects including MSA, DeepEP and DeepGEMM, alongside work used in inference frameworks such as SGLang and vLLM.
It points to Zhipu AI's February statement that its GLM-5 model supports inference across seven domestic accelerator types. Meituan said in June that LongCat 2.0 had been pre-trained entirely on Huawei Ascend 910B neural processing units, while a DeepSeek V4-related inference kernel for Huawei Ascend hardware was published after the company's April V4 paper, according to the report.
Those efforts do not eliminate China's need for leading-edge foreign hardware, particularly for the largest training workloads. They do show how software optimization, model architecture and domestic accelerator support can reduce dependence on a single chip supplier, even when access to the highest-performing GPUs is limited.
As Nvidia faces export scrutiny, explore how global crypto laws shape investor security and future-proof your compliance strategy.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
