EN / 中文

Jensen Huang’s Exit: The Power Struggle Shaping Nvidia’s AI Future

by kegongliliang·November 19, 2026

On November 18, Nvidia released the final quarterly report of its fiscal year under CEO Jensen Huang, and the performance was still incredibly strong.

In Q3 FY2027, Nvidia's revenue reached $115.8 billion, a year-on-year increase of 103%. Over 90% of this came from the data center business, totaling $107.6 billion, up 110% year-on-year. Net profit reached $65.3 billion, with a gross margin approaching 75%. According to Nvidia's projections, next quarter's revenue is expected to climb further to $130 billion.

Just two years ago, in FY2025, Nvidia's full-year revenue was only $130.5 billion. Riding the AI wave, Nvidia's current single-quarter revenue has nearly matched the full-year level from two years ago.

However, the leader who brought the company to its most profitable moment in history has been explicitly barred from returning.

On November 11, Jensen Huang resigned as CEO, President, and Director of Nvidia. But he clearly wasn't ready to truly leave. Just five hours before the earnings release, he posted his first public statement since resigning on social media platform X, admitting to serious mistakes in his personal life and offering a sincere apology. He then shifted tone, stating that if the board believed his return served the best interests of shareholders, employees, the company, and the United States, he would be honored to resume the role of CEO.

But the Nvidia board did not give him that opportunity. Hours later during the earnings call, new Chairman Harvey Jones left no room for ambiguity. He announced on the spot that the board had unanimously accepted Huang's resignation and had no intention of rehiring him as CEO.

The era belonging to Jensen Huang is largely over, and the meticulously crafted business path he built for Nvidia is also wobbling.

In recent years, Huang repeatedly lobbied Washington, hoping to persuade the US government to relax chip export restrictions to China. This was not due to any special affection for the Chinese market, but because Nvidia's advantage lies not just in chip performance, but in the fact that millions of developers worldwide use its computing platform. Abandoning the Chinese market meant losing a large customer base and potentially nurturing future competitors with one's own hands.

Six days after Huang stepped down, signs of a shift in Nvidia's operating logic emerged. On November 17, Nvidia announced the establishment of a National Security and Export Control Oversight Committee, responsible for supervising the company's export control compliance, national security policies, and global trade risks. The day before, the US Department of Commerce had just issued new regulations to further restrict Chinese enterprises from remotely accessing advanced US AI computing power via overseas data centers.

In the latest quarterly report, Nvidia also narrowed its expectations for the Chinese market. In Q3, the company sold $2.4 billion worth of Hopper architecture data center products to China with export licenses, but deliveries have now ceased. The projected revenue of up to $130 billion for the next quarter does not include income from China's data center computing business.

This controversy, superficially about Huang's personal scandal, is beneath the surface a battle of interests within the US AI industry regarding its future direction. The expansionist route represented by Huang is losing its most influential proponent. Meanwhile, forces wishing to keep advanced computing power and model capabilities firmly in American hands, such as Anthropic, are gaining new policy space.

Who Drove Huang Out?

Many speculate that Anthropic played a hand in Huang's downfall.

On November 10, New York Times reporter Tripp Mickle revealed on the podcast "The Daily" that as early as August 2025, he and his colleagues received a message from a contact in the tech industry. The source suggested they drop their current work and investigate Megaspeed thoroughly.

After Huang's resignation, some speculated on social media that the person who initially tipped off reporters might be linked to Anthropic, even pointing suspicion at Anthropic co-founder Jack Clark.

Another clue is that AnthroPAC, Anthropic's political action committee, provided political donations to some lawmakers involved in questioning Huang.

In April this year, Anthropic established AnthroPAC, beginning to support US congressional members involved in AI policy-making through voluntary employee donations. Some lawmakers demanding Huang's resignation had also received political donations from AnthroPAC or Anthropic employees.

Of course, these remain mere rumors. By comparison, the conflict between the two parties has indeed existed for a long time.

On November 17, Business Insider published an article with a blunt title: "Jensen Huang Was Building Anthropic's Nightmare, Then He Was Out." The report cited a CEO of a US AI startup saying that fierce commercial competition is occurring between Nvidia and top large model labs.

The core of this internal competition among US AI companies is who controls the profits of the AI industry. Anthropic relies on closed-source large models like Claude, charging customers via API calls, enterprise subscriptions, etc. Huang, however, hopes more enterprises can own their own AI models without having to pay minority model companies long-term.

Among the two current development paths for AI, profits from open models flow more to hardware and software vendors like Nvidia, while profits from closed-source models flow more to large model enterprises like Anthropic and OpenAI.

A subtle relationship thus emerges: Nvidia wants as many developers and models in the entire AI industry as possible, while closed-source model companies like Anthropic need to convince clients that the most advanced AI capabilities are still worth purchasing from them.

Nvidia had previously been advancing along the path laid out by Huang: planning to acquire the renowned open AI model community Hugging Face, continuously investing in its own Nemotron open model series, and investing in model development companies like Reflection AI, Thinking Machines Lab, and Poolside.

In the global open model community, Chinese large models have become unavoidable protagonists. Models like Alibaba's Qwen, DeepSeek, and Moonshot AI's Kimi are widely adopted by overseas developers. In the first seven months of this year alone, the Qwen series exceeded 2 billion downloads on Hugging Face, with over 150,000 derivative models.

In the Hugging Face acquisition documents submitted by Nvidia to the US Securities and Exchange Commission in September, a specific risk was listed: Many of the world's most popular open models come from China. If regulatory measures restrict Nvidia from providing relevant products and services to models from various regions including China, it could significantly impact the operational performance of both Hugging Face and Nvidia.

Furthermore, Nvidia's support for selling chips to China would also erode the market space for closed-source model companies like Anthropic. In January this year, at the World Economic Forum in Davos, Anthropic CEO Dario Amodei openly opposed the Trump administration allowing Nvidia to sell H200 chips to China, even comparing such deals to selling nuclear weapons to North Korea.

Huang, meanwhile, repeatedly warned Washington that chip bans were handing the Chinese market to domestic competitors. Last May, he estimated that China's AI chip market size could reach $50 billion that year. For Nvidia, continuing to sell chips to China also meant keeping domestic AI enterprises dependent on CUDA and Nvidia's computing ecosystem.

In September this year, the policy divergence between the two sides became fully public.

According to Axios, Anthropic announced its withdrawal from the Information Technology Industry Council (ITI) because ITI opposed incorporating three chip export restriction bills into the US National Defense Authorization Act. These bills involve strengthening congressional review of sensitive chip export licenses, tracking the flow of AI chips, and pushing allies like Japan and the Netherlands to further restrict semiconductor manufacturing equipment exports to China.

It is worth noting that Nvidia, Google, and OpenAI are all members of ITI.

However, in November 2025, Nvidia announced plans to invest up to $10 billion in Anthropic. By November this year, Anthropic's IPO prospectus still listed Nvidia as one of the investors planned to participate in private placements. This indicates that interest divergences among US AI companies do not prevent them from cooperating when there is money to be made. What the two sides are truly fighting for is pricing power and market dominance in the future AI industry.

Why Must Nvidia Protect the Global Market?

Although Nvidia is currently making a fortune, its revenue is highly dependent on a few large customers.

In Q3, of Nvidia's $107.6 billion in data center revenue, $60.1 billion came from hyperscale customers, primarily tech giants operating large cloud computing and AI infrastructure. The remaining $47.5 billion came from AI cloud services, industrial, and enterprise customers.

Over the past year, the former category grew 71%, while the latter grew 195%. Hyperscale customers still contributed 56% of data center business revenue, but this is lower than approximately 69% in the same period last year.

For Nvidia, this shows its strategy of expanding its customer base is working, but another tricky problem has surfaced: those most capable of buying GPUs are also the most powerful entities building their own GPUs.

Google is a typical example. Latest data shows Google has earned billions of dollars in revenue from selling its self-developed AI chips and plans to further expand deliveries in 2027. Last year, Google also reached a multi-billion dollar TPU computing supply partnership with Anthropic.

Microsoft is also accelerating the deployment of its self-developed chips. In October this year, Microsoft announced an expansion of the deployment scope of its Maia series AI chips. Maia 200 has already begun entering data centers in the US, Europe, and Australia.

On Amazon's side, in August, AWS announced plans to deploy an additional 2 million Nvidia GPUs between 2027 and 2028. Simultaneously, it is advancing its self-developed Trainium chips and preparing to use Nvidia's NVLink Fusion high-speed interconnect technology for the next generation of Trainium.

Purchasing heavily on one hand while developing alternatives on the other. The attitude of several major US tech giants is quite consistent. The competition Nvidia faces is no longer just traditional chipmakers like AMD, but also includes the large customers it nurtured itself.

This is why Huang never wanted to turn Nvidia into a company serving only a few US tech giants.

Nvidia's most core moat is the entire computing system built around GPUs. Over the past 20 years, more than 7 million software developers worldwide have used Nvidia's CUDA platform to write programs, train models, and develop various AI applications. Based on this, Nvidia integrates GPUs, CPUs, high-speed interconnect devices, and software into a unified computing system. The larger this ecosystem becomes, the harder it is for latecomers to challenge it.

But now, domestic US cloud providers are developing their own hardware, and in the Chinese market, export restrictions are forcing more and more customers to seek alternatives to Nvidia.

In the first half of this year, Nvidia's shipments in China's data center AI chip market fell 44% year-on-year, with market share dropping to 34%. Meanwhile, Huawei's Ascend chip shipments reached 572,000 units, accounting for 28% of the share, narrowing the gap with Nvidia significantly.

If more and more customers reduce purchases or even stop using Nvidia chips, its business model will inevitably be impacted.

Facing the dual pressure of customer self-development and maintaining growth, during the November 17 earnings call, UBS analysts directly asked: As large AI companies accelerate the adoption of self-developed chips, how long can Nvidia maintain its high pricing power and profit margins? Subsequently, Goldman Sachs analysts raised another question: How much money is Nvidia prepared to spend supporting customers in building data centers?

These two questions directly relate to how long Nvidia can maintain its current gross margin of over 70% and what price it is willing to pay to sustain future growth. Nvidia's interim CEO Colette Kress did not give a clear long-term commitment. Perhaps internally at Nvidia, there is still no unified answer to these issues.

Conclusion

Regarding Nvidia's future direction, the November 17 earnings call already revealed some signs.

First is chip supply. Facing tightening restrictions from the US government, the new management is clearly becoming more cautious. The board established a dedicated export control oversight committee, stopped delivering previously licensed Hopper data center products to China, and did not include China data center computing business revenue in next quarter's forecast.

But on open models, Nvidia has temporarily not changed Huang's route. Kress explicitly stated during the earnings call that the company will continue to support the open model ecosystem, allowing different countries, enterprises, and industries to develop their own AI. The previously announced Hugging Face acquisition plan remains unchanged, and investment in the Nemotron series continues.

This aligns with Nvidia's commercial interests. No matter which company develops the most popular AI model in the future, as long as these models run on Nvidia's computing platform, it can continue to make money. But if Nvidia wants global developers to use its computing platform while simultaneously losing Chinese customers under US policy constraints, cracks may appear in the global ecosystem it has painstakingly cultivated.

For the domestic AI industry, short-term pain is inevitable. Domestic GPUs still need time to catch up in terms of performance, software ecosystem, and large-scale cluster applications, and domestic model iterations will encounter bottlenecks. But as more and more enterprises begin using domestic computing power, chip manufacturers have more opportunities to accumulate experience and improve products in real business scenarios.

In the past, even if domestic enterprises wanted to use domestic chips, it was difficult to completely break free from the attraction of Nvidia's mature ecosystem. Now, the increasingly narrow channels are causing orders, developers, and application demands that might have flowed to Nvidia to gradually shift toward domestic vendors.

This is precisely the outcome Huang worried about most in recent years. He repeatedly reminded Washington not to help China cultivate its own competitors in the name of restricting China.

Currently, Nvidia can still maintain high growth relying on orders from the US and other markets, temporarily avoiding too great a cost for losing the Chinese market. But competition in the chip market has never been judged by just one or two quarters. As domestic enterprises gradually build their own computing platforms and developers get accustomed to new software ecosystems, Nvidia's proud core competitiveness may also be weakened bit by bit in the process of losing the Chinese market.