Jensen Huang Steps Down as NVIDIA CEO, Kress Appointed Interim CEO
On November 12, Jensen Huang resigned from his positions as NVIDIA's President, Chief Executive Officer, and Director, ending a 33-year tenure at the helm, with Executive Vice President and Chief Financial Officer Colette Kress appointed as interim CEO. The article notes that NVIDIA has publicly extended its Blackwell, Rubin, and Feynman product roadmap through 2028, and has opened up interconnect capabilities to MediaTek and other manufacturers through NVLink Fusion; meanwhile, major customers such as Google, Amazon, and Microsoft are accelerating their in-house ASIC development, with Microsoft's Maia 200 already in production. The article also mentions that Kress had previously raised objections to the high-risk investments and financing arrangements pushed by Jensen Huang.
Why read this
SelectedAI score 79
The article reviews NVIDIA's existing arrangements in technology roadmap, interconnect ecosystem, and compute financing after Jensen Huang's departure, as well as the pressure brought by major customers' in-house chip development.
Topics
Body · AI translation
The Jensen Huang era at Nvidia has come to an end.
On November 12, Jensen Huang resigned from his positions as Nvidia's president, chief executive officer, and director, ending a 33-year tenure at the helm. Executive Vice President and Chief Financial Officer Colette Kress took over as interim CEO until a formal successor is determined.
Since Nvidia's founding in 1993, Jensen Huang has been virtually synonymous with the company. Nvidia long employed a highly flat management structure, with roughly 60 subordinates reporting directly to Huang, and the company had no universally recognized second-in-command. Every strategic decision Huang made would set off a chain reaction across the entire AI industry.
This sudden seismic personnel change has confronted Nvidia with an unprecedented test.
For now, Huang has left Nvidia with a moat that will be difficult to surmount in the short term, but that alone is not enough to solve all the problems the company faces.
01 Something Completely Different
According to Vanda Analytics data, on the day Huang announced his resignation, retail investors bought approximately $111 million worth of Nvidia stock within half an hour of the U.S. market open, marking the largest retail buying volume in the stock's opening period on record.
Retail investors' confidence is not hard to understand: "Be greedy when others are fearful." Huang may be gone, but Nvidia's technological and ecosystem advantages remain.
First, Nvidia's technology roadmap for the coming years was already clearly planned.
As early as March 2025, Nvidia publicly disclosed a product roadmap extending through 2028. From Blackwell and Rubin to Feynman, the iteration directions for core technologies such as GPUs and interconnect networks have all been clearly laid out.
This means that, at least in the short term, Nvidia can still advance product development and iteration along its established technology roadmap.
Moreover, Nvidia's moat is not just leadership in chip performance.
On Nvidia's earnings call on August 26 this year, an analyst pressed Huang: OpenAI and Anthropic are accepting NVIDIA's investment and support while developing their own AI chips—how does Nvidia view this kind of competition?
Huang's answer was: "We're building something very different."
He also said he was "100 percent confident" that companies such as OpenAI and Anthropic would continue using Nvidia's computing platform for a long time to come.
A partnership deal in August this year explains Huang's confidence. MediaTek reached a cooperation agreement with Nvidia to use the latter's NVLink Fusion technology to develop custom AI chips for customers and connect them to Nvidia's data center systems. At the same time, Nvidia spent $3.5 billion to purchase convertible bonds issued by MediaTek.
▲As early as 2023, Nvidia and MediaTek announced a partnership in the automotive sector
At first glance, this partnership seems rather subtle. MediaTek is making a major push into the custom AI chip market and was originally coming to compete for Nvidia's GPU business—so why did it instead choose to cooperate?
The reason is that building a single high-performance chip and getting tens of thousands of chips to work together efficiently are two different things.
Large AI data centers need not only chips but also high-speed data exchange between chips and the ability to connect large numbers of computing devices into massive clusters. The industry calls these two processes Scale-up and Scale-out, respectively.
Because for AI data centers, what truly matters is not just the performance of a single chip, but how many computing tasks can be completed and how many tokens can be generated for the same amount of money and the same amount of electricity.
This is precisely where Nvidia's advantage lies. It uses NVLink and NVSwitch to achieve high-speed interconnection between chips, and then uses networking technologies such as InfiniBand and Spectrum-X to organize computing devices into large-scale AI clusters.
NVLink Fusion, launched in 2025, further opened this capability to other chip companies. In addition to MediaTek, companies such as Amazon, Fujitsu, Qualcomm, Arm, and Marvell have also joined this cooperative ecosystem.
Nvidia's calculation is that even if some GPU orders flow to custom chips in the future, as long as those chips still use Nvidia's interconnect and networking systems, it will have a chance to continue profiting from them.
This self-reinforcing cycle is the moat that is harder to replicate than mere chip performance leadership.
These competitive advantages built up over many years will not easily collapse with Huang's departure, but how long they can be maintained in the future is hard to say. Because every brick in this barrier bears the distinct imprint of Huang himself.
02 The Helmsman Is Absent, and Undercurrents Surge Beyond the Moat
Nvidia, having lost its helmsman, now faces a situation of greater complexity than ever before.
Nvidia's biggest customers are precisely the ones most eager to reduce their dependence on it.
Cloud giants such as Google, Amazon, and Microsoft, which purchase large quantities of GPUs from Nvidia every year, are accelerating their own ASIC (application-specific chip) development; AI model companies such as OpenAI and Anthropic have also begun developing their own chips, seeking to control everything from models all the way down to computing power.
In October of this year, Microsoft announced that its in-house chip Maia 200 had officially entered production and begun supporting businesses such as Microsoft Copilot, with plans to expand deployment in 2027. Amazon is also accelerating the deployment of its in-house chip Trainium. According to data disclosed earlier this year, Trainium2 already handles most of the inference tasks on the Amazon Bedrock platform.
For a long time, Nvidia's chips have been in short supply, giving it strong bargaining power—even tech giants have had to step in personally to secure supply.
Oracle founder Larry Ellison once recalled that at a dinner, he and Musk repeatedly begged Jensen Huang to sell them more GPUs, even saying again and again that they hoped Nvidia would charge more. In the end, he added, that dinner did indeed have an effect.
The continuous iteration of AI technology has already let the industry glimpse a possibility of challenging the moat of the CUDA ecosystem.
In November of this year, Anthropic disclosed a technological advance during its IPO roadshow. In the past, when the company onboarded a new chip platform, it often took dozens of engineers several months to port and optimize the underlying compute programs. But with its in-house Claude Mythos Preview model, the team needed only about 3 days to bring a model's prompt-processing task to usable performance on AMD MI355X chips, with nearly all of the relevant compute kernels ported or written by AI.
Compared with model training, inference tasks cover more diverse scenarios. Although Nvidia GPUs have the advantage of generality, for large-scale inference tasks with relatively stable demand, specialized chips can be optimized in a targeted way to complete computation at lower cost.
This means major customers have both the opportunity and the incentive to shift part of their computing demand to self-built platforms, gradually reducing their dependence on Nvidia.
03 Can the "AI central bank" keep loosening the taps?
Therefore, Nvidia not only sells shovels but also has to ensure that the miners have money to buy shovels.
The first step is to personally backstop customers.
Small and medium-sized cloud providers want to get into the AI infrastructure business but dare not stake their entire fortunes on GPUs. Buying equipment is easy, but no one can guarantee whether the computing power can be rented out. Nvidia's proposal is: go ahead and procure boldly and build data centers; if there is idle computing power in the future, I will lease it back or purchase it according to agreement.
CoreWeave is a typical case. Facing market doubts about computing power oversupply and high expenses, Nvidia offered long-term computing power purchase commitments, equivalent to adding a layer of insurance to its future revenue. Cloud providers thus find it easier to raise financing and get money to continue buying GPUs, and Nvidia also secures its orders.
But if the entire industry relies on Nvidia for backstopping, even the richest cash flow cannot sustain it. Hence the second step: bring Wall Street on board as well.
On August 10, 2026, Nvidia's official website formally announced the establishment of a computing power financing platform together with six major Wall Street asset management institutions, including BlackRock, Blackstone, and Goldman Sachs, planning to leverage $500 billion in third-party capital to lend to AI companies/cloud providers to purchase GPUs and build AI data centers.
Nvidia has become nothing less than an "AI central bank": it both sells core assets and provides credit for purchasing those assets. Even more cleverly, this support revolves around Nvidia's computing power platform; technological cooperation combined with capital support also ties Nvidia's interests more closely to those of downstream customers.
Moreover, the operation of this "computing power central bank" cannot be separated from Jensen Huang's own strategic will.
In April of this year, when discussing his investment logic on a podcast, he said that Nvidia does not intend to pick only winners but hopes to support everyone. This philosophy is not hard to understand. For Nvidia, as long as the entire AI industry continues to develop, no matter which company ultimately wins, it has a chance to become a customer of its chips and computing power platform.
What Huang values is not only the financial return of a single investment but also the long-term gains brought by the growth of the entire industry.
But this aggressive investment style is not without controversy inside Nvidia.
According to Reuters, the current interim CEO and former chief financial officer Colette Kress has repeatedly raised objections in the past to the high-risk investments and financing arrangements promoted by Huang, with disagreements between the two sides involving companies such as OpenAI and AI infrastructure company Lambda.
In the past, Huang, by virtue of his founder status and long-accumulated authority, was able to play a leading role in major investments. Now, does his successor still have enough boldness to continue this aggressive investment strategy?
After all, even Cook—the world's strongest professional manager, nominated by Trump as a new Nvidia director—failed to keep Apple leading the technological revolution after taking over Jobs' mantle.
AI Trends