The three major challenges Jensen Huang left for NVIDIA

In future NVIDIA GTC conferences, the familiar scene of "Jensen Huang" stepping onto the stage and waving to the entire audience may never be seen again.

On November 12 local time in the United States, Jensen Huang announced that he had left NVIDIA, the company he had led for 33 years. NVIDIA's stock price plunged by about 14% at one point during the intraday trading, then narrowed its decline, and finally closed down by about 4.3%.

Martin Casado, General Partner of Andreessen Horowitz, called it "a dark day".

Since Jensen Huang founded NVIDIA in 1993, he has almost led every major transformation of the company. From early graphics processing chips, to the CUDA computing ecosystem, and to today's AI data center business, NVIDIA has been able to seize several rounds of technological changes, which is closely related to Jensen Huang's judgment on technology direction and long-term investment.

Especially in this round of generative AI wave, Jensen Huang not only promoted NVIDIA's transformation from a chip supplier to an AI infrastructure platform, but also deeply participated in product roadmap formulation, supply chain coordination and large customer cooperation.

NVIDIA's current market position is largely built on this set of business strategies.

For NVIDIA, Jensen Huang's departure does not mean that the existing business will be immediately impacted. After years of development, the company already has a mature technical team, product roadmap and supply chain system, and these advantages will not disappear with the departure of a manager.

However, after losing Jensen Huang, the core decision-maker, can NVIDIA continue the rapid growth of the past few years? What new tests will this company, which is at the core of the global AI industry, face next?

01 The successor is not equal to the helmsman

At present, Colette Kress, who temporarily takes over as CEO of NVIDIA, has served as CFO for more than 13 years. She has participated in the whole process of NVIDIA's transformation from a gaming graphics card company to an AI infrastructure giant, and is familiar with the company's financial status, customer contracts, supplier arrangements and capital expenditures.

She has also attended earnings calls with Jensen Huang for a long time, and is a management team familiar to Wall Street.

John Vinh, an analyst at KeyBanc Capital Markets, believes that Kress's appointment as interim CEO is a positive signal, because she has a long tenure and in-depth understanding of NVIDIA's business and strategy. Josh Gilbert, lead analyst at eToro, also believes that Kress is the obvious candidate to stabilize the company. She is familiar with the business and can speak on behalf of the company when the market needs management to provide certainty.

However, this does not mean that Kress can immediately take over Jensen Huang's original decision-making responsibilities. According to a report by The Information, NVIDIA's interim CEO is not the leader of the company's major decisions, nor is she responsible for setting the agenda for the company's major decisions. The interim CEO is a completely different role from the long-term helmsman.

Previously, The Information, citing people familiar with the matter, reported that Kress is not the main promoter of NVIDIA's major decision-making agenda, and she is more inclined to control risks on some high-risk transactions, and the OpenAI credit support arrangement is an example.

Jensen Huang once promoted NVIDIA to provide large-scale credit support for OpenAI, and the initially discussed scale reached 250 billion US dollars to support its data center construction. Kress was cautious about this. Subsequently, the relevant arrangement was split, and the first part of NVIDIA's support was about 100 billion US dollars.

Stacy Rasgon, an analyst at Bernstein, believes that Kress's appointment helps stabilize short-term business, and he does not worry that this management change will immediately affect NVIDIA's operation. But he is more cautious about long-term issues: there is still great uncertainty about whether the new CEO can take over Jensen Huang's control over the product roadmap, as well as his network of relationships with customers, suppliers, employees and the entire industrial ecosystem. Rasgon therefore downgraded NVIDIA's stock rating and target price.

James Schneider, an analyst at Goldman Sachs, also pointed out that NVIDIA is not short of a mature functional management team, and there are experienced leaders in chip design, software, systems, finance and other fields. However, the technical vision, industry influence and cross-enterprise coordination capabilities that Jensen Huang possesses are difficult to be directly taken over by an ordinary professional manager.

Among the current internal candidates, the most concerned one is Ian Buck, a senior technical executive at NVIDIA. He joined the company in 2004, has long participated in the construction of GPU general-purpose computing and CUDA ecosystem, and is currently responsible for AI hardware and software products for hyperscale cloud vendors.

Among external candidates, Rene Haas, CEO of Arm, has been mentioned by many analysts. He once served as the head of the computing product business at NVIDIA, joined Arm in 2013, and later became CEO. In addition, Tim Cook, Executive Chairman of Apple, Johny Srouji, Head of Apple Hardware, and Nikesh Arora, CEO of Palo Alto Networks, were also named.

Investor Armina Rosenberg put forward a possibility: NVIDIA can consider letting Kress be responsible for operations and finance, and Buck be responsible for products and technology, forming a co-CEO-like management arrangement.

After all, according to reports, Jensen Huang used to directly manage more than 60 executives, and NVIDIA's decision-making system has long revolved around himself.

02 Can CUDA still hold its ground?

Looking back at NVIDIA's development history, one of Jensen Huang's most important roles is to promote the company to continue investing in technology directions that have not yet generated clear commercial returns.

The most representative case is CUDA.

In 2006, NVIDIA launched CUDA, enabling GPUs to handle general-purpose computing tasks beyond graphics rendering. In 2012, AlexNet trained with NVIDIA GPUs achieved a breakthrough in the ImageNet competition, and GPUs thus became an important computing platform for deep learning. Since then, NVIDIA has continued to improve the CUDA toolchain and computing libraries, and developers have gradually built model training, optimization and deployment processes around this system.

This is also the reason why competitors cannot catch up with NVIDIA simply by relying on chip performance. Even if enterprises purchase GPUs or dedicated chips from other manufacturers, they need to re-adapt existing codes, models and software tools. For customers who have invested a lot of resources in using CUDA, migration is not simply replacing hardware.

However, AI computing methods are changing, and CUDA must also expand accordingly. In the past few years, NVIDIA's advantages were mainly reflected in large model training; as the industry shifts to large-scale inference, customers are paying more attention to the computing cost per Token, GPU utilization and response speed. NVIDIA launched Dynamo 1.0 in 2026, which is specially designed for resource scheduling and request management for large-scale AI inference, trying to extend the competition from GPU performance to the entire inference system. It is also integrated with software such as TensorRT-LLM, and supports mainstream open source inference frameworks.

AI agents and robots have opened up new application space. Agents need to continuously call models, data and external tools, and robots need to process vision, spatial understanding and real-time control. NVIDIA is extending its underlying computing power to these scenarios through CUDA-X, the Isaac robot platform, the Omniverse simulation environment and the Cosmos model. In the future, the scope of CUDA may no longer be limited to data centers, but will also enter factories, automobiles and robotic devices.

However, this software barrier is not unshakable. At present, AMD is improving the ROCm ecosystem, and Google TPU and Amazon Trainium are also developing their own software systems. Open source tools such as PyTorch and Triton are reducing developers' dependence on specific hardware platforms. If more and more models can be adapted to different chips through a unified framework, the migration cost brought by CUDA may gradually decrease.

This is exactly the long-term problem that NVIDIA needs to face after Jensen Huang leaves. The success of CUDA comes from a technology investment with a very long return cycle, and inference, agents and physical AI may be the entrance for the next round of ecological expansion. Who will decide the investment priority of these directions, and who will promote the software platform to continuously adapt to new computing methods, will affect whether NVIDIA can continue the ecological advantages accumulated in the past.

03 How to fulfill the $2790 billion commitment?

After Jensen Huang left, another thing the market needs to pay attention to is the huge supply and production capacity commitments that NVIDIA has made.

NVIDIA's quarterly report as of July 26, 2026 disclosed that the company's total supply and production capacity commitments reached 2790 billion US dollars, compared with about 1190 billion US dollars in the previous quarter, an increase of about 1600 billion US dollars in a single quarter.

Distributed according to the performance time, the remaining time of fiscal year 2027 is about 920 billion US dollars, fiscal year 2028 is about 870 billion US dollars, fiscal year 2029 is about 880 billion US dollars, and the remaining about 120 billion US dollars are arranged in a later period.

This 2790 billion US dollars is neither the customer orders already received, nor the cash NVIDIA has paid. It mainly corresponds to the procurement and production capacity arrangements that the company has made with suppliers and production capacity partners to meet future demand. NVIDIA needs to lock in advanced manufacturing processes, packaging, memory and other key resources in advance to ensure that Blackwell and subsequent products are delivered as planned.

The problem here is that the commitments on the supply side may not be fulfilled in sync with customer demand.

Large-scale data centers require a long cycle from financing, site selection, power supply to construction and deployment. Even if customers have planned to purchase GPUs, they may postpone purchases due to delayed power access, blocked computer room construction, insufficient financing, or changes in the deployment rhythm of new generation products. NVIDIA also warned in its 10-Q filing that customers may delay or cancel purchases due to factors such as infrastructure, capital and technology adoption progress.

If the speed of customer demand fulfillment is lower than expected, but NVIDIA has already locked in upstream production capacity, it may face inventory backlog, cancellation costs and mismatch between supply and demand. Even if other buyers can be found in the end, product generation changes may also affect selling prices and gross margins.

Another easily confused figure is the approximately 360 billion US dollars AI cloud service agreement commitment disclosed by NVIDIA.

Such arrangements usually involve multi-year cloud service capacity cooperation, which is not the same caliber as the 2790 billion US dollars supply and production capacity commitment. NVIDIA obtains cloud computing capacity through partners, and may also resell part of the capacity to third parties; as the capacity is used, the relevant commitments will be reduced accordingly. However, if external customer demand is insufficient, NVIDIA may still bear certain procurement responsibilities.

These arrangements reflect the changes in NVIDIA's business model: the company no longer only purchases chips and delivers products according to orders, but increasingly participates in the resource allocation of the entire AI infrastructure in advance. Jensen Huang is good at promoting project implementation through customer relations and industrial influence, and the successor needs to strike a balance between maintaining growth and controlling financial risks.

04 The next growth engine?

In the past few years, NVIDIA has seized the generative AI wave and become the main beneficiary of global AI infrastructure construction. But for this company whose quarterly revenue is approaching 100 billion US dollars, how to maintain rapid growth has become more and more important. NVIDIA's most important source of growth is still the data center. In the second quarter of fiscal year 2027, the company's revenue reached 96.221 billion US dollars, a year-on-year increase of 106%; data center revenue was about 89 billion US dollars, a year-on-year increase of 117%, accounting for the vast majority of total revenue. The company's revenue guidance for the third quarter is 108 billion US dollars, and the growth momentum is still strong.

In addition, as large global technology companies continue to expand investment in AI infrastructure, the Blackwell platform is being deployed on a large scale, and the next-generation Vera Rubin platform has also entered the production and shipment stage.

Moreover, NVIDIA is also looking for other new growth opportunities.

Among them, AI inference is an important direction. As large models move from training to large-scale application, enterprises need to continuously call models to complete various tasks, and the popularization of AI agents may further increase the demand for computing power.

Compared with model training, the inference business pays more attention to computing cost, energy consumption and operation efficiency. NVIDIA is improving inference performance through new generation chips and whole-machine systems, hoping to continue to expand the business scale in this market.

Fields such as robotics, autonomous driving and industrial AI are also regarded by NVIDIA as future growth directions. For example, as more and more countries and regions hope to build self-controlled AI infrastructure, sovereign AI is bringing new chip procurement demands. Robotics and autonomous driving are expected to further expand the application scope of GPUs and related computing platforms.

But it still takes time to verify whether these new markets can reach the scale of the current data center business.

At the same time, the competition NVIDIA faces is also changing.

In the past, NVIDIA mainly competed with chip manufacturers such as AMD. Today, large customers such as Google, Amazon, and Microsoft are also developing their own AI chips, hoping to reduce computing power costs and reduce dependence on a single supplier. Broadcom participates in the chip R&D of large technology companies through its custom ASIC business. These products may not be able to fully replace NVIDIA GPUs in the short term, but they have provided customers with more choices in some computing tasks.

Jensen Huang led NVIDIA to seize the growth opportunities brought by generative AI. Next, under the changing competition in the AI industry, who will decide NVIDIA's next step and whether it can lead the company to seize the next round of growth opportunities will become a new test.

Special translator Jin Lu also contributed to this article

This article is from "Tencent Tech", author: Li Hailun, editor: Su Yang, published with authorization from 36Kr.