Nvidia’s Huang Wears Tang Suit

Nvidia CEO Jensen Huang made a splash at China’s Chain Expo on July 16th, donning a traditional Tang suit and delivering a speech in Mandarin, a rare move for the tech executive.

“Chinese developers and entrepreneurs are driving rapid AI innovation,” Huang stated. “There are now a million developers involved in this field. Companies like DeepSeek, Alibaba, MiniMax, and Baidu are developing world-class products, propelling the global development of artificial intelligence.”

Huang’s effusive praise for China’s AI industry and its players has been a recurring theme since his arrival in Beijing.

This charm offensive comes just days after Huang revealed that the U.S. government had approved sales of the previously banned H20 chip to China.

Given the “compute power anxiety” among Chinese AI firms, Nvidia’s $4.5 billion stockpile of AI chips, previously written down in earnings reports, is likely being urgently tallied and packaged for imminent delivery to Chinese customers.

However, almost in lockstep with Huang’s high-profile courtship of the Chinese market, U.S. Commerce Undersecretary Howard Leuters has thrown a curveball.

“Chinese companies won’t get the best chips, or even the second-best, or even the third-best,” Leuters bluntly declared in a U.S. media interview, articulating a strategy of selling China just enough AI chips to foster technological dependence on the U.S.

Leuters’ remarks underscored a clear position: under the current semiconductor restrictions, China’s downstream AI ecosystem will not have access to the top-tier products it desires, regardless of policy adjustments.

During his China visit, Huang noticeably downplayed the criticality of Nvidia GPUs, signaling a subtle shift in Nvidia’s narrative, at least within the Chinese market.

Nvidia’s Clock is Ticking

Multiple sources suggest that Leuters’ statement is largely an attempt at damage control, after China started retaliating with restriction on rare earth exports.

Additionally, since April of this year, the U.S. semiconductor industry has been lobbying and pressuring the Trump administration.

Nvidia has been particularly vocal, as the “China-specific” H20 chip has no other viable market. A continued sales ban would translate into substantial financial losses.

During a media briefing yesterday, Huang addressed the H20 issue: “Nvidia may not be able to fully recover the previous inventory write-down, but most of the assets are not permanently written off. The recovery rate may not reach 100%, but it will not be 0 either.”

Beyond the financial figures, what likely concerns Huang most is the potential vacuum Nvidia GPUs could leave behind.

Consider the developments in China’s AI chip industry over the past three months.

Take Huawei, a direct competitor to Nvidia. In May, Huawei officially launched the “CloudMatrix385 Ultra Node” compute platform at its Kunpeng Ascend Developer Conference. It utilizes 384 Ascend chips interconnected via the industry’s largest high-speed bus.

While Ascend chips may lag behind Nvidia’s in hardware specifications, Huawei, leveraging its telecommunications expertise, has surpassed Nvidia’s NVL72 in dense BF16 compute power and network interconnection bandwidth through its fully interconnected peer-to-peer architecture.

During the media briefing, when asked about competitors like Huawei, Huang acknowledged, “We’ve been doing this for 30 years, and they (Huawei) have been doing it for just a few. No company is putting more effort into building the AI ecosystem than Nvidia. The fact that Huawei can be mentioned in the same breath as us says something.”

Cambricon is another company capitalizing on the opportunity. According to its recently released Q1 earnings report, the company’s revenue surged from 26 million RMB in the same period last year to 1.111 billion RMB, achieving profitability in a single quarter for the first time. While this period doesn’t fully overlap with the Nvidia H20 export ban, considering downstream manufacturers’ inventory cycles, it is indicative of a trend.

Faced with the rising tide of domestic AI compute chips, Huang and Nvidia can ill afford to wait.

Furthermore, as reported by a Chinese media outlet, the migration of large language models to domestic compute platforms is a “one-way” process. Once deployed and migrated, whether from an operating cost or risk management perspective, it is unlikely that companies will revert to the Nvidia ecosystem.

This explains Huang’s eagerness to make statements and publicly announce Nvidia’s high-performance chip’s return to the domestic market.

Betting on China’s Embodied Intelligence

It is likely that the H20 chip will be discontinued after the current inventory is depleted, a point indirectly confirmed by Huang yesterday.

For China’s future market, Nvidia plans to primarily promote the “B30 chip” built on the Blackwell architecture. Notably, this is also a China-specific chip, and potentially a “reverse upgrade.”

According to reports, the rack-level solution built around the B30 will offer reduced performance compared to the H20 in some areas but is expected to have 30% higher energy efficiency and 40% lower procurement costs. This price reduction is based on switching the HBM chip to GDDR7.

Given that the large language model pre-training phase is largely complete, unless the B30 chip offers significant improvements in inference efficiency, its demand among Chinese manufacturers may be limited.

Huang did not disclose any details on the progress of this chip during his visit to Beijing. In stark contrast, his remarks on “humanoid robots” were exceptionally frequent, making it a primary focus of his trip.

When asked about the humanoid robot industry in China, Huang highlighted three key points:

First, the world is facing a severe labor shortage, with manufacturing facing a deficit of millions of workers. Increasing automation will significantly boost global GDP growth.

Second, the combination of humanoid robots and AI technology is happening at the right time, with both fields advancing simultaneously.

Third, China already possesses excellent AI technology, is doing a great job in mechatronics, and has a vast manufacturing base to deploy these robots.

“Therefore, I am very optimistic about the development of humanoid robots in China.”

However, one reason Huang may not have mentioned is that promoting Nvidia’s humanoid robot services in the Chinese market is a very “localized” strategy.

On the one hand, the humanoid robot supply chain has not yet been explicitly included in export restrictions to China. On the other hand, Nvidia has a large and complete technology stack in the humanoid robot field, allowing it to hedge its bets even in an uncertain geopolitical environment.

For example, for on-device computing, Nvidia has Jetson Thor. For simulation training, Nvidia has the Omniverse platform. To enable robots to learn new skills through imitation learning or trial-and-error with reinforcement learning feedback, Nvidia has Isaac Lab.

Even if the foundation model is a problem, Nvidia has GROOT N1.

If Nvidia was the “picks and shovels” provider in the generative AI era, then in the embodied intelligence era, Nvidia seems to be aiming to provide all the production tools.

In the booming embodied intelligence startup scene in China, Nvidia’s future market potential is self-evident.

Notably, at the past two GTC conferences, Huang showcased multiple humanoid robots from domestic manufacturers. At this Chain Expo, Nvidia brought new up and coming companies like “Accelerated Evolution” and “Zhi Ping Fang”, taking a neutral stance.

CUDA: Passively Compatible?

Huang made another statement at yesterday’s media briefing that was quite surprising.

When asked if Nvidia would consider open-sourcing CUDA, Huang replied, “If a platform is compatible with CUDA, I think that’s fine. In fact, CUDA itself is relatively open. You can review the detailed version of CUDA and then develop a compatible version based on it. So, in a sense, CUDA is already open-source.”

“This is different from the X86 architecture. If you develop products compatible with X86, they may be unhappy, but if you develop products compatible with CUDA, I wouldn’t mind at all.” Huang didn’t forget to indirectly call out Intel.

However, Nvidia has been quite sensitive on CUDA compatibility in the past.

For example, in March of last year, Nvidia added a clause to the End User License Agreement (EULA) of the CUDA 11.6 update: “You may not reverse engineer, decompile, or disassemble any portion of any software product generated using SDK components to convert such software product to run on non-Nvidia platforms.”

Have any companies actually done this in actual product development? Several companies have tried it.

For example, AMD ROCm is developing its own independent intellectual property tool library and technology stack based on maintaining consistency with CUDA at the API interface protocol and compiler level.

Some domestic manufacturers are even more direct, running unmodified binary files on third-party cards by calling CUDA functions, which undoubtedly harms Nvidia’s interests.

However, from Huang’s public statement yesterday, it is easy to see that Nvidia’s attitude on the key issue of “passive CUDA compatibility” has quietly changed.

This shift is driven by a dual reality: on the one hand, as mentioned earlier, Nvidia’s primary concern is that, given the U.S. government’s continued policy swings, it may be forced to cede the “market vacuum” in the promising Chinese market, creating opportunities for competitors.

On the other hand, as Huang emphasized, Nvidia is one of the few companies in the world that can achieve full-stack innovation from algorithm architecture, system software, hardware systems, network technology to chip design. Even if there are some losses on hardware sales, as long as it can maintain a firm grip on the core influence of the CUDA ecosystem, it can still realize commercial value through software licensing, technical services, and other channels.