AI News

Why China's Cloud and Model Assets Are Under Pressure First as Jensen Huang Steps Down Amid RASA Advance

HuxiuUpdated Nov 14 21:12

AI summary

Dolphin Research analysis argues that Jensen Huang stepping down as NVIDIA CEO, combined with the advancement of the Remote Access Security Act (RASA), could bring computing power that Chinese customers remotely rent through overseas data centers under export controls as well. The article states that China (including Hong Kong) as a share of NVIDIA's revenue has fallen from a peak of 28% to 8%, and New Street estimates that if remote demand is cut off, the impact could reach up to $40 billion, or about 10% of revenue. The author judges that in the short term, Chinese cloud vendors and model developers face pressure from incremental supply being choked off and insufficient domestic chip capacity and HBM, while in the long term the open source ecosystem may split into two hardware stacks: CUDA and Ascend.

Why read this

SelectedAI score 76

The article places Jensen Huang's resignation alongside RASA remote computing power controls, breaking down the transmission path through which Chinese cloud vendors and model developers face short-term pressure and the open source ecosystem may split.

Topics

Tags

Body · AI translation

[[**Jensen Huang steps down as NVIDIA CEO, compounded by RASA's plan to bring remote compute under export controls, potentially cutting off China's path to renting GPUs through overseas clouds—a short-term severe blow to Chinese cloud and model companies, and a long-term weakening of NVIDIA's open-source ecosystem.**

**Key Points**

**1. RASA closes the remote compute loophole**

RASA plans to treat Chinese customers' remote use of chips via overseas data centers as exports, handing restrictions to the Commerce Department; the "AI Overwatch Act" also requires the Commerce Department to draft detailed rules. Huang's departure may remove a key lobbying obstacle to this path.

**2. NVIDIA revenue faces a channel gap**

China (including Hong Kong) revenue share has fallen from a peak of 28% to 8%, but if remote demand invoiced through Southeast Asia, Taiwan, etc. is cut off, New Street estimates an impact of up to $40 billion, or about 10% of revenue; the market expects channel revenue to lose 10%-30% in fiscal 2027.

**3. Chinese cloud, IDC and model companies under short-term pressure**

Once the policy takes effect, it won't be "only domestic cards can be used" but "cards can't be used": existing stock keeps running, incremental supply is choked off, and gray channels become more expensive. Domestic card volumes are insufficient in the short term, and HBM this year is only enough for 250,000-300,000 finished cards, increasing pressure on cloud providers such as Alibaba and model companies.

**4. The open-source ecosystem may split into two hardware stacks**

Remote compute is a key channel through which NVIDIA "cultivates developers"; if cut off, the open-source world may divide into two stacks, CUDA and Ascend, benefiting Ascend, AMD and custom-chip customers; after the CFO takes over, long-term open-source investments such as Nemotron may also shrink. 2026-11-14 19:15 Old Huang out: NVIDIA loses its soul, but why do Chinese assets pay the bill? Dolphin Research quick take

This article is from the WeChat public account: Dolphin Research, author: Dolphin Jun

On November 12, Jensen Huang resigned as NVIDIA president, CEO and director, with CFO Colette Kress serving concurrently as interim CEO.

That day NVIDIA opened lower and closed higher, ultimately down 4.3%, but Alibaba fell 7.1%, GDS fell 11.4%, and Nebius fell 16.8%—falling even harder than the main subject.

Beyond the stylistic differences of the leadership change (aggressive Old Huang vs. conservative CFO), which may slow the expansion of AI investment.

A more serious expectation being played out is that the siege on Old Huang is just the starting gun, and the real purpose may be to clear obstacles for re-advancing a bill that has lain in the Senate for nearly a year, the Remote Access Security Act (RASA).

Here, bearish funds are calculating not just how much China revenue NVIDIA will lose, but the shrinking of the entire open-source ecosystem pie.

1. What is RASA?

First, let's understand RASA and exactly what it restricts.

Current export controls govern "whether chips can be shipped into China," i.e., procurement restrictions on physical equipment.

But the loophole is that chips can remain in overseas data centers while Chinese customers lease "remote use" through cloud services and bare-metal services. So what RASA aims to do is count such remote access as exports, hand it to the Commerce Department to regulate, and treat it as equivalent restrictions.

As of now, the RASA bill remains stalled, with differences between the Senate and House versions; the Senate version is harsher:

Although RASA did not come about because of Old Huang's scandal this time, this clearly gave it new political momentum. Therefore, it is hard not to suspect that the intent is elsewhere: first there is the willingness to push RASA, then Old Huang's scandal is exposed.

And at the same time Old Huang's scandal was exposed, more political forces are also probing:

1) Beyond RASA, restriction details have already advanced. According to The Information, in late October the White House demanded adding a clause to the "AI Overwatch Act" tasking the Commerce Department with closing the remote access loophole; this bill has been written into the Senate version of the National Defense Authorization Act and is expected to pass before January next year.

Dolphin Jun sorts it out: that is to say:

a. "RASA" grants the Commerce Department statutory authority to regulate overseas remote cloud compute (authorizing legislation);

b. The "AI Overwatch Act" goes further, requiring Congress to directly task the Commerce Department with formulating specific restriction rules (stipulating the necessity and details of regulation) and enforcing them.

2) There are already signs of NVIDIA making concessions. In early November, media reported that NVIDIA requires that among its cloud service provider customers, non-Chinese customers account for at least 51%. Although this is a supplier term set by NVIDIA itself and does not stem from direct legal constraints, it shows the company is already preparing for the worst case.

During the same period, Oracle was also reported by media to have canceled a 7 billion AI chip leasing order from Tencent.

3) Old Huang, the biggest opponent, exits the stage. Commerce Department officials once said NVIDIA "especially exerted itself" in persuasion and lobbying on RASA, while congressional consensus on restricting chips to China is strengthening. Therefore, Old Huang's resignation removes an important obstacle on the path to advancing remote compute regulation.

2. Will Chinese IDCs, major Chinese cloud providers, and NVIDIA all be badly hurt?

1. The short-term performance math: a not-small impact on NVIDIA revenue

The latest financial report shows that China (including Hong Kong) as a share of NVIDIA revenue has fallen from a peak of 28% to 8% (only legal revenue remains). But the key to this trouble is the "remote" AI compute used by Chinese customers in Southeast Asia and even Taiwan, Japan and other places:

(1) NVIDIA's billed revenue to Singapore once jumped from a 10% share in 2023 to 20%–22% at one point in 2025, while over the same period the revenue share from China (including Hong Kong) fell to 5%.

After the methodology adjustment in FY3Q26 (Q2 2025) (from "based on the customer's billing location" to "based on the end customer's location and shipping location"), Singapore became so small that it was merged into other regions.

NVIDIA's explanation at the time was that customers were merely using Singapore for centralized billing, and the products were almost all shipped elsewhere (99% of orders came from U.S. customers).

(2) The latest FY2Q27 (Q2 2026) financial report disclosed that revenue from China's Taiwan region was $26.985 billion, accounting for 28%, while the previous quarter was only 14.7%, a further jump of 14 percentage points.

The company's note was that more complete-system transactions are settled through Taiwanese ODMs (Hon Hai, Quanta, Wistron, etc.) rather than cloud providers purchasing directly from NVIDIA, and 86% of Taiwan revenue came from end customers in Europe and the United States.

Dolphin Research does a rough calculation: before the methodology change in 2Q26FY, the combined share of Singapore and China (including Hong Kong) basically fluctuated between 20-30%.

After China sales restrictions, although some revenue was lost, China's regional share fell from 20% to 6%, and most of that 14% share, about 10%, was recovered through the Singapore channel.

In addition, after the methodology change, Taiwan's share saw "intermittent" jumps, rising 6pp in 3Q26FY, then recovering, and then jumping 13pp again in 2Q27FY.

Here, the New York Times believes that Megaspeed obtained nearly $5 billion in chip purchases precisely through NVIDIA's server partner channels (purchasing $4.8 billion over the year and a half from FY25 Q2 to FY26 Q4), billing directly in Singapore in earlier years, for example, and later through Taiwanese ODMs.

According to New Street's forecast, if remote cloud computing demand is cut off, the impact on NVIDIA's revenue would not just be the less than $3 billion annualized obtained by Megaspeed (only 0.1% of revenue), but demand of up to $40 billion, or 10% of revenue.

The market expects a 10-30% channel revenue loss in fiscal 2027; taking the median of 20%, and assuming new demand can immediately fill half of it, this is actually roughly in line with Dolphin Research's rough estimate above of a 10-point China share.

This is significantly more severe than the estimated reduction of more than $6 billion in revenue under the assumption of tracing and cutting off commercial activities such as renting cards in Southeast Asia before RASA passes (New Street estimates that chip hardware transshipment plus offshore leasing accounts for 4%, i.e., of $16 billion in revenue, 60% is taken by cloud providers, leaving a 40% gap, i.e., 1.6%).

Dolphin Research believes that under the worst third-tier assumption: at a time when CoWoS is still tight and compute demand is exploding, the gap seems fillable by other demand, but this will inevitably affect the future demand trend, that is, NVIDIA's growth guidance:

With NVIDIA's Q3 financial report about to be released, will next year's 70% growth rate be adjusted? Even if it can be met in 2027, is it brought about by demand being pulled forward from after 2028?

On GPU server rental prices currently tracked by SemiAnalysis, B200 began to decline from the end of October, but B300 rents remain firm, especially in restricted China, where they rose 2.5 times in half a year.

3. The big account of the open-source ecosystem: the logic of "piracy cultivates users" breaks down

Using overseas remote cloud platforms to provide high-end compute services is the key core of the open-source ecosystem logic that NVIDIA tells.

This year, with the prosperity of the open-source ecosystem led by Chinese large models, Jensen Huang has repeatedly voiced support and personally contributed to the open-source ecosystem, especially with quite a tight schedule of related actions in 2026:

a. In July, posted on X publicly supporting open source;

b. In August, it was reported that a trillion-parameter open-source model Nemotron 4 is being developed;

c. In September, announced the acquisition of Hugging Face;

d. Most recently, in October, at a forum in Singapore, praised Chinese open-source models such as DeepSeek, Kimi, and GLM for their outstanding performance.

Promoting open source rather than closed source is beneficial to NVIDIA in the long run. Open source "turns scarce goods into cabbage prices"; the cheaper the models and the more customers, the greater the GPU demand.

Moreover, compared with closed-source leaders dominating the ecosystem, open-source models allow small and medium-sized enterprises and sovereign projects to train and deploy models themselves, dispersing NVIDIA's customer structure so that NVIDIA can maintain relative bargaining power.

Therefore, the open-source ecosystem will greatly alleviate NVIDIA's customer concentration and thereby reduce the long-term risk brought by customers developing their own chips.

In fact, Jensen Huang's strategic thinking is essentially somewhat similar to Microsoft's old logic of "piracy cultivates users"—Chinese users will use pirated copies anyway, so better to let them use Windows, cultivate the habit, and then slowly collect money. Therefore, the moat is not the genuine Windows disc, but the habit and the ecosystem.

NVIDIA's counterpart is GPU and CUDA: in the mainstream and on the surface, high-end GPUs are banned, but through "remote cloud services" another "unspoken" opening has been made for Chinese customers, keeping frontier training within NVIDIA's chip system. And through this opening, one can also clearly see the R&D progress of Chinese open-source models.

What Jensen Huang has repeatedly said over the past few years is essentially this: rather than driving Chinese developers into Huawei's ecosystem, it is better to let them keep using the American technology stack. Remote compute in overseas data centers is, objectively speaking, "cultivating developer habits."

(1) Most Chinese customers did not pay for Microsoft's pirated Windows, so their dependence was even stronger. But with NVIDIA's remote cloud compute, Chinese customers still have to pay a hefty fee, giving them a cost incentive to pursue domestic development.

(2) When Microsoft adopted that strategy, there was no compliance risk—only a business choice between short-term commercial revenue capability and long-term profit potential; whereas NVIDIA directly faces US national security pressure and bears political and legal risks.

(3) Microsoft's Windows has strong lock-in on file formats and software compatibility, but on the AI chip side, developers write PyTorch and use PyTorch to call CUDA, so the barrier to switching hardware is declining. Chinese customers, whether regulators or within the industry, are all intentionally continuing to push domestic substitution.

Therefore, the "cultivate users" playbook has a much shorter half-life in the AI era than in the Windows era. And the US government's push for RASA is an even more precise strike against this playbook.

English text is being prepared.

If even "remote use" is no longer allowed, then leading Chinese open-source models can only adapt to domestic chips faster and with higher priority. Although this is already happening, it must be acknowledged that Chinese AI chips, whether in performance or production capacity, are insufficient to meet the training and inference needs of frontier models:

(1) Zhipu GLM-5 has been adapted to 7 domestic accelerators from day one;

(2) DeepSeek open-sourced the inference kernel on Ascend for V4 Pro;

(3) Meituan said LongCat 2.0 was pretrained entirely on Ascend 910B.

Viewed separately, even if performance is beginning to catch up and narrow the gap, on the most urgent "volume," domestic cards simply cannot absorb it in the short term:

Southeast Asia's 790,000 H100s, converted at a 2.8x B300/H100 ratio, equal 280,000 B300 cards. If supply is cut off entirely, then under equivalent compute, the demand flowing back to Huawei Ascend 910C would be about 500,000–620,000 cards, close to Ascend 910C's planned output of about 600,000 units this year;

But the more serious bottleneck is HBM. Chinese memory makers' HBM this year is only enough for about 250,000–300,000 finished cards. In other words, the HBM gap required for these 300,000 finished cards will most likely be supplied by overseas HBM previously locked in or obtained through other channels.

Summary: Chinese CSPs + models may be "badly hurt" in the short term; NVIDIA's "position weakens" in the long term

Jensen Huang's departure + RASA's advancement: the short-term effect is not "China can only use domestic cards," but "China cannot use cards"—existing stock keeps running, incremental supply is choked off, and gray channels become more expensive.

Overseas channels are locked down by policy, and demand is squeezed back domestically, which is equivalent to domestic compute being assigned certain demand by policy. But for China:

a. For Chinese cloud vendors (such as Alibaba), as well as upstream IDC centers—capacity cannot be built, cloud business revenue and overseas paths are gone, and pressure on customer training services increases; domestic cards cannot even meet inference demand.

This affects Alibaba especially severely. When everything shifts to domestic procurement, leaving aside production capacity, it is hard to say how willing Huawei really is to supply Alibaba with how many chips.

b. For model makers, the narrative shifts accordingly—Chinese frontier model makers can no longer use CUDA operator optimization as a selling point, and may even have to further deny that they have NVIDIA cards, instead switching to talk of being native to the domestic stack, while also bearing the extra costs brought by inferior performance.

The ultimate outcome is that the open-source world is split into two hardware stacks, Chinese and American: one around CUDA, and one around Ascend and other Chinese compute platforms.

And for NVIDIA:

a. RASA's real loss is in the long term: NVIDIA uses open source to hedge against large customers' in-house chip development, and the most active force in open source is in China. Once the network cable is pulled, this feedback loop is broken, and the beneficiaries are Ascend, as well as AMD and those large customers developing their own chips (Google, Amazon, Meta, etc.).

b. A change of leadership further amplifies the risk. Open source is a typical founder-style long-term bet—unprofitable and hard to quantify. Only a commander like Jensen Huang, with absolute say and long-term strategic vision, could push it for years, just as with CUDA back then.

But CFOs tend to be conservative, and how much incentive would a professional manager brought in from outside as CEO have to advance the open-source ecosystem? The first to be hit would be the company's own open-source model Nemotron, which is currently both technologically behind and unprofitable, and may be the first to be cut.

NVIDIA without Jensen Huang is like a body without its most important soul; the risk and loss of value simply cannot be accurately assessed: not only the long-term growth vision under the open-source ecosystem described by Dolphin in this article, but also, from a short- and medium-term perspective, the special CEO-to-CEO cooperation along the industry chain (such as locking in core and critical HBM and CoWoS capacity, etc.). AI industry signal channel: Frontier Technology. This content comes from the internet, original link; the views represent only the author and do not represent Huxiu's position. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]

Source

Huxiu