Wall Street Analysis: Jensen Huang's Resignation Could Be an Opportunity for China's Domestic Compute
TOP Innovation District Research Institute, reviewing Wall Street views, states that after Jensen Huang's resignation, the channels through which major Chinese companies rent NVIDIA compute overseas may be blocked, and demand may flow back to domestic cards. The article mentions that Huawei's CloudMatrix 384 delivers about 300 PFLOPS, surpassing GB200 NVL72's 180 PFLOPS but consuming nearly four times the power, while electricity prices in western China are about half as low; in 2025, domestic vendors delivered about 1.65 million AI cards, accounting for 41% of the market. DeepSeek V4 has already added Ascend to its adaptation list, but Epoch AI estimates that Huawei's compute output is about one twenty-fifth of NVIDIA's, with HBM and production capacity still bottlenecks.
Why read this
AI score 67
The article places Jensen Huang's resignation alongside domestic compute substitution, laying out three threads: multi-card compute stacking, electricity price advantages, and ecosystem adaptation.
Topics
Body · AI translation
**Key points**
**1. The US blocks remote computing power, which instead drives domestic substitution**
After Jensen Huang's exit, the channel for major Chinese companies to rent NVIDIA computing power overseas may be closed, and the hope of waiting for stronger chips disappears, so demand may flow back to domestic cards.
**2. "More cards + cheap electricity" makes up for the per-card gap**
Huawei's CloudMatrix 384 delivers about 300 PFLOPS, exceeding the 180 PFLOPS of GB200 NVL72, but consumes nearly four times the power; electricity prices in western China are about half as low, making the stacked-card approach more feasible in China.
**3. Domestic AI chip shipments and profitability improve**
In 2025, domestic manufacturers delivered about 1.65 million AI cards, accounting for 41% of the market, with Huawei at about 810,000; Cambricon was profitable for the full year, and institutions expect NVIDIA's China share may fall to 8%.
**4. Ecosystem migration accelerates, but supply shortcomings remain**
DeepSeek V4 and others have adapted to Ascend and Cambricon, and the tighter coupling between models and domestic cards raises the cost of returning to CUDA; however, Huawei's computing output is estimated at only about 1/25 of NVIDIA's, and HBM, production capacity and utilization remain bottlenecks. 2026-11-16 08:37 Wall Street: Jensen Huang resigns, possibly China's "dream scenario" TOP Innovation Zone Research Institute quick overview
This article is from the WeChat public account: TOP Innovation Zone Research Institute, author: AI Research Group
On November 12, NVIDIA founder Jensen Huang resigned.
By conventional logic, this week has been terrible for Chinese people—
US intelligence agencies suspect that his extramarital affair partner was recruited by China's Ministry of State Security; both parties in the US Congress used the occasion to demand tighter chip export controls; and the person in Washington who most vigorously spoke for "selling chips to China" has thus been sidelined, although there was also news just now that he is preparing to return.
Let us say here that the intelligence itself is disputed; the Chinese side called it "purely baseless and slanderous," since after all our intelligence personnel do not quite take this high-profile route.
But the reaction of the capital market was the opposite.
Last week, A-share Cambricon rose about 22%. Dan Ives, one of Wall Street's best-known tech analysts, said that in the long run, Huang's departure may be a "dream scenario for China."
When a suspected intelligence operation is exposed, how does Beijing end up the winner instead?
First, the conclusion.
Rumors say that in 2025 NVIDIA will launch a special-supply version, the RTX 6000D, but major Chinese companies show little interest because they are hoping Washington will soon approve stronger chips.
Because the Secure Remote Access Act, which passed the House of Representatives in January this year by 369 votes to 22 but is stuck in the Senate, will most likely pass. And what this bill aims to block is precisely the channel through which Chinese companies "rent" NVIDIA computing power overseas. US House Speaker Johnson said that congressional consensus on restricting chip exports to China is expanding.
Nobel laureate Schelling once made a point: sometimes removing the retreat option makes people stronger.
Burning one's boats relies on exactly this.
The determination of the major companies that Beijing could not force out even if it wanted to, Washington delivered free of charge.
However, it is a fact that the per-card performance of domestic chips lags behind.
Fortunately, computing power can be "stacked."
Huawei's CloudMatrix 384 supernode uses 384 Ascend chips to achieve about 300 PFLOPS of computing power, surpassing the 180 PFLOPS of NVIDIA's flagship system GB200 NVL72. The cost is power consumption: the whole machine uses about 559 kilowatts, nearly four times that of NVIDIA's solution.
This calculation does not work in the United States, but it works in China, and the reason is electricity.
Electricity is what the infrastructure powerhouse has most in abundance.
By June 2026, China's total installed power generation capacity will exceed 4 billion kilowatts, of which renewable energy will exceed 2.4 billion kilowatts. The National Energy Administration estimates that from 2026 to 2030, computing power electricity use will increase by more than 100 billion kilowatt-hours per year on average, reaching about 800 billion kilowatt-hours by 2030.
Electricity in the west is even cheaper.
According to industry estimates, electricity prices in western China are about half lower than in the Beijing-Tianjin-Hebei region, and computing power costs are more than 60% cheaper. Tencent placed its largest intelligent computing center in Ningxia precisely because of electricity prices.
In Qingyang, Gansu, 11 data centers have been built over several years, with computing power exceeding 140,000 PFLOPS, including large-scale domestic GPU clusters.
This is the logic of "trading electricity for chips":
If a single card is a bit weaker, just plug in a few more cards—isn't it just a matter of electricity?
When analyzing Huawei's supernode, semiconductor research firm SemiAnalysis gave a blunt judgment: China has no power constraint.
Cards are still part of it, but NVIDIA's real moat lies in software.
For the past decade or so, AI developers around the world have been writing code on NVIDIA's CUDA. Chips can be replaced, but code written over ten years and habits formed over that time are hard to replace.
So NVIDIA has always used "special-supply" chips to keep Chinese developers on CUDA.
But the ecosystem that should be built still has to be built.
In April this year, DeepSeek released V4, writing Huawei's Ascend chips into its compatibility list for the first time. According to Reuters, Huawei said that part of the training of V4-Flash used Ascend chips, and V4 also supports supernode clusters based on Ascend 950.
Cambricon is also following up. It announced that it has completed adaptation to mainstream domestic open-source models such as GLM, DeepSeek, Tongyi Qianwen, Kimi and MiniMax.
Models adapt to domestic cards, and domestic cards in turn optimize around domestic models.
The tighter the two sides bite together, the higher the cost of going back to CUDA.
So this is a demand from an entire industrial chain, and what catches this wave of demand is an already-formed domestic roster, with astonishing scale and growth rates.
According to IDC data cited by Reuters, in 2025 Chinese domestic chipmakers delivered about 1.65 million AI accelerator cards, capturing 41% of the domestic AI server market, of which Huawei accounted for about 810,000.
At the same time, Cambricon achieved full-year profitability in 2025; Morgan Stanley called it the first domestic AI chip company to achieve sustainable profitability. In the first half of this year, revenue was about 6 billion yuan, up 108% year on year, with net profit of about 2.3 billion yuan. Moore Threads' first-half revenue was about 1.7 billion yuan, up 147% year on year, already exceeding last year's full-year figure. MetaX's first-quarter revenue grew 75% year on year.
Institutions are even more aggressive in their expectations.
Bernstein estimates that Huawei's share of China's AI chip market will rise from about 40% in 2025 to about 50% this year; NVIDIA's share in China could fall to about 8% in the coming years.
Morgan Stanley expects China's AI chip market to reach $67 billion by 2030.
These forecasts were all written before Jensen Huang was out.
At the time, the market was still calculating when NVIDIA would be let back in; although the probability was not high, now this variable is completely gone.
NVIDIA has lost its founder, and with no ready successor, only the chief financial officer could take over temporarily. During the board's deliberations on removal, NVIDIA's market value evaporated by more than $300 billion.
Ives estimates that without Jensen Huang, NVIDIA would be worth 25% to 30% less.
But the greater loss lies outside NVIDIA.
And awkwardly, U.S. intelligence agencies did not confirm that the woman was engaged in espionage for China. The Washington Post said that the wording of Secretary of State Rubio's memo was tougher than the CIA's original report. If the intelligence was exaggerated, the United States has, on the basis of disputed evidence, taken down the head of its own most valuable company.
Over the past few years, a considerable portion of Chinese big tech's computing power has been placed overseas. Hong Kong warehouses are reportedly holding about 600,000 NVIDIA chips; ByteDance, Alibaba and Tencent are reportedly renting NVIDIA computing power through data centers in Thailand, Malaysia and Japan.
According to data center research firm DC Byte, Malaysia has 17 hundred-megawatt-class data centers under construction and planned, with Chinese cloud demand being an important driver there.
It must first use domestic cards. Since 2025, newly built data centers receiving state funding have been required to use only domestic AI chips; Reuters counts that such projects have received more than $100 billion in state funding since 2021.
Domestic solutions consume more power, so electricity prices and power supply guarantees will move to the very front of site-selection criteria. Wood Mackenzie's analysis holds that training such latency-insensitive heavy loads will move toward the west, where renewable energy is abundant; latency-sensitive inference and financial applications will remain in the east.
In summary: the west gets training and storage, the east gets inference and applications.
Those places that can deliver "cheap power, domestic cards, and large-cluster operations and maintenance" as a package will win more orders in this round.
The shortcomings of domestic computing power remain obvious.
Research institute Epoch AI estimates that Huawei's total computing power output this year is about one twenty-fifth of NVIDIA's, and that by 2030 it will lag by about four years, with the tightest bottleneck in high-bandwidth memory. Huawei rotating chairman Xu Zhijun also admitted in September this year that Huawei's AI computing power capacity cannot even meet domestic demand.
Another hidden risk is "built a lot, used little."
A study by the Australian Strategic Policy Institute points out that under the "Eastern Data, Western Computing" initiative, many computing centers have low utilization rates; cheap electricity does not mean computing power can actually be used. The key is still whether they can take on large-model training/inference orders. Domestic chips ranked fifth to thirtieth, even once racked, will find it hard to have real commercial workloads.
Huawei's phones offered a reference price: listed on the Entity List in 2019, the Mate 60 Pro returned in 2023 with SMIC's 7-nanometer chip, taking about four years.
That time, Huawei held on.
Only one question remains:
How fast can domestic computing power run? Channel: International Hotspots. 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 in the Huxiu APP[ ]