China in HUGE Trouble! Nvidia Cuts Supply—Computing Power Runs Short, Large AI Models Collapse

China Observer
822粉絲
A shortage of computing power has repeatedly disrupted DeepSeek’s chat feature on both its website and app in recent days. Users have seen the message, “The server is busy. Please try again later.”
Join this channel to get access to perks:
https://www.youtube.com/channel/UCT2kPBcD6tXn8TP_aV7BmgA/join
#chinaobserver
All rights reserved.
⭐You can support us at: https://donorbox.org/china-observer-supporting-independent-news
⭐For business cooperation, please contact us: [email protected]
I might be one of the people hit hardest by this Deepsea crash.
I'd just gotten off the plane around noon today
when I got a bunch of messages and even some friend requests.
I knew something must be wrong with my app.
Sure enough, everyone was telling me that
Genji AI's automatic bookkeeping feature had stopped working.
I only found out later that DeepSeek had crashed.
My bookkeeping tool runs on DeepSeek's API and I don't have a backup.
So when it went down, I went down with it.
I knew I was in for another wave of bad reviews.
It was so embarrassing.
And it took several hours to come back.
DeepSeek really lost a lot of credibility this time.
This time.
This is a disaster for DeepSeek users.
A lot of people say they started getting limit warnings within minutes.
So far, three common issues have come up.
First, long chats stop working after 7 days of inactivity.
Second, a single message is now limited
to fewer than 40,000 characters.
Third, there is a daily chat limit.
Some people hit it in just over 10 minutes.
People on social media are up in arms.
Some say it feels like a breakup.
Others say it's like a virtual pet
that starves if you leave it alone too long.
Some users shared a way to rescue old chats.
Export the conversation first,
then split it into smaller files under 40,000 characters each.
Start a new chat and send one file at a time.
If you're still worried, give it a try.
A shortage of computing power has repeatedly disrupted
DeepSync's chat feature on both its website and app in recent days.
Users have seen the message.
The server is busy.
Please try again later.
In another screenshot published by IT Home, a user wrote,
I didn't think I'd hit the limit so fast.
I'd only been playing around with it for a few hours.
The user added that once the quota ran out,
none of my chats worked and I couldn't start a new one
either. Many Chinese commenters mocked how quickly DeepSeek seemed
to fall apart as soon as the
United States tied in restrictions.
They compared it to someone foaming at the mouth and rolling
They're rolling their eyes.
DeepSeek had already suffered several service disruptions,
with seven in October and three so far in November.
The incidents fell into two categories.
Most involved degraded performance, including slower responses,
long waits for texts to generate, and failed requests.
Others involved partial outages on the website or API.
DeepSeek makes money in part by selling API access to businesses
and developers and charging by the token.
For every 100 yuan in revenue, it makes 83 yuan in gross profit.
That margin is far higher than in most industries.
The website and app, meanwhile, are free to use,
so Deepsea covers the cost itself.
In mid-November 2026, NVIDIA founder
and CEO Jensen Huang abruptly resigned amid a scandal involving
customers with ties to Chinese espionage and his private life.
US controls on China's access to computing power
that moves quickly beyond bands on physical chip cells.
They expanded to remote access computing power leased through overseas
data centers in places like Southeast Asia and the Middle East.
Underlying hardware and software architecture
and mechanisms for remotely controlling or disabling chips.
The Wall Street Journal
and other outlets reported that while in office,
Jensen Huang pushed to set up commercial arrangements
that helped Chinese AI companies
such as ByteDance and Alibaba get around the hardware embargo.
The companies leased computing
power from high-end new video chips at data centres in Singapore
and elsewhere in Southeast Asia,
including through companies such as Megaspeed,
which was linked to Alice Huang,
the woman allegedly at the centre of his extramarital affair.
On November 16, the US Commerce Department's Bureau of Industry
and Security, or BIS,
issued new rules extending AI chip export controls
from physical cells to remote leases of overseas computing power.
The rules close off a route Chinese firms used
to access advanced U.S. computing power through third countries.
They designate Singapore, Malaysia, Thailand, Indonesia and Brazil
as high-risk countries for chip transshipment and computing services.
Exporters now need licenses to ship advanced AI chips
to those countries or provide related services there.
Southeast Asia has quickly become a global hub for data centers.
Malaysia and Thailand have
emerged as major investment destinations
while Indonesia's market is growing rapidly.
Chinese tech companies including Alibaba
and ByteDance have been releasing more computing
capacity and investing in infrastructure across
the region to train their AI models.
The new restrictions also cover Brazil.
In recent years, Brazil has promoted its data
center and AI industries
to strengthen what it calls digital sovereignty and attract investment
from U.S. and Chinese tech giants.
ByteDance is investing billions of dollars in a large
data center project in northeastern Brazil.
The region has abundant renewable energy
and is a major landing point
for South America's undersea fiber optic cables.
Meanwhile, new details have emerged about Alice Huang,
the former CEO of Singapore's Megaspeed.
Chinese outlet Caixin reported
that Singapore authorities began investigating her in February
3.20.25
A month later, on March 19,
she registered a company called Next Power Japan in Tokyo.
The company may have served as a channel for moving funds,
with 10 million yuan, about 1.4 million US dollars,
invested in setting up a server supply chain in Shanghai.
Japanese corporate records show Next Power Japan received
its corporate registration number that
same day, March 19.
It was registered in Kojimachi in Tokyo's Chiyoda Ward.
Japanese reporters visited
the site and found no sign that the company actually operated in Tokyo.
That led to speculation
that Alice Wong may have rerouted some funds
through Tokyo after Singapore authorities began
He began investigating her.
Chai-Sin also traced how Alice Wong used
the Tokyo company to gain control of a service
supply chain in China.
Records show that in June 2025, she used Next Power Japan to fund
the creation of SkyVault.
That December, through her wholly owned company Beihai
Rainbow Investment Consulting,
she brought Shanghai Kujan Technology into the group.
Shanghai Kezhang supplies server equipment
to major Chinese data centers.
The move gave her an end-to-end server supply chain.
U.S. outlet The Information had previously reported
that Alice Huang and her relatives
had already built a sprawling network across China and Singapore
to get around regulations and acquire advanced chips.
Japan is outside the five-country list,
which puts extra attention on Alice Wang's Tokyo-based
company, Next Power Japan, and its lack of active operations there.
The United States is also
seeking broader international cooperation.
On November 15, the United States signed the San
Francisco Declaration with partners including Japan,
Australia, Israel, the United Kingdom,
South Korea, India, Qatar, and the United Arab Emirates.
They agreed to share responsibility for protecting AI supply chains
and critical technologies. It's part of the US-led Paxilica initiative.
Jacob Helberg, the US Under-Secretary of State for Economic Affairs,
stressed that chip design,
manufacturing equipment and shipping ports are spread across
different countries.
A vulnerability at any point could put the entire supply chain at risk.
Over the past few years,
the United States has steadily tightened restrictions on exports of
high-end GPUs to China.
Chinese companies still had another way to get computing power.
U.S. chips went to third countries such as Singapore
and Malaysia where local data centers
bought GPUs such as Nvidia's H100 and H200
and leased the computing power to Chinese customers
through cloud platforms. The chip stayed overseas
but customers could submit computing jobs remotely
from China and send the results back once training was complete.
That created an awkward situation.
The United States could control where the chips physically went,
but had only partial control over who could use them.
CNBC reported that Chinese companies including ByteDance,
Alibaba and Tencent had used NVIDIA
chips remotely through data centres in places such as Malaysia,
Thailand and Japan.
JP Morgan just disclosed that about 3.8 gigawatts
of plant data center capacity in Malaysia
is linked to Chinese operators.
That's nearly 30% of the total capacity plant. Major AI companies
are tightening their restrictions on China.
Leading the way is Anthropic, seen as the top
ranked AI company right now
and among the most aggressive in restricting access from China.
Anthropix website states that its updated supported regions policy
officially took effect on November 12, 2026.
Many mainland Chinese users were alarmed by
what Claude was doing
and wonder if they'd have to rely on ByteDance AI assistance
for work. Claude is unavailable to users in mainland China,
Hong Kong and Macau.
The restrictions also apply to Americans while they're in those places
whether they're there for business, tourism or an academic conference.
Jensen Huang resigned amid rumors linking him
to Chinese Communist Party espionage
as U.S. President Donald Trump held talks
with Chinese Communist Party leader Xi Jinping.
In the run-up to the November 18-19 APEC summit in Shenzhen,
disputes over AI model distillation, intellectual property,
and controls on overseas computing power were becoming
a new focus of U.S.-China tech competition.
One especially awkward case involves Japanese entrepreneur
and tech blogger Makoto Yoshikawa.
He tried using Claude in Shenzhen
while attending the 2026 APEC meeting and Anthropic banned his
account. He had just landed
when he locked in and the system immediately flagged him for violating
its supported region's policy.
His appeal was rejected outright. Half joking, he urged APEC
Let's take leaders to discuss the issue.
As word spread many Chinese developers
and commenters joked that Claude's security system
checks your internet connection not your passport
and that foreign visitors finally got a taste of
what it's like to use Claude here.
Some joked they'd rather block 100 innocent users than
let one slip through and Claude will punish anyone
who goes to China without doing their homework.
There have been plenty
of similar cases since Anthropic explicitly listed mainland China,
Hong Kong and Macau as unsupported regions for November 12.
Trying to use cloud there
can trigger restrictions or an account ban.
Travelers have frequently run into trouble on the
For example, travelers from Taiwan, Singapore, Japan
and the United States have opened cloud in business trips
or vacations in Hong Kong or Shenzhen and immediately seen the message
cloud isn't available in your region.
Even some people using a Taiwanese eSIM or International Roaming
were flagged as being in an unsupported region.
The system can also ask for ID and selfie.
Users who failed verification may have their accounts banned outright.
Renowned theoretical physicist
and Fields Medal winner Edward Whitten recently gave an IAS
Distinguished Lecture on Black Hole Thermodynamics
at the Hong Kong University of Science and
Technology. While using his computer in front of the audience,
he opened Claude out of habit.
The projector immediately displayed the message,
Claude is available in your region.
Whitton developed M-theory and super string theory and is widely
regarded as a genius and one of the leading physicists of our time.
One mainland Chinese commenter joked that the one thing
to be thankful for was that the AI company
at least respected China's territorial integrity.
If it blocked China, it had to block Hong Kong too.
Another user getting caught up
in closed restrictions points to a water issue.
Some of the world's top scientists are experiencing
the effects of tech decoupling first hand.
Anthoping may see it as a routine, compliant decision.
For researchers and businesses around the world,
it sends a clear message.
Entering China can mean losing access
to the world's most advanced AI tools.
Alongside Anthropix's restrictions on users in China,
OpenAI explicitly says on its website
that accessing its services from an unsupported country or region
may result in an account ban or suspension.
Mainland China, Hong Kong
and Macau are all absent from its current list of supported regions.
The two companies' approaches
to account bans may get even more similar.
In early November, OpenAI's Shenzhen team poached Anthopix productly.
One Chinese commenter lamented
that the two companies seemed to be developing
a denial-of-service feature just for users in China.
People only wanted to chat, the commenter added,
and ended up as collateral damage
in the two countries raised to enforce AI rules.
If more international scholars,
entrepreneurs and executives have to worry about account bans,
service interruptions or losing access to their data,
they'll factor those risks into where they
hold conferences, who they work with and even where they invest.
Chip factories and data centres
are only part of the tech competition.
The bigger question is who can build an environment for
innovation where talented people from around the world want to come,
stay and work together.
As the United States tightened restrictions
on China's access to AI chips and advanced
computing power,
the effects on China's AI industry are becoming clearer.
After the new rules were announced,
shares of Chinese AI developers including Alibaba,
Tencent, Minimax and Dupu AI fell across the board.
Expectations for training costs
and cloud services revenue growth were revised downward.
In the short term, the rules affect the supply of computing power.
Over the longer term,
they affect how China can catch up technologically.
China used to be able to solve the problem by buying chips.
Later, it could lease computing power from overseas data centers.
Now even remote leasing is coming under strict limits.
As advanced computing power gets harder to obtain,
Chinese companies will have to spend more time
and money on the same research and development.
The United States is tightening access to chips
and targeting China's ability
to narrow the technology gap through distillation.
In its latest warning, the United States said
China's large-scale distillation efforts
could violate intellectual property rights.
Reuters reported on November 17
that it reviewed more than 300 Chinese academic papers,
patents, company disclosures, job listings and government documents.
It found that Chinese researchers
and companies had taken model distillation from lab research
to large-scale industrial use.
Reuters found that Chinese companies had automated data collection
and assigned dedicated staff to prepare training materials.
Government policies encourage industrial use of distillation
to cut the cost of training and deploying AI models
and expand their use in manufacturing.
In distillation a more capable model acts as a teacher.
It generates content or demonstrates tasks
and that material is used to train a smaller student model.
The technique itself is legal. The
dispute is over where ordinary research ends
and organized large-scale collection of competitors
model outputs begins, especially
when the goal is to extract propriety features and capabilities.
One example Reuters cited was a paper published
by Shanghai Jiao Tong University researchers
in November 2024.
They collected about 5,000 answers from OpenAI's O1 model,
then used GPT-40 mini to translate them.
That produced 10,750 bilingual training examples,
which they used to further train Alibaba's Qiwen 2.5-72b model.
Research involving Alibaba researchers and Meituan
also used GPT-4 and GPT-4-0 to generate training materials
or help AI agents carry out coding tasks.
Chinese government policy has
also supported the widespread use of distillation by Chinese companies.
In January, eight government agencies issued an AI
and manufacturing policy document
with guidance on using model training,
fine-tuning and distillation in industry.
Shanghai, Shenzhen and other cities have encouraged companies
to use these techniques to reduce their computing needs.
A public comment
from an OpenAI executive has drawn particular attention.
Dean Bull, a member of OpenAI's policy team,
recently wrote on social media that after
joining the company and learning more about what
was happening internally, he came to believe
model distillation could account for around 50%
of the progress China's AI industry had made
in closing the gap with the United States.
His earlier estimate was 30%. The figures reflect
His personal assessment,
they did not come from any official open AI analysis.
Ball put a number on a growing anxiety in the US AI industry.
People are starting to suspect that the pace
of China's AI progress depends increasingly
on what its developers can learn from leading American models,
as well as on how many GPUs
they can buy. For the past two years,
US AI policy toward China has rested on one premise.
If the United States controls the most advanced chips,
the gap between American and Chinese models
will gradually widen. So far, that prediction has only partly held up.
Despite limits on computing
power, Chinese models such as Deetsik, Qiwen,
Kimmy and GLN are still closing in on the world's
top models. The United States has
also raised the issue directly with China.
Reuters reported that U.S. Commerce Department officials met
with Chinese regulators in Chengdu
in July to raise concerns about Chinese companies collecting outputs
from American AI models
and warned that some of those activities
could violate intellectual property rights.
Chinese officials described distillation
as a standard technique and said they were
willing to address intellectual property violations.
They made no specific commitments.
The U.S. government has
also publicly accused Chinese companies of using the technique to
extract capabilities from American models.
On September 8th, the U.S. National Security Agency,
the FBI, and the Cybersecurity
and Infrastructure Security Agency issued a joint cybersecurity
advisory. It named six Chinese companies,
DeepSeek, Moonshot AI, Alibaba, MiniMax,
StepFun and DeepSeek AI.
The agencies accused the companies of obtaining billions of tokens
through millions of interactions with advanced American AI models,
including Claude, GPT, Gemini,
and Grok, dating back to late 2024.
US officials said the companies used large-scale distillation
to extract proprietary features and capabilities from American models.
The US concern is that Chinese companies may
be able to keep replicating the capabilities
of leading American models
and closing the gap at a fraction of what US companies spent.
If that approach works,
it could quickly erode the advantage American companies gained
by investing so much computing power, money and time.
A private sector entrepreneur Hu Luren,
familiar with the chip and data center industries,
said the new U.S. rules could significantly
limit China's AI development. He said a lot
of AI companies in China process
their data through computing providers overseas.
After this Jensen Huang incident, the U.S. started putting up barriers.
For the past few years, U.S. restrictions on chips
and technology going to China had basically no effect.
High-end U.S. chips made their way into China through third countries.
Hu Luren said overseas data centers had provided computing power
to China's AI data centers without interruption for years.
I used to work in this industry. A chip generally has a useful life
of five to eight years but it gets superseded much sooner.
In about two years a new generation
usually comes along with a big enough leap
in performance to make the old one obsolete.
NVIDIA, AMD and Intel put out new products every few months.
In the past countries trying to catch up
on advanced technology often had to spend years
on research trial and error and bringing products to market.
Large AI models changed that. The
capabilities of models that industry leaders spent tens
of billions of dollars training steadily
spread through research papers, open source code,
public tests, API services and everyday use.
If developers can use distillation to learn how advanced models behave,
they can bypass some of
the research and development spending
and reproduce those capabilities at a lower cost.
As China began using distillation to close the gap,
the contest shifted toward intellectual
property and how technology spreads.
The focus keeps moving. The rivalry increasingly looks
like a drawn-out struggle over who can gain time
and who can be made to lose it.
The United States is turning China's AI development from a roughly
110-yard sprint into a long-distance run with a heavy load.
Seen that way, falling behind is one risk in the AI era.
A bigger risk is having someone else decide
whether you even get to enter the race.
