Information Technology · Saturday, 22 August 2026

01 · Briefing · what happened

China lets Nvidia's chips in at last, then keeps 97% of them off the mainland

Information Technology 8 min 24 sources

ByteDance and Tencent each took about 10,000 Nvidia H200 AI chips - roughly 2.5% of the order book their companies were cleared to buy in January. Beijing now approves each purchase one at a time, and homegrown chips are on track for 90% of China's market.

2.5%

of the order book that reached the mainland

about 20,000 chips against 400,000-plus approved in January

90%

of China's chip market going to home-built designs

TrendForce forecast; Huawei and Cambricon lead

93.7%

how full SMIC's factories ran

up from 68% in early 2023, on a record $3bn quarter

15%

Samsung's price rise on advanced chipmaking

Chinese customers accepting the steepest increases

At a glance

ByteDance and Tencent each received about 10,000 Nvidia H200 chips - the first to reach mainland China since December.
That is roughly 2.5% of the 400,000-plus the big Chinese firms were cleared to buy in January; the rest must stay in Hong Kong, which lacks the data centres and power to run them.
Beijing built its own case-by-case approval system, quantity caps and end-location rules - the same instruments Washington used on it.
Chinese-designed AI chips are on track for 90% of China's market; Nvidia's share there fell from 66% in 2024 toward 8% this year.
The split supply chain is raising prices everywhere: SMIC raised its prices after a record $3bn quarter, Samsung raised advanced chipmaking prices up to 15%.
Japan's Ajinomoto, which makes about 95% of a film used in nearly every high-end chip package, is cutting supply to China by 30% - and Chinese substitutes are already qualifying.
Supermicro fired several employees after an outside investigation into $2.5bn of Nvidia hardware allegedly smuggled to China.
US rules control physical chips but not renting them remotely, so Chinese labs have reportedly used restricted chips through data centres in Southeast Asia.

Forces in play

US export controls High→

Every H200 sale needs an individual US licence and sends 25% of the price to the US Treasury; the best chips are still banned outright.

China's own controls Building↑

Beijing now approves each purchase one at a time and keeps most licensed chips off the mainland, in Hong Kong - the same tools used against it.

Home-built replacement High↑

Huawei and Cambricon are set to hold 90% of China's market; domestic high-end shipments are forecast to grow more than 83% this year.

Chip prices Building↑

SMIC, Samsung and the Japanese maker of a film in nearly every high-end chip package have all raised prices - the cost of running two supply chains instead of one.

Nvidia's China business Easing↓

Share of the Chinese market fell from 66% in 2024 to near zero, with half a million unsold H200s built for customers who no longer need them.

In play Nvidia — holds about half a million H200s built for a market that moved on Beijing's NDRC — approves every chip purchase individually, and keeps most imports out of the mainland Huawei and Cambricon — the domestic chip designers filling the gap the ban created SMIC — China's only advanced chip factory - record quarter, 93.7% full, raising prices Ajinomoto — Japanese maker of about 95% of a film needed in high-end chip packages, cutting supply to China by 30%

How it unfolded

2022 Washington starts capping how many advanced Nvidia chips China may buy
Last Dec Trump clears H200 exports, with 25% of each sale going to the US Treasury
January Chinese firms approved for 400,000-plus chips; US moves to case-by-case licences
Eight months Beijing says nothing; Nvidia's China share falls to zero
This week About 20,000 chips reach the mainland under China's new per-purchase approvals

Where this points

Watch whether Beijing raises the per-company mainland allocation above 10,000 units, or leaves the rest parked in Hong Kong - the first would mean it still wants the chips, the second that the home-built replacement is far enough along not to need them.

Full briefing

The door opened. Almost nobody walked through.

ByteDance and Tencent have each taken delivery of about 10,000 Nvidia H200 accelerators in recent weeks [1][2]. They are the first of these chips to reach mainland China since Washington cleared their export last December [1]. An accelerator is the specialised chip that does the heavy arithmetic behind AI training and answering. The H200 is at least two generations behind Nvidia’s best, which China still may not buy [2].

Set against the order book, the number is tiny. ByteDance, Alibaba and Tencent were collectively approved in January to buy more than 400,000 of these chips. What has actually reached the mainland is roughly 2.5% of that [1]. Most of each company’s US-licensed allowance - understood to be up to 100,000 units apiece - must stay outside the mainland, largely in Hong Kong, on Beijing’s instruction [2]. Hong Kong cannot host them: it does not have the data centres, and it does not have the power [2]. “Everyone needs the chips but struggles to find a way to use them,” one person familiar told the FT [2].

Nvidia is sitting on around half a million H200s built mainly for Chinese customers [2]. Jensen Huang told investors earlier this year that Nvidia’s share of the Chinese market had gone from 95% to zero [1].

Both sides now run the same machine

The striking part is not the volume. It is the shape of the control.

To let the chips in, Beijing built a new approval process from scratch at its National Development and Reform Commission, which signs off every purchase individually [1]. That mirrors the Commerce Department’s own case-by-case licence review, which Washington moved to on 16 January. By mid-May it had cleared about 10 Chinese firms, including Alibaba, ByteDance, Tencent and JD.com [1]. The 10,000-unit mainland allocations are quantity caps - the instrument US export rules have used since the first restrictions on Nvidia’s Hopper chips in 2022. The requirement to route imports through Hong Kong is an end-location condition, the mirror image of Washington’s demand that every chip transit US soil for inspection before re-export [1].

There is a price tag on the American side that rarely gets said out loud. Trump’s December approval sends 25% of every sale to the US Treasury [1].

Enforcement is getting sharper in the gaps. Supermicro said on Friday it had finished an outside investigation into the alleged smuggling of $2.5 billion of Nvidia hardware to China. It fired several employees, and said no current senior executive knew [7]. A co-founder was arrested five months ago over the same allegations [7]. A bigger gap is still open. US rules control physical chips, not remote access to them. So Chinese labs have reportedly rented the computing power of restricted GB300 chips through data centres in Southeast Asia [4]. A White House official accused Moonshot AI of doing exactly that through a facility in Thailand [4]. Legislation to close it is being discussed [4].

The market moved while the door was shut

Chinese-designed AI chips are set to take 90% of China’s domestic market, according to TrendForce, with Huawei and Cambricon the main winners [3]. Nvidia held 66% of that market in 2024, 40% in 2025, and was tracking toward 8% this year on Bernstein’s estimate [3]. China’s total market topped 4 million AI chips in 2025; Huawei alone shipped 812,000, Alibaba’s T-Head 265,000 [3]. Domestic high-end shipments are forecast to grow more than 83% this year [3]. Bloomberg’s own tally finds China now leading the US on usage and cost of AI models, on the strength of cheap, capable releases [6]. Nvidia is reported to be readying another China-specific chip for shipment by year-end [5].

Two of everything costs more than one of everything

The bill for splitting the world’s chip supply in half is now visible in prices, and it is not being paid only by the two governments involved.

SMIC is China’s only advanced chip factory - the kind that etches other companies’ chip designs into silicon. It posted its first $3 billion quarter, up 36.1% year on year, with net profit nearly tripling to $479.2 million [8]. Its factories ran at 93.7% of capacity. Co-CEO Zhao Haijun told analysts SMIC will charge more this quarter for the silicon discs its factories process, saying the company needs “fairer pricing” [8]. This is a firm that spent 2023 and 2024 discounting into spare capacity, its factories only 68% full [8]. Export controls did the work: they walled off Chinese AI demand from TSMC and Samsung at the leading edge and handed SMIC a captive market [8].

Samsung raised prices on new orders across its 4nm, 5nm and 8nm manufacturing lines in July, by up to 15% [9][10]. Chinese customers, cut off from advanced chipmaking tools, are accepting the steepest increases [9]. Samsung’s contract chipmaking arm has lost money since 2022; the same export curbs that pushed Chinese designers toward it are now turning it profitable [9].

The squeeze runs the other way too. Japan’s Ajinomoto is a food company that also makes about 95% of the world’s ABF. That is the insulating film used in nearly every high-end chip package. It has reportedly told mainland Chinese customers it will cut their supply by 30%, following Beijing’s own curbs on rare-earth exports [11]. China’s self-sufficiency in the film is thought to sit below 5%. Ajinomoto also pushed through a roughly 30% price rise this quarter. The film is about 30% of what the base a chip sits on costs to build, so that rise flows straight into every advanced package [11]. Goldman Sachs expects the gap between supply and demand for these packaging parts to widen from around 10% in late 2026 to 42% by 2028 [11]. Chinese substitutes are already qualifying: Huazheng’s alternative film reportedly comes out usable more than 85% of the time [11].

Everyone is piling into the pinch points. LG has signed its first contract to supply a maskless laser patterning tool for chip packaging, aimed at the high end, as TSMC’s advanced packaging stays constrained [12]. The boom is showing up in national accounts. Japan’s exports hit a record 11.5 trillion yen in July, up 23.2%, on chip demand [13]. Shipments of chipmaking equipment rose 49.1% by value [14].

Who is actually paying for the computers

Broadcom is in talks to raise more than $60 billion of debt to finance AI chips for Anthropic and others, Bloomberg reported. Apollo and Blackstone may take part, and the total could reach $100 billion [15]. The structure matters. A special-purpose vehicle - a separate company created to hold the assets and the borrowing against them - issues the debt. It sits on that vehicle’s books, not Broadcom’s. Anthropic does not buy the chips; investors buy them and lease the hardware to it [15].

Nvidia is doing a version of the same thing. It agreed to guarantee up to $105 billion so OpenAI can lease a data centre in Pike County, Ohio, built by SoftBank-owned SB Energy [17]. It will also invest $1.5 billion in SB Energy itself [16]. The credit supports 4.25 gigawatts of computing capacity initially, with an option for 3.75 more, coming online from 2028 [17]. Nvidia had been discussing a guarantee of up to $250 billion before cutting it [17]. Asked whether this is circular funding - the chipmaker financing its own customers’ purchases - Huang said it is not. He said Nvidia is using “its scale and long-term visibility” to help [16].

Also this week

Nvidia published research arguing that the harness matters more than the model for long-running tasks. The harness is the memory, tools and supervisor code wrapped around a model. With a custom harness, Claude Opus 5 scored 100% on the ARC-AGI-3 reasoning benchmark. Without it, it scored 30% - still the best of any model tested [18]. Chinese lab Z.ai released GLM 5.3, a free-to-download model it says nearly matches the best closed models at coding and security work. It also released a code-scanning service built on it - useful for defenders, and equally available to attackers [19].

GitHub gave more detail on its 17 August outage, which ran 7 hours 47 minutes and broke Actions, pull requests and Copilot. CTO Vladimir Fedorov wrote “if you were trying to ship software that day, we let you down.” Monthly commits have gone from 1.4 billion in April to 2.9 billion; Azure now carries about 58% of the platform’s load [20]. Microsoft will retire the licence-included version of Azure VMware Solution. That closes one of the last routes to VMware without buying Broadcom’s full bundle. Sales stop on 31 October 2026 and customers must move by 30 August 2027 [21].

India ordered Google to shut hundreds of Firebase accounts after finding a pattern of criminals using the app-building tool to impersonate banks. At least 57 sites and databases came down in August alone. Indians lost about $2.4 billion to cyber fraud in 2025 [22]. Apollo Global confirmed a July breach exposing names, dates of birth, addresses and social security numbers, part of a wave of phone-call-based attacks on US financial firms [23]. And CareCloud confirmed the theft of medical records for more than 3.75 million people - the fifth-largest health data breach of 2026 so far [24].

02 · Lesson · why it matters

You can hold a lever or you can pull it, but not both

Power over someone who depends on you is worth most unused. Every time you use it, you pay for the thing that replaces you.

How it works

You control something the other side needs
That control is worth most while unused
You use it - and they now must replace you
The squeeze gives them the reason and the budget
They build their own; your control is spent
Both sides end up paying for two of everything

The twist

Power over someone who depends on you can be held or spent, never both - every time you use it, you fund the thing that makes you unnecessary.

Where you've seen this

A sole supplier

the first big price rise is the moment the customer starts qualifying a second source

A landlord and a tenant

push the rent hard enough and the tenant learns the neighbourhood well enough to leave

An employer with a rare skill

squeeze the specialist and the company finally writes down how the job is done

A friendship where one side always decides

each use of the upper hand teaches the other side they can manage alone

The catch

It only works if a replacement is actually buildable - some things really have no substitute, and then the squeeze holds for a long time.

Full lesson

The door opened onto a room that had emptied

Beijing spent eight months saying nothing about Nvidia’s H200 chips. Then, in recent weeks, it let some in. ByteDance and Tencent each got about 10,000 - roughly 2.5% of what China’s big three had been cleared to buy in January. The rest must sit in Hong Kong, which has neither the data centres nor the electricity to run them.

Read that as a supply story and it is confusing. Read it as a story about a lever, and it is very simple.

What a lever is worth

Washington’s chip controls were built on a real fact: Nvidia made something China could not make. That is a lever. And a lever like this has a strange property. It is worth the most while it is only threatened.

Unused, it shapes behaviour every day. Pulled, it does one thing once, and then the other side goes to work.

You can see the price of pulling it in the numbers. Nvidia held 66% of China’s market for AI chips in 2024. Then 40%. This year, an estimated 8%. Jensen Huang put it more bluntly to investors: 95% to zero. Chinese-designed chips are now forecast to take 90% of China’s market.

None of that was Huawei out-engineering anyone. It was a rule, applied, that handed Chinese buyers a reason to switch and Chinese designers a guaranteed customer.

The squeeze pays for the substitute

Look at SMIC, China’s only advanced chip factory. In 2023 it ran at 68% of capacity, discounting to defend its share, profits falling. This quarter it ran at 93.7%, posted its first $3 billion revenue quarter, and told customers it needed “fairer pricing.”

The export rules did that. They cut Chinese AI demand off from the world’s best factories, and every one of those orders had to go somewhere. Washington did not just fail to slow SMIC. It filled its order book.

This is the mechanism, and it is not about chips. Every squeeze does two things at once. It denies the other side something now, and it funds their exit. The harder the squeeze, the better funded the exit.

Both hands now hold a lever, and both are spending

The tell is what Beijing built to let the chips in. A new office that approves each purchase individually. Caps of 10,000 units per company. A rule that the chips must sit in Hong Kong, not the mainland.

Those are not Chinese inventions. They are, item for item, the instruments Washington has used since 2022: case-by-case licences, quantity caps, conditions on where a chip may end up.

And it is happening a layer down too. Japan’s Ajinomoto makes about 95% of a film that sits inside nearly every high-end chip package. After Beijing restricted rare-earth exports, Ajinomoto reportedly cut supply to China by 30%. China’s own version of that film reportedly already comes out usable more than 85% of the time.

Each side is teaching the other exactly which dependency to remove next.

Why nobody puts the lever down

Because a lever never used looks, from the outside, like not having one. The only way to prove you hold it is to spend some of it. And once you have spent some, the other side has started building, which makes spending more feel necessary.

There is also a quieter reason. The December approval that opened this door came with 25% of every sale going to the US Treasury. A rule written for security has a revenue line inside it. That does not make it wrong. It does mean more than one interest is riding on the same rule, and interests that get paid rarely argue for restraint.

The people who never held a lever pay for the copy

Running two of everything costs more than running one.

SMIC has raised its prices. Samsung raised prices on its advanced lines by up to 15%, with Chinese customers accepting the steepest rises. The film Ajinomoto makes went up about 30%. It is about 30% of what a chip package costs to build, so that rise arrives inside every advanced chip made anywhere.

Nobody in Washington or Beijing pays that. It lands in the price of a phone, a laptop, a car full of chips, a cloud bill, a hospital’s scanner. You are downstream of a lever you have never touched, in an argument you were never part of, and the bill arrives as a slightly worse price on something unrelated.

The part worth holding loosely is how little anyone could see. In January, 400,000 approved chips looked like a hold over China. By August the same order book looks like inventory - half a million H200s built for customers who found another way. Both capitals are still playing the position they had two years ago. Nobody at either table can see the whole board, and the ones who are certain they can are the ones to watch.

03 · Lab · your turn

Spending the Lever

Rehearse the trade between denying a rival now and keeping any hold over them later.

04 · Hope · carry this

Every supply that got cut taught someone else to make the thing themselves, and that knowing does not get taken back. Nobody planned it, but more of the world now knows how.