domingo, 22 de noviembre de 2026

domingo, noviembre 22, 2026

Chips and cracks

How China continues to evade America’s tech-export rules

And is also developing a semiconductor industry to render them moot

Illustration: Daniel Stolle


“WE WILL NOT let anybody have them other than the United States,” declared Donald Trump last year of Nvidia’s most advanced chips—an essential tool in the race to develop superpowered AI. 

To enforce the pledge, his administration maintained rules restricting sales of certain types of semiconductor. 

Anything less, a Republican congressmen said, would be as foolhardy as handing Iran weapons-grade uranium.

Yet all the while, it turns out, a torrent of Nvidia’s nattiest chips was flowing to China or to middlemen who made them available to Chinese AI firms. 

Although America is now trying to plug the sieve of leaks, China’s efforts to obtain the latest technology are continuing by other means. 

What is more, China’s own chipmaking industry is advancing so fast that it may soon render America’s controls moot.

The scale of the leakage has been revealed by the downfall of Jensen Huang, Nvidia’s boss. 

While he was having an affair with a suspected Chinese agent named Alice Huang, the firm she worked at, Megaspeed, became the biggest buyer of Nvidia’s chips in South-East Asia. 

Many of these seem to have ended up in China, in breach of America’s regulations; others power data centres in the region used by Chinese clients—a set-up that had been legal, although this week American officials hastily revised their rules to curb it. 

Alibaba, a Chinese tech giant, is Megaspeed’s biggest customer and one of its biggest creditors.

Chunky chips

Since America first started restricting the export of advanced chips in 2020, Chinese firms have devised an ever-shifting set of workarounds to access “compute” (processing power). 

Huawei, a Chinese tech giant, was singled out by American sanctions that year, cutting it off from components made by TSMC, Taiwan’s biggest chipmaker. 

Yet when TechInsights, a research firm, dissected one of Huawei’s Ascend 910B processors, which are intended for use in AI, they found chips made by TSMC. 

According to SemiAnalysis, a consultancy, Huawei bought some $500m-worth of advanced TSMC chips through Sophgo, a Chinese chip-designer. 

(Sophgo says it had no dealings with Huawei.)

When the restrictions were broadened to all Chinese firms, many resorted to smuggling. 

In 2024 the Wall Street Journal detailed a “barely concealed” chip-trafficking network, including a student paid $600 to sneak Nvidia chips into China in his suitcase. 

But it was not just small-time bootlegging. 

America’s Department of Justice claims a co-founder of Supermicro, an American server-maker, and two other men conspired to sneak billions-worth of AI kit into China between 2024 and 2025 (the co-founder denies it). 

The trio allegedly ordered chips through a South-East Asian shell company, then repackaged them into unmarked boxes that were sent on to China. 

Security footage shows one of them placing fake equipment in the original packaging to make it seem like the chips had never left South-East Asia, to trick both compliance staff at Supermicro and American officials.

No one is quite sure how much smuggling there has been. 

SemiAnalysis thinks roughly 1.5 gigawatts of compute, which is equivalent to about 1m restricted Nvidia chips, have ended up in China. 

This stockpile is especially useful for Chinese firms since access to it “can’t be cut off”, notes a former China analyst at the CIA. 

But it is not big enough to satisfy China’s ever greater hunger for compute. 

Smuggled chips provide around 22% of its AI data-centre capacity, according to the Institute for Progress (IFP), a think-tank. 

Moreover, smuggling routes are becoming inconvenient. 

Since Malaysia tightened permitting rules for high-end semiconductor shipments last year, smuggling to China plummeted, Malaysian officials insist.

An alternative workaround is to exploit loopholes in the rules. 

AI processors depend on high-bandwidth memory (HBM), stacks of memory chips that China cannot yet make in large numbers. 

Before America banned their sale to China in late 2024, Huawei stockpiled some 13m HBM stacks, enough for 1.6m of its advanced AI chips, reckons SemiAnalysis. 

Since then, it says, Huawei has kept up supplies by getting HBM chips grafted onto other, outmoded chips. 

The combined product does not exceed the technical thresholds specified in America’s rules and so can be shipped to China legally, where it is disassembled again to get at the HBM chips.

An even bigger loophole has involved renting, rather than buying, computing power. 

Instead of trying to spirit restricted chips across borders, Chinese firms have simply been making use of them from afar. 

Remotely accessed capacity makes up a quarter of China’s total AI compute, estimates IFP. 

This, too, flows through South-East Asia. 

After Singapore limited data-centre construction for lack of space and power, projects gravitated to Malaysia, where data-centre capacity has grown 100-fold in the past five years. 

It now boasts 55 big data centres, with another 62 in the pipeline. 

Taiwan, the source of most of the world’s high-end chips, exports a majority of them to America and Mexico (where they are assembled into goods bound for America). 

Exclude those two countries, and Malaysia, Singapore and Thailand absorb more than half of the rest (see chart 1), an astonishing share of global demand that is clearly not catering simply to domestic consumers.


Oracle, an American cloud-computing firm, for instance, runs data centres in South-East Asia that work with ByteDance, the Chinese owner of TikTok, which also owns an advanced AI lab. 

A cottage industry of lesser-known data-centre firms has sprung up to sell remote access to Chinese customers. Authorities had felt little urgency to stop them. 

“Just the act of bringing in and using Nvidia chips is not illegal under Malaysian law,” says Raja Nushirwan Zainal Abidin, Malaysia’s director general of national security.

Many South-East Asian officials had believed the American authorities were aware that the rules were being skirted on a giant scale. 

After all, says another Malaysian official, it was American tech firms that were providing the computing power: “Everyone assumed the US was happy with it.” 

Singapore launched a probe into Ms Huang and Megaspeed, but says it got little co-operation from the American authorities.

America’s own treasury secretary, Scott Bessent, has stated that the Department of Commerce stymied efforts to ban Megaspeed from buying chips. 

Even after Nvidia stopped selling Megaspeed chips last year, it received “explicit permission” from the department to continue providing Megaspeed spare parts for repairs, an Nvidia spokesperson says. 

Several figures within the administration have long argued that imposing overly strict controls simply spurs the development of China’s indigenous chip industry, leaving America with diminished leverage.

If the Trump administration’s intention was to slow down the Chinese chipmaking industry, it has not worked. 

Indeed, China has worked just as hard if not harder to gain access to the wherewithal to make chips at home as it has to get hold of fancy foreign chips.

It started with a hunt for skilled engineers. 

Taiwan not only makes chips, but also has a big chip-design industry. 

After China launched a big investment fund to foster a chipmaking industry in 2014, mainland firms dangled salaries two or three times those in Taiwan, sometimes with school fees and subsidised flats thrown in. 

By 2019 some 3,000 Taiwanese engineers, nearly a tenth of the island’s semiconductor workforce, were employed on the mainland, according to the Taiwan Institute of Economic Research.

Many began drifting home during China’s covid-19 lockdowns. 

After 2022, when America limited exports of chips and chipmaking tools to most Chinese firms, a job on the mainland began to look like a career risk. 

Those laid off by Chinese firms found it “really impossible” to get work in Taiwan, says Shih Ping-an of the Science & Technology Law Institute, a think-tank.

Chipping in

But Chinese firms had already found a way round engineers’ reluctance to emigrate: employing them secretly in Taiwan. 

Chiang Min-yen of DSET, a Taiwanese think-tank, calls it “remote poaching”. 

Firms from the mainland need the Taiwanese government’s approval to operate or recruit on the island, so they hide behind local fronts. 

Typically, says Ms Shih, a shell company hires a single engineer, who is then asked to persuade colleagues to jump ship too. 

Sometimes an entire R&D team walks out at once.

Bitmain, a maker of chips for crypto-mining based in Beijing, was a pioneer. 

From 2018, prosecutors later alleged, it secretly ran two firms, IC Link in Hsinchu, the centre of Taiwan’s chip industry, and WiseCore near Taipei. 

The pair hid their Chinese ownership and hired more than 100 engineers from local chip-designers such as MediaTek, some at double their previous pay. 

The aim, prosecutors said, was to build expertise in AI chips. 

In 2021 four executives admitted wrongdoing and were fined. 

But the model spread.

Taiwan’s Ministry of Justice Investigation Bureau, a law-enforcement agency, has completed 166 investigations into illegal Chinese operations and talent-poaching since 2020, according to Rest of World, a tech publication. 

All 36 of the cases reviewed ended in convictions, of which 33 involved chip design or research. 

Among those accused is SMIC, China’s leading chipmaker, which is alleged to have run a Taiwanese subsidiary disguised as a company from Samoa. 

In August agents searched 64 sites and questioned 114 people at 17 more Chinese-funded firms.

Chen Kuan-ting, a legislator from the ruling Democratic Progressive Party, reckons poaching is now much harder. 

Local firms pay well. Engineers who want to emigrate are more interested in moving to America than the mainland. 

Mr Chiang is less sure. 

Chinese firms, he believes, have not given up: “They are taking more effort to hide their identity.”

Running factories that make cutting-edge chips also requires skilled engineers. 

Know-how stemming from decades of trial and error is as important as whizzy equipment. 

A young engineer’s knowledge of processes can be of limited use outside the fab (chipmaking plant) where he or she works, says a senior executive at a big Taiwanese chip company. 

Senior people, in contrast, understand the bigger picture and are “capable of creating real damage” if poached. 

They are also few in number. 

Chris McGuire of the Council on Foreign Relations, an American think-tank, says there are “probably 30 to 50 people at TSMC” who know things no one else does.

Just how much damage a handful of veterans can do is clear from a case in South Korea. 

In 2016 a 28-year veteran of Samsung, South Korea’s biggest chipmaker, joined CXMT, then a Chinese memory-chip startup. 

Prosecutors said he and other former Samsung staff handed over the firm’s recipe for making DRAM, a type of memory chip. 

One of the defectors copied hundreds of steps by hand into notebooks to dodge Samsung’s internal monitoring of electronic communications. 

In 2019 CXMT became the first Chinese firm to mass produce DRAM. 

In April a South Korean court jailed the veteran for seven years. 

When CXMT floated on Shanghai’s stock market earlier this year it became one of China’s most valuable firms.

TSMC, too, has been a target, of the Chinese state specifically according to Taiwanese prosecutors. 

In July they indicted a former deputy manager at the firm. 

They say he was recruited, through a go-between, by Ding Xiaohu, a Chinese national acting on instructions from China’s Central Military Commission. 

Between 2023 and 2024 the pair allegedly planned a chip-materials company on the mainland. 

The wayward manager drafted a plan to woo Taiwanese talent. 

He is accused of copying 21 documents containing trade secrets and taking them home. 

TSMC’s own investigators caught him, however, and prosecutors say all the files were recovered before any technology changed hands.

China is also on the hunt for chipmaking machines and especially extreme ultraviolet (EUV) lithography devices. 

Only one firm in the world makes them, ASML, which is based in the Netherlands. 

Since 2019 America has banned the sale of them to China. 

Without EUV, Chinese fabs struggle to make cutting-edge processors, smaller than seven nanometres, reliably. 

In June Howard Lutnick, America’s commerce secretary, told ASML’s bosses that one of the machines, or parts of one, may have reached China. 

ASML denies this, saying it knows the location of every EUV machine it has ever built.

Rumours persist that China is building its own. 

In December Reuters, a news agency, reported that a team in China including former ASML engineers had completed a prototype. 

But making a machine that can churn out chips quickly, with few defects, is far harder. 

Analysts think China will need 10-15 years to master that.


Even without EUV, China’s chipmakers have come a long way.

In 2021 Chinese firms supplied less than a tenth of the chipmaking equipment bought in China. 

This year they will supply more than a quarter, worth some $16bn, according to Bernstein, a research firm (see chart 2). 

For the most critical step, lithography, Chinese fabs still rely on ASML’s older deep ultraviolet (DUV) tools, and they have been buying them in bulk. 

By early 2026 Chinese-owned fabs had acquired some 340 of ASML’s DUV machines, estimates the Centre for Technology & Statecraft, an American think-tank, more than 180 of them in 2024 and 2025 alone. 

With these tools Chinese fabs can make chips of seven nanometres and a little below, but with more defects and at a higher cost.

Such cavils notwithstanding, China’s chip industry is clearly progressing rapidly. 

DeepSeek, China’s leading AI lab, reportedly plans to pre-train its next model, V5, on Huawei’s chips. 

If it succeeds, that would be the clearest proof yet that America’s controls on chip exports to China are a wasting asset.

Chip and pin

America has responded to China’s dogged pursuit of self-reliance with inconsistency, prevarication and bickering. 

Joe Biden, Mr Trump’s predecessor, steadily tightened restrictions on chip exports. 

The sweeping controls on sales to China came in 2022. 

A year later the rules were tightened to block chips that Nvidia had designed specifically to escape the earlier limits. 

Then in 2024 HBM chips were added to the list. 

Days before Mr Biden left office, his team tried to tighten the screws yet further, unveiling an “AI diffusion” rule that sorted the world into tiers, each with different access to American chips.

Mr Trump has been far more relaxed. 

His administration scrapped the diffusion rule before it took effect. 

Last year it told Nvidia it needed a licence to sell the H20, a chip made for the Chinese market, before agreeing a few months later to grant licences in exchange for a share of the revenue.

In December it went further, approving exports of Nvidia’s far more powerful H200, though China has been reluctant to let its firms buy it, since it wants to encourage use of domestic chips.

On November 16th the pendulum swung back. 

The Commerce Department announced new rules to stop China using “advanced computing power based on US technology”, including through remote access. 

Additional licences will be needed to move chips into and within Malaysia, Singapore and Thailand (and Brazil and Indonesia as well). 

Some in South-East Asia worry that cutting off Chinese demand for remote access could be a “game-over” moment for the region’s data-centre boom, stranding expensive investments and precipitating a painful bust.

Various bills before Congress would tighten the rules more. 

One would give lawmakers oversight of the Commerce Department’s decisions on licences for AI chips. 

Another would require chipmakers to make sure their chips are not diverted. 

A third would extend controls on EUV machines to DUV ones, too, further hampering Chinese chipmaking.

Clamping down on remote access is easier said than done, however, not least because Chinese firms are adept at hiding behind shell companies. 

Rather than a sudden stop, the new rules are more likely to split the market according to how much legal risk each data-centre operator will tolerate, says Ong Kian Ming, a former Malaysian deputy minister of trade.

Eliminating smuggling is harder still. Much of the policing falls to the firms themselves—something they say is harder than it sounds. 

Nvidia and AMD, another big chipmaker, sell mostly through server-makers such as Dell and Supermicro, which in turn sell to cloud firms and other middlemen. 

Customers are cagey about sites and plans for commercial reasons. 

Refusing to sell to suspect ones can invite complaints of anti-competitive behaviour.

Moreover, impeding smuggling and remote access will do little to slow the development of China’s indigenous industry. 

Kyle Chan of the Brookings Institution, a think-tank in Washington, thinks further controls will yield diminishing returns. 

American restrictions have already spurred a drive for self-sufficiency. 

“That horse is sort of out of the barn,” he says. 

But that does not mean that all is lost: to build a cutting-edge industry, he argues, China needs people with the know-how to make and run the most critical components. 

In the end, it is all about talent. 

The front line of the chip wars may lie not in Silicon Valley or in Singapore, but in Hsinchu.

Next
This is the most recent post.
Entrada antigua

0 comments:

Publicar un comentario