For years, NVIDIA tried to thread the needle. Each time the U.S. government tightened export controls, NVIDIA engineered a new, deliberately hobbled chip to stay compliant and keep its foothold in China. The A100 got restricted, so the H800 appeared. The H800 got restricted, so the H20 emerged. When the H20 was banned, NVIDIA modified it again. The pattern was predictable: restrict, redesign, repeat. That game is now over. NVIDIA's fiscal 2027 outlook includes zero data center compute revenue from China. Jensen Huang has acknowledged the company's share of the Chinese AI chip market has shrunk to nothing. The most consequential technology decoupling of our era is no longer a slow drift, it's a clean break.

The long road to zero

The timeline of NVIDIA's retreat from China reads like a case study in escalation. In October 2022, the Biden administration imposed the first round of semiconductor export controls targeting China, restricting sales of the A100 and H100. NVIDIA responded by designing the H800 and A800, chips tuned just below the performance thresholds set by the Bureau of Industry and Security. When those were restricted too, the H20 arrived, a chip powerful enough for inference tasks but deliberately crippled for large-scale training. The Trump administration took things further. In April 2025, it banned H20 exports outright, costing NVIDIA a $4.5 billion inventory write-down and $2.5 billion in lost Q1 revenue. NVIDIA tried yet again, modifying the H20 in May 2025 with plans to ship a downgraded version by July. By August, the White House reversed course and issued export licenses for the H20, but Beijing responded by urging domestic firms to stop buying it. NVIDIA's CEO said he was "disappointed" as China effectively de facto banned purchases of the chip. The saga continued with the H200. In December 2025, the U.S. approved exports of its most advanced Hopper-generation chip to China, subject to a 25% revenue-sharing arrangement. Chinese companies placed orders for over two million units. Then in January 2026, Chinese customs blocked the shipments. Suppliers paused production. It wasn't until March 2026 that Beijing finally approved H200 sales, but by then the damage was done. NVIDIA has generated zero revenue from China chip sales, and its FY2027 outlook assumes no China data center revenue at all. Every attempt to stay in the market was met with a new restriction, from Washington, from Beijing, or from both.

Decoupling creates competitors, not dependence

The conventional wisdom behind export controls is that cutting off access to advanced chips will slow China's AI progress. The reality is more complicated. Restrictions have accelerated China's domestic chip push in ways that may ultimately produce real competitors. Huawei's Ascend 910C, fabricated on SMIC's 7nm nodes without access to EUV lithography equipment, now delivers roughly 60% of the inference performance of NVIDIA's H100. That's not parity, but it's not insignificant either, especially at scale. DeepSeek researchers have validated this figure, and Huawei already holds roughly 50% of China's AI chip market in 2026. Behind Huawei, a wave of domestic players is emerging. Cambricon targets 500,000 AI accelerator shipments in 2026, including 300,000 units of its flagship Siyuan 590 and next-generation Siyuan 690. Alibaba's T-Head, Baidu's Kunlunxin, and Hygon all contribute to an ecosystem that captured 41% of China's AI accelerator server market in 2025, according to IDC data. Chinese chip firms posted record-high revenue for 2025, driven by both AI demand and the self-sufficiency mandate. China's chip industry leaders candidly admit a five-to-ten-year lag in data center chips compared to NVIDIA's cutting edge. But "good enough" alternatives at scale, backed by government procurement mandates, can sustain a parallel AI ecosystem. SMIC has warned that rushed capacity expansion could leave some data centers idle, but the investment momentum is undeniable. The historical pattern holds: countries cut off from foreign technology don't simply wait. They build.

The weaponization of supply chains

NVIDIA's exit from China is not an isolated event. It's part of a broader pattern in which the U.S. government is using technology supply chains as instruments of statecraft, on multiple fronts simultaneously. Consider what's happened in parallel. Anthropic, one of America's leading AI companies, was designated a "supply chain risk" by the Pentagon in March 2026, the first time an American company has ever received that label. The designation came after Anthropic refused to remove safety guardrails from its AI models for military use. A federal court denied Anthropic's motion to lift the label, and the standoff has exposed deep fissures in the relationship between AI companies and the U.S. government. The message from Washington is clear: technology access is contingent on alignment with government priorities. For NVIDIA, that meant navigating export controls that changed every few months. For Anthropic, it meant facing retaliation for maintaining safety policies. For Chinese firms, it meant being cut off entirely. This is not just chip policy. It's the emergence of a new framework where computing power, AI models, and semiconductor supply chains are treated as strategic assets to be controlled, withheld, or leveraged. U.S. senators have pushed to suspend NVIDIA's export licenses entirely. The AI Overwatch Act would give Congress 30-day review authority over any advanced chip exports to adversaries. The Bureau of Industry and Security, which manages these controls, has lost nearly 20% of its licensing staff, creating bureaucratic bottlenecks on top of policy uncertainty. The rules of engagement for the global technology industry are being rewritten in real time.

ASEAN as the neutral broker

As the U.S. and China split into two chip ecosystems, a third geography is quietly positioning itself in the middle: Southeast Asia, with Singapore and Malaysia at the center. Singapore has emerged as a preferred "gray zone" for Chinese firms seeking overseas computing power. With U.S. export controls making it harder to operate advanced chips on Chinese soil, Chinese cloud providers and AI companies have turned to data centers in Singapore and Malaysia to access the hardware they can no longer import. The region now hosts the most Chinese-owned data centers outside China. The numbers are significant. Southeast Asia has 370 data centers and growing, with data center demand projected to increase 20% annually through 2028. GITEX AI Asia 2026 in Singapore showcased what organizers called a $78 billion inflection point in Asian AI and quantum computing infrastructure. Singapore's National AI Strategy, first launched in 2019, positions the city-state as both a regulatory leader and an infrastructure hub. But this neutral-broker role comes with risks. A New York Times investigation revealed that at least one Singapore-based company, Megaspeed, allegedly smuggled NVIDIA chips into China while also leasing compute to Chinese companies from Southeast Asian data centers. Malaysia has begun reining in data center growth partly in response to concerns about export control circumvention. U.S. authorities are watching closely. The opportunity for ASEAN is real: being the neutral ground where both ecosystems can operate. The risk is equally real: becoming the channel through which export controls are undermined, inviting the same restrictions that drove NVIDIA out of China in the first place.

The training cost gap will widen, until it doesn't

With NVIDIA's exit, the cost of training frontier AI models in China is about to increase substantially. The H200 offers roughly six times the performance of the H20, and NVIDIA's Blackwell and Vera Rubin architectures push even further ahead. Chinese alternatives can't match that performance for large-scale training, even if they're competitive for inference. This matters because the race to build frontier models depends heavily on available compute. NVIDIA projects AI data center revenue from its latest platforms will reach $1 trillion by 2027. The companies with access to those chips, overwhelmingly in the U.S. and allied nations, have a structural advantage in training the next generation of models. But China has a track record of finding workarounds. DeepSeek demonstrated that algorithmic efficiency can partially compensate for hardware limitations. The Ascend 910C's 60% inference performance relative to the H100 is being achieved on SMIC's 7nm process, without the EUV tools that produce NVIDIA's most advanced chips. And reports of chip smuggling, from Super Micro's co-founder being charged to Megaspeed's alleged operations in Southeast Asia, suggest that hardware finds its way to where demand exists, regardless of policy intent. The training cost gap will widen in the short term. Whether it stays wide depends on Chinese firms' ability to innovate around constraints, something they've done repeatedly.

What happens next

For developers and companies building on AI infrastructure, NVIDIA's departure from China concentrates GPU supply further into Western markets. Cloud computing costs outside China may stabilize or even decrease as NVIDIA redirects inventory. Inside China, a parallel ecosystem of domestic chips, domestic cloud, and domestic AI frameworks will continue to grow, potentially incompatible with the tools and standards most of the world uses. For Singapore and the broader ASEAN region, the next few years will determine whether the neutral-broker position is sustainable or whether it becomes a liability. If the region can maintain credible export control compliance while offering infrastructure to both ecosystems, it stands to capture an outsized share of global AI investment. If enforcement concerns mount, the window could close. For the U.S. and China, the AI chip decoupling is now complete, but the competition is just beginning. Export controls didn't stop China's AI industry. They reshaped it. The question is no longer whether two separate AI ecosystems will exist, but how they'll interact, compete, and coexist in the years ahead.

References