The strategy was straightforward: restrict advanced American chips to China, and the US would maintain an unbeatable lead in artificial intelligence. Instead, the policy triggered the exact opposite outcome. NVIDIA, which once controlled between 66% and 95% of China's AI accelerator market, now holds zero percent. Meanwhile, Huawei has emerged as the dominant player in a parallel, fully domestic Chinese AI hardware ecosystem that is already printing billions in revenue.

What Happened to NVIDIA's China Strategy?

In October 2022, the Biden administration implemented sweeping export controls on advanced computing chips, adding NVIDIA's flagship H100 and A100 processors to the Commerce Control List. The logic seemed airtight: starve Chinese AI development of American silicon, and American technological dominance would be preserved by default.

NVIDIA responded by engineering the H20, a deliberately throttled version of its high-end hardware designed to stay within legal performance thresholds for export to China. The company invested heavily in this middle-ground approach, betting it could maintain market access while complying with US policy. In April 2025, the Trump administration restricted the H20 anyway. NVIDIA took a $5.5 billion inventory charge that quarter, effectively writing off its China strategy.

"Conceding an entire market the size of China probably does not make a lot of strategic sense, so I think that has already largely backfired," stated Jensen Huang, CEO of NVIDIA.

Jensen Huang, CEO at NVIDIA

Huang's assessment was blunt and direct. The policy that was supposed to cripple Chinese AI development had instead forced Beijing to build its own alternative. Markets do not tolerate vacuums. With NVIDIA sidelined, every Chinese hyperscaler, AI startup, and state research lab pivoted to domestic alternatives almost overnight.

How Did Huawei Become the Dominant Player So Quickly?

Huawei's Ascend chips do not match NVIDIA's Blackwell processor on raw training performance, and the company makes no pretense that they do. Instead, Huawei made a strategically shrewd bet: optimize for inference, not training. This distinction matters enormously because inference, the process of running already-trained models to generate outputs, dominates AI compute spending the moment models leave the research lab.

The results have been staggering. Huawei is now positioned to capture roughly 60% of the Chinese AI chip market by the end of 2026. AI-chip revenue at Huawei is forecast to reach approximately $12 billion in 2026, up 60% year-over-year, driven by orders for the Ascend 950PR processor that entered mass production in March 2026.

To put this in perspective, Cambricon, another Chinese chip manufacturer, posted $423 million in revenue in a single quarter during the first quarter of 2026, with 160% year-over-year growth and net profit up 185%. The domestic Chinese chip ecosystem is not a future story; it already generates substantial cash.

Why CUDA's Dominance Is Eroding in China

NVIDIA's real moat was never just the hardware. It was CUDA, the software framework that made NVIDIA chips the default choice for AI developers worldwide. CUDA created a powerful network effect: more developers wrote for CUDA, which made NVIDIA hardware more valuable, which attracted more developers.

That network effect is now breaking down in China. Every Chinese developer who can no longer buy NVIDIA hardware is forced to write for CANN, MUSA, or other non-CUDA targets. Each developer who switches represents one less person reinforcing the network effect that made NVIDIA's gross margins defensible. Over time, this erodes NVIDIA's competitive advantage in the world's second-largest economy.

Steps to Navigate the Bifurcating AI Hardware Market

• Design for Hardware Portability: Tightly coupling your AI stack to a single chip vendor is now a strategic vulnerability disguised as an optimization. Companies should architect systems to run on multiple hardware platforms from day one, ensuring they can adapt if regulatory restrictions change or new competitors emerge.
• Invest in Hardware-Agnostic Infrastructure: Tools that compile models efficiently across heterogeneous hardware are becoming critical infrastructure. RadixArk recently raised $100 million at a $400 million valuation specifically on this thesis, with NVIDIA itself as a co-investor, signaling that even NVIDIA recognizes the market is shifting toward portability.
• Build Compliance Functions: The interaction of US export controls and Chinese domestic procurement mandates is now complex enough to require dedicated teams. Rules changed in 2025, changed again in 2026, and will continue changing. Companies need either internal expertise or external consultants to navigate this landscape.

What Does This Mean for Global AI Infrastructure?

The global AI hardware market is now splitting into two distinct ecosystems with separate software stacks, separate foundry supply chains, and separate sets of model weights. This bifurcation is not a temporary disruption; it is a structural reordering of where AI capital expenditure flows.

McKinsey estimates that 30 to 40 percent of all global AI spending will be influenced by sovereignty concerns by 2030, representing a $500 billion to $600 billion market. Over $100 billion is already committed globally to sovereign AI infrastructure in 2026. This is not a niche government procurement story; it is a fundamental shift in how companies evaluate hardware investments.

The battleground has moved from training to inference. Companies evaluating AI hardware must now consider not just raw performance on benchmark tests, but also where their models will actually run in production. A chip that excels at training but struggles with inference optimization may be the wrong choice for a company focused on cost-effective deployment.

The export control regime is unlikely to be reversed. Industry experts and investors are now positioning for the fracture to deepen rather than heal. The next three years of go-to-market strategy, hiring decisions, and capital allocation should assume the bifurcation continues, not that the two ecosystems eventually reunify.