Somewhere in the sprawling semiconductor fabs of mainland China, production lines are running hot. Huawei’s Ascend 910C processor, the company’s most powerful AI training chip, is being manufactured at a pace the company plans to double by the end of 2026, according to Bloomberg reporting citing people familiar with the matter. The chip has become the default option for Chinese organizations that need high-end AI hardware, filling a vacuum left after Washington effectively locked Nvidia out of the market it once dominated.
Huawei claims the Ascend 910C delivers 2.8 times the performance of Nvidia’s H20, the chip Nvidia specifically designed to comply with earlier U.S. export rules. That figure has circulated widely in Chinese tech media since Huawei began promoting the chip, but no independent benchmark lab or academic institution has published results confirming it. The number should be understood as a manufacturer’s assertion, not a proven fact.
What is proven, however, is that the competitive landscape for AI chips in China has shifted dramatically, and the forces driving that shift are regulatory, financial, and industrial all at once.
Nvidia’s $5.5 Billion Problem
The turning point came in April 2025. Nvidia disclosed in a regulatory filing with the SEC that the U.S. government would require a license for exports of H20-class chips to China, including Hong Kong and Macau. The company estimated the financial damage at up to $5.5 billion, covering inventory write-downs, purchase commitments, and related reserves.
That filing transformed the H20 from Nvidia’s last legal bridge into the Chinese market into a stranded asset. The chip had been carefully engineered to slip beneath previous export-control thresholds, a product born of regulatory compromise. When Washington moved the goalposts, billions of dollars in inventory lost its destination overnight.
The policy framework behind that decision is now formalized. The U.S. Commerce Department’s Bureau of Industry and Security published a revised license review policy for semiconductors exported to China, explicitly naming Nvidia’s H200 and AMD’s MI325X accelerators. A related Federal Register notice established a case-by-case review framework with national security requirements for advanced computing components. Washington has not banned chip exports outright, but it has forced every shipment through an approval bottleneck that, in practice, has frozen most high-end AI chip sales to Chinese buyers.
As of mid-2026, Nvidia has not publicly detailed any successful large-scale H20 shipments to China under the new licensing regime.
Huawei Steps Into the Vacuum
With Nvidia sidelined, Huawei has moved aggressively. The Ascend 910C is now being marketed to Chinese cloud providers, state-backed research institutions, and enterprises building large language models. Bloomberg’s reporting, based on sources described as familiar with Huawei’s plans, indicates the company intends to double Ascend 910C output in 2026 to meet surging domestic demand.
That production ramp is the strongest available signal that Chinese buyers are not waiting for export-control relief. They are building AI infrastructure around domestically produced silicon.
The scale of adoption, however, remains opaque. Neither Huawei nor any major Chinese cloud provider, including Alibaba Cloud, Baidu, Tencent, or ByteDance, has published purchase volumes, order backlogs, or procurement contracts for the Ascend 910C. Industry watchers have noted deployments at several state-linked AI research labs, but detailed disclosures are scarce. Whether the production doubling is driven primarily by organic commercial demand, directed government purchasing, or a combination of both is an open question.
Beijing’s official stance on AI chip procurement strategy has not been laid out in any public document. The contrast with Washington is stark: U.S. policy is spelled out in granular regulatory language, while China’s response is visible mainly through corporate actions and supply-chain signals rather than formal government statements.
The Performance Claim Deserves Scrutiny
The 2.8x performance figure at the center of Huawei’s pitch requires careful context. Nvidia’s H20 was never meant to represent the company’s best technology. It was a deliberately constrained product, engineered to fall below U.S. export-control thresholds on metrics like chip-to-chip interconnect bandwidth and memory bandwidth. Outperforming the H20 is a lower bar than matching Nvidia’s unrestricted accelerators, such as the H200 or the company’s newer Blackwell-generation chips, which remain unavailable in China.
In other words, even if Huawei’s claim proves directionally accurate, it tells a narrower story than it first appears: Huawei may have built a chip that beats a product Nvidia intentionally held back, not one that rivals Nvidia’s best available hardware.
There are also system-level questions that chip specifications alone cannot answer. Large AI training jobs depend on interconnect bandwidth between chips, software maturity, compiler optimization, and ecosystem support. Huawei’s Ascend platform runs on its own software stack, called CANN, and its own cluster networking solutions. Independent evaluations of how Ascend-based systems perform end-to-end against Nvidia-based clusters are scarce. A chip that looks strong on paper can underperform in a 10,000-GPU training cluster if the software and networking layers are not equally mature.
Manufacturing Constraints Loom
Huawei’s ability to sustain its production ramp faces a fundamental constraint: fabrication technology. The Ascend 910C is widely believed to be manufactured by SMIC, China’s most advanced chipmaker, which operates without access to ASML’s extreme ultraviolet (EUV) lithography equipment due to separate export restrictions imposed by the Netherlands and the United States.
Without EUV, SMIC relies on older deep ultraviolet (DUV) lithography techniques, using multi-patterning workarounds to achieve smaller feature sizes. Industry analysts have questioned whether these fabrication lines can achieve the yields needed to support a production doubling without significant cost inflation or quality trade-offs. No public data from Huawei or SMIC confirms or denies those concerns.
This is the bottleneck that could determine whether Huawei’s ambitions match reality. Doubling chip output is straightforward if yields are high and wafer supply is abundant. It becomes far more expensive and uncertain when the underlying manufacturing process is pushing against its physical limits.
What This Means for the Global AI Race
The picture that emerges by mid-2026 is not one of settled technological dominance by either side. It is a story of rapid market reconfiguration under geopolitical pressure.
Three things are established with high confidence. First, U.S. export controls have severely curtailed Nvidia’s ability to sell high-end AI accelerators in China, costing the company billions. Second, Washington has codified a restrictive licensing regime that adds friction and uncertainty to every advanced chip shipment. Third, Huawei is investing heavily to scale up a domestically produced alternative, and the demand signal is strong enough to justify a production doubling.
What remains unresolved is just as important. The exact performance gap between the Ascend 910C and Nvidia’s export-limited chips has not been independently measured. The durability of Huawei’s manufacturing ramp, given yield and tooling constraints, is uncertain. The degree to which Chinese state entities are orchestrating procurement versus letting the market operate is unclear. And the system-level competitiveness of Huawei’s full AI training stack, from chip to software to networking, has not been rigorously tested against Nvidia’s ecosystem in any public forum.
For now, the balance of evidence points to a Chinese AI chip market that is decoupling from American suppliers faster than most forecasts predicted, with Huawei as the primary beneficiary. Whether that decoupling produces technology that genuinely competes with Nvidia’s best, or simply fills a gap with a good-enough alternative, is the question that the next year of benchmarks, deployments, and earnings reports will begin to answer.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
