On 13 January 2026, the US Commerce Department revised one of the key choke points in the AI supply chain, changing the rules governing exports of Nvidia’s H200 and AMD’s MI325X and comparable AI chip to China from a presumption of denial to case-by-case licensing conditional on US supply assurances, recipient security procedures, and independent third-party testing. The change was designed to reopen a narrow, compliance-heavy channel for sales.
The market reaction has been to treat the change as permission without predictability. Reporting described H200 sales to China as effectively stalled while a national-security review remains unresolved and buyers avoided placing orders until they can see the conditions and assess approval odds. The events followed closely the precedent set by the H20: in April 2025 Nvidia disclosed that H20 exports to China would require an authorization tied to bandwidth thresholds for an ‘indefinite’ period. Then, sales partially resumed under an inherently undependable and politicized licensing regime.
China’s workaround for Nvidia is increasingly systems-level rather than chip-for-chip. Instead of trying to ‘beat Nvidia per accelerator,’ Huawei is attempting to replicate the function of the Nvidia stack: a tightly coupled, multi-rack machine that behaves like a single computer, paired with a software toolchain that reduces the friction of using domestic hardware. Huawei explicitly attempted to offer a domestic, rack-scale substitute for Nvidia’s GB200 NVL72 by launching the Atlas 900 A3 SuperPoD. This AI cluster scales-out the company’s CloudMatrix 384 super-node with 384 Ascend 910C accelerator chips across 16 racks operating as ‘a single computer.’
Despite the appearance of direct competition with Nvidia, each Ascend 910C can deliver only about 40% of the older H100 chips. Considering that, depending on the benchmark, the H100 achieves 45-61% of the performance of the newer B200 chips with which the GB200 is equipped, the per-chip gap is undeniable. Yet the swarm of lower-grade chips in the Atlas 900 A3 brute forces performance outstripping Nvidia’s GB200, albeit at the cost of materially worse performance-per-watt. While power-consumption ranks as a top concern for AI hyper-scalers in the US and is holding back Europe, the Atlas 900 A3 is a tolerable bargain for Beijing, whose immediate strategic objective is not best-in-class efficiency, but a domestically designed and produced AI compute. Huawei embraced this ‘good-enough’ approach when it announced the Atlas 950 and the 960 SuperPoDs equipped with 8,192 and 15,488 AI chips, respectively, to compensate for individually weaker processors.
Historically software and interconnect have been this approach’s weakest spots. Unreliable compilers, kernel bugs, and slow inter-node fabrics make it harder to achieve acceptable tokens-per-dollar from less capable accelerator chips, especially for training. Huawei’s response is to co-design the machine and the toolchain. The CloudMatrix384 architecture centers on a high-bandwidth ‘Unified Bus’ network intended to enable all-to-all communication and dynamic resource pooling across 384 chips. On the software side, Huawei offers CANN as a fully ‘open-source’ toolchain designed to replace Nvidia’s CUDA computing platform. The case of Chinese AI developer DeepSeek publishing resources to make their models immediately deployable on Huawei hardware using CANN shows the first sign of ecosystem-level synergies in the Chinese AI stack. Even if the migration remains partial and training still appears more fragile than inference, this combination of hardware scale, viable interconnect, and open-source tooling can shift engineering effort and operational ‘muscle memory’ away from the US tech stack and increase China’s AI self-reliance.
RAM ‘Made in China’ as Enabling Layer
A domestic accelerator can be ‘good enough’ for serious AI deployment only if the surrounding memory supply is also reliable. If memory remains imported, then the system-level workaround simply trades one external choke point (chips) for another (dynamic random-access memory, or DRAM and packaging throughput). Recent US export-control moves have explicitly treated high-bandwidth memory and the tools needed to manufacture it as strategic constraints, which is exactly why Beijing has pushed to indigenize memory alongside compute.
China’s industrial policy has been explicit about this objective for a decade. The ‘Made in China 2025’ plan, published in 2015, set self-reliance targets across strategic technologies, semiconductors included. The ‘National Integrated Circuit Industry Investment Fund’ (‘Big Fund’) translated those priorities into state-directed capital formation. According to the Chinese company database, memory producer ChangXin Memory Technologies (CXMT) was established with a capital of RMB 24bn in 2017. Authorities at various levels of government invested an estimated RMB 220bn in industrial scaling aimed at making CXMT a viable venue for domestic DRAM production. In 2020, CXMT was the first Chinese company to complete its own DRAM fab.
The exact figures vary by source and quarter, but CXMT’s trajectory is clear. In 2024, four years after it first emerged as a challenger to the SK Hynix-Samsung-Micron oligopoly that has long dominated DRAM production, CXMT recorded a roughly 5% market share. In 2025, CXMT maintained its position and even improved it to an estimated 8% of the market in the last quarter, Last year, it manufactured 270,000 silicon wafers per month, marking a 68% increase year-on-year and a 13-fold jump compared to 2020. The rapidity of CXMT’s expansion led analysts to forecast Chinese overcapacity due to scale driving cost and learning effects. The key point is not that domestic supply is sufficient on its own, but that it is now large enough to be strategically relevant.
Quality and integration are the harder step. For commodity DRAM, credibility is built through systematic qualification by large equipment manufacturers (OEMs) and designers (ODMs); for AI accelerators, it is built when domestic DRAM can evolve into HBM-grade output with acceptable yields, stacking, and packaging. The US Department of Defense’s decision to include CXMT to the Section 1260H list of entities identified as Chinese military companies in January 2025 raised compliance and reputational frictions for US-linked supply chains, even if it is not an automatic blanket ban. At the same time, market insiders indicate that major OEMs (including Dell and HP) have explored qualifying CXMT memory while major suppliers of DRAM like Kingston, Thermaltake, Lexar, and Adata are sourcing CXMT chips. Validation and wider use will benefit CXMT’s scale and process maturity even if they cannot, by themselves, lessen the challenge of ‘jumping’ from standard DRAM to high-bandwidth memory (HBM) for AI.
HBM is where the ‘good-enough’ AI stack either becomes real or remains a political pet project. In 2024, CXMT began mass-producing old-generation HBM2 earlier than previously expected but yields and scalability remain uncertain. CXMT is actively developing previous-generation HBM3 and current-generation HBM3E for AI despite reportedly operating dated lithographic machines only capable of 19nm production processes. As export controls on advanced equipment tighten, CXMT risks being trapped by the same dynamic that engulfed another Chinese tech champion, YMTC. China’s only escape hatch in DRAM production is increasingly becoming scale growth paired with the development of domestic lithographic machinery.
Reality Check: Export Controls Both Delay a Chinese AI Stack and Make It Inevitable
Export controls have curbed China’s access to the US AI stack, but they also make a domestic stack a rational inevitability. The shift, in January 2026, to case-by-case licensing for the export of Nvidia H200, AMD MI325X, and comparable accelerators did not restore predictability. On the contrary, it formalized a discretionary, compliance-heavy political gatekeeping that makes the US technological stack an unreliable input for Chinese AI firms. Over time, its effect seems to be less ‘China cannot buy AI chips’ than ‘China cannot reliably build on the US stack’; and that reliability problem is precisely what industrial policy programs such as Made in China 2025 were designed to mitigate.
In the near term, the strategic logic is ‘good enough’ rather than hardware leadership in frontier AI. Beijing wants a domestic AI stack under the Communist Party’s control that can deliver acceptable tokens-per-second at a workable latency and a functional software ecosystem. Huawei’s CloudMatrix384 is best understood, in this frame, as a ‘production-grade’ supernode capable of meaningful hardware–software co-optimization. The multi-thousand-chip scale-out promised by the Atlas 950 and 960 can improve the current generation’s lead over US technology in inferencing workloads, even if Nvidia will maintain an edge in model training, power efficiency, and overall ecosystem maturity.
The longer-term question is whether ‘good enough’ can become a robust ecosystem for large-scale training, rather than focusing on inference and fine-tuning. That requires enabling layers where China’s AI stack has developed unevenly. First, memory and advanced packaging at high yields: HBM-grade output is the binding constraint for top-end accelerators, not merely the commodity DRAM that CXMT is rapidly learning to manufacture at scale. Second, a toolchain that can be widely adopted without prohibitive porting costs and based on steady open-source development. Third, scale-out networking and datacenter-level performance-per-watt that approaches Nvidia’s current-generation offering: brute force can close gaps on raw throughput, but it taxes power budgets and raises total cost of ownership.
On that score, CXMT’s rapid scaling strengthens Beijing’s position, but it does not resolve the hardest substitution problems by itself. Similarly, CloudMatrix384 demonstrates credible progress in inference serving, although Huawei hardware and the CANN ecosystem are still underperforming for the most demanding workloads. The practical reality check is therefore two-sided: export controls can slow the development of China’s AI stack, but they also make sustained indigenization the dominant strategy, because the alternative is structural dependence on a structurally unreliable supplier.
