The most consequential admission in semiconductor industry executive commentary in 2026 came not from an earnings call but from an interview Jensen Huang gave in May. Asked about NVIDIA's position in China, the CEO said the company had "largely conceded" the Chinese AI chip market to Huawei. The statement was a recognition of what had already happened over the preceding 18 months: the cumulative effect of four rounds of US Bureau of Industry and Security export controls, the closure of the H20 sales channel into China in October 2024, and the scale-up of Huawei's Ascend production at SMIC's domestic 7nm fabrication capacity. China's AI infrastructure market is no longer a market NVIDIA participates in at the high performance tier. It is a market Huawei dominates.
The consequence is that the world now has two AI compute stacks running in parallel. The Western stack runs NVIDIA, AMD, and hyperscaler custom silicon on TSMC's most advanced nodes with SK Hynix and Samsung HBM. The Chinese stack runs Huawei Ascend, Alibaba T-Head, Baidu Kunlun, and Cambricon on SMIC's domestic 7nm capacity with CXMT-produced HBM at increasing volume. These are no longer technology generations apart. They are running in production simultaneously, serving overlapping use cases at different price-performance points, and the gap between them is closing on every dimension except the most leading-edge process nodes.
How Huawei got here in two years
The trajectory from Huawei's near-extinction under the 2019 US sanctions to its current position as the dominant AI silicon supplier in China is genuinely extraordinary, and worth retracing because the lessons are not yet fully internalised by the rest of the industry.
The HBM constraint is the real ceiling
The technical constraint on Huawei's expansion in 2026 is not Ascend chip die manufacturing. SMIC can produce die for more than one million Ascend chips annually using its current 7nm capacity. The constraint is HBM supply. Each Ascend 910C package requires multiple HBM stacks, and the HBM stacks China has been using have come primarily from stockpiled Samsung product accumulated before 2024.
The Beijing directive in September 2025 that tightened HBM export controls to China, combined with Korean and US enforcement of the controls through 2025, has substantially depleted that stockpile. CXMT (ChangXin Memory Technologies) is China's leading DRAM company and the primary domestic HBM development programme. Industry estimates suggest CXMT will produce approximately 2 million HBM stacks in 2026 — sufficient for approximately 250,000 to 300,000 Ascend 910C-equivalent packages. That number is below the die production capacity SMIC can deliver, making HBM the binding constraint on Huawei's 2026 output.
This is the variable that determines whether Huawei's USD 12 Billion 2026 revenue projection is conservative or aggressive. If CXMT exceeds its HBM3E roadmap targets, the constraint loosens and Huawei can ship closer to one million Ascend packages. If CXMT struggles with through-silicon via yield — which has been the primary technical challenge for every HBM producer entering volume production, including Samsung — Huawei output may fall closer to 400,000 packages. The range of plausible outcomes is wide enough that the 2026 revenue number could be plus or minus USD 3 to 4 Billion versus the published projection.
The software ecosystem is where the gap actually closes
The hardware performance comparison between Huawei Ascend 910C and NVIDIA Hopper H100 has been documented extensively: the Ascend 910C delivers approximately 780 TFLOPS of dense BF16 compute at 350W package power, trailing the H100 on both peak throughput and power efficiency, but offsetting the gap by scaling out through the CloudMatrix 384 cluster architecture. What has received less attention is the much harder problem: the software stack.
NVIDIA's competitive moat was never CUDA alone. It was the cumulative result of fifteen years of CUDA being the default development target for every major AI research lab, every machine learning framework, and every GPU-accelerated scientific computing application. Migrating an existing AI workload from NVIDIA hardware to Huawei Ascend requires not just porting CUDA kernels to Huawei's CANN software stack, but also re-validating model accuracy, re-tuning hyperparameters, and re-qualifying production inference latency targets on the new hardware. The work is non-trivial, and a substantial part of the hyperscaler procurement decision is whether the cost-per-inference improvement justifies the engineering investment.
Inside China, that calculation is increasingly favourable to Huawei. Chinese cloud providers — Alibaba Cloud, Tencent Cloud, Baidu AI Cloud, Huawei Cloud, and ByteDance's Volcano Engine — have all completed CANN migration for their primary AI workloads. The DeepSeek model architectural innovations published in early 2025 specifically optimised for the lower-bandwidth memory architecture available on Chinese domestic AI hardware. If a similar level of investment from Chinese AI labs flows to making CANN a first-class development target — which Chairman Xu's May 2026 commentary suggests is the explicit strategy — the software gap will narrow on a multi-year horizon.
The hardware gap between Huawei Ascend and NVIDIA Hopper-class accelerators is real, measurable, and approximately one to two GPU generations wide on peak compute performance. The hardware gap will not close before 2028 in the absence of Huawei gaining access to EUV lithography, which the current export control regime forecloses.
The software gap is different. The CANN ecosystem is being actively built by Huawei plus a coordinated effort from Chinese cloud providers, AI labs, and research institutions. It has the scale of investment and the coordination of intent to reach competitive parity with CUDA for the subset of AI workloads that matter most to Chinese AI infrastructure customers — even if it never reaches the breadth of CUDA's coverage across academic and scientific computing.
What "bifurcation" actually means for non-Chinese, non-US customers
The strategic question that most analyses of the US-China semiconductor decoupling avoid is what happens in the rest of the world. India, Brazil, Indonesia, Saudi Arabia, the UAE, Vietnam, and Mexico are large enough AI infrastructure markets that the choice of which stack to deploy is consequential. None of these countries is structurally aligned with either the US or Chinese ecosystem to the point where the choice is automatic.
The economic argument for Huawei Ascend at the volume tiers these markets are deploying is real. Chinese AI accelerator pricing is consistently 40 to 60% below NVIDIA H100-equivalent pricing on a total cost of ownership basis. The political argument is more complex. The US has applied secondary sanctions pressure to discourage Huawei procurement by US-allied countries, but enforcement against sovereign infrastructure decisions in non-aligned markets has been inconsistent. Saudi Arabia's PIF and UAE's G42 have both publicly committed to NVIDIA-based infrastructure at scale, partly under US diplomatic pressure. Other markets have made less public commitments either direction.
The likely 2026 to 2028 pattern is that AI infrastructure procurement in non-aligned markets becomes hybridised: critical national infrastructure on NVIDIA or AMD silicon to preserve compatibility with Western AI services, plus commodity inference and lower-tier training workloads on Huawei Ascend at substantial cost saving. This pattern requires customers to operate two parallel AI software stacks, which is operationally expensive but achievable. It also creates a market position for hyperscaler cloud services that abstract the underlying hardware — customers buy AI inference capacity from Google Cloud, AWS, or Azure without needing to know which silicon runs the workload.
What CFOs at multinational companies should hold in mind
For corporate strategy teams at multinational companies with operations in both the US and Chinese markets, the bifurcation creates concrete procurement obligations that did not exist three years ago. AI infrastructure deployed in mainland China cannot run on NVIDIA hardware above defined performance thresholds. AI infrastructure deployed in the US to serve Chinese mainland customers cannot legally run inference for Chinese government, military, or defined dual-use applications. The compliance burden is non-trivial, the audit requirements are accumulating, and the cost of getting it wrong is rising as enforcement actions become more frequent.
The Huawei Sophgo case — the USD 1 Billion TSMC fine for unwitting fabrication of Huawei dies through a front company — is the precedent every supply chain compliance team is now operating against. The lesson is that the BIS Entity List does not control access through second-order intermediary structures, and that downstream supply chain participants are exposed to fines and reputational damage even when their own conduct was reasonable at each individual transaction.
The conservative posture — applied increasingly by procurement and legal teams at large US and EU enterprises — is to treat any AI infrastructure purchase that could plausibly involve Chinese-origin silicon as requiring explicit supply chain documentation, end-use certification, and shipment tracking. The administrative overhead is real, the cost is meaningful, and it is now a fixed feature of doing business in the AI infrastructure category for the foreseeable future.
The 2028 question
The bifurcation is structural and is unlikely to reverse on any policy horizon that current US or Chinese governments are operating on. The 2028 question is whether the gap between the two stacks stabilises at the current width, narrows further, or widens. Three variables will determine the answer.
First, whether CXMT closes the HBM gap to Samsung and SK Hynix faster than projected. The 17nm DDR5 milestone in September 2024 was three to four DRAM generations behind leading practice. The HBM3E milestone targeted for end-2026 would close that gap to approximately two generations. If CXMT's roadmap holds, China's HBM constraint becomes substantially less binding by 2028.
Second, whether the export control regime tightens further or relaxes under different US administrations. Both directions are politically conceivable, and both would substantially change the trajectory. The current trend through 2025 and 2026 has been progressive tightening, but the economic costs are accumulating in ways that may produce political pressure to relax specific provisions.
Third, whether Chinese AI lab innovation — of which DeepSeek's late-2024 and 2025 model architecture work is the leading example — continues to produce algorithmic improvements that disproportionately benefit Chinese hardware. If Chinese AI research continues to optimise for the memory bandwidth and interconnect characteristics of domestic silicon, the effective performance gap on the workloads that matter to Chinese customers shrinks faster than the underlying hardware gap.
None of these three variables are observable from outside the supply chain with sufficient resolution to forecast confidently. What is observable is that all three are moving, all three are moving in directions that benefit the Chinese AI stack rather than constraining it further, and the cumulative effect over a five-year horizon is likely to be material. The semiconductor industry's institutional assumption through 2024 was that US export controls would maintain a structural Western technology lead in AI infrastructure for at least a decade. That assumption is no longer well-supported by the evidence emerging in 2026.