Huawei is forecasting that its artificial-intelligence chip Revenue/">Revenue will rise by at least 60 per cent in 2026, reaching about $12bn, as China's largest cloud, internet and telecoms groups pivot away from Nvidia toward home-grown silicon. The forecast, disclosed in recent investor and customer briefings, marks one of the clearest signals yet that US export controls and the rapid maturation of Chinese alternatives are reshaping the world's second-largest market for high-end accelerators.

The catalyst for the surge is twofold. The first is the deepening US restriction regime around Nvidia's most capable chips, which has constrained the volumes of H200, B200 and Blackwell-class hardware that can ship to mainland Chinese customers and added regulatory and pricing friction even to the export-tailored H20 part. The second is the launch of DeepSeek's V4 large language model, which has been engineered to run efficiently on Huawei's Ascend 950 family and which has prompted a wave of orders from Chinese hyperscalers that had previously hedged between domestic and US silicon.

For Huawei, the implications are strategic as well as financial. After years in which the company's most advanced chips were dismissed as a generation behind Nvidia, Ascend 950 is now being treated by Chinese cloud customers as a credible primary supplier rather than a backup. For Nvidia, the loss of Chinese order flow at this scale challenges a long-held investor assumption that the company's near-Monopoly/">Monopoly on AI Training/">Training silicon would remain effectively untouchable in every major market.

From export restrictions to a domestic incumbent

The shift now playing out across China's data centres has been years in the making. Successive rounds of US export controls — from the original 2022 restrictions targeting the A100 and H100, through the 2023 update that introduced the China-specific H800 and A800 parts, to the more recent rules constraining H200, Blackwell and HBM3e shipments — have steadily narrowed the slice of Nvidia's roadmap available to Chinese customers. Each tightening has given Beijing fresh incentive to back domestic alternatives, and Huawei has been the principal beneficiary.

Industry surveys put Huawei's share of accelerator orders at large Chinese cloud groups well above half for 2026, up from less than 15 per cent two years ago. Alibaba Cloud, Tencent Cloud, Baidu's AI cloud, ByteDance's Volcano Engine, China Mobile and China Telecom have all materially increased their planned spend on Ascend hardware over the past 18 months, according to people familiar with the procurement plans of each company. Several have signed multi-year Supply/">Supply agreements that lock in capacity through 2027.

That dynamic has been reinforced by Beijing's broader policy push for self-sufficiency in advanced semiconductors. Procurement guidance from state-owned enterprises and from local-government sponsored AI projects increasingly favours domestic accelerators, and a series of national-level standards for AI infrastructure have been written in ways that play to the strengths of Ascend hardware and the associated CANN software stack.

Inside the Ascend 950 ramp

The chip at the heart of the story is the Ascend 950PR, which entered mass production in March 2026 and which Huawei is targeting to ship in volumes of around 750,000 units across the year. The 950PR has captured the bulk of Chinese AI accelerator orders for 2026, but Supply/">Supply is expected to fall well short of Demand/">Demand, with output forecast to ramp materially in the second half of the year. An upgraded variant, the 950DT, is scheduled to enter mass production in the fourth quarter.

On a like-for-like basis, the 950PR is reported by Chinese cloud customers to outperform the Nvidia H20, the China-tailored part that Nvidia has been able to ship freely under the current export-control framework, while still trailing the more advanced H200 on raw Training/">Training throughput. For inference workloads, however, where energy efficiency, memory bandwidth and software optimisation often matter more than peak floating-point performance, the gap is reported to be much narrower.

Huawei's software stack, CANN, has matured considerably since 2023. PyTorch and other major frameworks now run on Ascend with materially better operator coverage and performance than the company's roadmap of two years ago suggested, and a growing community of Chinese model developers — including the team behind DeepSeek — has invested in writing custom kernels and optimisations that further extend the platform's effective capability.

DeepSeek V4 and the inflection point

The release of DeepSeek's V4 large language model has functioned as a catalysing event for the Chinese accelerator market. By demonstrating that a frontier-class Chinese model could be trained and served at competitive cost on domestic hardware, V4 effectively retired one of the strongest remaining arguments for buying Nvidia in China. Cloud customers that had previously been willing to pay an Nvidia premium for ecosystem assurance found that they could match or exceed model quality on Ascend at materially lower total cost of ownership.

DeepSeek's choice of platform reflected both technical and strategic considerations. On the technical side, the team had spent years optimising their Training/">Training pipeline for Ascend silicon and had reported repeatedly that the platform delivered competitive performance for their specific workloads. On the strategic side, building on domestic infrastructure insulated the project from any future tightening of US export rules — a consideration that has weighed increasingly heavily on Chinese AI groups since 2024.

Other Chinese model developers have followed DeepSeek's lead. Several of the most actively used domestic foundation models are now trained primarily or exclusively on Ascend, with knock-on effects on tooling, talent and the broader ecosystem. Once a critical mass of model groups is committed to a platform, the gravitational pull on cloud-customer procurement decisions tends to be self-reinforcing.

What this means for Nvidia

For Nvidia, the loss of Chinese order flow at this pace is significant but not yet existential. China was at one point estimated to account for more than 20 per cent of the company's data-centre Revenue/">Revenue. Most analysts now model that share to fall meaningfully below 10 per cent over the next two years if current trends continue, with some forecasting an even sharper drop if export controls tighten further or if the H20 part is restricted.

The company has so far been able to absorb the squeeze because Demand/">Demand for Blackwell and Blackwell Ultra parts in the United States, Europe, the Middle East, Japan and South Korea has been strong enough to soak up any capacity displaced from China. Hyperscale cloud groups, sovereign-AI projects and a new wave of corporate AI factories have between them more than offset the lost Chinese Volume/">Volume. But the shift does erode Nvidia's claim to a truly global moat, and it crystallises a competitive threat that bulls had previously dismissed as theoretical.

Nvidia chief executive Jensen Huang has acknowledged repeatedly over the past year that Huawei is a serious competitor, and the latest comments out of Beijing only reinforce that assessment. The challenge for Nvidia's investor narrative is that if Chinese customers can build globally competitive AI models on Huawei silicon, the argument that Nvidia is uniquely indispensable to advanced AI development becomes harder to maintain.

Implications for the global AI Supply/">Supply chain

The bifurcation of the AI accelerator market between a US-aligned ecosystem built around Nvidia, AMD and a growing roster of custom silicon from cloud hyperscalers, and a China-aligned ecosystem built around Huawei and a handful of smaller domestic players such as Cambricon, Biren and Moore Threads, has profound implications for the broader semiconductor industry. Foundry capacity at Taiwan's TSMC, advanced packaging at Amkor and ASE, high-bandwidth memory production at SK Hynix, Samsung and Micron, and lithography support from ASML are all being asked to serve two increasingly distinct customer bases.

For Huawei, the most acute Supply/">Supply-chain question is whether SMIC, China's leading foundry, can produce enough advanced-process wafers to meet Ascend 950 Demand/">Demand. SMIC's progress on its own equivalents of leading-edge nodes has been impressive but constrained by the absence of access to ASML's most advanced extreme-ultraviolet lithography tools. Industry observers expect Huawei to pursue a portfolio of fabrication strategies, blending SMIC capacity with creative packaging approaches that combine chiplets and stacked memory to extract more performance from less advanced silicon.

For Western policymakers, the rapid maturation of the Chinese AI hardware stack raises a familiar dilemma. Tighter export controls are designed to slow Chinese capability, but they also accelerate the build-out of an indigenous alternative ecosystem that, once mature, may compete with US hardware in third markets and may be much harder to dislodge from China itself.

Investor takeaways and market reaction

The Huawei Revenue/">Revenue forecast has had ripple effects across global Equity/">Equity markets. Nvidia and AMD shares have been more sensitive to commentary on Chinese Demand/">Demand for high-end accelerators in recent quarters, and analysts have begun to embed a more conservative path for China Revenue/">Revenue into their models. UK and European investors with exposure to ASML, ARM, Infineon and STMicroelectronics have also been watching closely, given how deeply the European semiconductor Supply/">Supply chain is woven into the global AI build-out.

Within China, listed companies in the AI infrastructure ecosystem — including Cambricon, Innolight, Eoptolink and several memory and packaging-related names — have seen their share prices respond to the wave of Ascend-related news flow. Liquidity/">Liquidity, sentiment and policy support combine to produce significant moves in this segment of the Chinese market, and the latest Huawei guidance has reinforced the broader 'AI self-sufficiency' theme that has powered domestic AI equities through much of 2026.

Bond and Credit/">Credit investors are also paying attention. Huawei does not have publicly listed Equity/">Equity, but the company's bond programme and the Credit/">Credit profiles of its major suppliers and customers are sensitive to the trajectory of its core businesses. Stronger AI chip Revenue/">Revenue strengthens the financial profile of a company that, only a few years ago, was widely seen as fighting for survival under the weight of US sanctions.

Risks and uncertainties

There are real risks to the Huawei forecast. Production yields on the Ascend 950PR remain commercially sensitive and could underwhelm if SMIC's advanced-node throughput is not as high as the company hopes. Demand/">Demand from any single very large cloud customer could be revised down if the broader Chinese AI build-out proves to be less front-loaded than current procurement plans suggest. And US export controls, while already restrictive, could be tightened further to capture more components of the AI hardware stack, including high-bandwidth memory or advanced lithography accessories.

Conversely, there is a plausible bull case in which Huawei's growth materially exceeds the 60 per cent guidance. Several Chinese state-owned enterprises and local government bodies have not yet placed their full 2026 orders, and a wave of additional procurement could push Revenue/">Revenue above the $12bn target. Demand/">Demand from non-Chinese customers in markets such as the Middle East and Southeast Asia, where geopolitical alignment has shifted in recent years, could also become a meaningful incremental driver.

The political risk cuts both ways. Any escalation in US-China tension — whether through tariffs, export-control tightening, secondary sanctions or restrictions on cross-border technology transfer — could simultaneously hurt Nvidia's China Business/">Business further and constrain Huawei's access to inputs. The net impact on each company's Revenue/">Revenue trajectory is genuinely uncertain.

A new competitive landscape

The combination of Huawei's growth forecast, the maturation of DeepSeek and other Chinese frontier models, and the steady tightening of US export controls has crystallised a new structure for the global AI hardware market. China is increasingly buying Chinese silicon to train Chinese models running on Chinese cloud infrastructure. The rest of the world continues to buy a US-anchored stack, with strong contributions from European, Korean and Taiwanese suppliers. The two ecosystems are interoperable today but are diverging steadily.

For investors, the practical implications are clear. The simple narrative of Nvidia as the sole gatekeeper of the AI revolution is becoming harder to sustain in a market with more than one critical economic geography. The simple narrative of Chinese chip companies as perennial laggards is also becoming harder to sustain. The next decade of AI infrastructure will be built across two stacks, and Capital/">Capital allocation decisions are likely to look very different as a result.