Huawei readies Nvidia challenger as Ascend 960 enters the spotlight

A chip launch with platform ambitions

Huawei’s next AI-chip move is no longer just a component story. It is a test of whether China can turn a sanctioned semiconductor stack into a credible computing platform.

The company is expected to present its upcoming Ascend 960 AI chips at its annual Huawei Connect summit in Shanghai on Thursday, September 17, with commercial availability planned for 2027 [1]. The timing matters because Huawei is not merely introducing another accelerator; it is trying to convince Chinese cloud providers, model developers and infrastructure buyers that a domestic alternative can become a dependable long-term platform in a market where Nvidia’s most advanced Blackwell processors remain unavailable directly in China [1].

The immediate pitch is clear: if Chinese AI companies cannot reliably buy the best Nvidia silicon, Huawei wants Ascend to become the default local answer. The harder question is whether the new chip, and the systems around it, can support the workloads that matter most: frontier model training, high-volume inference, multi-chip networking and software compatibility at scale.

The Ascend 960 is the headline, but the system is the product

The latest reporting says rotating chairman Wang Tao will present the Ascend 960, a successor to Huawei’s current Ascend 950 family [1]. The 950DT, a member of that family, is described as using faster and higher-capacity memory for demanding AI inference workloads, and it has already become important enough that DeepSeek is planning to deploy at least 160,000 Huawei Ascend 950DT chips at an Inner Mongolia data center for running its models [1].

That distinction between training and inference is central. Chinese frontier labs still rely heavily on Nvidia chips for training large models, while shifting more inference activity to domestic semiconductors [1]. In practical terms, Huawei can win meaningful share before it fully matches Nvidia at the highest end of training. But to become more than a substitute under pressure, Ascend must demonstrate that it can support the full AI pipeline, not just serve responses after a model has already been trained elsewhere.

Huawei’s answer is increasingly system-level. The company has introduced a SuperPod architecture designed to link large numbers of Ascend chips in clusters, and it uses a UnifiedBus networking protocol to accelerate data transmission between chips [1]. That strategy acknowledges a reality Huawei itself cannot escape: single-chip performance remains a challenge against Nvidia’s top offerings, so the company is trying to narrow the gap through cluster design, networking, packaging, memory and software-hardware co-optimization [1].

Export controls created the opening, but not the victory

The political backdrop is inseparable from the technical one. Huawei’s push comes after years of U.S. export controls aimed at limiting China’s access to advanced semiconductors and semiconductor-manufacturing capabilities [1]. Those restrictions created a huge protected market for domestic accelerators, but they also raised the bar for Chinese suppliers: customers need volume, stable supply, compatible software and competitive total cost, not just patriotic procurement.

Memory is one of the most visible pressure points. Huawei and other Chinese AI-chip makers have been cut off from the latest AI memory products from major foreign suppliers including SK Hynix, Samsung Electronics and Micron as part of U.S. export controls [1]. The Ascend 950 chips are reported to be Huawei’s first AI accelerators powered by memory chips it designs in-house, though the company has disclosed little about production details [1].

That is important because modern AI accelerators are only as useful as the memory systems around them. High-bandwidth memory affects how quickly models can move data, how efficiently clusters can serve requests and how economically companies can deploy large fleets. Huawei may have a domestic-demand tailwind, but if memory supply stays tight, the cost and timing of deployments will remain a constraint.

Demand is real, and so is scarcity

The clearest sign of demand is pricing. Huawei recently raised the price of the Ascend 950DT by 60 percent, citing tight component supply in communications with customers [1]. That kind of move suggests customers are lining up, but it also signals that Huawei cannot yet satisfy all demand at a stable cost.

The same report cites Morgan Stanley’s estimate that China’s AI-chip market could reach 646 billion yuan, or about 96 billion dollars, by 2030 [1]. That scale explains why Huawei is moving aggressively. Every major Chinese cloud, internet and model company needs access to compute, and the more foreign supply is restricted or uncertain, the more valuable a domestic platform becomes.

But scarcity cuts both ways. If Ascend accelerators are too expensive or delivered too slowly, customers may ration them for inference, continue using Nvidia where possible, or optimize models around mixed hardware. A real Nvidia challenger is not just the fastest chip in a slide deck. It is the platform customers can buy, program, network, cool, maintain and upgrade across years.

Huawei is also selling a view of the AI future

The chip launch sits inside Huawei’s broader argument that AI demand is about to explode. On September 16, the company released its “Intelligent World 2035: Turning Vision into Action” report, identifying ten technology directions for the agentic AI era, including AGI, computing clusters, storage and memory systems, intelligent connectivity, chip design, Agent OS, device intelligence, intelligent driving, AI data-center power and cooling, and security and privacy [2].

Huawei’s report says agentic AI will shift the digital world from application-centric to agent-centric, with agents continuously perceiving, reasoning, making decisions, invoking tools and interacting with humans in real time [2]. It predicts that by 2035 global annual token consumption will grow 100,000-fold, with agents accounting for more than 90 percent of token traffic [2].

Reuters separately reported Huawei’s expectation that autonomous agents will dominate AI token traffic by 2035 and require huge increases in computing power, alongside new systems to keep independent software secure and controllable [3]. Guo Ping, chairman of Huawei’s supervisory board, told employees in remarks released Tuesday that Huawei views AI as its “biggest opportunity” and wants its computing infrastructure to become China’s equivalent of Nvidia [3].

That last phrase is the core of the strategy. Huawei is not only aiming at chip substitution. It wants to own the infrastructure layer for a future in which AI agents generate constant compute demand.

The CUDA problem has not disappeared

The main obstacle is not only transistor density or memory bandwidth. Nvidia’s position rests on CUDA, developer habits, optimized libraries, networking, systems integration and years of software maturity. Huawei’s reported goal is for Ascend chips to operate every AI model in China and beyond, and Guo has argued that the company is closing the gap through chip architecture and software optimization [1].

That ambition is necessary, but it is also the most difficult part of the plan. Model developers do not want to spend engineering cycles rewriting kernels, debugging unstable clusters or accepting slower training because the sanctioned alternative is politically favored. If Huawei can make Ascend feel routine to use, it can erode Nvidia’s moat inside China. If it cannot, Ascend risks becoming a constrained-market workaround: valuable, necessary and strategically important, but still not a full platform replacement.

What to watch next

The launch to watch is not just a specification sheet. The decisive signals will be whether Huawei discloses credible performance data for the Ascend 960, whether it explains memory supply and production scale, and whether it shows software compatibility that gives Chinese AI labs confidence beyond inference.

For now, Huawei has the opening it needed: restricted foreign competition, huge domestic demand and political support for self-sufficiency. It also has the burden that comes with being China’s best answer to Nvidia. The Ascend 960 can be a milestone, but the bigger contest is whether Huawei can turn chips, clusters and software into an ecosystem strong enough to make CUDA feel less inevitable.

Sources from the last 72 hours

1. [1]Huawei set to unveil China’s best answer to Nvidia AI chip reignSep 17, 2026, 12:44 AM UTC
2. [2]Huawei's latest report proposes 10 key directions to Intelligent World 2035Sep 15, 2026, 4:00 PM UTC
3. [3]China's Huawei forecasts billions of agents will dominate AI traffic by 2035Sep 16, 2026, 12:53 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.