SHANGHAI, Sept. 24, 2026 - Huawei’s latest AI hardware push is not really about one new chip.
It is about a change in strategy.
For years, the central question around China’s AI computing industry was whether domestic companies could build an accelerator powerful enough to replace restricted foreign hardware.
Huawei is now trying to move the argument away from individual-chip comparisons.
Its answer is a vertically integrated system: Ascend processors, high-bandwidth memory alternatives, a proprietary interconnect, open-source software, large SuperPoD systems and clusters that can connect hundreds of thousands of processors.
At HUAWEI CONNECT 2026 in Shanghai, the company unveiled a faster Ascend roadmap, a new Peerium computing architecture and the Atlas 960E SuperPoD. Huawei says Peerium is designed to make processor systems at the million-chip scale behave like one large computer.
That claim is ambitious.
The broader strategy behind it is easier to verify.
U.S. export controls have restricted China’s access to advanced AI chips, high-bandwidth memory and semiconductor manufacturing tools. Instead of depending on one imported component, Chinese technology groups have been pushed toward building more of the stack domestically.
Huawei has become the clearest example of that response.
The latest chip roadmap is moving faster
Huawei first laid out a multi-year Ascend roadmap in 2025.
At that time, it said the Ascend 950DT would arrive in the fourth quarter of 2026, the Ascend 960 in the fourth quarter of 2027 and the Ascend 970 in the fourth quarter of 2028.
The 2026 update brings parts of that schedule forward.
Huawei now says the Ascend 960DT will become available in the first quarter of 2027, three quarters earlier than its original roadmap.
The Ascend 960PR is scheduled for the third quarter of 2027.
Huawei also says Ascend 970 is planned for 2028 and Ascend 980 for 2029.
The company is targeting a one-generation-per-year cadence.
Huawei says future generations will continue increasing compute, memory bandwidth, memory capacity and interconnect bandwidth.
These are company roadmaps, not guarantees.
Semiconductor schedules can slip because of fabrication yield, packaging, memory supply, software readiness or system integration.
But the acceleration is strategically important because AI infrastructure is moving quickly enough that a one-year delay can change competitive position.
The bigger shift is from chips to systems
Huawei’s central argument is that the performance of an AI data centre cannot be judged only by the capability of one accelerator.
Large AI models already require thousands of processors working together.
As models, context windows and agentic workloads grow, system-level bottlenecks become increasingly important.
Memory capacity matters.
Memory bandwidth matters.
Interconnect speed matters.
Network latency matters.
Storage throughput matters.
Software scheduling matters.
Huawei is therefore building around the idea of the SuperPoD.
A SuperPoD is a large group of physical machines connected so tightly that they can operate more like one logical computing system.
The company’s 2025 Atlas 950 design was built around as many as 8,192 Ascend NPUs.
Huawei’s newer architecture goes further.
Peerium is Huawei’s million-processor bet
On September 17, Huawei introduced what it calls the Peerium Computing Architecture.
The design uses nested parallelism, unified memory addressing and peer interconnects.
Huawei says the architecture is intended to allow up to one million processors to work together as one computing system.
The key technology underneath it is UnifiedBus.
UnifiedBus is designed to connect CPUs, NPUs, memory, SSDs, network interfaces and switches through one high-speed protocol.
The strategic idea is simple.
If access to the very best individual process technology remains constrained, Huawei can try to compensate by connecting more processors more efficiently.
That is a different engineering path from simply chasing the highest performance on a single chip.
It does not eliminate the importance of semiconductor fabrication.
A system built from less efficient chips can require more power, more cooling, more space and more hardware to deliver the same useful compute.
But if the interconnect and software are strong enough, system architecture can narrow some of the effective gap.
Atlas 960E pushes the architecture into a product
Huawei also announced the Atlas 960E SuperPoD.
The company describes it as the first SuperPoD to use near-packaged optics, or NPO.
Near-packaged optics moves optical connectivity closer to the processors and switching hardware, reducing some of the electrical-distance limitations that appear as systems become larger.
Huawei says the system is designed for training and inference involving models with as many as 10 trillion parameters.
It also says the updated TaiShan 950 SuperPoD can support up to 4,096 NPUs through all-optical UnifiedBus networking.
Those performance claims come from Huawei.
They should not be read as independent benchmark results.
The more important fact is that Huawei is moving optical interconnect technology closer to the compute layer because conventional electrical links become harder to scale as clusters grow.
That is an industry-wide physical problem, not a marketing problem.
A 256,000-card cluster is already being deployed, Huawei says
Huawei’s September architecture announcement states that a 256,000-card Atlas 950 SuperCluster is already being deployed.
That is one of the most consequential claims in the current roadmap.
A cluster with hundreds of thousands of AI processors moves the engineering challenge far beyond chip design.
Failure rates become important.
Network topology becomes important.
Power distribution becomes important.
Cooling becomes important.
Job scheduling becomes important.
Even a tiny failure probability at the component level can produce constant faults when a system contains hundreds of thousands of parts.
The economic question is therefore not how many processors can be connected.
It is how much useful model training or inference can be delivered per unit of power, capital and time.
That is the benchmark that will matter for customers.
Export controls helped create this strategy
China’s domestic AI-computing push did not begin with U.S. export controls.
Beijing had already identified semiconductor self-reliance as a strategic objective.
But controls introduced and expanded since 2022 changed the urgency.
The U.S. Bureau of Industry and Security imposed restrictions on advanced computing chips and semiconductor manufacturing equipment intended to limit China’s access to high-end computing capabilities.
Those rules were expanded over time.
In late 2024, U.S. controls were extended to additional semiconductor equipment and certain high-bandwidth memory products.
In May 2025, BIS issued specific guidance concerning advanced Chinese computing chips.
The guidance listed Huawei Ascend 910B, 910C and 910D devices and warned that use of certain advanced Chinese chips could create export-control compliance risks because their development or production may have involved U.S.-origin technology subject to the Export Administration Regulations.
That action showed how broad the technology conflict had become.
The issue was no longer simply whether a U.S. company could directly sell one high-end accelerator into China.
Rules were reaching into manufacturing equipment, design technology, memory and foreign-produced chips connected to controlled U.S. technology.
The controls have also changed over time
The export regime has not remained static.
In January 2026, the U.S. government changed its licensing policy to permit case-by-case review of applications for Nvidia H200, AMD MI325X and similar chips for approved Chinese customers, subject to security and compliance conditions.
That did not end export controls.
It changed the boundary.
This matters because China’s semiconductor strategy is being shaped by both restriction and uncertainty.
Companies cannot assume that access to a specific foreign chip will remain permanently open.
They also cannot assume it will remain permanently closed.
That uncertainty itself increases the strategic value of a domestic alternative.
Huawei is trying to replace more than the accelerator
The hardest part of challenging an established AI computing platform is not only silicon.
It is software.
Developers need compilers, libraries, kernels, debuggers, inference engines, distributed-training software and integration with machine-learning frameworks.
Huawei’s software layer is CANN, or Compute Architecture for Neural Networks.
The company has moved CANN toward sustained open-source development.
Huawei says external developers now account for 61% of CANN developers and that the community has more than 5,200 monthly active developers.
It also says more than 40 models have been natively pre-trained on Ascend and CANN.
Ascend now supports more than 90 third-party open-source projects, including PyTorch, Triton, vLLM and veRL.
Huawei also says Ascend is officially supported as a PyTorch accelerator backend.
Again, these developer numbers are company-reported.
But they identify the correct battlefield.
A chip without a usable software ecosystem is not a complete AI platform.
Why software may be harder than hardware
AI infrastructure buyers do not want to rewrite every model for a new accelerator.
They want existing frameworks and training code to move with minimal modification.
That is why mature software ecosystems create powerful lock-in.
If Huawei wants Ascend to become a true alternative inside China, hardware availability is only the first step.
The company has to reduce migration cost.
It has to support popular frameworks.
It has to make debugging predictable.
It has to optimize inference engines.
It has to create tools that allow developers to move models from research to production without large engineering teams.
The faster CANN and its surrounding ecosystem mature, the more credible Ascend becomes.
China’s AI demand is creating pressure before the roadmap is complete
Huawei says China now consumes around 500 trillion inference tokens per day.
That figure is company-reported and should be treated as an estimate rather than an independently audited market statistic.
The directional point is less controversial.
Inference demand is rising rapidly as AI moves into search, enterprise software, coding, agents, devices and industrial systems.
This changes the semiconductor problem.
Training frontier models is extremely compute-intensive, but inference can become even larger in aggregate once millions of users and software agents run models continuously.
That creates demand not only for maximum training performance but for lower-cost inference.
Huawei’s product roadmap reflects both.
Memory is one of the hardest bottlenecks
AI accelerators depend on high-bandwidth memory because large models constantly move data between memory and compute units.
Export controls on advanced HBM therefore matter independently of restrictions on processors.
Huawei has been working on its own memory architecture and packaging approaches around Ascend.
Its earlier Ascend 950 roadmap referenced proprietary memory systems designed to increase capacity and bandwidth.
This is strategically important because a powerful accelerator without sufficient memory bandwidth cannot deliver its theoretical compute performance.
China’s AI self-reliance effort therefore has to solve multiple linked semiconductor problems at once.
Compute logic alone is not enough.
More processors can compensate for some limits, but at a cost
Huawei’s architecture strategy raises an important economic question.
If one processor is less powerful or less energy-efficient than the leading alternative, can a much larger cluster close the gap?
Technically, sometimes.
Economically, it depends.
Doubling the number of processors can increase capital cost.
It can increase electricity consumption.
It can increase cooling requirements.
It can increase networking complexity.
It can increase failure probability.
It can also reduce utilization if software and interconnects cannot keep all processors busy.
This is why Huawei’s challenge will ultimately be measured in useful work per dollar and useful work per watt, not only theoretical FLOPS.
The company has published aggressive performance claims for its SuperPoD systems.
Independent large-scale comparisons will matter more as deployments expand.
The challenge to Nvidia is strongest inside China
The immediate competitive effect is geographically uneven.
Inside China, Huawei has several structural advantages.
Its hardware is locally controlled.
Its roadmap is designed around domestic supply constraints.
Its systems can be integrated with Chinese cloud providers, model developers and state-supported infrastructure projects.
Its software ecosystem is being developed specifically to reduce dependency on foreign AI platforms.
And domestic customers have strong incentives to qualify alternatives because access to U.S. hardware remains regulated.
Outside China, the challenge is harder.
Export-control compliance can complicate international use of some advanced Ascend products.
Huawei also has to compete with a much more mature global software ecosystem.
That means the near-term story is less “Huawei replaces Nvidia worldwide” and more “China is building a second AI computing stack that can become increasingly self-sufficient.”
That distinction is essential.
The strategy could fragment global AI infrastructure
If Huawei succeeds, the AI industry could become more divided.
One ecosystem could remain centered on U.S.-designed accelerators, software and networking.
Another could grow around Chinese processors, Chinese interconnect technology, domestic memory solutions and locally optimized software.
Models may still move between ecosystems.
Open-source frameworks may still provide compatibility.
But the underlying infrastructure could become increasingly regional.
This would have implications for cloud providers, model developers, semiconductor equipment suppliers and governments.
AI infrastructure could start to resemble telecommunications, where technical standards and supply chains sometimes follow geopolitical boundaries.
The biggest risk is manufacturing
Architecture cannot fully solve manufacturing constraints.
Advanced AI processors require sophisticated fabrication, packaging and memory.
Yield matters.
Volume matters.
Supply-chain reliability matters.
Huawei can publish a roadmap every year, but customers need chips in quantity.
If production cannot scale, large SuperPoDs become difficult to build no matter how strong the design is.
This is one reason supply availability is as important as benchmark performance.
The global AI race is increasingly a manufacturing race.
The second risk is power efficiency
AI data centres are becoming constrained by electricity.
A system that delivers competitive performance by using substantially more chips may be harder to deploy in locations where grid capacity is limited.
Power density also increases cooling complexity.
This makes efficiency a strategic variable.
Huawei is investing in data-centre power and cooling alongside compute.
That vertical integration may help.
But the energy economics of its largest clusters will need to be proven in production.
The third risk is software maturity
Developers tolerate hardware limitations more easily than software friction.
If a training job fails unpredictably, if an operator is missing, or if moving a model requires substantial code changes, total cost rises.
Huawei’s push to open CANN and connect Ascend to mainstream projects is therefore not secondary.
It may determine whether customers view Ascend as a strategic necessity or a genuinely preferred platform.
What changed most after export controls
The largest strategic change is that China is no longer responding only by attempting to reproduce individual restricted chips.
The response has become systemic.
Huawei is building processors.
It is building interconnects.
It is developing memory approaches.
It is expanding compiler and framework support.
It is constructing SuperPoDs.
It is designing clusters at hundreds of thousands of processors.
It is working on a million-processor architecture.
That does not mean China has eliminated its semiconductor constraints.
It means those constraints have changed the direction of engineering investment.
The strict conclusion
Huawei’s 2026 AI announcements show that the contest over AI hardware is becoming a contest over complete computing systems.
The Ascend roadmap is accelerating.
The Ascend 960DT is now scheduled for the first quarter of 2027 and the 960PR for the third quarter.
The Atlas 960E introduces near-packaged optical connectivity.
Peerium is designed around scaling toward one million processors.
UnifiedBus is intended to connect compute, memory, storage and networking under one architecture.
CANN is being pushed into a larger open-source ecosystem.
These pieces are more important together than separately.
Export controls were designed to restrict China’s access to advanced computing technology.
They have imposed real constraints, particularly around leading-edge semiconductors, manufacturing equipment and memory.
They have also strengthened the incentive to build a parallel domestic stack.
The result is not yet technological independence.
Nor is it proof that Huawei has matched the global leader on every measure.
It is something more strategically important: China now has a serious, increasingly integrated AI computing architecture that is being designed specifically for a world in which access to foreign technology cannot be assumed.
The next phase will be decided by manufacturing volume, software quality, energy efficiency and real-world cluster performance.
Those are harder tests than a product launch.
They will determine whether Huawei’s AI chips remain primarily a domestic substitute or become the foundation of a genuinely separate global computing ecosystem.
Reader questions
Frequently asked questions
What new AI chips has Huawei announced?
Huawei’s current roadmap includes the Ascend 960DT for the first quarter of 2027 and Ascend 960PR for the third quarter of 2027, followed by Ascend 970 in 2028 and Ascend 980 in 2029.
What is Huawei's Peerium architecture?
Peerium is Huawei’s new computing architecture designed to connect processors, memory, storage and networking through unified memory addressing and peer interconnects, with a stated goal of scaling to systems containing up to one million processors.
What is the Atlas 960E SuperPoD?
The Atlas 960E is Huawei’s latest large AI-computing system using near-packaged optical connectivity. Huawei says it is designed for training and inference workloads involving models at the 10-trillion-parameter scale.
How have U.S. export controls affected Huawei's AI strategy?
Export controls have restricted access to advanced chips, high-bandwidth memory and semiconductor manufacturing technologies, increasing the incentive for Huawei and other Chinese companies to build domestic processors, software, memory systems and large-scale computing architectures.
Can Huawei replace Nvidia in AI computing?
Huawei is becoming a more credible domestic alternative in China, particularly because customers want supply security. Globally, the challenge is harder because software maturity, manufacturing capacity, power efficiency, export-control compliance and ecosystem scale remain important competitive factors.
What is CANN?
CANN, or Compute Architecture for Neural Networks, is Huawei’s software platform for Ascend AI processors. Huawei is moving it toward sustained open-source development and expanding compatibility with widely used AI frameworks and inference tools.
Are Huawei's performance claims independently verified?
Many of the SuperPoD specifications and performance figures currently available are Huawei-reported claims. Independent large-scale benchmarks will be important for comparing real-world training efficiency, inference throughput, reliability and energy use.
Nexuswild welcomes factual corrections. Email [email protected] with evidence and the article URL.
