NextFin News - Huawei has moved up the launch of its next-generation Ascend 960DT AI chip to the first quarter of 2027, three quarters ahead of its original roadmap, as the Chinese technology group sharpens its challenge to Nvidia's dominance in AI computing. The acceleration, announced Thursday by rotating chairman David Wang at the HUAWEI CONNECT 2026 conference in Shanghai, is the tactical headline. The more consequential shift is the architecture behind it: Huawei is betting that its own interconnect standard, not raw chip speed, will decide who wins the race to million-processor AI superclusters.
The Situation: A Chip Pulled Forward, and a System Built Around It
Huawei said the Ascend 960DT will be ready in Q1 2027, and the Ascend 960PR in Q3 2027 — the latter one quarter ahead of its prior schedule. A company spokesperson confirmed the 960DT had previously been planned for Q3 2027, meaning the launch has been pulled forward by roughly nine months.
"The Ascend 960DT is expected to be ready in Q1 2027. Ascend 960 chips are launching ahead of schedule, doubling performance and advancing year by year," the spokesperson said.
Wang, Huawei's deputy chairman of the board and rotating chairman, framed the move around a single constraint: the industry is running out of time. Foundation model parameters are approaching 10 trillion and are projected to exceed 100 trillion by 2030, he said, while daily inference token consumption in China alone has surged to around 500 trillion and is expected to reach quintillions by the end of the decade. On-device models for smartphones have expanded from three billion parameters in 2024 to 30 billion today. The demand curve is not bending; the hardware has to catch up.
The hardware answer is not just a faster chip. Huawei unveiled the Atlas 960E SuperPoD, the industry's first SuperPoD built on near-packaged optics, or NPO, which can scale to 4,096 NPUs and deliver 8 exaflops of FP8 compute with up to 1 petabyte of high-bandwidth memory. The system relies on Hi-ONE, Huawei's NPO optical engine with a built-in light source, which the company says delivers 7.2 TB/s of transmission capacity and is the first such product ready for mass production. By using 5,500 Hi-ONE units, the SuperPoD eliminates the 48,000 800G optical modules that would traditionally be needed to connect all the NPUs — cutting power consumption by more than 550 kilowatts and doubling fault-free operating time to 99.8% system availability.
Beyond the single pod, Huawei's agentic SuperCluster, interconnected through its UnifiedBus technology, can scale up to one million NPUs. The company said it has already shipped more than 1,000 such systems to more than 370 customers. And in a move that extends beyond its own product line, Huawei said a standard proposal for a "12.8 TB/s NPO Module," jointly submitted with the China Academy of Information and Communications Technology and other companies, was officially initiated by the Optical Internetworking Forum in May 2026 with support from more than 40 global industry-chain vendors.
The announcement lands days before U.S. President Donald Trump and Chinese President Xi Jinping are scheduled to meet on September 24 in Washington, and it underscores a central tension: the United States has restricted China's access to advanced semiconductor technology, yet Huawei is compressing its roadmap rather than stretching it. The question the rest of this piece answers is whether that compression reflects genuine capability — or a roadmap trimmed to fit what sanctions allow.
Why the Interconnect, Not the Chip, Is the Real Battleground
The market's first-order read of this news is straightforward: Huawei is catching up to Nvidia on chip performance, and a faster Ascend means more competition for the world's most valuable company. That read is not wrong, but it stops at the surface. The transmission mechanism that actually matters runs through the interconnect, not the processor.
At the scale of modern AI training, individual chip speed is no longer the binding constraint. Once a cluster reaches tens of thousands of accelerators, the system's effective performance is determined by how quickly those chips can talk to each other and to a shared memory pool — what the industry measures as Model FLOPs Utilization, or MFU. A fast chip in a slow network is a fast chip that spends most of its time waiting. Huawei's own numbers make the trade-off explicit: the Atlas 960E SuperPoD's headline claim is not a clock speed or a transistor count, but the elimination of 48,000 optical modules and a 550-kilowatt power saving. The bottleneck it is attacking is communications overhead, not compute density.
This is why the NPO standards push is more strategically significant than the nine-month acceleration. By taking a "12.8 TB/s NPO Module" proposal through the OIF with more than 40 vendors behind it, Huawei is attempting to do something it could not do at the chip level: set the rules of the road. If NPO becomes the accepted interconnect standard for AI superclusters, then Huawei's Hi-ONE — already in mass production and already deployed in the Atlas 960E — becomes a reference implementation that competitors and customers must design around. That is a different kind of moat from process-node leadership, and it is one that export controls cannot easily reach.
The logic extends to the software stack as well. Huawei said Ascend now spans more than 90 leading third-party open-source projects and is officially supported as a PyTorch accelerator backend, while its Kunpeng ecosystem has attracted 4.16 million developers. Compute is only useful if developers can run their workloads on it; the interconnect is only useful if the software can schedule across it. Huawei is building both halves of that equation at once.
There is also a commercial logic to the timing. Demand for Huawei's AI accelerators is running ahead of supply: reports this month said the company raised the price of its Ascend 950DT accelerator by about 60% over the prior three months, citing component tightness, while Nvidia has reportedly told some of its largest customers that server prices will rise by more than 15% — a claim Nvidia has not confirmed. A double-digit price increase on the incumbent, arriving alongside a faster and more power-efficient alternative in the world's second-largest AI market, is the kind of opening a challenger spends years waiting for. Huawei does not need to match Nvidia specification-for-specification to win share; it needs to be good enough, cheap enough, and available enough for customers who have no other option.
The Counter-Thesis: A Roadmap Compressed Because It Was Trimmed
The strongest argument against reading this as unalloyed strength is the one that comes from Huawei's own prior roadmap. China tech analyst Rui Ma pointed out that Huawei had previously said its Atlas 960 SuperPoD would scale to 15,488 Ascend 960 chips, while this week's announcement referred to a system with 4,096 chips.
"The chip itself is coming WAY earlier, but the SuperPoD they announced is much smaller than what they originally laid out," Ma wrote.
The gap is material: the announced pod is roughly a quarter of the scale Huawei had outlined. This discrepancy matters because it points to the one constraint Huawei cannot engineer its way out of with architecture alone: sanctioned manufacturing. Huawei's performance claims are self-reported, and no independent third-party benchmarking organization has verified them — a caveat that applies to every Ascend generation. If the 960-series chips cannot be manufactured at the yields, volumes, or power envelopes Huawei originally assumed, then the rational response is exactly what we are seeing: bring the chip to market earlier at a smaller system scale, and sell the acceleration as a victory while quietly resizing the ambition. The acceleration, on this reading, is not a sign of strength but a hedge against it.
The counter-thesis has a named authority behind it in Ma's scaling observation, and it attacks the core claim at its foundation — that Huawei is closing the gap on Nvidia's ability to build giant AI systems. It deserves more than a dismissal. The honest answer is that both things can be true: Huawei's interconnect strategy is a genuine structural advance, and its manufacturing base remains the fragile link in the chain. The 4,096-chip pod is still a serious system — 8 exaflops of FP8 and a petabyte of HBM are not trivial — but it is not the million-processor machine the SuperCluster vision promises. Until the 15,488-chip configuration ships at scale, the structural claim rests on a prototype, not a deployed fleet.
There is a second, quieter counterpoint: the market did not treat the news as an existential threat to Nvidia. Shares of Nvidia rose modestly on the day, and the stock is up about 15% year to date — even as the broader semiconductor index has far outpaced it. Investors are effectively pricing in the view that Huawei's gains are largely confined to a Chinese market that Nvidia cannot fully serve anyway. If that assessment holds, then Huawei's acceleration changes the map inside China without redrawing the global balance of power.
What This Means: Cyclical Acceleration on Top of a Structural Shift
The right way to separate the two forces at work is by time horizon. The nine-month pull-forward of the 960DT is cyclical and tactical: it is a response to a demand surge, a competitor's price increase, and a political moment. Cyclical moves mean-revert — roadmaps slip, yields disappoint, and the next generation does not always double performance. Huawei has outlined an annual cadence beyond the 960-series, with the Ascend 970 planned for 2028 and the Ascend 980 for 2029, a rhythm the company has tied to the Tau (τ) Scaling Law that HiSilicon president He Tingbo presented in May 2026. A two-year cadence that delivers consistent doublings is a claim, not a track record.
The interconnect and standards play, by contrast, is structural. Once a cluster architecture and an interconnect standard are deployed across hundreds of customers, they create switching costs that outlast any single chip generation. If NPO becomes the industry's way of linking AI processors, Huawei's first-mover position in mass-produced NPO engines is a regime-level advantage, not a quarterly one. That is the part of this story that will not revert on its own.
So the judgment is: a cyclical acceleration layered on top of a structural bet, with the structural bet carrying more weight than the headline suggests. Huawei is not simply launching a chip early; it is trying to lock in an architecture while the window is open.
What to Watch
Three signals will separate the structural case from the cyclical one. First, whether the 15,488-chip Atlas 960 SuperPoD configuration ships at commercial scale by late 2027 — if it does not, the scaling thesis is weakened. Second, whether the OIF "12.8 TB/s NPO Module" standard picks up vendors outside the Chinese supply chain; a standard with only domestic support is a domestic standard, not a global one. Third, whether Huawei's self-reported performance numbers survive independent benchmarking; a single credible third-party test would do more for the company's credibility than any roadmap slide.
The falsifying signal for the structural thesis is specific: if, by the end of 2027, Huawei has not shipped a SuperPoD configuration above 5,000 chips in volume, or if the OIF NPO standard fails to attract more than a handful of non-Chinese vendors, then the architecture moat is narrower than this announcement implies, and the story reverts to a contained regional competitor.
The base case is that Huawei consolidates its position as the default AI infrastructure supplier inside China, where U.S. restrictions already limit Nvidia's reach, and uses the 960-series to hold that ground through 2027. The upside case is that NPO becomes a globally adopted interconnect standard and Huawei's Hi-ONE becomes a reference design for customers worldwide, turning the company from a sanctioned challenger into a rule-setter. The downside case is that manufacturing constraints keep system sizes capped, shipments fall well below the roughly 1.6 million dies analysts have projected for the broader Ascend line, and the acceleration proves to be a roadmap trim rather than a capability leap.
Huawei's rotating chairman Eric Xu put the stakes plainly when he argued that Chinese companies need to advance further before they can fully understand and address the risks posed by more powerful AI systems — a rebuttal to calls from some U.S. industry leaders to slow AI development over safety concerns. Ma framed the broader dynamic more bluntly:
"I think it's futile to stop China's development in semiconductors because the stakes for self-sufficiency are just too high at this point."
The acceleration of the Ascend 960DT is the news of the week. The architecture behind it is the news of the decade — if Huawei can build it at the scale it has promised. Until then, the market is right to treat this as a contained threat: a formidable regional champion, not yet a global rule-setter.
Data in this article are as of September 18, 2026.
Explore more exclusive insights at nextfin.ai.
