While Washington debates AI regulators, the other superpower is quietly building the muscle underneath it all. At its Huawei Connect 2026 conference in Shanghai on Thursday, the Chinese tech giant made a claim that would have sounded absurd two years ago: it cannot make AI chips fast enough to meet China's own demand, and its next-generation Huawei AI chips are arriving ahead of schedule.
01."Demand for AI chips outstrips supply"
Huawei's rotating chairman Eric Xu told reporters that the company doesn't have enough manufacturing capacity to satisfy Chinese demand for its Ascend AI computing systems — which is why it isn't expanding into international markets "in a fully-fledged way" and is only shipping limited volumes to select countries. Huawei even raised the price of its Ascend 950DT chip by 60 percent, citing tight component supply, according to Straits Times reporting.
Xu also made a bold market claim: reliable data on Nvidia's share of China's AI chip market is hard to get, but "I think Ascend market share should be bigger than Nvidia's." Huawei didn't provide data to back the estimate — so treat it as a confident founder's boast rather than audited fact. His framing, though, was dead serious: "We cannot accept a destiny where we cannot control our fate being determined by others in terms of willingness to sell chips to China or not."
02.A chip every year: 960DT in 2027, 970 in 2028, 980 in 2029
The headline announcement: Huawei's flagship Ascend 960DT chip will be ready in the first quarter of 2027 — three quarters earlier than previously planned — with the 960PR variant following in Q3 2027. Rotating chairman David Wang said each new generation roughly doubles the computing power of its predecessor, and the roadmap now targets the Ascend 970 in 2028 and Ascend 980 in 2029. That is Nvidia-style annual cadence from a company that was supposed to be choked off by US export controls.
03.Clusters over chips: the Atlas 960 SuperPoD
Here's the thing nobody's talking about: Huawei can't yet match Nvidia's best single chips, so it's changing the game instead. The new Atlas 960 SuperPoD — the computing cluster built on the new architecture — promises a 2.3-fold improvement in training and 2.5-fold in inference for 10-trillion-parameter models compared to its Atlas 950 predecessor. Huawei says a single Ascend 960 "supernode" can connect up to 4,096 AI processors, and multiple supernodes can be linked into clusters of hundreds of thousands — with its new Peerium architecture eventually designed for up to 1 million processors working as one system.
The logic is simple. Huawei says communication between machines can eat up more than 40 percent of training time in conventional server systems, leaving processors idle. Its UnifiedBus interconnect — aimed squarely at Nvidia's NVLink — tries to kill that bottleneck, squeezing more work out of every scarce chip. Huawei says it has already deployed more than 1,000 systems using its earlier Ascend 910C processors, over 5,200 developers are active monthly on Ascend software, and more than 40 AI models have been trained directly on its platform.
04.The numbers that explain the urgency
Morgan Stanley estimates China's AI chip market could reach 646 billion yuan — roughly $96 billion — by 2030. US export controls keep Nvidia's most advanced Blackwell processors out of China, and labs like DeepSeek still largely train on Nvidia chips while running inference on domestic hardware. Nvidia also keeps its biggest moat: the CUDA software ecosystem that most AI developers grew up on. So Huawei's bet is that systems-level engineering plus a captive home market can close a gap that sanctions were designed to widen. Chinese AI labs, for their part, now have a reason to train on Ascend 950DT systems as early as next year, according to Xu. And as we wrote when Google, Nvidia and Anthropic teamed up on AI's electricity problem, compute — chips, power, cooling — is now the real currency of AI progress. And for everyday builders, the cheapest compute is still the compute you don't waste — our GPT-6 Astra — Zero to Hero Guide has a full cost-optimization chapter on exactly that.
05.Why India should be watching closely
For Indian readers, this story has an extra layer. As we covered when PM Modi inaugurated Semicon India 2026 in Delhi, India has just committed ₹1.27 lakh crore under Semicon 2.0 and is targeting 200 chip design startups — with Applied Materials announcing a $5 billion India investment on the same day. The Huawei-Nvidia fight shows exactly why countries are racing to build their own AI compute: whoever controls the chips controls the pace of AI development. India's bet on design talent and packaging is a different playbook from Huawei's full-stack sovereignty push, but both start from the same anxiety about depending on someone else's silicon. And with the US and China both pouring billions into compute, the pressure on India to scale its own AI infrastructure just went up another notch.
Related reads
- Apple AI Server: M8 Ultra Chips and the 2029 Plan — Apple is building its own AI server chips too.
- Nvidia Local AI Push: DGX Spark and RTX Laptops — the local-AI hardware push from the other side.
- Google, Nvidia, Anthropic Launch AEMA for AI Data Centers — why data centres are scrambling for supply.
Key Takeaways
- Huawei says AI chip demand in China outstrips its supply, so overseas sales stay limited
- Ascend 960DT lands in Q1 2027 (three quarters early); 960PR in Q3 2027; new generations every year through 2029
- Atlas 960 SuperPoD: up to 2.3x training and 2.5x inference gains for 10-trillion-parameter models
- Peerium architecture targets up to 1 million processors as a single system
- China's AI chip market could hit 646 billion yuan (~$96 billion) by 2030, per Morgan Stanley
Frequently Asked Questions
What did Huawei announce about its AI chips?
At Huawei Connect 2026 in Shanghai, Huawei said demand for its AI chips in China outstrips supply. The Ascend 960DT chip now arrives in Q1 2027 — three quarters earlier than planned — with a new chip generation every year after.
How is Huawei compensating for weaker individual chips?
Its new Atlas 960 SuperPoD delivers 2.3x training and 2.5x inference gains over its predecessor, and its answer to weaker individual chips is giant clusters of up to 1 million processors working as one system.
What This Means for You
Most of us in India won't ever buy an Ascend chip, but the price of AI compute affects everything you do — the cost of running ChatGPT-class models, the salary of anyone who can build with them, the viability of AI startups in Bengaluru and Hyderabad. A genuine second supplier of AI silicon means more competition, better prices and faster model releases globally. In the meantime, the smartest move is still to master using today's AI efficiently — prompt caching, cost optimization and agent-building are skills that pay off no matter whose chip is under the hood. Our GPT-6 Astra — Zero to Hero Guide covers exactly that, with a full chapter on cutting your AI costs by up to 90% using techniques most beginners have never heard of.