Gangsta AI
Huawei Just Pulled Its Next AI Chip Forward Three Quarters — and Unveiled a Plan to Wire One Million of Them Together. Chip for Chip, It Still Loses to Nvidia. It Doesn't Care.
The Machines

By Arnold Schwarzenegger · 2026-09-18 · 4 min read

Listen to me. When America banned Nvidia's best chips from China, the plan was simple: no muscle, no machine. On Wednesday in Shanghai, Huawei walked on stage and said: fine. We will grow our own. And then it moved the schedule up.
At Huawei Connect 2026, rotating chairman David Wang announced that the company's next flagship AI accelerator, the Ascend 960DT, is arriving in the first quarter of 2027 — three quarters ahead of schedule. Nobody in chips ships early. Chips slip. This one did not slip. It did a sprint. Its sibling, the 960PR for inference, moves up a quarter to Q3 2027. Then the Ascend 970 in 2028 and the 980 in 2029, one generation a year, each roughly doubling the last. Huawei even has a name for the cadence: "Tao's Law." I respect a man who names his own law.
And then Wang unveiled the big one: an architecture called Peerium, built on Huawei's UnifiedBus interconnect, designed to lash together up to one million processors into a single logical computer. One million. I have been in movies with smaller robot armies.
The specs, no steroids
Here is where I am honest with you, because a champion knows his real numbers. Per The Register's breakdown:
- Ascend 960DT: 288 GB of Huawei's own custom memory, 2 petaFLOPS at FP8, 4 at FP4. Against Nvidia's B300 that is similar memory, but about half the FP8 compute and a third of the FP4.
- Nvidia's Rubin: 35 to 50 petaFLOPS at FP4. One Rubin is roughly ten of these. That is not a rivalry, that is a weight class.
- The 960 SuperPoD: 4,096 chips in one node, 16 exaFLOPS at FP4, and the industry's first use of near-packaged optics — 48,000 pluggable transceivers collapsed into 5,500 units, saving about 550 kilowatts.
- The million-chip cluster: The Register calls it "only theoretical, not something the company has actually proven in the wild," and no independent benchmarks exist at that scale.
So chip against chip, Huawei loses. It knows it loses. DeepSeek's own founder reportedly told investors in July that the previous Ascend trailed Nvidia by about four-to-one and two years.
“Huawei's answer is not a better bicep. It is more arms.”
If you cannot build the strongest chip, you build the biggest system: more chips, homemade memory, homemade optics, a homemade interconnect, all in one rack you control. And here is the line that matters: neither Rubin nor AMD's best is sold in mainland China. Huawei does not have to beat Nvidia. It has to beat nothing. Demand in China already exceeds what Huawei can make, and DeepSeek has reportedly ordered 160,000 of the current-generation Ascend 950DT for a gigawatt-scale data center in Inner Mongolia. Bernstein estimates Nvidia's share of China's AI chip market this year at about 8%.
What this means for the machines you actually use
Follow the muscle to the models. DeepSeek, Qwen, Kimi, GLM — the Chinese models that keep showing up near the top of the leaderboards at a fraction of the price — are about to be trained and served on a domestic stack that gets a new generation every twelve months. Two separate hardware worlds now, each pumping out frontier models on its own schedule. The number of serious contenders is not shrinking. It is doubling, like Tao's Law.
Which means the question "which AI is best?" now has an expiration date measured in weeks, and the honest answer is: it depends on the question, and nobody wins every round. ChatGPT, Claude, Gemini, Grok, DeepSeek — each one lifts heavy in one event and drops the bar in another. So do not marry one machine. Put your question to ChatGPT, Claude, Gemini, Grok and 30 more at the same time and take the single cross-checked, cited verdict they produce together. That is what Gangsta AI is: the whole gym, spotting each other, so no one model's bad rep lands on you.
The chip war just got a second superpower. See which models are winning the actual competition right now on our best AI models rankings. I'll be back — and according to Huawei, three quarters early.
Sources / Receipts
- The Register — Huawei's next-gen Ascend NPUs could become China's best option (full specs, Nvidia comparison, million-NPU caveat)
- TrendForce — Huawei speeds up AI chip roadmap, pulls Ascend 960DT forward three quarters to 1Q27
- Electronics Weekly — Huawei launches ten AI chipsets (David Wang keynote at Huawei Connect)
- Tech Times — Huawei pulls Ascend 960 three quarters early, initiates global AI interconnect standard (SuperPoD scale, DeepSeek order, market share)
- Seoul Economic Daily — Huawei moves up AI chip launch, vowing to become Nvidia
- Hero photo: Huawei headquarters campus in Shenzhen, by RG72 — Wikimedia Commons (CC BY-SA 4.0). Byline avatar: Arnold Schwarzenegger photo by Bernhard Holub, Wikimedia Commons (CC BY-SA 4.0)
More: Best AI models · Compare all AI · Frontier Models · All articles