Alibaba Unveils China's Most Powerful AI Chip

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Alibaba Unveils China’s Most Powerful AI Chip · Channel: Fahd Mirza · 08:03 · Watch on YouTube

Topics: AI infrastructure and hardware, Markets and economy, AI models

Summary

The video discusses Alibaba’s Xuantie V900, presented at the Apsara Conference as a next-generation AI chip with three times the performance of its predecessor. The host highlights its reported memory, chip-to-chip bandwidth, low-precision support, and ability to interconnect up to 1,024 cards, framing it as part of Alibaba’s in-house computing, storage, and networking stack. The launch is presented as a response to China’s limited access to foreign chips, alongside Alibaba’s plans for its own AI models and data-center infrastructure. The host cautions that the specifications have not been independently benchmarked and that mass production is not expected to begin until the first quarter of 2027.

Key points

Alibaba introduced the Xuantie V900 at its Apsara Conference and said it offers three times the performance of its predecessor.
The video identifies T-Head, also known as PingTouGe, as Alibaba’s in-house chip design unit.
The V900 is described as supporting both AI training and inference.
Reported specifications include 216 GB of memory, 1,200 GB/s of inter-chip bandwidth, and FP8 and FP4 precision.
Alibaba says the chip can support up to 1,024 fully interconnected cards.
The host describes the V900 as part of an in-house stack that includes chips, switches, and storage controllers.
The video connects the chip’s development to restrictions and obstacles around Nvidia chip sales to China.
Alibaba also discussed training Quan models and targeting models with 5 to 10 trillion parameters.
The host notes that the claims have not been independently benchmarked and mass production is not due to begin until Q1 2027.
The video raises open questions about price, CUDA, and whether models trained on the chip can be ported to Nvidia or AMD GPUs.

Chapters

Start	Chapter	Summary
00:01	Alibaba announces the Xuantie chip	The host introduces Alibaba’s claim that its next-generation AI chip is China’s most powerful and outlines the significance of the announcement.
01:43	Reported V900 specifications	The video describes the V900 as a training-and-inference chip and reviews its reported memory, bandwidth, precision support, and scaling capacity.
03:49	Alibaba’s in-house infrastructure	The host presents the chip alongside other Alibaba silicon and explains the company’s ambition to build its computing, storage, and networking stack in-house.
05:23	Chip restrictions and Alibaba’s broader AI plans	The video relates the chip’s development to restrictions on Nvidia hardware in China and discusses Alibaba’s plans for its models and data-center capacity.
07:01	Caveats and open questions	The host emphasizes that the chip’s claims remain unverified, notes the expected production timeline, and raises questions about software compatibility and portability.

Quotes

The most powerful AI chip China has ever produced. — 00:36

Support for 1,024 cards, fully interconnected at full bandwidth. — 02:45

This is Alibaba building an entire computing, storage, and networking stack in-house. — 03:49

it’s all Alibaba’s own claim. Not yet independently benchmarked, by the way. — 06:28

We don’t need Nvidia anymore. — 07:01

Concepts and entities

Alibaba, T-Head, PingTouGe, Xuantie V900, ICN switch, Yitian, Quan 4.5, Quan 5, NVIDIA, CUDA, Eddie Wu

Full transcript (original language)

00:01 Let’s start this video with a short clip from Alibaba’s Apsara Conference. Alibaba introduced the Xuantie 900, its next generation AI chip that it said delivers three times the performance of its predecessor, the Xuantie May. This chip is being said to be the most powerful AI chip ever from China. Every year Alibaba throws open the doors on its biggest stage of the year, the

00:36 Apsara Conference in Hangzhou, and this year they used it to drop a bombshell. On September 22, 2026, their CEO Eddie Wu stood up there and called it plainly, “The most powerful AI chip China has ever produced.” Within minutes, Alibaba’s Hong Kong shares jumped over 5%. This wasn’t a footnote in a keynote. This was Alibaba telling the world and telling Washington in particular that it no longer needs anyone’s permission to

01:09 build frontier AI. And that’s a big, big deal. So, who actually built this thing? Not Alibaba the shopping app, not Alibaba the cloud giant, but a quieter, sharper arm of the company called T-Head, also known as PingTouGe. This is Alibaba’s in-house chip design unit, the team that has been grinding away since 2018 on exactly one mission, breaking China’s dependence on foreign silicon. T-Head didn’t just show up with one

01:43 product, either. They walked on stage with a full stack, chips, interconnects, storage controllers that all designed under one roof. This chip has a name that sounds almost mythological, Xuantie, and today we are going to unpack exactly what it can do. This is Fahad Mirza, and I welcome you to the channel. Recently, I also have started this free weekly AI newsletter, which you can subscribe at fahadmirza.substack.com.

02:14 Now, this is the actual product page, and here is where the numbers get real. The Xuantie V900 is what they call a training and inference AI chip, which means that one chip can handle both building the model and running it, instead of needing separate hardware for each job. >> [clears throat] >> Look at the advantages in Chinese which they have listed here. You can take the screenshot and get it translated from any AI. So, it is fully self-developed, from the hardware all the way up through

02:45 the parallel computing architecture, high performance with 216 GB of memory, 1,200 GB/s of inter-chip bandwidth, and support for FP8 and FP4 precision. And this line right here is the one that should stop you. Support for 1,024 cards, fully interconnected at full bandwidth. That’s over 1,000 chips acting as a single brain. And that is where I think

03:16 they get unblocked in training huge models like Claude or, you know, OpenAI’s model. And that is why I think that this is a big big deal. So, there’s a lot of other information here, but let me take you to some of the screenshots which I have taken from X from Alibaba’s different sources. So, this photo tells the real story better than any spec sheet, as you can see. That’s the Xuantie chip on stage, flanked by two other pieces of Alibaba

03:49 silicon, the ICN switch and the Yitian chip. The banner above them says it plainly, the foundation for AI infrastructure. This is not one chip trying to be a hero. This is Alibaba building an entire computing, storage, and networking stack in-house. So that when they scale a cluster to, let’s say, 500,000 chips, every single piece of that system, right down to the switch connecting them, is something they own and control.

04:20 Now, this is another image which I took from social media. I don’t know who owns it. That is why I couldn’t really credit it, but this is from X. And this is the chip itself, by the way, sitting in its display case with the numbers spelled out plainly. 216 GB of memory, 1,200 GB per second of chip-to-chip interconnect, FP8 and FP4 low-precision support. This is not a marketing fluff. This is literal spec card Alibaba will put next to the physical silicon at their own

04:52 conference. Three times the performance of its predecessor, the M890, which itself only launched back in just May, 4 months. Even less. That’s how fast this road map is moving. Now, here is why any of this matters. Back in January 2026, the US approved Nvidia’s H200 for export to China with a 25% tariff attached. Days later, Beijing blocked it at customs anyway.

05:23 Chinese officials quietly told domestic companies not to buy the chip unless absolutely necessary. Nvidia’s China data center revenue effectively collapsed to zero. So, picture the position Alibaba was in. The best chips on earth sitting right there, and they were not allowed to touch them. The Xuantie V900 isn’t happening in spite of that. It’s happening because of it. That is just my own thinking. And the chip is only half of Alibaba’s bet. On the same stage,

05:57 they confirmed Quan Forest training right now with Quan 4.5 and Quan 5 targeting 5 to 10 trillion parameters, up to four times the size of today’s flagship Quan 3.8 max. Chips built in-house, models built in-house, and a promise to push data center capacity past 20 gigabytes by 2032. This is a full loop. Silicon, software, and infrastructure all coming from one company. And you know that there is just

06:28 not one company in China. There are, you know, it’s raining AI companies in China at the moment. And you can see on my channel that every day there is a new company releasing model. Anyway, so step back for a moment. Everything you just saw, the specs, the demo, case, the roadmap, it’s all Alibaba’s own claim. Not yet independently benchmarked, by the way. So we have to be realistic, too. And mass production doesn’t even begin until the first quarter of 2027.

07:01 But the direction is unmistakable. China’s biggest tech company just stood on a stage and said out loud, “We don’t need Nvidia anymore.” Whether the V900 actually delivers is a question for 2027. That they built it at all is the story for today, in my opinion. Yes, I know there are a lot of questions which need answering still. The price, CUDA angle, which is, I think, I believe is the, you know, biggest question mark on

07:32 AI training and inference at the moment. Would it be acceptable to port these models if they get trained on these chips to, let’s say, Nvidia GPUs or AMD GPUs because that thing matters a lot. Anyway, I just wanted to share this information because I think this is a big deal and lot of things are going to get sprouted from here. Let me know your thoughts in the comments. Again, please subscribe to this newsletter. Um it is free. I won’t spam you, only

08:03 once a week. And if you want to help out the channel, please become a member. Just go to the home page or to the first print comment. Thank you for all the support.