Did Alibaba Just Solve China's AI Chip Problem?

0:00 Alibaba just called its new chip China's most powerful AI chip and at launch it didn't publish a single number for how fast that chip actually computes. Not at any precision, not even the formats it

0:10 was built for. That blank matters. If you pick hardware or rent GPUs, you need to know which of these numbers are published specs and which are just claims. Same if you'd rather not repeat

0:21 a headline that falls apart next month. Because on paper, this chip does beat Nvidia's H200 in two places. How much memory it holds and how fast chips talk to each other. How fast it actually

0:32 computes isn't one of them. At least not in anything Alibaba's published. Here's one worked example. Alibaba's own flagship model, Qwen 3.8 Max, has 2.4 trillion parameters. How many of these

0:44 new chips does it take just to hold that model versus Nvidia's H200? The answer's simple arithmetic and it favors Alibaba. What it doesn't tell you is whether the chip's any faster. It launched on

0:56 September 22nd at Alibaba's Apsara Conference in Hangzhou, where its chip unit T-Head unveiled the Zhen Ru 900. Reuters reported three times the performance of the previous chip, the

1:08 M890, and clusters of up to 500,000 chips. And that phrase, China's most powerful AI chip, AP reported it as Alibaba's own description. Now sort the spec sheet into two piles, what Alibaba

1:21 disclosed, 216 GB of memory per chip, 1.2 TB per second between chips, native FP8 and FP4, those 500,000 chip clusters, and the three pies claim. What it held back, flops, memory

1:36 bandwidth, process node, foundry, and power draw. Hardware Busters put that gap right in its headline. Even the first pile is Alibaba's word, passed on by the press with no independent

1:47 benchmarks and no ML perf results yet. Then there's that three pies. It's measured against the M890, which launched on May 20th with 144 GB of memory and 800 GB per second between

1:59 chips, the M890 also shipped without published flops or ML perf results. Three times an unknown number is still an unknown number, just a bigger one. For what it's worth, the specs Alibaba

2:11 did publish went up 1 and 1/2 times, 144 to 216 GB, 800 to 1,200 GB per second. So, whatever the three times measures, it isn't either of those. Okay, the worked example. Qwen 3.8 Max, 2.4

2:28 trillion parameters. The V900 supports FP8, where each parameter takes 1 byte. So, the weights alone come to 2.4 trillion bytes, about 2.4 terabytes. Picture that as 2,400 1 GB blocks that

2:42 need to sit in the chip's memory before the model can answer a single prompt. Pour them into V900S at 216 GB each. 11 chips gets you to 2,376, just short. So, you need 12. Now, H200S

2:58 at Nvidia's official 141 GB each. 17 gets you to 2,397, 3 GB short. So, you need 18. 12 versus 18. That's a real advantage on paper. A third fewer chips to hold the same

3:12 model. One caveat on my own math, it's weights only. A running model also needs memory for its KV cache, activations, and redundancy. And Alibaba hasn't said what precision Qwen 3.8 Max actually

3:25 runs at. So, treat and 18 as floors, not a deployment plan. Now, the second question. Does needing fewer chips make it faster? Think of memory as a water tank and memory bandwidth as the pipe

3:36 out of it. When a model generates text, it typically has to read those weights out of memory again and again, so the pipe often sets the pace. A bigger tank doesn't widen the pipe. Nvidia publishes

3:47 the H200's pipe, 4.8 terabytes per second. Alibaba hasn't published the V900s at all. A huge memory number with no bandwidth figure is like a car ad that only lists trunk space. The other

4:00 number Alibaba did publish is the link between chips. Split a model across 12 chips and they're passing results to each other constantly, so that link matters. The V900s is 1.2 terabytes per

4:11 second. That beats the H200s NVLink at 900 GB per second. It trails Nvidia's Blackwell HGX platforms, which list 1.8 terabytes per second GPU to GPU. So, the accurate

4:24 version of beats Nvidia is much narrower. On paper, it beats the H200 in two numbers, memory capacity and chip-to-chip speed. Against Blackwell's link, it's behind. On compute and memory

4:36 bandwidth, there's nothing to compare yet. What about China's most powerful? The obvious domestic rival is Huawei. It's Ascend 950 DT is due in the fourth quarter of 2026.

4:48 It lists 144 GB of memory, 4 terabytes per second of memory bandwidth, and 2 terabytes per second of interconnect. And the Ascend 950 series targets 1 petaflops at FP8. The V900 has more

5:01 memory. Huawei's link is faster on paper. And Huawei published bandwidth figures and a compute target where Alibaba published neither. So, Alibaba's most

5:10 powerful claim can't be checked. It might well be true. Right now, it's a claim sitting next to an empty scorecard. So, why build a chip like this at all? Go back a year.

5:19 At Apsara 2025, Alibaba laid out a roadmap to artificial superintelligence in three stages. First, emergent intelligence, then autonomous action, then self-iteration. That last one's

5:31 where AI connects to raw data from the physical world, learns on its own, and in Alibaba's words, ultimately surpasses humans. It came with a three-year AI infrastructure plan worth 380 billion

5:42 yuan, about 53.2 billion dollars. This year, Alibaba CEO Eddie Wu pushed further. He said the truly groundbreaking products of the machine intelligence era

5:52 have not yet arrived. Qwen 4 is in training. Four variants, Max, Flash, Plus, and 27B, were named on stage with no release date, price, or weights. After that, Qwen 4.5 or 5 is projected

6:06 at 5 to 10 trillion parameters. Run the same arithmetic. 10 trillion parameters at FP8 is 10 terabytes, so at least 47 V900S just for the weights. Those are plans, not products. Still, they make

6:19 memory per chip an obvious number for Alibaba to care about. Then there's scale. Alibaba's target is more than 20 gigawatts of capacity by 2032, and the V900 is built to link in clusters of up

6:31 to half a million chips. Wu was blunt about the bottleneck. The industry's mid- to long-term demand far outpaces our supply capabilities. Designing your own chip is one way to ease that, though

6:42 V900 mass production doesn't start until early 2027. The other half of the story is the models themselves, and here the independent data is strong. Hugging Face's state of open models report from

6:53 mid-August counts about 2.05 billion Qwen downloads so far in 2026, and 151,448 models built on Qwen. That's the most derivatives of any organization on the

7:05 hub, ahead of Google's 82,506 and 2.6 times Meta's total footprint. Alibaba's own numbers are bigger, more than 3 billion downloads and over 300,000 derivatives. Part of that gap is

7:18 real because Hugging Face can't see model scope, API traffic, or private deployments. Still, Alibaba's derivative count is roughly double Hugging Face's. When you're already first on the

7:28 scoreboard, bringing your own scoreboard is a bold choice. I'd quote Hugging Face's figures. They make the case on their own. So, back to the question from the start. 12 V900S versus 18 H200S

7:40 tells you the V900 can hold a giant model on fewer chips. It doesn't tell you it's faster because the numbers that decide speed, flops, and memory bandwidth weren't published. What

7:50 Alibaba showed on paper, more memory than the H200 and Huawei's 950DT, and a faster chip-to-chip link than the H200. What it only claimed, the three bucks, China's most powerful in any

8:03 sense that it beats Nvidia overall. My verdict, if you're an ML engineer or a team choosing hardware, treat the V900 as a memory capacity story until flops, memory bandwidth, or ML perf results

8:14 show up. It isn't in mass production until early 2027 anyway. If you build on open models, Quen's lead as the base people build on is real, and hugging faces own data backs it. The question

8:26 I'd watch, when Alibaba finally publishes the V900's memory bandwidth, does it land anywhere near the H200's 4.8 terabytes per second?