tphuang

Battery makers & their top customers in Aug: Interesting to see eHDT being more prominent now. Sinotruk is #4 customer of BYD Batteries & #2 for CALB. XCMG is 4th for CALB. Biggest CATL customers are Geely, ChangAn, Tesla, Xiaomi & IO Leapmotor is top OEM for Gotion & Eve https://t.co/N8PSyQ13GR

Alibaba Cloud says its 1st Zhenwu M890 Cluster in Ulanqab is in service w/ 130k card & it will start operation w/ 2nd such SuperCluster in Ningxia Q4. Comes as ODM Huaqin tech express it's ramping up SuperNode type of product in Q3 & Q4. Consider 2026 the Yr 0 of China's AI SuperNode. Production will ramp up significantly over next 2 yr as Alibaba will build its V900 SuperNode & 500k-card cluster after it gets launched in 2027Q1. Alibaba's compute should scale significantly in 2027/28.

Alibaba said V900 is the strongest domestic AI chip. Just how accurate is this claim? Based on HBM size + likely bandwidth, it's in the same class as Ascend-960DT & Blackwell. Will be part of 1024-card SuperNode & 500k-card Cluster vs 128 + 122k for M890. So, the full V900 cluster is a huge upgrade over the M890 cluster. roughly 12x more powerful w/ 6x more HBM. But is it actually the best in China? Well, it was 2nd place w/ 16% share vs 49% for Ascend amongst Chinese AI chipmakers in 2025 according to IDC. So while it's clearly growing rapidly in total compute & memory, it likely falls behind that of HW & Ascend team, who is not a major competitor to other AI labs. Alibaba's AI cluster is likely to just serve its own cloud computing team. It's entire model is to get corporate customers to use Alibaba Cloud, whether that's through using Qwen models directly or Qwen work or running their models on Alibaba cloud (like Moonshot & Minimax).

Alibaba now says its Cloud compute will grow to 20GW globally by 2032. So it should probably be adding 4-5GW/yr by 2030-2032 range. Using all its own chips here including CPUs. On top of its successful Yitian-710 project, it's also launching 192-core Yitian-720 & 32-core SMT2 Yitian-730 by next Q3 followed by Yitian-750 later. This allows it to conserve capital on all these expansion projects.

Alibaba T-Head is releasing the Zhenwu V900 in Q1 (1 yr after M890) followed by J900 in late 2028. They claim 3x performance boost each generation. It will come as part of new SuperNode, that seems to reach 1024-nodes (so 16x cabinets) supported by self developed ICN Switch & NIC card chips + eSSD chips. All the important chips you need in an AI cluster.

Last yr, I spoke w/ @GlennLuk about Alibaba being the google of China w/ its full stack AI approach & they are showcasing it again this yr @ its APSARA conference in Hangzhou. Keep in mind here, when it says 5-10T sized model. That's for Qwen & Kimi. Once Chinese labs don't lack data & compute, they naturally will train much larger models.

Last yr, I spoke w/ @GlennLuk about Alibaba being the google of China w/ its full stack AI approach & they are showcasing it again this yr @ its APSARA conference in Hangzhou. Keep in mind here, when it says 5-10T sized model. That's for Qwen & Kimi. Once Chinese labs don't lack data & compute, they naturally will train much larger models.

Horizon Robotics continue to be on a row. It now signed deal to put its J6H chip & HSD 2.0 product on ID. Aura T6 pure electric SUV. Its rapidly going to displace Nvidia in China's ADAS market. https://t.co/c0CovxfQ4R

As models get larger, larger SuperPoD allows for higher MFU vs cluster built w/ 8-card server or smaller SuperNodes. If 4k is 2.75x & 352 card is 2.06x, then 4k is effectively 1/3 higher in MFU, since less time spent on memory transfers & networking. https://t.co/EtglXyJCVB https://t.co/Z7MRJh7W82

Ascend Timeline - 2025/9 vs 2026/9 1Q+ jump in 960 availability Huge change in focus from FP8 -> FP4 FP4 used to be 2x FP8, now its 3-4x FP8 Big jump from 970 to 980 in memory speed (14.4 -> 38.4 TB/s) & interconnect speed (4.4 -> 8 TB/s) ofc the real big change is when 990 comes along in 2030 w/ logic folding