Huawei dropped the Ascend 950 in Q1. Two variants: 950PR for inference, 950DT for training. Both use Huawei’s self-developed HBM (高带宽内存), which is the real headline here. Before this, only SK Hynix and a handful of others could mass-produce HBM. Now Huawei has its own.
The 950PR runs on SMIC’s N+3 node — that’s 5nm-equivalent without EUV, using multi-patterning DUV. I’ve seen this before. The yield numbers are never what the press releases claim. But the fact that Huawei can ship it at scale, with domestic HBM, means the “independent kingdom” (独立王国) that Nikkei wrote about is no longer a metaphor. It’s a supply chain.
Here’s the spec sheet, per 证券时报 and Huawei’s own Connect 2025 keynote: 950PR hits 1 PFLOPS at FP8, 128GB HBM at 1.6 TB/s bandwidth. 950DT bumps that to 144GB and 4 TB/s. The interconnect, called Lingqu (灵衢), pushes 2 TB/s — faster than NVLink 5.0’s 1.8 TB/s. Single-card, Huawei admits they trail NVIDIA’s Blackwell. But cluster-to-cluster? The Atlas 950 SuperPoD scales to 8,192 cards. Huawei says that beats NVIDIA’s NVL576, which doesn’t ship until 2027.
I don’t buy all the benchmark claims. What I buy is that ByteDance, Baidu, and Zhipu are already deploying these clusters. When your alternative is an export-controlled H100 or a marked-up Blackwell, a 30-40% cheaper domestic card with decent software support starts looking like a bargain. Morgan Stanley predicted Huawei plus Cambricon would ship over 1 million AI chips in 2026. That’s double 2024.
Meanwhile, SMIC is throwing 54.4 billion RMB at expansion this year — 94% of its 2024 revenue. The goal: push “advanced” (7nm and below) wafer output from under 20,000 wafers per month to 100,000 within two years. That’s a 5x jump. TechNews reported that SMIC, Hua Hong, and Huawei-linked fabs like Pengxin Micro are all copying the same N+2/N+3 playbook across multiple sites. Beijing is replicating what Shanghai proved.
The memory side is where the real money is moving. DRAM contract prices jumped 55-60% in Q1 2026. NAND flash is up 33-38%. A server DIMM that cost $255 in Q3 2025 now lists at $900. SK Hynix, Samsung, and Micron have sold out their 2026 capacity. They’re diverting everything to HBM and server DDR5, starving the consumer market.
This is killing the low-end phone business. Counterpoint Research says 2026 global smartphone SoC shipments will drop 7%, with sub-$150 models taking the worst hit. Transsion’s profit halved last year because it couldn’t pass memory costs to African buyers. Xiaomi, OPPO, and vivo have all cut their 2026 shipment forecasts by 15-20%.
But the pain is creating opportunity for domestic memory. Changxin Storage (长鑫存储) is running its Hefei and Beijing fabs flat-out. It’s planning an HBM line by end of 2026. Yangtze Memory (长江存储) started its third Wuhan phase in September 2025 and is already installing equipment. Both companies filed for IPO in May — Changxin on the STAR Market, Yangtze with CITIC Securities. Changxin’s Q1 revenue was 50.8 billion RMB, up 719% year-over-year. Net profit hit 24.8 billion RMB, up 1,688%. Those are not typos.
On the EV front, the March numbers from CPCA are out. Tesla Model Y still leads with 55,856 units. But look down the list: Geely Xingyuan (吉利星愿) is second at 40,613. Li Auto’s new i6 hit 24,198. Xiaomi YU7 slipped 33% month-over-month to 13,558 — supply constraint or demand fatigue, hard to tell. NIO ES8 is up 3,000% year-over-year, though that’s from a comically low base of 514 units last March.
BYD sold over 210,000 NEVs in January, its 56th consecutive monthly crown. Full-year 2025 was 4.6 million vehicles. Overseas sales crossed 1 million for the first time, up 145%. The company is now a top-five global automaker by volume. I’ve been to the Shenzhen headquarters. The vertical integration there — batteries, motors, chips, software — is what lets them undercut everyone else and still make money.
Huawei’s Mate 80 Pro Max 风驰版 launched March 23 at ¥8,499. The full lineup includes the Enjoy 90 series starting at ¥1,299. HarmonyOS 6 is the default now. Huawei reclaimed the 2025 China smartphone crown with 46.7 million units, just ahead of Apple’s 46.2 million. The gap between first and fifth place was only 3.3 million units. This market is a knife fight in a phone booth (电话亭里的刀战).
What’s next? SMIC’s 5nm-equivalent “N+3” is already in risk production. Huawei’s roadmap shows Ascend 960 in Q4 2027 and 970 in Q4 2028. The Shanghai Changxin fab, a 345 billion RMB project, breaks ground this year with 400,000-600,000 wafers per month planned. If even half of that hits HBM3 production, the AI training cost curve in China bends downward fast.
The U.S. sanctions were supposed to slow China down. Instead, they forced a complete rebuild of the supply chain from design tools to fab equipment to memory. It’s messy. Yields are lower. Costs are higher. But it’s working. The question is no longer whether China can make advanced chips. It’s whether they can make them cheap enough to matter globally.