SMIC posted its Q1 2026 numbers last week. Revenue hit 176.17 billion yuan, up 8.1% year over year. Net profit? A measly 13.61 billion yuan, basically flat at 0.4% growth. The foundry is running at 93.5% average capacity utilization, yet margins sit at 20.1%. I’ve seen factories in Shenzhen with better margins selling USB cables.
But here’s the thing: SMIC’s Q2 guidance calls for 14-16% revenue growth quarter over quarter, with gross margin climbing to 20-22%. The company says it is “more optimistic” about the full year. My sources in Shanghai say the 12-inch expansion is absorbing depreciation faster than expected, and the mix is shifting toward 55nm and 40nm automotive nodes. Don’t expect EUV-level logic here. SMIC is still printing chips the hard way, and the yield numbers on anything below 7nm remain a state secret.
Meanwhile, Huawei dropped the Atlas 350 at its Shenzhen partner conference in March. The card runs the new Ascend 950PR processor, and Huawei claims 1.56 PFLOPS of FP4 compute. That is 2.87 times the throughput of Nvidia’s H20, according to their slides. IT之家 reported the price at roughly 111,000 yuan, or about $16,000.
I don’t buy the FP4 comparison. Hopper-era cards don’t even support FP4 natively, so Huawei is comparing its new format against a product that can’t run it. That is marketing 101. What I do buy is the 112GB of in-house HBM, the 1.4TB/s bandwidth, and the 128-byte memory access granularity. Those are real architectural changes. The LingQu interconnect at 2TB/s is 2.5x faster than the Ascend 910 generation. Huawei is building its own CoWoS workaround because TSMC is off limits. The chip pulls 600W, which is 1.5x the H20’s TDP. You pay for that performance in heat and power bills.
On April 24, Huawei announced that its Ascend super-node lineup now fully supports DeepSeek V4. The timing is not accidental. DeepSeek’s V4 preview dropped that same morning, and Huawei had the compatibility press release ready within hours. That is the kind of coordination you get when the government, the cloud providers, and the chip vendor are all reading from the same script. 36kr has been reporting on this symbiosis for months.
Over in mobile, MediaTek launched the Dimensity 8500 and 9500s on January 15. The 8500 is a modest refresh: 7% CPU uplift, 25% GPU gain over the 8400, still on TSMC N4P. The 9500s gets the flagship treatment with a Cortex-X925 prime core, TSMC second-gen 3nm, and 290 billion-plus transistors. REDMI Turbo 5 Max will launch it first. CounterPoint predicts that in 2026, one-third of all smartphone chips will use 2nm or 3nm nodes. Samsung’s Exynos 2600 is already there. Qualcomm and MediaTek are racing to catch up.
But the real story is automotive. A 36kr report from March says China’s domestic auto chip self-sufficiency rate is still only 25%. Seventy-five percent of the chips in Chinese EVs come from abroad. For high-function-safety SoCs and MCUs, the localization rate is below 5%. Over 200 domestic chip companies exist, but more than 70% have fewer than five product lines. Most are playing in the shallow end of the pool.
The numbers get worse when you look at memory. DRAM prices for automotive applications have tripled in the past three months, according to UBS. Samsung, SK Hynix, and Micron are prioritizing HBM for AI data centers, so mature-node DRAM supply is getting squeezed. 36kr warned that 2026 could bring serious cost pressure and even production interruptions for Chinese EV makers. When your battery management system and your infotainment stack are fighting for the same memory wafers, someone loses.
Huawei’s HiSilicon leads the domestic pack with about 6% market share in automotive chips. Horizon Robotics is next. In advanced driver-assistance SoCs, Huawei holds roughly 23% of the high-end NOA market, trailing Nvidia at 50-63%. The gap is real, and it is not closing fast.
Xiaomi, for its part, spent 331 billion yuan on R&D in 2025 and now holds over 45,000 granted patents globally. Its ESG report claims 99.97% smartphone hardware yield and a quality complaint rate down 18.2%. Those are factory-floor numbers, and they matter. But Xiaomi’s custom SoC plans, rumored for a 2025 launch on TSMC 4nm, have gone quiet. I’ve heard the project is alive but delayed. When you are building cars and phones and AI models at the same time, something slips.
So where does this leave us? SMIC is profitable but not exciting. Huawei is making bold claims it can’t fully verify, but the hardware is real and the ecosystem is tightening. MediaTek is executing on schedule. And the automotive chip gap is a 75% import dependency that no amount of press releases will fix. The foundries matter. The packaging matters. The memory matters. And right now, China only controls one and a half of those three.