TLDR
- Huawei launched the Atlas 350 AI accelerator card, powered by its Ascend 950PR chip
- Huawei claims the Atlas 350 delivers 1.56 petaflops of FP4 computing power — 2.8x more than Nvidia’s H20
- Nvidia CEO Jensen Huang confirmed the company restarted H200 manufacturing for China last week
- Analyst KeyBanc estimates China could generate $30 billion in Nvidia revenue this year from ~1.5 million H200 sales
- Nvidia stock was up 1.6% in premarket on Monday but remains range-bound, down 3.28% on the session
Huawei has launched a new AI chip card it says can outperform Nvidia’s best China-approved hardware. The timing is pointed — Nvidia just restarted manufacturing its H200 chips for the Chinese market last week.
Huawei Unveils 950PR AI Chip…delivering performance 2.87x that of the H20
At the Ascend AI Partner Summit held during the "Huawei China Partner Conference 2026" on March 20, Ma Haixu, Huawei Vice President and President of ICT Portfolio Management & Solutions, officially… pic.twitter.com/kUOl5F1uFN
— Jukan (@jukan05) March 21, 2026
The Atlas 350 is an accelerator card built around Huawei’s Ascend 950PR chip. It’s designed for AI inference tasks — the process of generating output from trained models. That covers things like search recommendations, large language model responses, and multimodal content generation.
Huawei VP Ma Haixu unveiled the card at the firm’s China Partner Conference on Friday. Zhang Dixuan, head of Huawei’s Ascend computing business, put the headline number at 1.56 petaflops of FP4 computing power.
That’s a 2.8x improvement over Nvidia’s H20 — the chip Nvidia is currently permitted to sell in China under US export restrictions.
Nvidia’s China Play Just Got More Complicated
Nvidia CEO Jensen Huang said last week the company had restarted manufacturing of its H200 processors for sale in China and had already received orders. The H200 sits above the H20 in performance, but below Nvidia’s current-generation Blackwell chips.
Huang has previously described China as a $50 billion annual market for AI infrastructure, growing at 50% per year. KeyBanc analyst John Vinh estimates Chinese companies could buy around 1.5 million H200 chips this year if sales are permitted at scale — that’s roughly $30 billion in potential revenue.
That’s a big number Nvidia needs. The stock has been stuck in a range and is looking for a catalyst.
Nvidia was up 1.6% in premarket trading on Monday, carried by broader market gains after President Trump said the US and Iran were in talks over a potential deal. But on the session, NVDA closed down 3.28%.
How Real Is the Huawei Threat?
Direct chip comparisons are tricky. Different configurations can produce wildly different benchmarks, and Huawei’s previous AI server systems have drawn criticism for high power consumption compared to Nvidia’s hardware.
Still, a credible domestic alternative matters in China. Companies that might otherwise buy Nvidia chips have a local option — one that isn’t subject to US export policy risk.
Huawei has been building toward this. The Ascend 950PR was unveiled in September 2025 as part of a three-year road map. The company has been developing its AI chip stack without American technology following US sanctions.
On the storage side, Huawei said it plans major upgrades to its OceanStor Dorado and Pacific 9926 systems this year, and will launch the FusionCube A1000 cabinet for SME AI deployment.
“While the first half of the AI era focused on computing power, the second half will be defined by data,” said Yuan Yuan, president of Huawei’s data storage product line.
Nvidia’s H200 restart for China and KeyBanc’s $30 billion revenue estimate are the most recent concrete data points on the table.







