China's AI Chip Moment: Cambricon Doubles Revenue, Huawei's 950DT Arrives, and Nvidia's China Share Hits Zero

China's AI Chip Moment: Cambricon Doubles Revenue, Huawei's 950DT Arrives, and Nvidia's China Share Hits Zero

For the past two years, the dominant narrative around Chinese AI has focused on the software layer — the model releases, the benchmark scores, the open-weight licensing debates. Cambricon Technologies changed that conversation on August 7, 2026, when it filed its semi-annual results with the Shanghai Stock Exchange and revealed that its revenue had more than doubled in the first half of the year. The same week, Huawei debuted its Ascend 950DT AI accelerator on its cloud platform ahead of schedule, with DeepSeek identified as a potential early adopter. And Nvidia CEO Jensen Huang, speaking to investors, confirmed what the data had been suggesting for months: his company's market share in China's AI accelerator sector has fallen to zero.

Taken together, these three data points mark a structural inflection point. The hardware layer of Chinese AI independence — long dismissed as aspirational, perpetually "three to five years behind" — is arriving faster than most Western analysts predicted.

Cambricon's Numbers Tell the Story

The headline figure from Cambricon's H1 2026 filing↗ is striking: ¥5.996 billion (approximately US$890 million) in revenue for the first six months of 2026, a 108% year-on-year increase. Net profit attributable to shareholders reached ¥2.311 billion, up 123% from the ¥1.038 billion recorded in the same period of 2025. Stripping out non-recurring gains, adjusted profit grew by 137% to ¥2.166 billion.

The quarterly breakdown is equally revealing. Second-quarter revenue alone hit ¥3.1 billion, slightly exceeding the ¥3 billion consensus estimate from a Bloomberg analyst poll. The company's stock, trading at ¥1,199.93 on August 7, commands a market capitalization of approximately ¥753.91 billion — a 52-week gain of roughly 154%.

"Cambricon's growth is fundamentally linked to Beijing's push for domestic AI hardware substitution. The company is one of the few firms on the government-approved AI hardware procurement list, and that position is translating directly into revenue." — Implicator.ai analysis of the H1 filing↗

The driver is not mysterious. U.S. export controls have effectively blocked Nvidia's most advanced accelerators from the Chinese market, and Beijing has been applying direct pressure on domestic AI companies to replace foreign hardware with locally produced alternatives. Cambricon, whose Siyuan 590 and next-generation Siyuan 690 processors are manufactured on SMIC's N+2 7nm-class process, has been the primary beneficiary of that policy environment.

The ByteDance Dependency — and Its Implications

The concentration of Cambricon's customer base is both its greatest strength and its most significant risk. According to Robonaissance's deep-dive on the company↗, ByteDance has historically accounted for approximately 80% of Cambricon's revenue, with the social media giant reportedly pre-ordering around 200,000 units of the Siyuan 590. The top five clients collectively represent 94% of revenue.

That concentration creates a structural vulnerability. If ByteDance shifts procurement toward Huawei's Ascend series — which it is already evaluating — or develops its own custom silicon (a path DeepSeek and Zhipu AI are reportedly exploring), Cambricon's growth trajectory could stall abruptly. The company's own stock-incentive plan, filed on July 28, 2026, sets a full-year 2026 revenue threshold of ¥13.5 billion for full vesting — meaning Cambricon needs to roughly match its entire H1 performance again in the second half, while facing intensifying competition.

The production side adds further complexity. Tom's Hardware's reporting on Cambricon's 2026 targets↗ notes that the company aims to ship 500,000 AI accelerators this year — a figure that would more than triple its 2025 output. But yield rates on SMIC's N+2 process have been reported at approximately 20%, meaning four out of every five silicon dies are unusable. Inventory carrying value reached ¥8.248 billion by the end of June 2026, representing 45% of total assets — a figure that warrants close attention.

Key metrics from Cambricon's H1 2026 filing:

• Revenue: ¥5.996 billion (US$890 million), up 108% year-on-year
• Net profit: ¥2.311 billion, up 123% year-on-year
• Adjusted profit (ex-non-recurring): ¥2.166 billion, up 137%
• Q2 revenue: ¥3.1 billion, slightly above consensus
• R&D investment: ¥700 million, up 29.63% year-on-year
• Inventory as % of total assets: 45% — a concentration risk worth monitoring
• 2026 full-year revenue target (for stock vesting): ¥13.5 billion

Huawei's 950DT: The Training Chip China Has Been Waiting For

While Cambricon's earnings dominated the financial headlines, the more technically significant development of the week was Huawei's accelerated deployment of the Ascend 950DT on its cloud platform. The chip was originally slated for a fourth-quarter 2026 commercial launch, but Huawei Central confirmed↗ that Vice President Chen Lin announced the August debut at the Huawei Cloud 2026 INSPIRE Creators Event.

The 950DT is architecturally distinct from its sibling, the Ascend 950PR. Where the 950PR is optimized for inference prefill and recommendation workloads — the tasks that dominate day-to-day model serving — the 950DT is engineered for the more demanding work of training and inference decode. Its specifications reflect that focus:

• Memory: 144 GB of HiZQ 2.0 high-bandwidth memory (HBM), delivering 4 TB/s of memory bandwidth
• Interconnect: 2 TB/s interconnect bandwidth — critical for scaling across multi-chip clusters
• Data format support: FP8, MXFP8, MXFP4, and Huawei's proprietary HiF8 format
• Manufacturing: SMIC's N+3 node (7nm-class)
• Architecture: Da Vinci 3.0, with significantly improved vector computing power over the Ascend 910C

"The 950DT's high bandwidth capacity is critical for decode-heavy tasks, placing it in a performance tier comparable to NVIDIA's H200 in specific memory-bound scenarios." — TrendForce analysis↗

DeepSeek as the Proving Ground

The relationship between Huawei and DeepSeek has become the central case study for China's hardware-software co-design ambitions. TrendForce identified DeepSeek's V4.2 series↗ as a potential early adopter of the 950DT, and the engineering history between the two organizations is already substantial.

According to a detailed technical account on Dev.to↗, porting DeepSeek V4 to the Ascend 950 platform required approximately 30 person-years of effort, including rewrites of over 200 core CUDA operators to align with Huawei's CANN (Compute Architecture for Neural Networks) framework. The result was Day 0 compatibility — DeepSeek V4 was available on Huawei Cloud and SuperNode clusters at launch, with subsequent optimizations bringing the Ascend 950PR to competitive efficiency on FP4 compute and MoE inference tasks.

That engineering investment is now paying dividends. As Huawei Central reported↗, the collaboration has validated Huawei's hardware-software co-design approach and positioned the Ascend ecosystem as the de facto training and inference platform for China's frontier AI labs — at least for those operating under domestic compute constraints.

Nvidia's China Share Hits Zero — What That Actually Means

The most geopolitically charged data point of the week came not from a Chinese company but from Nvidia's own CEO. Jensen Huang confirmed to investors that Nvidia's market share in China's AI accelerator sector has fallen to zero — a remarkable admission from the company that, as recently as 2023, supplied the vast majority of the compute powering Chinese AI development.

The trajectory is documented in Morgan Stanley data cited by OfficeChai↗: China's domestic AI chip self-sufficiency has risen from approximately 20% in 2023 to over 40% by mid-2026, with projections suggesting it could approach 85% by 2030 if current investment trends — backed by an estimated $150 billion in state subsidies — continue.

Huawei's own revenue projections tell the same story from the supply side. The company expects its AI chip revenue to reach $12 billion in 2026, up from $7.5 billion in 2025 — a 60% increase driven almost entirely by domestic demand from Alibaba, Baidu, ByteDance, and Tencent, all of which have pivoted away from Western hardware as an operational necessity. Tom's Hardware's reporting↗ notes that Huawei aims to produce approximately 750,000 units of the Ascend 950PR this year alone.

The Gaps That Remain

None of this means China has achieved parity with the global frontier. The Substrate's analysis of China's AI chip supply chain↗ identifies several structural constraints that will not be resolved quickly:

• Manufacturing equipment: China remains unable to access EUV lithography, keeping SMIC's most advanced process at roughly 7nm-class — three to five years behind TSMC's leading edge
• High-bandwidth memory: CXMT is pursuing HBM3-class memory, but domestic capacity remains insufficient for large-scale training, and the company lags approximately three to four years behind SK Hynix and Samsung
• Software ecosystem: Huawei's CANN platform has been open-sourced, but the global CUDA developer base dwarfs the CANN community by orders of magnitude, requiring custom kernel rewrites for advanced architectures
• Yield rates: Cambricon's reported 20% yield on SMIC's N+2 process compares unfavorably to the 90%+ yields TSMC achieves on comparable nodes, making domestic production significantly more expensive per functional chip

The performance gap is real. Domestic chips are estimated to be several generations behind current global standards, and their competitive advantage in the Chinese market is currently defined by availability and "good enough" performance in a restricted environment rather than technological superiority.

What This Week's Data Means for the Global AI Landscape

The convergence of Cambricon's earnings, the Ascend 950DT launch, and Nvidia's zero-share admission is not coincidental. It reflects a structural shift that has been building for three years and is now becoming visible in financial statements and product roadmaps simultaneously.

For developers and enterprises building on Chinese AI models — whether Qwen, DeepSeek, GLM, or Kimi — the hardware story matters because it determines the compute economics that underpin model pricing, availability, and future capability. A Chinese AI ecosystem that can train and serve frontier models on domestic hardware is one that is structurally insulated from Western export controls in a way that was not true even eighteen months ago.

For Western AI companies and policymakers, the data presents a more uncomfortable picture. The export control strategy that was designed to slow Chinese AI development by restricting access to advanced compute has, in practice, accelerated the development of a parallel domestic compute stack. The window during which hardware restrictions could serve as a decisive lever is narrowing — and this week's numbers suggest it may be closing faster than official assessments have acknowledged.

Cambricon's next milestone is clear: it needs to sustain its H1 momentum through a second half that will test both its production capacity and its customer diversification. Huawei's 950DT deployment will be watched closely by every major Chinese AI lab as the first real-world test of whether the training-optimized chip can match its specifications under production workloads. And the broader question — whether China's domestic AI chip ecosystem can close the remaining gaps in manufacturing, memory, and software before the next generation of Western hardware widens them again — remains genuinely open.

What is no longer open is whether the effort is succeeding. The revenue numbers say it is.