China's AI chip market remains a seller's market, with capacity and system-level capabilities determining competitive differentiation

Overview

This report summarizes a China AI chip market expert call hosted by Deutsche Bank on July 15, 2026. The expert came from the senior R&D management of a leading domestic AI chip company. The core view is that China’s domestic AI chip market remains in a seller’s market with strong demand, manufacturing capacity is the most critical bottleneck, localization continues to improve, and the supply-demand gap is likely to persist through 2026 and 2027.

Core views

The report argues that the competitive landscape of China’s AI chips will shift from strong demand-driven growth to divergence based on supply capability and system capability. Huawei is in the first tier; T-Head, Cambricon, and Kunlunxin are in the second tier; Hygon and the 'four little dragons' such as MetaX, Moore Threads, and Iluvatar are in the third tier; and other emerging smaller vendors are in the fourth tier. Domestic vendors’ hardware specifications are broadly converging, and the real differentiators are supply chain assurance, software compatibility, interconnect capability, and the stability of large-scale customer delivery.

Analysis framework

The report uses an expert call summary format to synthesize changes in China’s AI chip market from dimensions including supply and demand, shipments, localization rate, wafer foundry capacity, vendor tiering, comparisons with NVIDIA, differences between inference and training workloads, and customer procurement strategies.

Methodology notes

• Expert interviewExpert call

  Industry expert call

  Through discussions with senior R&D executives at a domestic AI chip company, the report obtains qualitative judgments on demand, supply, competitive landscape, and technology gaps, and combines these with key indicators such as shipment volume, domestic share, and customer procurement thresholds to form an industry view.

• Competitive landscapeTiering framework

  Vendor tiering

  Domestic AI chip vendors are divided into four tiers based on market presence and influence, to assess supply chain capabilities, customer recognition, and potential share changes.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Huawei / Ascend 950

  Representative of the first tier of domestic AI chips

  Strengths

  Relatively leading in compute power and has begun to partially cover AI factory-like system-level solutions.

  Weaknesses

  Compared with NVIDIA, there is still a gap in overall system-level integration and ecosystem.

  Comparison

  In some metrics such as memory capacity, it can benchmark against NVIDIA H100/H200, but it still lags in compute power and overall system solutions.

  Risks

  Constrained by foundry capacity, process-node generations, software ecosystem, and validation in large-scale training scenarios.

• T-Head (Alibaba)

  Second-tier domestic AI chip vendor

  Strengths

  Has begun partial deployment of system-level solutions and benefits from the backing of a large internet ecosystem.

  Weaknesses

  Compared with globally leading solutions, its complete AI factory architecture is still in catch-up mode.

  Comparison

  Ranks relatively high among domestic vendors, but still trails NVIDIA’s full-stack solution.

  Risks

  Customer validation, supply chain assurance, and software compatibility capabilities still need to be continuously proven.

• Cambricon

  Second-tier domestic AI chip vendor

  Strengths

  Has market presence and a domestic substitution beneficiary thesis.

  Weaknesses

  The report does not provide specific evidence of its leadership in compute power, software, or supply chain.

  Comparison

  Listed in the second tier together with T-Head and Kunlunxin, below Huawei.

  Risks

  Under hardware homogenization, its competitive advantage may be weakened if supply chain and customer delivery capabilities are insufficient.

• Kunlunxin (Baidu)

  Second-tier domestic AI chip vendor

  Strengths

  Backed by Baidu’s application and cloud ecosystem, with potential scenario synergies.

  Weaknesses

  The report does not clearly specify its technical advantages relative to Huawei or NVIDIA.

  Comparison

  Listed in the second tier together with T-Head and Cambricon.

  Risks

  Needs to prove performance and delivery capability in large-customer procurement benchmarks and multi-vendor strategies.

• SMIC and Hua Hong

  Key capacity constraints for domestic AI chips

  Strengths

  They are the main carriers of manufacturing capacity for domestic AI chips and are expected to expand capacity gradually.

  Weaknesses

  Current constraints in capacity, process technology, and technology generations are causing the supply-demand gap to persist.

  Comparison

  Compared with globally advanced foundry capabilities, domestic AI chip supply remains constrained by local manufacturing capacity.

  Risks

  Slower-than-expected capacity expansion would limit growth in domestic AI chip shipments and gains in localization share.

• NVIDIA H100/H200

  Global leading benchmark products

  Strengths

  Provides not only accelerator cards but also a more complete AI factory architecture and system-level capabilities.

  Weaknesses

  In the China market, it faces customer procurement diversification amid local substitution and supply constraints.

  Comparison

  Domestic chips can approach it on some memory-capacity metrics, but still lag in compute power, interconnect, and system integration.

  Risks

  If domestic supply chains continue to improve, some inference scenarios may see accelerated penetration by domestic substitution.

Key data

• Total shipments in 2025about 4 million unitsThe expert said total shipments in 2025 were about 4mn units.
• Expected shipments in 2026about 5 million unitsThe expert expects shipments to expand toward 5 million units in 2026.
• Domestic shipment CAGR over the next 2-3 years>30%The expert expects domestic AI chip shipments to grow at a compound rate of over 30% in the next 2-3 years.
• Change in domestic sharefrom about 40% to above 50%The report mentions a significant localization trend, with domestic share expected to continue rising.
• Installed-base ratio of inference to trainingabout 10:1The installed base of domestic AI chips is clearly skewed toward inference workloads.
• CSP single-vendor delivery thresholdmore than 50,000 unitsInternet cloud service providers usually require vendors to have large-scale delivery capability.

Impact & implications

For investment and industry judgment, the opportunities for domestic AI chips come from strong demand growth, domestic substitution, and customers’ multi-vendor strategies; the main constraints come from advanced-node capacity, training workload capability, interconnects, and system-level solutions. In the short term, vendors with stable supply chains, software ecosystems, and large-customer delivery capabilities are more likely to benefit; in the long term, it will be necessary to watch whether domestic vendors can upgrade from single-chip competition to AI factory-level system architecture competition.

Risks

• The supply-demand gap may persist through 2026 and 2027, limiting shipment realization.
• There is uncertainty around the pace of foundry capacity expansion at SMIC and Hua Hong.
• Domestic chips still have capability shortfalls in training, especially in large-model pre-training scenarios.
• Converging hardware specifications may push competition toward price, supply chain, and software ecosystem, putting pressure on weaker vendors’ market share.
• Compared with NVIDIA, domestic solutions still lag in compute power, interconnect, and AI factory-level system integration.
• Although customers’ multi-vendor strategies reduce procurement risk, they may also lengthen certification and volume ramp-up cycles.

What to watch

• Whether domestic AI chip shipments in 2026 can approach 5 million units.
• Whether domestic share can rise from about 40% to above 50%.
• The progress of capacity expansion at SMIC and Hua Hong and improvements in process capabilities.
• Whether vendors such as Huawei and T-Head can expand their system-level solution capabilities.
• Validation progress of domestic chips in training and large-model pre-training scenarios.
• The actual share allocation under multi-vendor strategies in procurement by internet CSPs, telecom operators, and SOEs.