China AI chip design industry: demand is growing rapidly, but supply bottlenecks will determine the short-term winners

Overview

This report discusses the domestic AI chip fabless industry in China across demand, supply, customer structure, and valuation frameworks. The core view is that against the backdrop of uncertainty in overseas high-end AI chip supply and rapid growth in China's AI inference demand, domestic AI chip demand will grow quickly; however, advanced-node wafer supply, yield, and capacity ramp-up are the short- to medium-term constraints. Industry differentiation will depend on whether vendors can secure supply, optimize inference efficiency, and win design-ins from leading CSP customers.

Core views

The report believes unit demand for domestic AI chips in China can achieve roughly 40% CAGR over 2026-2030. Tier-1 suppliers, thanks to stronger single-chip performance, foundry capacity support, and relationships with leading CSPs, are more likely to replace U.S. AI chips first; Tier-2 suppliers have a lower revenue base and are currently more focused on government projects and application-oriented projects, but if they have sufficient supply and a product strategy tailored to CSP inference scenarios, they may achieve faster sales growth. In the short term, high inventory and deliverability are more preferred by customers; in the long term, localized supply chains, P/D-disaggregated inference architecture adaptation, and CSP customer breakthroughs are key.

Analysis framework

The report derives unit AI chip demand from token consumption forecasts, and combines assumptions on processor throughput, inference service levels, compute die performance improvements, AIDC utilization, process nodes, and die size to estimate the domestic AI chip supply and AI server sales potential. At the same time, the report classifies vendors into Tier-1 and Tier-2, compares their customer structure, supply chain, inventory, product strategy, and valuation multiples, and uses P/S multiples as the primary valuation method.

Methodology notes

• Demand forecastAI chip demand model driven by token consumption

  Derive AI chip unit demand based on China token consumption, inference workloads, service levels, and processor throughput assumptions.

  The report assumes that rapid growth in token consumption is mainly driven by inference, and estimates chip demand using tokens processed per second per TFLOPS, output speed, batch size, KV Cache length, AIDC utilization, and compute die performance improvement.

• Industry segmentationTier-1 and Tier-2 domestic AI chip fabless classification

  Segment suppliers by revenue scale, CSP customer exposure, chip performance, and capacity support.

  Tier-1 includes Huawei, Cambricon, Kunlunxin, and T-Head, with stronger performance, capacity, or CSP relationships; Tier-2 includes Biren, MetaX, Iluvatar CoreX, Moore Threads, and Enflame, which are currently more focused on government projects and vertical applications.

• Valuation methodsP/S multiple valuation

  Because the industry has a short history and most companies are still loss-making, the market mainly uses forward P/S multiples to value domestic AI chip companies.

  Using Cambricon and domestic peers as references, the report assigns Iluvatar CoreX a 20x forward P/S and MetaX a 30x forward P/S, reflecting growth potential, A/H share risk premiums, and relative valuation levels.

Asset mapping & comparison

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

• Iluvatar CoreX - H (9903.HK)

  The report's top-pick Tier-2 domestic AI chip fabless name.

  Strengths

  It leverages both domestic and international foundry partners, resulting in lower supply disruption risk; its products are well-suited to P/D-disaggregated inference architectures; if it enters the supply chain of leading CSPs, its growth optionality is substantial; the report expects sales CAGR of 102% in 2026-2028 and profitability in 2027.

  Weaknesses

  Its market capitalization remains relatively small, execution uncertainty persists around the industry supply chain and customer design-ins, and its valuation depends on high growth being delivered.

  Comparison

  Relative to MetaX, the report places greater emphasis on its CSP design-win potential, inference efficiency, and upside after valuation discount, which is why it is the top pick.

  Risks

  Further tightening of U.S. controls on foundry services, lower domestic foundry yields, CSP capex coming in below expectations, and rising R&D expense ratios.

• MetaX - A (688802.SH)

  A constructive Tier-2 domestic AI chip fabless name, but rated Neutral.

  Strengths

  It has relatively ample inventory and stronger short-term delivery capability; its supply chain is fully localized, which supports government-backed projects; its product strategy is more prudent, and the N+1 dual-die C600 is expected to meet medium-term demand from 2H26, with a long-term plan to upgrade to N+2.

  Weaknesses

  The report believes its 40x 2027E P/S valuation is somewhat high, and it is currently better suited to waiting for a better entry point.

  Comparison

  Relative to Iluvatar CoreX, MetaX has an advantage in government projects and localized supply chains, but it is less attractive in terms of CSP customer breakthroughs and valuation.

  Risks

  Poor domestic foundry yields affecting performance or delivery, CSP capex coming in below expectations, better availability of overseas chips weakening demand, and rising R&D expense ratios hurting profitability.

• Cambricon (688256.CH)

  An uncovered name, but used as a benchmark for the domestic AI chip leader and for valuation/share-price performance.

  Strengths

  It has benefited from earlier inventory and ramp-up in domestic foundry capacity, resulting in strong revenue growth; its share price has significantly outperformed the CSI 300 since its IPO.

  Weaknesses

  The strong growth may partly reflect prior inventory and capacity allocation, and future growth is still constrained by capacity and yield.

  Comparison

  The report uses Cambricon as the anchor for P/S valuation and market-share gains among domestic AI chip companies.

  Risks

  Supply chain, capacity, yield, and volatility driven by overly high market expectations.

• China AI chip fabless industry

  The core industry theme covered by the report.

  Strengths

  Driven by China's token consumption, inference demand, AI infrastructure investment, domestic substitution, and rising CSP procurement.

  Weaknesses

  Insufficient advanced-node foundry capacity, constraints on key equipment and materials, and higher inventory dependence among Tier-2 vendors.

  Comparison

  Tier-1 is closer to U.S. chip substitution and leading CSP customers; Tier-2 has greater growth optionality but also more pronounced differentiation.

  Risks

  Changes in overseas supply restrictions, slow domestic advanced-node ramp-up, CSP demand volatility, and uncertainty around policy and export controls.

Key data

• Growth in domestic AI chip unit demand in Chinaabout 40% CAGRDerived by the report from 2026-2030E token demand and processor throughput assumptions.
• China token consumption forecastabout 330% CAGRJPM Internet team forecasts rapid growth in China token consumption from 2025 to 2030.
• Domestic AI chip supply forecast2 million / 3 million / 5 million units in 2026E / 2027E / 2028EIncludes GPU and AI ASIC, based on 7nm or N+2, 700mm2 equivalent logic die, and single-die packaging assumptions.
• Corresponding China AI server salesRmb230bn / Rmb326bn / Rmb532bnCorresponding to 2026E / 2027E / 2028E, taking into account stable memory prices and a rising share of supernode solutions.
• Potential ByteDance AI chip demandmore than 0.5 million unitsEstimated based on a 700mm2 equivalent die assumption in a scenario where U.S. AI chips are unavailable.
• Iluvatar CoreX rating and target priceOW; HK$620 Jun-27Based on 20x forward P/S, the report names it as the top pick.
• MetaX rating and target priceNeutral; Rmb550 Jun-27Based on 30x forward P/S; the report views fundamentals as solid but valuation as slightly high.
• Valuation reference for the sectoraverage around 27x 2027E P/SBased on consensus valuations for domestic AI chip companies.

Impact & implications

For investors, the main opportunities in the domestic AI chip industry come from rapid demand growth, constrained overseas supply, and accelerating domestic substitution; however, stock selection should not simply bet on industry demand, but rather identify companies that can secure capacity, maintain deliveries, adapt to inference efficiency, and enter the supply chains of leading CSPs. The report is more constructive on Iluvatar CoreX because of its supply flexibility, benefits from P/D disaggregation, and potential CSP breakthrough; MetaX has strengths in government projects, localized supply chains, and inventory, but its current valuation limits the risk-reward profile.

Risks

• The U.S. further tightens restrictions on Chinese or China-linked entities' access to advanced foundry services, key equipment, materials, and high-end AI chips.
• Domestic advanced-node wafer capacity ramps slowly, yields remain low, or the learning curve is longer than expected, leading to disappointing delivery, performance, and costs.
• CSP capex is below expectations, or end-market AI demand is weaker than expected, reducing AI chip procurement.
• Improved availability of overseas AI chips may reduce the urgency of domestic AI chip substitution.
• New chip designs and tape-outs bring higher R&D expense ratios, affecting profitability.
• If Tier-2 vendors fail to secure design-ins from leading CSPs, revenue growth may remain largely limited to government and vertical projects.
• High valuations depend on strong sales growth being realized; if orders, deliveries, or market share fall short of expectations, share-price volatility could be substantial.

What to watch

• The ramp-up of domestic leading-edge capacity after 2H26 and the actual shipment progress of locally produced chips by Tier-2 vendors.
• Whether Iluvatar CoreX secures design-ins from leading CSPs and the pace of commercialization of its P/D-disaggregated inference architecture.
• The delivery progress of MetaX's N+1 C600 dual-die chip, its N+2 upgrade progress, and the realization of government project orders.
• Changes in the procurement mix and capex of leading CSPs such as ByteDance, Alibaba, and Tencent for domestic AI chips.
• Changes in U.S. export controls, TPP/PD restrictions, HBM, and advanced-node-related policies.
• Whether AI server sales, AIDC utilization, token consumption, and inference workload growth are in line with the report's assumptions.