China AI chip fabless enters commercialization stage, with supply bottlenecks and CSP breakthroughs driving divergence

Overview

This is a J.P. Morgan research report on the China AI chip fabless sector, focusing on commercialization opportunities for domestic GPUs and AI ASICs under AI infrastructure investment, U.S. AI chip supply uncertainty, shifts in CSP procurement, and domestic advanced-node capacity constraints. The report argues that after the past 3 to 4 months of intensive IPOs by Chinese AI chip companies, the industry has entered a commercialization and industrialization stage for domestic GPUs and AI ASICs. The near-term key constraint is supply chain and advanced-node capacity, while the long-term growth driver comes from CSP customer breakthroughs, product strategy, and inference-efficiency optimization.

Core views

The report's core view is: first, demand for domestic Chinese AI chips will grow rapidly, with unit demand expected to expand at about a 40% five-year CAGR through 2030; second, tier-1 suppliers have advantages in performance, capacity, and CSP relationships, and may be the first to replace U.S. AI chips; third, tier-2 suppliers have a lower revenue base but higher optionality, and whether they outperform depends on stable foundry capacity, sufficient inventory, access to top CSP customers, and whether they can offer better TCO for inference workloads; fourth, prefill-decode separation will create opportunities for domestic chips suited to specific inference stages. At the company level, the report is more constructive on Iluvatar CoreX's supply flexibility and CSP breakthrough potential, while taking a relatively cautious stance on MetaX.

Analysis framework

The report analyzes the industry along four lines: demand, supply, customers, and valuation. On the demand side, it derives unit chip demand from China's token consumption growth, inference service levels, and AI infrastructure utilization; on the supply side, it evaluates U.S. export controls, constraints on access to international advanced foundry services, the pace of domestic capacity release, inventory buffers, and process-node migration; on the customer side, it focuses on the procurement shares and design-in opportunities at CSPs such as ByteDance, Alibaba, and Tencent; and on the valuation side, it uses P/S multiples, with domestic AI chip companies such as Cambricon as the market anchor.

Methodology notes

• Demand ForecastingDeriving chip demand from token demand

  Use token consumption, inference throughput, and service-level assumptions to derive AI processor demand.

  The report uses throughput assumptions per processor and inference service-level settings, such as 20 tokens/second output, batch size 128, and KV cache length 4,096, to estimate domestic AI chip unit demand in China, leading to a 2026-2030E CAGR of about 40%.

• Supply AnalysisFoundry capacity and inventory buffer framework

  Assess short- and medium-term shipment capacity by looking at advanced-node availability, domestic capacity ramp, inventory levels, and yields.

  The report argues that advanced wafer manufacturing capacity is the current key bottleneck; second-tier suppliers rely on existing inventories in the near term, and leading domestic process capacity is expected to support large-scale shipments of local chips from 2H26 onward.

• Customer BreakthroughCSP design-in and TCO framework

  Supplier win probability depends on CSP procurement share, mass-delivery capability, system efficiency, and total cost of ownership.

  The report believes China's leading CSPs will consume more than half of domestic AI chips and may further increase domestic procurement; suppliers that can guarantee large-scale shipments and improve system efficiency and TCO are more likely to win design-in opportunities.

• Valuation methodsP/S relative valuation

  Use sales multiples to value high-growth domestic AI chip companies that are not yet fully profitable.

  The report notes that the market widely uses P/S multiples to value domestic AI chip companies and uses Cambricon as the anchor; Iluvatar CoreX's target price is based on 20x forward P/S, and MetaX's target price is based on 30x forward P/S.

Asset mapping & comparison

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

• Iluvatar CoreX (9903.HK)

  The report initiates coverage and assigns an OW rating, Jun-27 target price HK$620.0.

  Strengths

  Its product mix can leverage both local and international foundry services, so the supply bottleneck is relatively limited; it has CUDA compatibility; customized products are well suited to prefill-decode separated inference infrastructure; the report expects 2026-2028E sales CAGR of 102% and profitability in 2027.

  Weaknesses

  It is still subject to advanced-node capacity constraints, CSP design-in timing, and execution risk on ramping new chips; the target multiple needs to reflect the A-share/H-share market risk premium.

  Comparison

  Compared with other second-tier peers, the report believes Iluvatar CoreX is more likely to enter top CSP accounts thanks to ample supply and products suited to inference optimization.

  Risks

  Tighter U.S. restrictions on foundry services, domestic foundry yields below expectations, weaker-than-expected CSP capex, recovery in overseas chip supply, and rising R&D expense ratio.

• MetaX (688802.SH)

  The report initiates coverage and assigns a Neutral rating, Jun-27 target price Rmb550.0.

  Strengths

  It has sufficient inventory in the near term, with end-2025 inventory at about Rmb1.5bn; it uses a fully domestic supply chain; the N+1 dual-die C600 chip is expected to ramp in 2H26; and it has opportunities to gain government-supported AIDC, state-owned enterprise, telecom operator, and CSP customers.

  Weaknesses

  The valuation looks somewhat stretched at about 40x 2027E P/S; CSP customer breakthroughs still require faster or more substantive progress; medium- to long-term performance and power-efficiency improvements depend on subsequent process upgrades such as N+2.

  Comparison

  Compared with Iluvatar CoreX, MetaX has a higher degree of supply-chain localization and a more conservative strategy, but the report does not see the current share-price entry point as attractive enough.

  Risks

  Domestic foundry yield affecting performance or delivery, CSP capex coming in below expectations, improved availability of overseas chips, rising R&D expense ratio, and slower-than-expected CSP adoption.

• Cambricon (688256.CH, not covered)

  Discussed as the valuation and growth anchor for domestic AI chips.

  Strengths

  Its share price has risen more than fivefold since IPO, benefiting from rapid revenue growth, market-share gains, inventory buildup, and greater domestic foundry capacity allocation.

  Weaknesses

  The report does not cover the company, and the discussion is mainly used as an industry reference and valuation anchor.

  Comparison

  The market tends to use Cambricon as the P/S valuation anchor for domestic AI chip companies; its forward P/S is about 27x, versus a historical average valuation of about 51x.

  Risks

  The valuation is highly sensitive to revenue growth, market-share gains, and supply-demand disruptions.

• China AI chip industry

  The report believes the industry has entered a commercialization and industrialization stage.

  Strengths

  Heavy AI infrastructure investment, growth in token consumption, expanding inference demand, and uncertainty in U.S. AI chip supply are increasing demand for domestic substitution.

  Weaknesses

  Insufficient advanced-node foundry capacity, slow domestic yield and capacity ramp, and a limited CSP customer base for second-tier suppliers.

  Comparison

  Tier-1 suppliers have advantages in performance, capacity, and CSP relationships; tier-2 suppliers have lower revenue bases but higher potential growth, with supply chain and customer breakthroughs being the key.

  Risks

  Changes in export controls, recovery in overseas chip supply, lower CSP capex, slow domestic capacity release, and weaker application demand.

Key data

• Domestic AI chip demand growth rateabout 40% CAGRBased on token-demand forecasts, referring to AI chip unit demand from 2026-2030E.
• Domestic AI chip supplier shipment forecast2026E/2027E/2028E: 2mn/3mn/5mn unitsIncludes GPU and ASIC, based on assumptions such as 7nm or N+2 nodes and single-die packaging.
• Corresponding China AI server salesRmb230bn/Rmb326bn/Rmb532bnFor 2026-2028E, assuming stable memory prices and a higher mix of supernode solutions.
• ByteDance China AI chip demand2026E more than 0.5mn unitsIn a scenario without available U.S. AI chips, ByteDance could become the largest AI chip buyer.
• Tier-1 supplier 2026E revenue scaleclose to Rmb10bnTier-1 suppliers include Huawei, Cambricon, Kunlunxin, and T-Head.
• MetaX inventoryabout Rmb1.5bnAs of end-2025, the report believes this can support near-term sales and improve visibility.
• Iluvatar CoreX sales growth forecast2026-2028E CAGR of 102%The report expects the company to turn profitable in 2027.
• MetaX sales growth forecast2026-2028E CAGR above 70%The report expects the company to reach breakeven by end-2026.
• Iluvatar CoreX rating and target priceOW, Jun-27 PT HK$620.0Based on 20x forward P/S, the target multiple is 30% below the domestic peer average.
• MetaX rating and target priceNeutral, Jun-27 PT Rmb550.0Based on 30x forward P/S, the report believes the current valuation of about 40x 2027E P/S is somewhat stretched.

Impact & implications

For investors, the main theme in the domestic AI chip sector is not simply demand growth, but the combination of supply realizability, CSP customer adoption, and inference-scenario fit. In the short term, companies with inventory buffers and stable foundry resources are more likely to convert revenue; in the medium term, domestic advanced-node capacity, N+1/N+2 migration, and yields will determine shipment pace; in the long term, higher domestic procurement by leading CSPs, the rollout of prefill-decode separation architectures, and the expansion of AI IaaS, PaaS, SaaS, and MaaS will determine whether domestic chip companies can sustain market share gains. The report points to industry divergence: tier-1 suppliers are closer to U.S. chip substitution, while selected tier-2 suppliers may offer greater revenue upside.

Risks

• Further tightening of U.S. restrictions on China's or Chinese entities' access to advanced semiconductor products and foundry services, leading to supply disruptions.
• Domestic advanced-node capacity ramping more slowly than expected or yields coming in below expectations, affecting chip performance, delivery, and revenue realization.
• CSP capex falling short of expectations, possibly due to weaker end-demand for AI or renewed availability of overseas AI chips.
• Second-tier suppliers failing to secure design-in from top CSP customers, or design-in scale and speed coming in below expectations.
• New chip design, tape-out, and process migration causing the R&D expense ratio to rise and weighing on profitability timing.
• Valuations having already priced in high growth expectations, so if sales momentum or market-share gains disappoint, share prices may come under pressure.

What to watch

• The procurement share and design-in progress of domestic AI chips at leading CSPs such as ByteDance, Alibaba, and Tencent.
• The pace of domestic advanced-node and N+1/N+2 capacity release, yield improvement, and mass-production timelines.
• The pace of inventory drawdown, restocking ability, and shipments of domestically produced chips after 2H26 for second-tier AI chip fabless players.
• Deployment progress of prefill-decode separation in inference infrastructure, and the actual efficiency performance of domestic chips at prefill or decode nodes.
• The mass-production progress of Iluvatar CoreX's next-generation GPU in 2H26, CSP customer breakthroughs, and profitability realization.
• MetaX C600 shipment ramp, the stability of its fully domestic supply chain, government-supported AIDC, and growth from state-owned enterprise customers.
• Changes in U.S. AI chip export rules, TPP and PD restrictions, and the availability of overseas chips such as H20 in the China market.
• China AI server sales, the mix of supernode solutions, memory prices, and changes in AIDC utilization.