The Chinese Market Panorama Under Gartner's $2.6T AI Spending Framework

Gartner's May 2026 edition provides a figure: total global AI spending of approximately $2.59 trillion. Gartner divides this spending into eight layers: AI Infrastructure ($1,366B), AI Services ($589B), AI Software ($452B), AI Cybersecurity ($51B), AI Platforms ($31B), AI Models ($26B), AI Application Development ($8B), and AI Data ($3B). The eight layers sum to approximately $2,526B; the gap with the $2.59T total stems from Gartner's cross-category reconciliation. Of the eight layers, infrastructure alone accounts for 54%, exceeding the combined total of the remaining seven.

China accounts for approximately 15-20% of global AI spending—at face value, merely a share question. But decomposed into the eight-layer framework, China's spending distribution diverges sharply from the global pattern: infrastructure exceeds 70%, versus a global average of 53%. The services and software layers are correspondingly thinner. Where the money flows determines which segments of the supply chain get fed—and which remain chronically starved.

This article focuses on the four largest and fastest-shifting dimensions across the eight layers: compute, storage, networking, and platforms (encompassing cloud services and the model layer). For each, we dissect the Chinese market's player landscape, share dynamics, product roadmaps, and core divergences from the international market. The judgment offered for each layer: where does China stand, where is the chokepoint, and what will happen in 2026-2027.

---

I. Compute

1.1 AI Chips: The Domestic Substitution Window from 95% to 50%

China shipped approximately 4 million AI accelerator cards in 2025. NVIDIA's share collapsed from approximately 95% in 2023 (industry consensus estimate) to approximately 55% in 2025; Bernstein forecasts further contraction to 8-12% in 2026. The vacated share was absorbed by three players: Huawei ShengTeng (昇腾) pushed from 20% toward 50%, Cambricon (寒武纪, Hánwǔjì) rose from 3% to 9%, and Hygon (海光, Hǎiguāng) climbed from 2% to 8%.

The pace of share transfer is outstripping most expectations. NVIDIA's collapse in China is not a technology problem—it is export control. After the H20 was restricted, the Blackwell and Rubin series are essentially unable to ship to China. The residual 8-12% share in 2026 comes primarily from legacy data center maintenance and gray channels.

Huawei is filling not NVIDIA's void, but the space created by policy. The ShengTeng 910B (FP16 256-376 TFLOPS) is roughly A100-class (FP16 312 TFLOPS, tensor core dense). vs H100 (FP16 tensor core dense 989 TFLOPS, FP8 tensor core dense 1,979 TFLOPS), the gap is 2.6-3.9x (FP16) to 5-8x (FP8). The bigger bottleneck is memory bandwidth: 910B HBM2e ~400GB/s vs A100 2,039GB/s (5x) vs H100 3.35TB/s (8.5x); late HBM3e versions narrow to 2-3x. The 910C uses a 2×910B die package (similar to B200 dual-die design), achieving FP16 ~800 TFLOPS total. Per-chip still trails, but Huawei's strategy is not to chase single-chip specs: the CloudMatrix 384 supernode connects 384 ShengTeng chips into a single training cluster, using system-level interconnect to compensate for per-card density., but Huawei's strategy is not to chase single-card specs: the CloudMatrix 384 supernode connects 384 ShengTeng chips into a single training cluster, using system-level interconnect to compensate for insufficient per-card density. This changes the metric of competition: from FLOPS to MFU (Model FLOPs Utilization), from single-card performance to cluster system efficiency.

Cambricon and Hygon represent an alternative path. Cambricon achieved its first full-year profitability in 2025, with revenue of ¥6.497B (up 453% YoY) and net profit of ¥2.059B; the Siyuan 590 has entered internet companies' procurement lists on the inference side. Hygon pursues CUDA syntax compatibility (via an HIP translation layer, requiring recompilation)—its DCU Deep Compute IV can run most CUDA code, offering the lowest migration cost. For customers who can neither buy NVIDIA nor want to be locked into Huawei's ecosystem, Hygon is the only compromise.

Gartner forecasts 80% domestic AI chip localization in China by 2030; the current figure is approximately 20%. 2026 is the inflection year: NVIDIA's effective embargo, ShengTeng 950DT's training-side breakthrough, and Cambricon and Hygon entering volume production cycles. The core divergence from the international market: NVIDIA's 80%+ global monopoly rests on CUDA's 17-year ecosystem moat; China relies on policy-mandated substitution plus supernode clustering to bypass the single-card gap. These two paths will not converge in the near term.

1.2 AI Servers: Inspur Leads, Air Cooling Still Dominant

The top three combined account for approximately 50%—lower concentration than the global market (Dell + HPE + Supermicro combined ≈55%). Inspur ranks fourth globally in x86 servers (5.9%), Lenovo fifth (5.2%), but these positions are built on general-purpose server volumes; AI servers are a different competitive axis.

Cooling form factors are migrating:

Liquid cooling penetration rose from 30% in 2025 to 43% in 2026. The incremental volume comes almost entirely from new high-density training clusters. Among liquid-cooling server vendors, Inspur exceeds 50%, XFusion approximately 10%, Nettrix approximately 8%, H3C approximately 7%, Lenovo approximately 5%.

Market growth remains rapid: ¥160.7B in 2024, approximately ¥250B in 2025, projected ¥350B in 2026 (Gartner / IDC estimates). But ASP is declining—from ¥140K to ¥95K, a 32% drop. Internet customer gross margin is only 2.8%; Inspur Information's overall gross margin in 2025 Q3 was 4.91%. Compared with Dell and HPE's 20-25%, Chinese server vendors have chosen scale over profit.

Air cooling will not disappear. The PCIe and OAM air-cooled installed base exceeds 70%, and these machines are still in their depreciation cycles. More importantly, domestic chip power consumption is approaching the limits of conventional air cooling. The ShengTeng 910C is approximately 400W TDP, Cambricon Siyuan approximately 300W—air cooling can still support them, but with little headroom. Liquid cooling growth is concentrated in NVIDIA H100/B200-class import clusters (legacy) and Huawei CloudMatrix-class domestic supernodes (new builds).

Inspur's "one machine, multiple chips" (一机多芯, yī jī duō xīn) platform is a product form unique to China: a single server compatible with NVIDIA, AMD, ShengTeng, and Cambricon accelerator cards. Internet customers can deploy training tasks across different chips within the same data center, reducing idle rates. This product has no equivalent in the global market, because the global market does not need to solve the problem of unstable chip supply.

---

II. Storage

2.1 Storage Chips: YMTC Is the Sole Bright Spot, HBM Is the Biggest Gap

The global semiconductor storage market was approximately $171.3B in 2025, projected at $189.7B in 2026 (Fortune BI semiconductor storage market report, covering DRAM/NAND/SRAM, etc.). Including enterprise storage systems and services, the broader storage market is larger. A single AI server consumes 8-10x the storage of a traditional server—AI doesn't only buy GPUs; it simultaneously devours DRAM, NAND, and HBM.

DRAM is an oligopoly:

The top three exceed 90% combined. CXMT's 3% means DRAM domesticization is just beginning, and is dominated by consumer-grade DDR4/LPDDR4; server-grade DDR5 volume production timeline remains undetermined.

NAND tells a completely different story:

Counterpoint predicts YMTC will surpass Kioxia and Micron to reach global #3. The Xtacking architecture is the key: by bonding CMOS wafer and NAND wafer at the wafer level, YMTC skipped certain intermediate generations in the layer-count race. Its 232-layer NAND is in volume production, with competitive density and cost. This is China's only genuine leapfrog in the storage chip domain.

HBM is the opposite extreme.

HBM is a complete triopoly, and export controls target it precisely. A single HBM3E die has 24GB capacity and 819GB/s bandwidth—an essential component of AI accelerator cards. Without HBM, there are no high-performance training chips. CXMT has only just begun exploring HBM2, trailing the international HBM3E/HBM4 by 2-3 generations. This means that no matter how advanced Huawei ShengTeng 950's architecture is, HBM supply remains a physical ceiling. Cambricon and Hygon face the same problem.

NAND has YMTC; DRAM has CXMT at the starting line; HBM is nearly zero. The storage chip gap is not evenly distributed—it is concentrated at the most critical link.

2.2 AI Storage: The Five Vendors' Intelligent Compute Data Foundation

IDC's 2025 annual data shows China's enterprise external storage market reached $7.44 billion, up 7.4% YoY, accounting for 22.5% of the global market. All-flash array grew 20.8% to 31.2% share. The global market is accelerating: IDC Q1 2026 global external storage grew 22.9%, with AI as a significant driver.

The AI era demands entirely new storage capabilities: high-throughput parallel file systems for checkpoint writes and data loading during training, and low-latency KV Cache persistence and Agent Memory access during inference. Each vendor's AI storage product is evolving from a "data container" into an "intelligent data platform that participates in computing."

Traditional enterprise centralized storage (SAN/NAS arrays) serves more as a "data foundation" than a "training front line" in AI scenarios: data preprocessing, model archiving, and version management reside on traditional storage, while training-side data loading and checkpoint writes are handled by the AI storage products listed above. Each vendor's traditional storage offerings (Huawei OceanStor Dorado V6, Inspur HF18000G7, H3C CF22000 G2) still see steady demand in enterprise and government markets, but the incremental AI growth is concentrated in the AI product lines on the left.

2.3 Distributed Storage: The Data Highway for AI Training

In 2025, China's distributed storage market reached ¥25.85B, up 30.4% YoY (Zhiyan Consulting). AI training, autonomous driving, and large-scale data labeling are the primary growth engines. Distributed file storage accounts for 44.9%, block storage 31.3%, and object storage 23.8%. IDC data shows China's enterprise SSD market reached $6.25 billion in 2024 (up 187.9% YoY), projected to reach $9.1 billion by 2029, with AI as the core driver.

Huawei OceanStor A310 (AI data lake) and A800 (AI cluster dedicated) are AI-optimized variants of distributed storage; Inspur's A9000 series (supporting GPU-Direct Storage and native KV Cache) targets AI inference scenarios. These products belong to the AI subcategory of distributed storage, distinct in positioning from traditional centralized SAN/NAS.

AI training imposes two core storage requirements: high-throughput sequential writes (checkpoints) and high-concurrency random reads (data loading). Traditional NFS/CIFS protocols become bottlenecks at thousand-GPU scale; parallel file systems are a necessity.

Global parallel file system landscape: Lustre (open source, HPC mainstay), IBM Storage Scale / GPFS (commercial, strong mixed workloads), WekaFS (cloud-native AI storage), BeeGFS (Germany, small-to-mid clusters), DAOS (Intel, high performance). China primarily pursues two paths for parallel file systems: in-house development (Huawei OceanStor Pacific, Inspur AS13000, Alibaba Cloud CPFS) and open-source secondary development (based on Lustre/Ceph).

Alibaba Cloud CPFS (Cloud Parallel File System) is China's most mature AI cloud-native parallel file system. It features end-to-end optimization for large model training: data loading throughput supporting Tbps-class, sub-millisecond latency, and deep integration with Lingjun (灵骏) intelligent compute clusters. In March 2026, Alibaba Cloud raised CPFS pricing by 30%—a representative case of the current cloud storage price hike cycle, reflecting supply-demand imbalance in AI storage.

Training checkpoints are the flashpoint where storage pressure concentrates. A single checkpoint of a trillion-parameter model can reach tens of TB, written once per hour during training. If storage throughput is insufficient, GPUs idle, wasting expensive compute. Inspur's DataTurbo acceleration engine claims to reduce preparation time by 40%; Huawei OceanStor A800 improves AI cluster bandwidth availability by 30%. Behind these numbers is the same underlying problem: storage cannot keep pace with GPU growth.

2.4 KV Cache Storage Tiering: The New Battlefield of the Inference Era

AI inference is shifting from "single-turn Q&A" to "multi-turn dialogue + agent pipelines." A single agent task may involve dozens of model inferences, each requiring the loading of prior context. KV Cache (key-value cache) stores the attention state during model inference, avoiding redundant computation—it is the critical data structure for inference performance.

The problem is that KV Cache growth far outstrips GPU memory capacity. Calculated at BF16 precision, a 175B-parameter model at 128K context has approximately 40GB of KV Cache per user; at 1M context, this reaches 320GB—no current GPU can hold even a single user's complete KV Cache in on-die memory.

At CES in January 2026, NVIDIA released the Inference Context Memory Storage Platform (ICMSP/ICMS), defining a five-tier KV Cache storage hierarchy:

G3.5 is the key innovation. It uses the NVIDIA BlueField-4 DPU to build a shared flash pool at the rack level, enabling multiple GPU nodes to share KV Cache as if accessing local memory. This is particularly critical in Prefill-Decode disaggregated architectures: after Prefill nodes (compute-intensive) compute KV tensors, they write KV blocks to G3.5 via the NIXL transport library; Decode nodes (memory-bandwidth-intensive) then read from G3.5. KV Cache is no longer bound to a single node—it becomes shareable "context infrastructure."

Quantified impact: in a 128K context, 90% prefix-reuse scenario, loading KV blocks from NVMe takes approximately 0.8 seconds, versus 11 seconds for recomputation—approximately 13x speedup. In production workloads with high prefix reuse, overall inference throughput can improve 5x. But for single-turn short-context inference with no prefix reuse, the NVMe tier adds overhead instead.

The Chinese market's KV Cache storage is still in its early stages. Inference frameworks such as vLLM and SGLang support CPU DRAM offloading (G2 tier), but multi-tier storage management at G3 and above has no mature Chinese solution yet. NVIDIA's ICMS / BlueField-4 ecosystem is restricted by export controls domestically, and the domestic substitution path remains unclear. Potential directions include: CXL 2.0-based memory pooling (CXL-attached memory as a G2.5 tier), domestic DPUs (such as Alibaba Cloud CIPU) serving the G3.5 role, and RDMA-based distributed KV Cache pools (similar to the Mooncake architecture). This is an emerging market; the first products are expected to land in 2026-2027.

NAND has YMTC; DRAM has CXMT at the starting line; HBM is nearly zero. The storage chip gap is not evenly distributed—it is concentrated at the most critical link. On the storage systems side, China has self-developed capabilities and market share in both distributed and centralized storage, but AI-native KV Cache storage tiering is nearly a blank slate—this is the next gap to close.

---

III. Networking

3.1 Data Center Switches: Huawei Leads, Ethernet Replaces InfiniBand

China's data center switch market reached ¥211.5B in 2024 and approximately ¥226.8B in 2025, with AI compute networking contributing over 45% of incremental growth (Zhiyan Consulting). The global Ethernet switch market reached approximately $55.1B in 2025 (+31.5%), with the datacenter segment at approximately $32.5B (+53.5%). In Q1 2026, the total market reached approximately $15.4B (+39.8%) and the datacenter segment approximately $10.0B (+61.0%), with AI as a significant driver (IDC Quarterly Ethernet Switch Tracker).

The global landscape is shifting dramatically. IDC Q1 2026 data shows NVIDIA, leveraging Spectrum-X, became the #1 vendor by revenue in global datacenter Ethernet switching for the first time ($2.1B, +192.7% YoY, 21.5% share), surpassing Cisco and Arista. Huawei's global switch revenue reached $895M (+27.2%), with 5.8% market share. However, in the Chinese market, domestic vendors (Huawei + H3C + Ruijie + ZTE) collectively hold over 80%, with a localization rate far exceeding the chip layer.

Port speed iteration is rapid: 400G is the 2025 mainstream (38% share), 800G is in volume production, and 1.6T is expected to commercialize in 2026. Gartner forecasts 65% of AI clusters will be Ethernet-based rather than InfiniBand by 2029. China is further along the Ethernet path—Huawei's Xinghe AI Fabric 2.0 has fully shifted to open Ethernet architecture, abandoning InfiniBand.

But switch chips are a different story. Import dependency for 400G+ high-end switch chips exceeds 90%.

China can build switch chassis; it cannot build the most critical chip inside. This is the upstream chokepoint of the networking layer, and one of the slowest areas for domestic substitution progress.

3.2 Optical Modules: China's Only Segment with Absolute Global Leadership

China is the absolute center of the global optical module industry. Zhongji Innolight (中际旭创) reported 2025 revenue of ¥38.24B (+60%), net profit of ¥10.8B (+109%), 21.09M units shipped, and 42.6% gross margin. Eoptolink (新易盛) reported 2025 revenue of ¥24.84B (+187%), net profit of ¥9.53B (+236%), 16.03M units shipped, and 47.8% gross margin. The two companies derive 86% and 94% of their revenue from overseas, respectively—they earn primarily in dollars.

The global optical module market reached $23B in 2025 (up 50% YoY), with AI cluster optics approximately $5B; 2026 projected at $26B (up 60%), with AI cluster optics exceeding $10B (LightCounting). Breakdown by speed:

800G+ share of total shipments (by unit count) jumped from 19.5% in 2024 to 60%+ in 2026 (TrendForce). Silicon photonics (SiPh) penetration is 50-60% at 800G and 60-70% at 1.6T. CPO (co-packaged optics) penetration is approximately 2-3% in 2026, expected to reach 20-32% by 2030 (LightCounting / Nomura). Innolight's NPO path targets 2027 mass production; Eoptolink has released 1.6T DR4, 6.4T NPO, and 12.8T XPO.

Optical modules are the only segment in the entire AI supply chain where China holds absolute global leadership. But the upstream core optical chips (EML lasers) are dominated by Coherent, Mitsubishi, and Sumitomo Electric, with a supply gap of 5-15%. Downstream assembly and packaging are where Chinese companies lead globally; upstream core components remain import-dependent.

A leading optical module manufacturer summarizes it bluntly: "The domestic market is extremely cutthroat on pricing—customers' target prices are basically at cost. Overseas has demand volume, and pricing is relatively acceptable." This explains why Zhongji Innolight and Eoptolink have such high overseas revenue ratios: it's not that they don't want to serve the domestic market, but the domestic pricing model cannot support R&D investment.

3.3 Copper Cables and Interconnects: The Physical Layer of Scale-up

AI cluster interconnect can be divided into three layers: Scale-up (intra-rack GPU interconnect), Scale-out (inter-rack network interconnect), and Scale-across (cross-datacenter interconnect). Scale-up is dominated by copper cables, Scale-out by optical modules, and Scale-across by DWDM coherent modules.

Copper cables fall into three categories based on whether they contain signal-processing chips:

The global AOC+DAC+ACC+AEC market was approximately $2.7B in 2025, projected to reach $10.7B by 2034 (16.5% CAGR, Market Growth Reports). LightCounting forecasts the combined market will exceed $2.5B by 2028, with AI-driven 800G/1.6T AEC growing the fastest.

Global key players (Note: Amphenol/Molex/TE shares are for the total connector market; Wolong/Luxshare/Zhaolong shares are for the high-speed DAC/AEC subsegment—different scopes, not directly summable):

Chinese vendors' position in the supply chain: Primarily in supporting and OEM roles. Wolong is China's largest high-speed copper cable manufacturer with approximately 25% global market share (#2 worldwide), mainly supplying Amphenol. Luxshare has self-developed 112G PAM4 DAC/ACC. Zhaolong is Credo's core copper partner with orders from Amazon and Microsoft. Bochuang Tech covers the full 25G-800G DAC/ACC/AEC portfolio. Xinya Electronics accounts for approximately 60% of Amphenol's procurement from China.

NVIDIA NVL72 case study: A single rack uses 5,184 high-speed differential copper cables totaling over 3,200 meters. The copper solution costs approximately $93,000 per rack vs $557,000 for the equivalent bandwidth optical solution—a 6x cost savings, while also saving approximately 20kW of power per rack. Amphenol custom-engineered the NVLink Spine Cartridge for this, the highest-value interconnect component per rack.

Notably, NVIDIA's Rubin NVL144 (Kyber) architecture transitions to a cableless design using a PCB midplane instead of cable backplanes. This poses a structural risk to Amphenol's NVLink copper revenue, but external 800G/1.6T copper and connector demand continues to grow—the net effect is a category shift rather than total contraction.

China's supernodes follow an independent path: Huawei CloudMatrix 384 uses OAM interconnect + copper backplane, and Alibaba Panjiu AL128 is also copper-based. The physical limitation of copper (effective transmission distance <10m) makes it irreplaceable inside the rack, and determines the necessity of optical modules for Scale-out scenarios.

---

IV. Platforms

4.1 AI Cloud / IaaS: Five Leaders with Vertical Integration

Pricing is turning. In 2024, across-the-board price cuts—with reductions up to 97%. In 2026, across-the-board price increases: at the cloud services level, Tencent Cloud raised some products by 5%, Alibaba Cloud raised AI compute/storage products by 5-34% (T-Head Zhenwu 810E compute cards, CPFS storage, etc.). At the model API level, Tencent Hunyuan HY2.0 Instruct model input price rose from ¥0.0008/1K tokens to ¥0.004505 (up 463%), and output price from ¥0.002 to ¥0.01113. From price war to price hike, the cycle took less than two years. The reason is straightforward: they can't afford to keep losing money. Infrastructure investment is astronomical, model training costs have no economies of scale, and the share captured through price cutting faces churn risk when prices rise.

In the global market, AWS/Azure/GCP compete in layered separation, each operating in its own lane. China's five leaders universally pursue self-developed chips + self-developed models + cloud platforms—a trinity that is more deeply vertically integrated than their international counterparts. Alibaba has T-Head + Qwen, Baidu has Kunlun Xin + ERNIE, Huawei has ShengTeng + Pangu. The advantage is a domestic closed loop with high synergy efficiency. The disadvantage is ecosystem lock-in: migrating a training workload from Alibaba Cloud to Huawei Cloud means changing chips, frameworks, and models—migration costs are extreme. This vertical integration is also suppressing the development of cross-platform toolchains.

4.2 Large Model APIs: Open-Source Dominance, Price War Transitions to Price Hikes

Frost & Sullivan data: In 2025H1, China's enterprise large model daily token call volume reached 10.2 trillion.

Qwen's 17.7% was not won through price wars. Its 300+ open-source models span the full size range from 0.5B to 110B, with cumulative HuggingFace downloads ranking #1 globally. Open source is Alibaba's core strategy: models are free; monetization comes through the cloud platform and compute. DeepSeek took a different path: API pricing at 25% of the industry average, using extreme low pricing to drive call volume while prioritizing adaptation to domestic chips (ShengTeng, Hygon) to reduce inference costs. The 10.3% share is underpinned by a real economic calculation: if inference costs can be compressed to one-quarter of competitors', price warfare is rational.

Zhipu (智谱, Zhìpǔ) is the only company that has cracked the to-B paid model. 240K paid developers, ARR up 60x, listed on the HKEX main board on January 8, 2026 (02513.HK) at an offer price of HK$116.2, with an IPO market cap of approximately HK$51.1B. Total shares outstanding: 446 million. As of July 7, 2026, the stock price was HK$1,825, with a total market cap of approximately HK$814B. The GLM series leads in penetration of government/enterprise and financial scenarios; B-end customers' willingness to pay for model capability is higher than C-end.

In the international market, OpenAI/Anthropic/Google maintain a closed-source triopoly with high-price harvesting. China has taken the opposite path: open-source dominance, price wars for volume, and entry into a price-hike cycle in 2026. Chinese model companies are also doing something their international peers do not: deep binding to domestic chips. DeepSeek V4 is adapted for ShengTeng 950; Zhipu GLM-Image is trained end-to-end on ShengTeng and topped the HuggingFace leaderboard. Chip-model co-adaptation in China is not optional—it is a survival strategy.

Chinese users' AI payment willingness is approximately one-quarter to one-third that of US users (estimated). This explains why Doubao, despite 345M DAU, only began monetization in May 2026, and why model companies are far more fixated on the to-B market than to-C.

---

V. Structural Assessment

5.1 Three Advantage Segments

Optical modules (assembly + packaging). Zhongji Innolight is #1 globally, with a generational lead at 1.6T. Eoptolink's shipment volumes are exploding. T&S Communications is positioned at the core of the CPO supply chain. China's lead in optical module packaging was not policy-engineered—it is the accumulation of over a decade of contract manufacturing experience and engineer dividends.

Servers (system integration). Inspur ranks #4 globally; "one machine, multiple chips" is a product form unique to China. Inspur's liquid-cooling share exceeds 50%. System integration is a traditional strength of China's hardware industry; AI servers are an extension of this capability.

Open-source models. Qwen is #1 globally by downloads; DeepSeek is disrupting the global pricing system through API pricing. China cannot catch up to OpenAI/Anthropic on closed-source models, but its open-source strategy is effective: free models capture developer mindshare, while domestic chip adaptation lowers deployment costs.

5.2 Three Chokepoint Segments

HBM. Samsung, SK Hynix, and Micron maintain a complete triopoly; export controls enforce a precise blockade. The domestic gap is 2-3 generations; CXMT's HBM2 is still in exploration. HBM is the physical prerequisite for AI accelerator cards—without HBM, there are no high-performance training chips. This is the largest single ceiling in China's AI supply chain.

EML laser chips. Coherent, Mitsubishi, and Sumitomo Electric dominate, with a supply gap of 5-15%. China leads globally in downstream optical module assembly, yet is constrained upstream on core components. This contradiction has no near-term solution.

High-end switch chips. Broadcom leads with Tomahawk series at approximately 58% monopoly; Tomahawk 5 (51.2T, 5nm) widely deployed, Tomahawk 6 (102.4T, 3nm) entered volume shipment in March 2026. NVIDIA Spectrum-X at approximately 18% ($1.2B in order bookings for 2025). China's Centec at only approximately 5% concentrated in mid-to-low-end (2.4T-12.8T). Import dependency for 400G+ switch chips exceeds 90%. Switch chassis can be built; the core chip inside cannot—mirroring the GPU situation.

5.3 Core Divergence: Technology-Moat Monopoly vs. Policy- and Ecosystem-Driven Substitution

The two paths may see a substantive shift in competitive dynamics in 2027-2028. "Shift" here does not mean technological parity or market share crossover, but rather large-scale domestic chip substitution on the inference side, continued model-layer price advantage output, and consequent reshaping of the global AI supply chain's pricing system. This is not a deterministic forecast but a hypothesis worth tracking. Three key variables: whether domestic HBM progress breaks through HBM2, whether ShengTeng 950DT training-side MFU reaches 50%+, and whether export controls tighten further.

Gartner's eight-layer tower, in China, is not eight layers—it is three: one layer building roads (infrastructure), one layer building vehicles (models + platforms), and one layer not yet under construction (applications + services). Roads are being built fastest and most aggressively; vehicles are transitioning from assembly to self-development; the services layer remains distant. But precisely because roads are being built so fast, those building vehicles need not worry about having nowhere to drive: infrastructure oversupply will continue to depress model training and inference costs until, one day, the application layer catches up and absorbs the compute. When that day arrives is a harder question than chip market share.

---

Disclaimer: This article is based on Gartner's May 2026 AI spending forecast ($2.59 trillion, eight-layer framework), Bernstein's China AI chip research (2026 forecast), Counterpoint's NAND market tracking (Q1 2026), Fortune BI semiconductor storage market data, IDC China Enterprise Storage Systems Tracker (2025 annual), IDC Global Enterprise Storage Systems Tracker (Q1 2026), IDC Quarterly Ethernet Switch Tracker (Q1 2026), LightCounting optical module and AOC/DAC market data, Goldman Sachs 800G/1.6T shipment forecasts, Frost & Sullivan's China enterprise large model token share report, and public financial disclosures and information from the respective companies. It does not constitute investment advice. Data herein is as of July 7, 2026.

<!-- china-ai-market-2026-20260709-en -->