Inside China’s Domestic AI Chip Push: What Overseas Buyers Need to Verify Before Partnering

1. Why 2026 Is a Turning Point for Chinese AI Silicon

Walk the floor of the 2026 World Artificial Intelligence Conference (WAIC) in Shanghai this July, and you could feel the ground shift. For the first time since U.S. export controls on advanced GPU sales to China tightened in late 2022, a credible cohort of domestic chip designers was not just presenting roadmaps — they were showing running silicon, racked clusters, and paying customers. A few weeks later, on September 11, 2026, China’s State Council formally ratcheted up the policy signal at an executive meeting, naming the national integrated computing network as one of six strategic infrastructure grids and earmarking roughly RMB 4 trillion (≈USD 560 billion) of direct investment through the 15th Five-Year Plan.

The commercial logic is straightforward. TrendForce estimates that domestic solutions will capture close to 90% of China’s AI accelerator market in 2026, with NVIDIA’s China share collapsing from over 90% in 2021 to single digits. Domestic AI chip shipments are projected to approach 3 million units for the year, headlined by Huawei’s Ascend line (around 750,000 units shipped and targeted for 2026) and a “second tier” of Cambricon, Hygon, and a fast-closing “new forces” group — Moore Threads, Biren, MetaX (沐曦), Iluvatar (燧原), Tianshu Zhixin, and a 2024-founded dark horse named Oriental Compute Chip (东方算芯) — that together are expected to ship nearly 1 million cards.

For overseas enterprises, law firms, system integrators and funds that need to source, partner with, invest in, or compete against Chinese AI-chip vendors, this is simultaneously a remarkable opportunity and a remarkably easy market in which to get burned. The same structural conditions that produced the boom — state procurement preference, a flood of venture and policy capital, urgency from hyperscalers locked out of NVIDIA supply — have also generated an unusually wide gap between “product announced” and “product delivered”. This article gives you a practical framework to separate the two.

2. WAIC 2026 Scoreboard: Five Launches That Mattered

Before we talk about due diligence, it is worth grounding the discussion in what actually shipped or was credibly demonstrated at WAIC 2026 (July 17–20). Five launches captured industry attention, each representing a different answer to the question of how Chinese vendors are working around advanced-process and HBM constraints.

A few patterns jump out. First, the competitive unit is no longer the single GPU card but the super-node — a rack-or-larger scale-up domain that compensates for per-card performance gaps with aggressive interconnect engineering (optical, cable-less orthogonal backplanes, unified-memory protocols). Open-source Securities estimates the domestic super-node market will reach RMB 341.4 billion by 2028, a 194% CAGR from 2026. Second, software-defined and near-memory architectures (Oriental Compute Chip being the most aggressive example) are being used as architectural detours around HBM and EUV access. Third, the gap between “sample on stage” and “volume in customer racks” still spans quarters, and in some cases years.

3. The Launch-vs-Ship Gap: Why Demos Mislead Buyers

The single most expensive mistake an overseas buyer can make in this market is to equate a keynote announcement with production readiness. The history of Chinese AI silicon is littered with chips that posted breathtaking peak-FLOPS numbers at launch events but never reached the volumes, yields, or software stability required for commercial deployment.

There are structural reasons for this. A modern AI accelerator is not a single product but a tightly coupled stack: silicon die, interposer and advanced packaging, HBM or domestic equivalent memory, PCB and board-level design, driver and compiler software, framework adaptation (PyTorch, MindSpore, vLLM, DeepSeek-MoE), cluster-management middleware, and after-sales support. At 64- or 1,024-card scale, system-level linearity — the fraction of theoretical FLOPS you actually get when running a real model — becomes the dominant metric, and it is rarely demonstrable at a booth. MetaX CTO Peng Li told reporters at WAIC that “a 100-card cluster often delivers only ~90% of theoretical compute; getting to 95%+ at 10,000-card scale is the real moat.” Huawei executives have privately warned that at 520,000-card SuperCluster scale, the total cable run exceeds 1,600 km and the system integrates roughly 3 billion discrete components.

The rest of this article walks through the four verification pillars any serious buyer — whether procuring chips, investing in a vendor, or building a joint-venture AI infrastructure play in China — should walk through before signing.

4. Pillar 1 — Independent Third-Party Certification

China’s AI-chip market reached an important institutional milestone on May 27, 2026, when the China Information Technology Security Evaluation Center and the National Secrecy Science and Technology Evaluation Center jointly published the Safety and Reliability Evaluation Results Announcement (2026 No. 2). For the first time, the catalogue — historically limited to CPUs, servers and OS products — included “AI training and inference chips” as a standalone category, with nine chips from seven vendors receiving Class-I (highest) certification:

• HiSilicon (Huawei): Ascend 310, Ascend 910
• Alibaba T-Head: Zhenwu M530, Zhenwu M890
• Biren Technology: Bili 166 (壁砺166)
• Hygon Information: DCU-3G
• Tianshu Zhixin: KCC-V100X
• MetaX: MXC600
• Moore Threads: PH100

Why this matters to you. The Class-I certification is effectively the admissions ticket for government, state-owned-enterprise, financial, healthcare, energy and other Critical Information Infrastructure Operator (CIIO) procurement pipelines. If you are partnering with a Chinese chip vendor to sell into those sectors — or if you are a foreign fund underwriting a vendor that claims those pipelines as revenue — absence from this list is a material disclosure item. Conversely, presence on the list does not guarantee commercial performance; it is a necessary, not sufficient, condition.

Other third-party signals worth requesting directly from the vendor include: China Compulsory Certificate (CCC) for the server SKU, MLPerf or China AIPerf benchmark submissions with audited results, Open Compute Project (OCP) or ODCC acceptance, and — for export-facing deals — confirmation that the chip’s EDA/IP provenance does not trigger U.S. Entity List or Foreign Direct Product Rule complications for your own downstream customers.

5. Pillar 2 — Proven Volume Production & Delivery Records

A chip company can show a live demo at WAIC with fewer than 100 functional die. A revenue-generating product line requires tens of thousands — and the wafer, packaging, test and board-level supply chain to build them reliably. Here, the Chinese market is deeply bifurcated.

Huawei’s Ascend is in a category of its own. The Ascend 950PR entered volume production in Q1 2026; the company is targeting 750,000 Ascend 950-series shipments for 2026 and has adjusted card ASPs upward by 20–50% in response to demand, with the 950DT card reportedly priced above RMB 250,000. DeepSeek, the model developer, is publicly reported to be deploying at least 160,000 Ascend 950DT chips at a gigawatt-scale datacenter in Inner Mongolia. Sugon’s 100,000-card Dengfeng cluster, inaugurated July 10, processed 150,000+ jobs on its first day of operation — hard evidence that a real workload is running, not a benchmark.

For second-tier and new-entrant vendors, the evidence is thinner. Oriental Compute Chip — founded only in May 2024, now 500+ employees — told WAIC reporters that DF1000 “has entered mass production and delivery” and planned 100,000-unit shipments in H2 2026; those are ambitious numbers for a 26-month-old company running on a 14nm domestic process. Biren’s 1,024-card NPO super-node is a reference design, with 16-card and 128-card electrical variants not scheduled for commercialization until H1 2027. MetaX’s S600 has been launched but ships through OEM partners, so direct end-customer volume is harder to verify.

6. Pillar 3 — Supply-Chain Single-Point-of-Failure Risk

The entire point of China’s domestic AI-chip push is supply-chain resilience — but many “domestic” chips still rely on concentrated single points of failure that an overseas buyer must map.

Manufacturing. Advanced 7nm-class logic is effectively a single-foundry market in China (SMIC). Mature 14nm/28nm is more diversified but still concentrated. Memory — especially HBM-equivalent — is the tightest bottleneck: Huawei’s HiZQ 2.0 “Zhuque” HBM at 4 TB/s on the 950DT is a rare in-house solution; other vendors rely on domestic DRAM stacks (CXMT) or on architectural workarounds such as Oriental Compute Chip’s 3D hybrid-bonded DRAM-Logic wafer stacking that deliberately sidesteps HBM. Co-packaged and near-packaged optics (CPO/NPO) — the technology behind Biren’s and Lightelligence’s super-nodes — depend on a small group of domestic photonic suppliers (e.g., Sicoya, Eoptolink, Accelink) whose volumes are still ramping.

Customer concentration. Several of the “GPU Four Little Dragons” (壁仞/沐曦/燧原/摩尔线程) have historically relied on one or two anchor customers — often Tencent, Alibaba, or a state-backed智算中心 — for the majority of early revenue. Iluvatar CoreX (燧原), which debuted on the STAR Market on September 11, 2026 at an IPO price of RMB 142.18/share, lists Tencent as its largest shareholder and largest customer. That is a perfectly acceptable growth strategy, but an overseas partner or investor should know whether “strong revenue growth” is broad-based demand or a single-account concentration risk.

Geographic exposure. Vendors on the U.S. Entity List face restrictions on access to EUV, certain EDA tools, and — increasingly — TSMC and third-party packaging. Your own downstream use may be affected if your product embeds or re-exports such chips.

7. Pillar 4 — Software Stack & Ecosystem Maturity

The hardest problem in Chinese AI silicon is not the chip — it is the software. A GPU with impressive peak TFLOPS but a half-finished compiler will sit idle in customer racks.

Huawei is furthest ahead: the CANN (Compute Architecture for Neural Networks) stack now supports core CUDA APIs, and the April 24, 2026 DeepSeek-V4 release was fully adapted to Ascend SuperPoDs out of the box — a milestone that drew public comment from NVIDIA CEO Jensen Huang about the strategic risk of CUDA lock-in eroding. The 950-series additionally introduces SIMD/SIMT dual-mode programmability and native FP4 support, reducing — though not eliminating — porting cost for models originally written for NVIDIA hardware.

MetaX’s MXMACA stack claims compatibility with mainstream AI frameworks and ships with 80+ pre-adapted scientific-computing packages (VASP, LAMMPS, OpenFOAM, etc.), important for AI-for-Science workloads. Hygon’s DCU uses a CUDA-compatible path (a legacy of its AMD Zen/DTJV licensing history) and has been deployed in the Sugon 100,000-card cluster. Most other vendors offer some CUDA-porting layer, but real-world porting effort for frontier MoE, long-context and multimodal models varies from days to months.

8. How to Pull the Official Records Before You Sign

None of the four pillars above can be validated from a press release. What you need is primary-source documentation from Chinese government registries, IP offices and procurement platforms — most of which are either Chinese-language-only, require real-name verification with a Chinese national ID, or are structurally difficult for a non-Chinese-speaking compliance team to navigate remotely.

This is where a systematic Chinese corporate-verification workflow pays off. Before negotiating volume terms, you should be able to answer the following questions from primary records, not from the vendor’s data room:

1. Corporate registration & good standing. Does the company’s registered business scope (经营范围) actually cover IC design and semiconductor sales? Has it been flagged in the National Enterprise Credit Information Publicity System (国家企业信用信息公示系统) for administrative penalties, operating anomalies, equity freezes, or tax arrears? Has it had recent shareholder or legal-representative changes that suggest instability?
2. Intellectual property footprint. Do the chip’s key patents actually belong to the company, to a VIE entity offshore, or to a founder’s previous employer — a recurring source of IP-litigation risk in Chinese semiconductors? Are the core trademarks registered in Class 9 (semiconductors) and Class 42 (SaaS/software), and do they match the product brands being marketed internationally? Cross-checking a vendor’s patent families against the claims in its marketing deck is one of the fastest ways to catch overstated capabilities. ChinaBizInsight’s Chinese intellectual property verification service pulls certified records directly from CNIPA, the National Copyright Administration, and the Trademark Office, in English.
3. Government procurement & bidding record. Which provincial or central-government智算中心 projects has the vendor actually won, versus merely announced “strategic cooperation” MOUs that never convert to revenue? These records are public on platforms such as ccgp.gov.cn and provincial public-resource trading centers but require Chinese-language navigation and careful entity-name disambiguation.
4. Executive backgrounds and risk. Have the founders, CTO or key technical leads previously been associated with a failed or sanctioned chip venture? Do they hold concurrent positions at competing entities? The 董监高 (director/supervisor/executive) layer is the most predictive place to look for undisclosed related-party transactions.
5. Financials and tax behavior. For un-vendors that are not yet public (Oriental Compute Chip, for example, is still private), you will not find audited financials on a stock exchange — but tax-invoice flows, high-tech enterprise certification status, and social-insurance headcount trends (all retrievable from the official business credit file) will tell you whether a 500-person headcount claim matches what the company is actually paying into the system.

For overseas buyers, the practical shortcut is to commission an official Professional Enterprise Credit Report that pulls the official AIC filing, cross-references IP, litigation, tax and bidding records, and delivers a single English-language dossier on the target vendor — typically within three to five business days.

9. Pre-Partnership Verification Checklist

Use this ten-point checklist before signing any supply, investment or distribution agreement with a Chinese AI chip vendor. Each item should be evidenced by an independently sourced document, not a vendor slide.

China’s domestic AI-chip market is the most dynamic semiconductor story of 2026, and for overseas partners that approach it with disciplined due diligence it is also one of the most commercially exciting. Treat every headline number as a question, not an answer — and ground that question in primary Chinese registry and IP records before you commit capital. The chip that wins in China over the next 24 months will be the one that ships in volume, runs real customer workloads, and shows up clean in the official filings. Everything else is just a demo.

10. References

1. State Council executive meeting (国务院常务会议) deployment on computing-network construction, September 11, 2026. Gov.cn.
2. China Information Technology Security Evaluation Center & National Secrecy Science and Technology Evaluation Center. Safety and Reliability Evaluation Results Announcement (2026 No. 2), May 27, 2026.
3. “Biren Unveils NPO Optical Super-Node with up to 1,024 GPUs at WAIC 2026.” NetEase Technology / TMTPost, July 17–21, 2026.
4. “MetaX Launches Xijing S600 Super-Node at WAIC 2026.” 10JQKA / Global Times / International Finance Report coverage, July 17–25, 2026.
5. “Oriental Compute Chip Unveils DF1000 3D Near-Memory AI Chip at WAIC 2026; Awarded SAIL Prize.” IT之家 / Jiemian / Red Star Capital, July 17–23, 2026.
6. “Huawei Ascend 950DT Chip: Specifications, Timeline & DeepSeek Deployment.” Baidu Baike / Bloomberg, September 4–10, 2026; Huawei Connect 2025 roadmap disclosures.
7. “Sugon 8000 ‘Dengfeng’ — China’s First All-Domestic 100,000-Card AI Super-Cluster Inaugurated.” Guangming Daily Tech / Guizhou Media / China Science Daily, July 10 – August 31, 2026.
8. Open-source Securities, Guosheng Securities & Guotai Haitong research notes on domestic super-node market sizing, July 2026; TrendForce 2026 China AI accelerator market share forecast; Baidu Baike “Domestic AI Chips” entry, updated September 11, 2026.