Hygon Information / Iluvatar CoreX / Enflame: Integrated Research on Second-Tier Domestic Compute Vendors

1. Introduction

1.1. Positioning of the Integration Research

This research consolidates three second-tier domestic compute vendors into a single study, on the grounds that the three together constitute specimens of exactly three technology routes of domestic AI chips:

1. Hygon Information: the ecosystem-compatibility school — a GPGPU route where DCU is compatible with the ROCm/"CUDA-like" ecosystem, trading the lowest migration cost for market access;
2. Iluvatar CoreX: the fully self-developed training school — China's first 7nm GPGPU mass-production enterprise, with a fully self-developed instruction set and software stack;
3. Enflame: the DSA inference school — a dedicated-architecture route that was the first to achieve scaled commercial volume shipping with inference products.

In a market window where NVIDIA is constrained by export controls and Huawei Ascend is in short supply, the three share a common backdrop: they absorb the "second-tier domestic compute" demand; their common landing ground is government, finance, and telecom-carrier centralized procurement, plus smart computing center construction.

1.2. Basic Information Card for the Three

1.3. Industry Background: IDC 2025 Market Landscape

Cross-referenced between the IDC 2025 China Cloud AI Accelerator Market Report (as relayed by the Tencent Cloud developer community, grade B) and data cited in Enflame's prospectus (grade B):

Per CIC (Consulting) metrics (cited in Enflame's prospectus): China's AI accelerator card market grew from ¥122.54 billion in 2020 to ¥2164.77 billion in 2024, and is expected to reach ¥11076.46 billion by 2028 (2024–2028 CAGR 50.40%), by which time the China market is expected to account for about 30% of global demand.

1.4. Position in the AI Harness System

For all three vendors, the visible capabilities across Harness's six layers are concentrated at L1/L2 (migration toolchain and inference execution), with orchestration and governance capabilities weaker than Huawei's full stack. Their Harness value proposition is "entering the domestic compute pool at the lowest software migration cost" — Hygon pursues CUDA compatibility, Enflame pursues TopsRider all-in-one out-of-box deployment, and Iluvatar pursues DeepSpark community model verification.

2. Glossary

3. Feature Overview and Product Lines

3.1. Hygon Information: DCU and DTK

1. Product line: Shensuan-series DCU — Shensuan No.2 (single-card FP16 approx. 256 TFLOPS, INT8 1 PFLOPS, third-party metric, approaching 70–80% of the A100/H100), Shensuan No.3 (industry-press metric: performance reaching 90% of the A100, already entering the operator replacement sequence, alongside Ascend and Cambricon);
2. Software: the DTK software platform is compatible with the ROCm/"CUDA-like" ecosystem, with CUDA compatibility rated by third parties at approx. 90% (grades B/C) — this is Hygon's core commercial selling point: the lowest migration cost for existing CUDA code;
3. Thousand-card cluster: measured efficiency approx. 65% (third-party metric);
4. Synergy: "chip + server" synergy with Dawning Information (Hygon chips enter Dawning's server supply system).

3.2. Iluvatar CoreX: Tiangou / Zhikai / DeepSpark

The following product positioning follows Iluvatar CoreX's official website (grade A, reconfirmed this round at www.iluvatar.com):

1. Tiangou 100 (released 2021-03): a general GPU training product, China's first fully self-developed, GPU-architecture cloud training chip; the system architecture, instruction set, core operators, and software stack are all self-developed; 295 TOPS INT8 (industry-press metric); serving education, finance, autonomous driving, and healthcare industries;
2. Zhikai 100 (released 2022-12): a general GPU inference product based on the second-generation general GPU architecture; supports FP32/FP16/INT8 multi-precision mixed inference, multiple video decoders, and mainstream domestic and international deep-learning frameworks; applied to intelligent voice, finance, healthcare, education, and vehicle-road collaboration;
3. DeepSpark open-source community: positioned around "co-building the platform, sharing the ecosystem, and mutual industry benefit," building an AI and general-computing application development and evaluation platform (a hundred-application open platform), curating hundreds of open-source algorithm models tied to industry scenarios and constructing a multi-dimensional evaluation system;
4. Thereby, Iluvatar becomes the only domestic vendor that simultaneously offers a complete cloud-edge collaboration and train-infer combined solution under a GPU architecture (official website claim).

3.3. Enflame: Yunxun Series and TopsRider

1. Product evolution: released the first-generation cloud training chip Siusi 1.0 and the Yunxun T10 accelerator card in 2019; iterated 4 generations of architecture and 5 cloud AI chips over 8 years (prospectus/media claim); self-developed the GCU-CARE acceleration compute unit and GCU-LARE high-speed inter-chip interconnect;
2. Software: self-developed the "Yusuan TopsRider" full-stack AI computing and programming software platform, lowering the development difficulty and migration cost of mainstream AI models;
3. Product focus: training products explored gradually, inference products advanced rapidly (company's self-stated route) — during the reporting period, over 80% of AI accelerator card and module revenue came from inference products; the S60 inference card saw large-scale commercialization in search/recommendation/advertising/voice/image-classification scenarios;
4. Intelligent computing center footprint: Qingyang, Wuxi, Yichang, Chengdu, Tianjin, Guizhou; China's first ten-thousand-card inference cluster; cross-region compute scheduling (east-west collaboration, the "East Data West Computing" project);
5. All-in-one appliance: DeepSeek all-in-one series (early 2025), supporting domestic CPU platforms.

4. Platform Architecture

4.1. Comparison of Three Technology Routes

Figure 8-1 | Three technology routes of second-tier domestic compute: ecosystem compatibility / full self-development / DSA inference

Data sources: Hygon Information 2025 H1 report metrics (as relayed), Iluvatar CoreX official website, Enflame prospectus metrics (as relayed via Phoenix Tech/East Money); the diagram is based on this research's analysis.

4.2. Smart Computing Centers as a Shared Landing Ground

Over 50 smart computing centers nationwide (under construction/completed) with a total of 30 EFLOPS FP16 compute (Guanyan metric, grade C) form the shared demand pool for the three: Enflame has already positioned itself in Qingyang, Wuxi, Yichang, Chengdu, Tianjin, and Guizhou; Hygon has entered the operator replacement sequence; Iluvatar relies on DeepSpark for application-evaluation adaptation. The government background of these centers makes Xinchuang compliance (L6) a veto factor in procurement decisions, which explains why all three ground their commercial narratives in "self-control and trustworthiness."

5. Harness Design

5.1. Overview of Six-Layer Capabilities

Merged assessment of the three (per layer, taking each vendor's strongest form, with attribution noted):

5.2. L1/L2 Context and Tool Layers

L2 is the layer where the three's visible capabilities concentrate: Hygon DTK's CUDA migration path reduces development-and-migration cost through toolchain-level compatibility (third-party compatibility rating approx. 90%); Enflame positions TopsRider as a full-stack programming platform and offers an "integrated software-hardware, out-of-box deployment" form via the DeepSeek all-in-one — front-loading L2 adaptation to manufacturing. Neither has a public mechanism at L1.

5.3. L3 Orchestration and Control Layer

Enflame's cross-region compute scheduling platform (Qingyang and other east-west collaboration, corresponding to the "East Data West Computing" project) is its distinctive L3 form: the scheduling target is a cross-regional pool of smart computing center compute rather than a single cluster; Hygon/Iluvatar have no public orchestration mechanism.

5.4. L4/L5 State and Observation Layers

Iluvatar DeepSpark's community model verification system (500+ algorithm models, a multi-dimensional evaluation-system standard) carries an "ecosystem-level L5" function: community evaluation replacing single-vendor benchmarks — the most distinctive practice among the three in the evaluation-and-observation layer.

5.5. L6 Governance and Security Layer

Owing to its x86 licensing history (the AMD licensing route), Hygon holds a dual compliance position in the domestic-context sense of both "x86 software ecosystem compatibility" and "domestic identity"; Iluvatar/Enflame enter via full self-development / Xinchuang adaptation. The domestic-content procurement policies in government and finance scenarios are the shared underlying driver of the three's orders; references to specific policy document names were not verified against original text, so this article does not cite specific policy entries.

6. Actual Cases

6.1. Hygon Information: Revenue and Centralized Procurement

1. Financials: 2025 H1 revenue ¥54.64 billion (+45.21%), gross margin above 60% (grade B/C relay) — the largest in scale of commercialization among the three (its revenue includes the CPU business, of which DCU is only a part, so note the basis of accounting);
2. Centralized procurement: operator/finance/energy centralized procurement orders exceeding ¥11 billion (industry-press metric); Shensuan No.3 has entered the operator replacement sequence, alongside Ascend and Cambricon (industry-press metric).

6.2. Iluvatar CoreX: Industry Deployment Cases

1. China Merchants Bank intelligent risk control, SANY Heavy Industry quality inspection: industry-press metric (grades B/C) — Iluvatar enters via a "training + inference dual products + industry ISV" model;
2. Capability surface confirmed by the official website (grade A): Tiangou 100 widely supports traditional machine learning, mathematical computation, encryption/decryption, and digital signal processing; Zhikai 100 supports multi-precision mixed inference computing and mainstream deep-learning frameworks;
3. Explicit gap noted: official 2025–2026 revenue and shipment figures for Iluvatar are missing, and this article does not cite any third-party self-media shipment estimates.

6.3. Enflame: Prospectus Metrics and STAR Market Listing

Enflame is the most transparent of the three in data disclosure (prospectus metrics, grade B relay; cross-checking against the exchange's original text is recommended):

Case interpretation: Enflame is a specimen of the "DSA inference route + single large customer" model — its inference products were the first to cross the commercialization threshold (66,000 units), but customer concentration (over 80% to Tencent) and continued losses constitute its two core tests after listing.

7. Summary

Strengths (the three combined):

1. Complete route coverage: from CUDA compatibility (Hygon) to full self-development (Iluvatar) to DSA inference (Enflame), forming a route spectrum for second-tier domestic compute;
2. Clear divergence in commercialization validation: Hygon has the largest revenue, Enflame the fastest shipment growth, Iluvatar the most complete ecosystem;
3. Xinchuang compliance dividend: government/finance/operator centralized procurement and smart computing center construction provide a sustained demand pool.

Weaknesses:

1. Small market shares: Enflame 1.7% and Cambricon 2.9% versus Huawei 20.3% and NVIDIA 55%, showing a clear scale gap in the second tier;
2. Missing public benchmarks: none of the three has MLPerf-grade third-party results, and "percentage of the A100" style metrics all come from industry-press relays;
3. Shallow Harness capability: of the six layers, only L2/L6 have visible mechanisms, while L1/L4 are largely blank;
4. Governance-structure risks: Enflame's Tencent dependence (over 80% of revenue), Iluvatar's unlisted funding channel, and Hygon's mixed revenue basis including CPU.

Applicability boundary: Hygon suits existing-CUDA-code migration and operator replacement scenarios; Enflame suits inference-intensive internet scenarios (recommendation/search/advertising) and east-west smart computing collaboration projects; Iluvatar suits scenarios needing train-infer integration plus a Xinchuang self-developed instruction set.

Selection recommendation: any "X% versus an international peer of the same generation" statements from the three should be replaced by actual measured results; migration-type projects should prioritize Hygon (DTK compatibility), new inference clusters should prioritize Enflame (all-in-one + the ten-thousand-card inference cluster precedent), and choose Iluvatar when self-development control is the top priority; also watch for the possibility of Enflame's L600 (fourth generation) breaking through in large-model training scenarios.

Information Gap Statement

1. The original source text for Shensuan No.3's official specs and the "90% of the A100" figure is missing (industry-press relay only; this article cites it only as a weak-basis figure);
2. Official 2025–2026 revenue and shipment figures for Iluvatar CoreX are missing;
3. Enflame's STAR Market IPO-related financial data comes from media relays of the prospectus (Phoenix Tech, East Money, Sohu, etc.); review against the original prospectus disclosed by CNINFO/SSE is recommended;
4. The original text of policy documents on domestic-content procurement ratios was not verified, and this article does not cite specific policy entries;
5. The original annual reports/prospectuses of the three companies should be double-checked before formal citation.

8. References

1. Compute determines the ceiling, scenarios determine the floor: in-depth analysis of China's AI Agent industry chain (domestic chip vendors layout table) — Guanyan Report Net, 2025. https://m.chinabaogao.com/detail/814803.html
2. Rise of domestic AI chips and scenario-based selection (IDC data + each vendor's route) — Tencent Cloud Developer Community, 2026. https://cloud.tencent.cn/developer/article/2686867
3. Enflame's smart computing center layout achievements unveiled at WAIC — Tencent Net/Enflame press release, 2025-07. <https://gu.qq.com/resources/shy/news/detail-v2/index.html?t=1#/index?_tentrees_trans=0&id=SN20250728163736a6c597c0>
4. 2025 ranking of China's most promising compute chip industry enterprises — AskCI Net, 2025-04. https://m.askci.com/news/20250407/110624274399518242572737.shtml
5. Hygon Information official website — Hygon Information, 2025. https://www.hygon.cn
6. Enflame official website — Enflame, 2025. https://www.enflame-tech.com
7. Iluvatar CoreX official website (Tiangou 100 / Zhikai 100 / DeepSpark) — Iluvatar CoreX, 2025. https://www.iluvatar.com/
8. DeepSpark open-source community — initiated by Iluvatar CoreX. https://www.deepspark.org.cn/
9. "A big meat-sign again": earn 160,000 yuan per lot (Enflame prospectus financial and shipment metrics) — Phoenix Tech, 2026. https://tech.ifeng.com/c/8wKVQvS4a9M
10. Enflame listed on the STAR Market, opening up over 188% — East Money Net (Xinhua Finance), 2026-09-11. https://finance.eastmoney.com/a/202609113871802756.html
11. Enflame IPO: "Tencent dependence" and profitability become the test — Huacheng Client, 2026-09. https://huacheng.gz-cmc.com/pages/2026/09/04/717c3530d75148b488c98af9a51f9166.html