Hygon Information Technology: The Trojan Horse of Chinese Silicon

I. Introduction & Episode Roadmap

In April 2016, AMD faced severe financial strain. Its stock traded in the low single digits, and the company had sold and leased back its corporate headquarters. Ryzen had not yet left the laboratory in Austin, and the Zen microarchitecture — the design that would eventually restore the firm's competitiveness — was still eighteen months from commercial release. Into that environment arrived a consortium of Chinese investors offering $293 million in cash plus ongoing royalties in exchange for a license to AMD's x86 and system-on-chip intellectual property, restricted to server processors sold inside China.¹

AMD accepted the terms. In doing so, it granted a Chinese joint venture access to high-performance x86 server processor technology, a commercial capability held by only two other companies globally.

Ten years later, the entity resulting from that transaction — 海光信息技术股份有限公司 Hygon Information Technology — has become one of mainland China's highest-valued semiconductor companies. Its shares trade on the 上海证券交易所 Shanghai Stock Exchange STAR Market under ticker 688041.SS. In May 2026, amid a broader rally in AI hardware equities, Hygon's market capitalization reached an intraday peak of approximately RMB 826 billion before adjusting to roughly RMB 275 per share by early August.² Annual revenue for 2025 reached RMB 14.38 billion, representing a 56.9% year-over-year increase, while net profit attributable to shareholders reached RMB 2.54 billion.³ In the first half of 2026, revenue rose an additional 66.5% year over year to RMB 9.10 billion.⁴

While these audited figures reflect rapid top-line growth, a closer review reveals operational and strategic nuances behind the headline performance.

The core narrative arc. Hygon's trajectory involves four key developments:

First, the licensing deal: in 2016, a consortium tied to the 中国科学院 Chinese Academy of Sciences industrial ecosystem acquired access to AMD's first-generation Zen microarchitecture and x86 instruction set, roughly three years before tighter export restrictions took effect.

Second, the trade constraints: in June 2019, the U.S. Bureau of Industry and Security placed Hygon, its two joint-venture entities, and its largest shareholder on the Entity List.⁵ AMD halted further technology transfers, leaving Hygon to develop and iterate independently on its existing single-generation architecture.

Third, the product lineup: Hygon operates a dual-chip model combining general-purpose x86 server CPUs with GPGPU AI accelerators, which the company designates as Deep Computing Units, or DCUs.³

Fourth, the terminated corporate integration: in May 2025, Hygon announced a plan to absorb its largest shareholder, server manufacturer 中科曙光 Dawning Information Industry (Sugon), through an all-stock reverse merger.⁶ On December 9, 2025, after six months of preparation, both boards called off the transaction, citing significant shifts in the market environment and immature conditions for major asset restructuring.⁷³

Examining market assumptions against disclosure filings. Several common assertions regarding Hygon warrant examination against official filings:

On x86 licensing longevity. Domestic financial media often characterize Hygon as possessing a permanent x86 instruction-set authorization.² However, AMD Chief Executive Lisa Su stated publicly in June 2019 that THATC was a single-generation technology license with no additional licenses granted.⁸ The distinction between an ongoing right to use existing architecture and access to future IP defines the company's long-term technical runway.

On DCU product disclosures. Hygon's FY2025 annual report does not break out DCU revenue as an independent line item. Instead, it discloses a consolidated "high-end processors" category totaling RMB 14.36 billion.³ Consequently, external estimates of AI accelerator revenue rely on figures outside the primary financial statements.

On customer concentration and related-party sales. In 2025, sales to a single related party reached RMB 8.14 billion, representing 56.68% of total revenue — up from RMB 3.68 billion in 2024.³ Overall, Hygon's top five customers generated 90.28% of total sales. This concentration was highlighted during the company's 2022 IPO process, when the Shanghai Stock Exchange focused extensive inquiries on related-party dependencies.⁹

These operational characteristics frame Hygon's business model, which remains closely tied to public policy, concentrated customer channels, and external foundry supply chains. Examining the verified record provides a clearer view of the company's trajectory as it confronts core architectural and manufacturing constraints.

↑ Back to Top

II. The x86 Imperative & China's Enterprise Compute Conundrum

Picture the core banking system of a large Chinese state-owned commercial bank. It is not one program. It is a sedimentary rock formation of software: a mainframe-era transaction engine, a commercial relational database, three decades of stored procedures written by retired engineers, middleware that nobody fully understands, and a compliance layer bolted on after each successive regulatory cycle. Nobody has a complete map. Nobody wants one.

Now imagine a government directive requiring that this system run on domestically designed silicon.

This is the problem that made Hygon possible, and it is worth dwelling on, because it is more of a linguistic problem than an engineering one. A processor's instruction set architecture is the vocabulary in which compiled software speaks to hardware. Software compiled for x86 — the vocabulary Intel invented and AMD extended — cannot run on an ARM processor any more than a French novel can be read aloud by someone who only speaks Japanese. You can translate. Translation costs money, introduces errors, and, in the case of binary translation layers, costs performance.

For consumer software, this is a solvable problem: recompile from source and ship a new build. For the enterprise stack, it frequently is not. The source code may be owned by a foreign vendor who will not certify a new architecture. The application may be certified by a regulator against a specific binary. The team that wrote it may no longer exist. Industry estimates for the share of enterprise server software assuming x86 run north of 90%, and while the exact figure is not independently audited, the direction is clear.

This "software friction tax" is the single most important structural fact in Hygon's favor.

The alternatives, and what each of them costs. China's domestic processor landscape offers four other roads, and every one of them charges a toll at the software layer. 华为鲲鹏 Huawei Kunpeng and 飞腾 Phytium both build on ARM, which means recompilation and revalidation for anything that is not already ARM-native. 龙芯中科 Loongson went further still, developing LoongArch, an independent instruction set — offering maximum sovereignty alongside maximum ecosystem isolation. The Sunway line, derived from Alpha, powers supercomputers but effectively nothing in commercial enterprise IT.

Hygon's pitch to a bank CIO is therefore not "our chip is faster." It is "your software will boot." That is a far more powerful sales argument in a procurement cycle driven by a compliance deadline, because it converts a multi-year migration project into a hardware refresh.

The policy tailwind, and its double edge. The mechanism turning this technical advantage into orders is 信息技术应用创新（信创）Xinchuang — the IT application innovation program, which mandates substitution of foreign technology across critical national infrastructure: government, finance, telecommunications, energy, the state grid, and transport. Hygon's FY2025 report describes its CPUs as already deployed "at scale" across telecoms, finance, internet, education, and transport.³

Investors should be precise about what a policy tailwind is and is not. It is a demand guarantee; it is not a pricing guarantee or a market-share guarantee. Xinchuang procurement runs through tender committees at state-owned enterprises — buyers with concentrated purchasing power, explicit multi-vendor mandates, and a structural interest in ensuring no single domestic supplier becomes indispensable enough to raise prices. The same policy that creates Hygon's addressable market also creates Huawei's, Phytium's, Loongson's, and Cambricon's. A protected market with five protected suppliers is not a monopoly; it is an allocation exercise.

Hygon's own risk disclosures acknowledge this buyer-side concentration directly: because the server industry is dominated by a handful of players, "customer concentration is high," and if major customers run into difficulty and the company fails to broaden its base in time, it faces material operating risk.³ That is management describing the bargaining-power problem in its own words.

The CAS lineage. The strategic insight that led to the AMD deal came from inside China's state research apparatus. Sugon — 曙光信息产业股份有限公司, the commercial arm that grew out of the 中国科学院计算技术研究所 Institute of Computing Technology at CAS — had spent two decades building supercomputers and enterprise servers, which meant it understood the software-migration problem from the customer side rather than the chip side. Its team had watched customers refuse to move.

The conclusion they reached was pragmatic to the point of heresy: the fastest route to sovereign compute was not to build a new instruction set, but to rent an old one. Buy the vocabulary, and the twenty-year software library comes free.

That decision was, depending on the perspective, either the most capital-efficient technology transfer in modern Chinese industrial history or a strategic dependency that has quietly persisted for a decade. The 2016 transaction is where both readings begin.

↑ Back to Top

III. The 2016 AMD Joint Venture Alchemy & IP Mechanics

A specific corporate-structuring challenge sat at the center of the 2016 transaction, and its mechanics explain both Hygon's corporate architecture and its ongoing financial obligations.

AMD does not own the x86 instruction set outright. It operates under a cross-licensing agreement with Intel that allows each company to utilize the other's patented technologies. That agreement contains a key restriction: AMD cannot transfer x86 licensing rights to an independent third party. It can, however, extend rights to an entity over which it retains corporate control.

The dual-JV architecture. The Chinese investment consortium was established as 天津海光先进技术投资有限公司 Tianjin Haiguang Advanced Technology Investment (THATIC). Rather than forming a single joint venture, the parties created two distinct entities with inverted ownership structures:

成都海光微电子技术有限公司 Chengdu Haiguang Microelectronics (HMC) — Majority-owned by AMD, holding the licensed intellectual property and managing front-end wafer fabrication.

成都海光集成电路设计有限公司 Chengdu Haiguang Integrated Circuit Design (HICD) — Majority-owned by the Chinese consortium, managing customer relationships and selling finished processors in domestic markets.

By placing the licensed technology within an AMD-controlled vehicle, the structure satisfied the technical requirements of the Intel cross-license agreement. Simultaneously, commercial revenues flowed primarily through the Chinese-controlled design entity. Three years later, U.S. regulatory authorities determined that this arrangement bypassed the intent of export control rules.

A structural detail frequently overlooked in external summaries, drawn directly from Hygon's FY2025 consolidated financial statements, is that this split-ownership arrangement remains active. Hygon holds 49% of HMC and 70% of HICD. In 2025, non-controlling minority interests in these two subsidiaries were credited with roughly RMB 1.07 billion in net profit, compared with RMB 2.54 billion attributable to Hygon's controlling equity shareholders. In effect, nearly 30% of consolidated net earnings accrued to minority equity partners rather than public investors.³ By year-end, minority equity inside the design subsidiary reached RMB 3.26 billion.

Furthermore, in its related-party notes, Hygon disclosed RMB 585.5 million in intellectual-property licensing fees paid during 2025 to a counterparty identified only as "Company G," described as a minority shareholder in one of its subsidiaries — an increase from RMB 429.4 million in 2024.³ While Hygon does not formally name the entity, market participants infer who receives a growing nine-figure annual royalty from a product line derived from AMD's x86 core. The disclosure satisfies regulatory reporting requirements while remaining substantively opaque, sitting in contrast to management's narrative of full technological self-reliance.

Initial output: the Dhyana processors. First-generation silicon arrived in 2018 under the Dhyana name within the Hygon 1000 and 3000 processor lines. Technical evaluations revealed chips that were functionally identical to AMD's first-generation EPYC "Naples" server processors — to the extent that Linux kernel maintainers supported them with modest patch sets, with the primary visible differences limited to vendor ID strings and family identifiers.[^13][^14] The initial silicon was not an independent domestic redesign of Zen, but a licensed production build wearing a different brand.

For AMD, the arrangement provided critical cash and royalty streams from a market segment it could not directly serve at a moment of corporate vulnerability. For the Chinese consortium, it compressed roughly a decade of clean-room CPU development into two years.

Thesis claim, and the historical falsification pass. A central assertion surrounding the 2016 transaction — frequently repeated by corporate management, sell-side analysts, and financial media — is that the joint-venture architecture established a permanent, self-sustaining foundation for Chinese x86 processor development.

Evaluating this claim against official disclosures highlights key boundaries.

The original license was strictly limited to AMD's first-generation Zen microarchitecture. It did not include Zen 2. In June 2019, AMD Chief Executive Lisa Su clarified on the record that the agreement was "a single-generation technology license," with no further technology transfers planned.⁸ In the years since, AMD launched Zen 2 with modular chiplet architecture and a redesigned front end, followed by Zen 3, Zen 4, and Zen 5. These subsequent microarchitectures introduced broader instruction-set extensions, including AVX-512 variants and low-precision data types tailored for AI workloads. Hygon's licensed baseline predates all of these architectural iterations.

Consequently, the thesis holds true only in a narrower form: Hygon acquired permanent rights to a specific architectural vintage, rather than an ongoing development seat within the x86 ecosystem. It holds perpetual rights to a 2017 microarchitectural baseline in an instruction-set language that has continued to evolve. Whether that baseline remains sufficient depends on the target workload: for traditional enterprise database and transaction processing, legacy x86 compatibility meets operational requirements; for frontier AI model training, the architectural gap presents greater challenges.

The empirical test for long-term independence lies in whether Hygon can advance its roadmap through proprietary instruction-set extensions rather than relying on historical AMD compatibility. Management has framed its strategy around this transition, branding its microarchitecture "C86" and asserting an independent development path.³ Demonstrating that independence will ultimately depend on published benchmarks and enterprise adoption rates rather than branding terms.

What the public record confirms is that the window for external x86 technology transfers closed definitively in June 2019.

↑ Back to Top

IV. The 2019 Entity List Shockwave & The Forced R&D Pivot

On June 21, 2019, the U.S. Department of Commerce announced that five Chinese entities involved in high-performance computing would be added to the Entity List, with the final rule published in the Federal Register on June 24.⁵ The designated entities included Sugon, the Wuxi Jiangnan Institute of Computing Technology, Higon (Hygon), Chengdu Haiguang Integrated Circuit, and Chengdu Haiguang Microelectronics Technology.¹⁰ The stated regulatory rationale cited military end-use concerns surrounding high-performance computing.

The practical impact was immediate. Entity List placement prohibits the export of U.S.-origin technology without a license, with license applications subject to a presumption of denial. AMD immediately suspended engineering support, design file deliveries, and technical consultations. Every future iteration of the microarchitectural baseline that Hygon relied upon became unreachable overnight.

The Financial Times noted the severity for Sugon, which had constructed much of China's supercomputing infrastructure and was suddenly cut off from the components powering it.¹¹ The Wall Street Journal described the enforcement action as opening a second front in the technology trade conflict, coming weeks after restrictions on Huawei.¹² For Hygon, the operational challenge was even more acute than for Huawei: while Huawei had spent years developing internal chip-design capabilities at HiSilicon, Hygon's flagship server product remained a licensed derivative of an external architecture, supported by an engineering team that had not yet designed or shipped an independent high-performance core.

The native iteration crucible.

Evaluating Hygon's subsequent technical execution requires separating corporate messaging from empirical disclosures.

Following the severance of external support, Hygon's engineering team — composed largely of researchers from the CAS Institute of Computing Technology and former Sugon engineers — began modifying the frozen architectural baseline. The company's disclosed progression proceeded across multiple silicon generations: a second generation focused on memory controllers, cache hierarchy, and on-die interconnects; a third generation, which served as the commercial volume product during its IPO period, reworking security engines and instruction dispatch; and a fourth generation expanding core counts and multi-die packaging, alongside a fifth generation in development.

The financial commitment required to sustain this iterative cycle is detailed in Hygon's FY2025 annual report. Research and development expenditures reached RMB 4.57 billion in 2025, a 32.6% year-over-year increase that consumed nearly a third of total revenue (31.78%).³ Staffing allocation reflects the same focus: out of 3,333 total employees at year-end, 2,766 — or 82.99% — served in R&D roles, with average annual R&D compensation reaching approximately RMB 817,000.³

This sustained spending produced a growing intellectual property portfolio. As of December 31, 2025, Hygon held 1,101 granted invention patents and 358 integrated-circuit layout design registrations out of 3,326 cumulative IP applications.³ Disclosures also detail heavy capital commitments across active design pipelines: three in-progress programs — a new-generation general-purpose processor, a next-generation general processor, and a co-processor program — represented cumulative investments of roughly RMB 4.66 billion, RMB 773 million, and RMB 893 million, respectively.³

What "iterating on a frozen baseline" actually means.

To evaluate these technical efforts, it is necessary to separate core microarchitecture from peripheral processor subsystems.

A modern server processor consists of an execution core surrounded by extensive supporting infrastructure, including memory controllers, cache hierarchies, on-die interconnect fabrics, I/O interfaces, power management systems, security engines, and advanced multi-die packaging. The execution core represents the most complex design element, which Hygon acquired through the 2016 licensing agreement. The surrounding subsystems offered immediate scope for internal engineering.

Hygon directed its engineering resources toward these peripheral subsystems. Upgraded memory controllers increased effective memory bandwidth without altering the core architecture. Redesigned interconnects enabled higher core counts to share a single package coherently. Multi-die packaging increased total core count per socket without requiring larger monolithic dies — a key operational adaptation given domestic foundry yield constraints. Redesigned hardware security engines addressed specific compliance mandates required by domestic enterprise and government buyers.

While these architectural modifications improved system-level throughput and created competitive server products for domestic procurement, they did not alter single-thread core IPC performance. Hygon optimized the architectural elements within its engineering control to meet customer and regulatory requirements, while the underlying execution core remained tied to its 2017 baseline.

Thesis claim, and the historical falsification pass.

The central assertion examined here is that internal R&D has fully decoupled Hygon from foreign intellectual property and established architectural parity with global CPU leaders.

Financial disclosures from Hygon's official filings provide key evidence regarding this claim.

First, royalty obligations persist. A company that had fully decoupled from underlying foreign IP would not continue paying substantial licensing royalties to a minority shareholder in its operating subsidiary. In 2025, Hygon paid RMB 585.5 million in IP licensing fees to that entity — a 36% increase from 2024.³ This financial cash flow confirms that commercial scaling remains tied to the underlying licensing agreement.

Second, toolchain and supply-chain dependencies remain active. Hygon's official risk disclosures state that its supply chain relies on third-party wafer foundries, packaging and test providers, IP licensors, and EDA software vendors. Management explicitly notes that supplier concentration is high, that certain inputs are "scarce and proprietary," and that geopolitical actions or Entity List status could cause suppliers to suspend access, resulting in high switching costs.³ Microarchitecture iteration does not eliminate reliance on foreign-origin electronic design automation (EDA) software and licensed IP building blocks.

Third, the architectural scaling path reflects physical constraints. While third-party benchmark data remains limited, Hygon's disclosed engineering roadmap focuses on adding CPU cores, expanding die surface area, utilizing multi-die packaging, and increasing power envelopes. This design trajectory is characteristic of scaling system capacity around a core architecture that cannot be updated as rapidly as competing global designs. While multi-chip packaging is an established strategy to scale aggregate throughput, it increases die area, packaging complexity, and power consumption — all of which are constrained factors for a sanctioned fabless designer.

The empirical record indicates that the broad claim of global architectural parity is unverified. However, the evidence supports a narrower conclusion: Hygon has established a functional engineering organization capable of iterating around a baseline architecture to meet domestic enterprise computing requirements, while remaining dependent on foreign design tooling and paying ongoing IP royalties. Falsification of this capability would require evidence of tape-out failures or generational delays caused by supplier cutoffs; confirmation of full parity would require independently audited single-thread benchmark results demonstrating parity with contemporary global processors. Neither event has occurred.

The primary structural outcome of U.S. trade restrictions was the creation of a protected domestic market for Hygon's processors. That market dynamic leads directly into an examination of the product lineup itself.

↑ Back to Top

V. Product Anatomy, Segment Economics: CPU vs. DCU

In Hygon's FY2025 annual report, the segment disclosure contains a single line item for a business of its scale and complexity: "high-end processors," reporting revenue of RMB 14.36 billion, cost of sales of RMB 6.06 billion, and a gross margin of 57.78%.³ Geographic reporting is similarly consolidated into a single domestic category covering mainland China, Hong Kong, Macau, and Taiwan. Despite selling two distinct categories of silicon—general-purpose CPUs and AI-focused GPGPUs—into different end markets, Hygon reports their financial performance as a single unified figure. This consolidation presents a fundamental disclosure challenge for assessing segment-level profitability and valuation.

What the company does tell you.

Hygon categorizes its CPU line into three main series: the 7000 series for enterprise servers and advanced computing, the 5000 series for mainstream general-purpose servers, and the 3000 series for workstations, industrial controls, and edge computing. In marketing filings, the company highlights performance, system compatibility, and security as core selling points, though x86 ecosystem compatibility remains its primary commercial differentiator.³

Its GPGPU accelerator line, designated as the 8000 series Deep Computing Units (DCUs), features full-precision support targeted at scientific computing, AI training, inference, and large-scale data processing.³ Hygon does not disclose specific DCU product SKUs in its annual report, detailing only the broader product family and generational progress. Industry coverage of the FY2025 performance notes that the DCU roadmap has advanced to a fourth generation, while earlier iterations remain in commercial deployment.¹³

The software lever, explained plainly.

The primary commercial vector for Hygon's GPGPU accelerators against market incumbent Nvidia is software compatibility rather than raw silicon performance. Nvidia's competitive moat rests on CUDA—a mature library, tooling, and kernel ecosystem accumulated over nearly two decades. Alternative silicon providers must either build an independent software ecosystem from scratch or offer a translation layer compatible with existing CUDA codebases.

Hygon adopted the second strategy. Its proprietary software stack, DTK (Developer Toolkit), is structured around AMD's open-source ROCm/HIP framework, enabling CUDA source code to be ported through automated translation rather than manual rewriting. Hygon claims DTK represents one of mainland China's most complete GPGPU development environments, stating that it "substantially reduces application migration difficulty."³

To support this positioning, Hygon reported that by year-end 2025, DTK had achieved full adaptation and joint fine-tuning across 365 mainstream large language models, including DeepSeek, Qwen3, Hunyuan (混元), and Zhipu (智谱). Management asserts this coverage encompasses 99% of global open-source AI models.³ Surrounding this software layer is the Guanghe Organization (光合组织), an ecosystem initiative boasting over 6,000 partner entities and more than 15,000 completed adaptation projects across government, finance, and energy sectors.³

In September 2025, Hygon expanded its integration efforts by launching Hygon System Link (HSL), a proprietary bus interconnect protocol designed to link third-party accelerators, I/O hardware, operating systems, and OEM servers directly to a Hygon CPU base.³ Having shifted away from complete server manufacturing following its origin with Sugon, Hygon is positioning HSL as an interconnect architecture to unify domestic heterogeneous compute hardware around its processors.

Who Hygon actually competes with, and on what.

Competitive dynamics differ sharply across Hygon's CPU and GPGPU product lines.

In general-purpose server CPUs, competition splits between foreign incumbents and domestic peers. Against Intel Xeon and AMD EPYC, Hygon does not compete on raw performance per watt or architectural benchmarks; its market share gains are driven primarily by domestic Xinchuang procurement mandates. Conversely, against domestic CPU designers such as Huawei Kunpeng (华为鲲鹏), Phytium (飞腾), and Loongson (龙芯中科), Hygon's main advantage is x86 binary compatibility. While ARM and LoongArch alternatives require enterprise software recompilation and revalidation, Hygon enables legacy enterprise workloads to run without software modification—a decisive factor in strict compliance procurement.

In AI accelerators, the competitive landscape presents greater headwinds. Huawei Ascend (华为昇腾) serves as the domestic benchmark, offering integrated compute stacks that combine silicon, interconnects, liquid cooling, and software frameworks designed for frontier model training.¹⁴ Publicly traded Cambricon (中科寒武纪) and venture-backed Biren Technology compete in standalone AI accelerators, while Nvidia remains restricted by trade controls in high-performance segments.

Hygon's DCUs primarily target enterprise inference, academic research, and custom clusters where customers migrate existing CUDA codebases rather than building dedicated training environments. The company's principal advantage in this segment stems from its dual-processor portfolio: existing enterprise customers utilizing Hygon CPUs can integrate DCU accelerators within established vendor relationships, software stacks, and security accreditations.³

Now the financial trajectory, and what it actually says.

Hygon's top-line revenue expanded from RMB 6.01 billion in 2023 to RMB 9.16 billion in 2024, reaching RMB 14.38 billion in 2025.³ Net profit attributable to shareholders followed a matching trajectory, rising from RMB 1.26 billion in 2023 to RMB 1.93 billion in 2024, and RMB 2.54 billion in 2025.³ Annual operating cash flow more than doubled in 2025 to RMB 2.10 billion.³ Momentum continued into the first half of 2026, with six-month revenue generating RMB 9.10 billion—nearly matching full-year 2024 results—and net profit reaching RMB 1.80 billion.⁴

Behind these aggregate growth metrics, however, key operational margin and working capital shifts emerged.

Gross margin is falling, and management has explained why.

Hygon's consolidated gross margin contracted by 5.92 percentage points year-over-year in 2025 to 57.78%.³ Compression continued in the first half of 2026, dropping an additional 4.94 percentage points to 55.21%.¹⁵ This contraction reflected cost inflation outstripping top-line expansion: cost of sales grew 82.8% in 2025 against 57.2% revenue growth, while first-half 2026 operating costs increased 87.2%.³¹⁵

Management attributed the margin pressure to rising upstream material prices, tight semiconductor packaging and testing capacity, higher foundry fees, and elevated unit manufacturing costs.³ Intermediary financial disclosures for early 2026 also cited rising costs for memory and storage components.¹⁵

This margin trend highlights structural limits on Hygon's pricing power. Despite operating within a protected policy environment, the company was unable to pass input-cost increases on to its end customers, absorbing a 10.86 percentage point decline in gross margin over eighteen months. This dynamic indicates that supply-chain bargaining power remains concentrated upstream with foundries, packaging providers, and memory manufacturers rather than fabless chip designers.

Concurrently, Hygon's sales channel mix underwent a significant structural shift. Distribution revenue fell 55% in 2025 to RMB 627 million, accompanied by a 23.8 percentage point drop in distribution gross margins, while direct enterprise sales grew 76.8% to RMB 13.73 billion.³ Management attributed this transition to growing customer accounts migrating from distributor channels to direct sales contracts.³ Consequently, consolidated gross margins now reflect pricing terms negotiated with a concentrated group of major direct buyers.

The working capital picture turned in 2026, and it deserves attention.

Hygon's balance sheet structure shifted markedly between late 2025 and mid-2026. At year-end 2025, inventories reached RMB 6.41 billion (17.98% of total assets), accounts receivable totaled RMB 4.03 billion (11.32% of total assets), and contract liabilities—representing customer prepayments—more than doubled to RMB 2.02 billion.³¹³

By mid-2026, inventory levels rose further to RMB 7.52 billion, while prepayments to suppliers increased to RMB 4.77 billion, up from RMB 3.44 billion at the end of the first quarter. Conversely, contract liabilities fell 99% to RMB 19.4 million, and operating cash flow swung from positive RMB 2.18 billion in the prior-year period to negative RMB 428 million.¹⁵

This working capital shift reflects dual possibilities. It may indicate proactive capacity reservation, with Hygon securing foundry and packaging allocations ahead of anticipated supply constraints and price increases. Alternatively, it suggests profit absorption into inventory and supplier advances amid declining upfront customer prepayments. Subsequent financial reporting will clarify whether these inventory accumulations convert into fulfilled shipments or represent working capital drag.

Thesis claim, and the historical falsification pass.

The assertion examined here is that Hygon's DCU serves as a seamless drop-in replacement for Nvidia hardware across both large-scale AI training and inference.

Factual disclosures and competitive benchmarks highlight clear limitations to this claim.

First, financial filings omit standalone performance metrics for the accelerator business, bundling DCU sales into a single high-end processor reporting line without disclosing separate unit volumes, revenue breakdown, or gross margins.³

Second, competitive deployments in large-scale cluster training show a performance gap. At the World Artificial Intelligence Conference in July 2026, Huawei demonstrated the Atlas 950 SuperPoD, a 1,024-card supernode architecture scaling up to 8,192 Ascend 950DT accelerators, 1 EFLOPS of FP8 compute, 256TB of unified memory, and three-microsecond interconnect latency, targeted for commercial release in Q4 2026.¹⁴ Its predecessor, the Ascend 384 SuperPoD, was deployed across more than 750 installations in telecom, finance, internet, and research sectors, and was cited as the only domestic supernode system to complete training on a state-of-the-art frontier model.¹⁴

Large-scale model training relies heavily on multi-node cluster topology and high-speed interconnect bandwidth. While Hygon introduced its HSL bus interconnect protocol in September 2025, HSL remains a system specification rather than an integrated, commercially proven multi-thousand-card cluster architecture.³

Consequently, the drop-in replacement thesis holds for localized inference and small-cluster fine-tuning, where CUDA source translation through DTK provides functional software compatibility across adapted open-source models. However, the claim is unverified for frontier-scale AI training, where cluster-level interconnect performance dominates system capability and Huawei Ascend serves as the primary domestic reference architecture. Domestic cloud hyperscalers—including Alibaba (阿里巴巴), Tencent (腾讯), and ByteDance (字节跳动)—maintain multi-vendor accelerator deployment strategies to mitigate single-supplier dependency.

Verifying Hygon's competitive standing in AI hardware requires dedicated financial reporting of DCU revenue alongside sustained commercial adoption outside policy-mandated procurement channels.

Which raises the question of customer composition—and the concentrated sales channels detailed in company filings.

↑ Back to Top

VI. Governance, Management Credibility, & The Sugon Mega-Merger

On May 26, 2025, trading in shares of both Hygon and Sugon was suspended on the Shanghai Stock Exchange. The announcement that followed detailed one of the more unusual corporate restructurings in Chinese capital markets: Hygon, the fabless chip designer, proposed issuing new A-shares to absorb Sugon, the server manufacturer that was also its largest shareholder with a 27.96% equity stake.⁶³

In effect, the subsidiary proposed swallowing its parent entity. Trading remained halted for up to ten days while transaction terms were structured.⁶

The strategic logic was straightforward and promoted enthusiastically. Vertical integration would fuse semiconductor design directly with server chassis manufacturing, liquid cooling infrastructure, supercomputing cluster integration, and enterprise sales channels. The transaction aimed to create a full-stack domestic compute champion at a time when national policy actively encouraged integration. For public shareholders, it offered a potential structural solution to Hygon's long-standing customer concentration by absorbing its primary related party into a wholly owned operating division.

Hygon's board approved the transaction framework on June 6, 2025, and trading in the stock resumed on June 10.³ Monthly progress updates followed through the summer and autumn on July 9, August 7, September 6, October 1, October 30, and November 29.³

Then, on December 9, 2025, the board voted to terminate the merger.³

What the termination actually said. Official disclosures cited the transaction's complexity, the number of participating entities, and lengthy negotiations, stating that "the market environment has changed significantly since the transaction was initially planned," and that "the conditions for implementing this major asset restructuring are not yet mature."³ Both companies affirmed they would continue working together across the computing supply chain.⁷ Sugon general manager Li Jun (历军) — who also serves on Hygon's board — offered the post-hoc rationale that remaining independent allowed Hygon to serve as a neutral chip supplier to the broader domestic server market.⁷³

Market reaction was immediate. Sugon's stock opened at its limit-down floor the next trading day and closed at RMB 90.12 per share, while Hygon shares fell 0.36% to close at RMB 218.50. Financial media coverage attributed the deal's collapse primarily to Hygon's substantial share-price appreciation during the six-month negotiation window, which rendered the agreed valuation and share exchange ratios difficult for both boards to reconcile.

From a governance standpoint, the transaction highlighted communication and execution challenges. Management announced a transformational merger, suspended equity trading, conducted six months of advisory work, issued six progress reports, and ultimately cancelled the transaction after its own valuation increased. The official explanation — that market conditions were "not yet mature" — provided minimal detail regarding the specific points of disagreement. Furthermore, the post-termination assertion that operating independence was preferable directly contradicted the strategic rationale presented when the deal was announced in May. When management's justification for abandoning a transaction reverses its original rationale without detailed disclosures, it raises questions regarding capital allocation planning and disclosure discipline.

The related-party problem the merger was supposed to solve, and did not.

Customer concentration remains a central structural vulnerability in Hygon's business model. In 2025, sales to related parties within the corporate group designated in financial filings as "Company A" — entities controlled by or sharing board representation with the controller of a 5%-plus shareholder — doubled to RMB 8.14 billion, up from RMB 3.68 billion in 2024.³ Hygon's single largest customer, an affiliated party, generated 56.68% of total annual revenue, while the top five customers combined represented 90.28% of total sales.³

Management explained that special project procurement requirements drove this surge in volume from its top customer.³ While operational demand may explain the single-year expansion, the filings confirm that Hygon's record top-line growth in 2025 remained heavily reliant on a single affiliated buyer.

Regulatory scrutiny over this arrangement dates back to the company's public listing. During IPO inquiries in March 2022, the Shanghai Stock Exchange focused the third section of its second-round review on customer concentration and related-party sales. Regulators noted that related-party transactions accounted for 100.00%, 87.39%, 56.24%, and 66.88% of revenue across prior reporting periods, requiring Hygon to document pricing fairness, verify that credit terms aligned with non-affiliated customers, clarify intermediary relationships, and outline formal "measures to reduce related-party transactions" over time.⁹

Four years after that regulatory review, related-party sales stood at 56.68% of revenue and were expanding.³

The empirical record shows that related-party reliance did not normalize following listing. After initially declining post-IPO, the proportion of affiliated sales expanded sharply in 2025, the same year the corrective merger was proposed and abandoned. While the transaction was meant to eliminate related-party friction, its termination leaves the original commercial structure intact. In absolute terms, affiliated revenue in 2025 exceeded Hygon's entire 2024 revenue base. Although auditor reviews produced standard unqualified opinions confirming arm's-length accounting,³ the concentration level indicates that Hygon's top-line performance remains tied to procurement decisions by affiliated entities.

The control structure, which is genuinely unusual. Hygon operates without a controlling shareholder or ultimate actual controller. Company filings state that no single equity holder commands more than half of board seats or holds sufficient voting rights to dictate shareholder resolutions.³ Ownership is distributed among major shareholders, including Sugon with 27.96%, Tianjin Haifu Tianding (天津海富天鼎) with 10.81%, Chengdu Industrial Investment Group (成都产业投资集团) with 7.21%, and Chengdu Lanhai Qingzhou (成都蓝海轻舟) with 6.09%.³ Both Sugon and CAS Holdings (中国科学院控股) submitted formal undertakings at IPO agreeing not to seek corporate control.³

This diffuse ownership model prevents any single investor from extracting outsized private benefits at the expense of minority shareholders. However, it can also create decision-making friction during complex corporate restructurings, where no single lead shareholder holds authority to finalize terms. The terminated merger illustrates how a cap table structured to avoid concentrated control can complicate strategic execution.

Management, and how to read them. Executive leadership saw key adjustments in 2025. Meng Xiantang (孟宪棠) retired as chairman during the year as part of broader board changes.³ General Manager Sha Chaoqun (沙超群), a director and legal representative who signed the FY2025 financial disclosures, serves as operational lead.³ Sha's career mirrors the company's corporate lineage: he joined Sugon in 2002, directed its product center, served as vice president of technology from 2011 to 2018, and held senior vice president roles before transitioning to Hygon.¹⁶ Xu Wenchao (徐文超) serves as director, deputy general manager, chief financial officer, and board secretary.¹⁶

Executive incentives align with long-term performance through established equity programs. An employee incentive plan distributed stock awards across 878 staff members, with Sha holding approximately 0.53% of total equity.³¹⁶ The board declared an FY2025 dividend of RMB 1.50 per 10 shares, totaling RMB 556.7 million including interim payments, representing 21.88% of net profit attributable to shareholders.³ This payout ratio reflects a conservative capital distribution strategy appropriate for a fabless designer requiring substantial working capital prepayments for foundry allocations, though it requires shareholders to fund ongoing growth and absorb inventory risks.

The accounting judgment worth watching. Hygon capitalizes a portion of its research and development costs. In 2025, the company capitalized RMB 423 million out of RMB 4.57 billion in total R&D expenditures — a capitalization rate of 9.26%, representing a noticeable decrease from 17.04% in 2024.³ Management attributed this lower rate to expensing newly initiated processor design and forward-looking research projects, while previously capitalized co-processor development programs reached completion.³ This reduction reflects a more conservative accounting posture.

R&D capitalization practices previously drew regulatory scrutiny during Hygon's IPO process. In second-round inquiries, regulators questioned capitalization thresholds, technology readiness definitions, and whether amortizing self-developed intangible assets belonged in R&D expenses or cost of goods sold.⁹ Hygon adjusted its accounting policies by reclassifying amortization of commercialized self-developed intangibles into main operating costs, which lowered reported gross margins.⁹ In the FY2025 audit, BDO China Shu Lun Pan (立信会计师事务所) flagged R&D capitalization alongside revenue recognition as key audit matters, issuing an unqualified opinion.³ The balance of self-developed intangible assets remains significant enough that management explicitly identifies impairment as an operating risk if underlying technologies become obsolete.³

These accounting choices do not indicate reporting non-compliance. They demonstrate that key profitability indicators — including gross margins and reported R&D intensity — are influenced by capitalization rules previously reviewed by regulators, making the R&D capitalization rate an important metric for tracking underlying earnings quality.

The cancelled merger leaves Hygon's operating structure unchanged: a fabless chip designer selling the majority of its production to an affiliated server manufacturer within a policy-supported market. Evaluating the sustainability of this business model requires dedicated analytical frameworks.

↑ Back to Top

VII. Playbook: Business & Investing Lessons

Strip away the geopolitics, and Hygon presents a core question that every investor in a protected industry eventually confronts: is this a self-sustaining business, or is it a regulatory license?

The distinction is critical. A business generates returns because customers repeatedly choose its products. A license generates returns because competitors are legally barred from entering the market. Both models can produce strong financial performance for years, but only one survives if regulatory rules shift.

Hamilton Helmer's 7 Powers framework helps illustrate this boundary. Applied to Hygon, three of the seven powers warrant detailed examination, yielding mixed conclusions.

Counter-positioning — powerful, but not Hygon's own. Counter-positioning occurs when an incumbent cannot copy a challenger's model without damaging its existing business. Intel, AMD, and Nvidia cannot sell advanced silicon to sanctioned Chinese entities because U.S. export controls forbid it, allowing Hygon to fill the resulting supply gap.

However, the mechanism driving this dynamic is regulatory rather than strategic. Foreign incumbents are not abstaining from competition to protect existing profit margins; they are legally prohibited from competing. This exclusion stems from foreign trade policy rather than proprietary strategic positioning, and its effects are symmetric. The same trade restrictions that created Hygon's addressable market also created market openings for domestic rivals like Huawei, Cambricon, and Loongson. Consequently, Hygon holds a share of a protected domestic market rather than a unique competitive moat within it.

The defensible core of this positioning is narrower: within China's protected market, Hygon remains the sole domestic vendor enabling enterprise buyers to bypass software recompilation.

Switching costs — the genuine article. High switching costs represent Hygon's most durable advantage. For a major state-owned bank, replacing an x86 server with another x86 system requires no modifications above the hardware abstraction layer. In contrast, migrating to an ARM architecture requires software recompilation, code revalidation, vendor recertification, and extensive regression testing spanning multiple quarters. The primary switching cost is not hardware expenditure, but engineering labor and operational delay.

Hygon's customer disclosures illustrate this dynamic in practice. Its processors are deployed at scale across telecommunications, finance, education, and transport—the precise enterprise segments bound to legacy codebases.³ Furthermore, the company's ecosystem initiative has completed more than 15,000 adaptation projects spanning chips, systems, and enterprise software applications, reflecting substantial third-party capital and engineering commitment to its platform.³

However, switching costs primarily defend existing legacy workloads. They offer little protection in greenfield AI deployments, where software stacks are built without legacy x86 dependencies—the very market segment driving current industry growth and market valuations.

Cornered resource — moderate, and decaying by design. In its FY2025 annual report, Hygon highlights its position as holding "China's only self-developed 'C86+GPGPU' product matrix," reflecting a carefully framed competitive claim.³ Commercial x86 rights remain exceptionally scarce within mainland China, with the alternative lineage—Shanghai Zhaoxin—occupying a lower performance segment.

However, an intellectual property asset that cannot be updated functions as a depleting resource rather than a permanent advantage. The single-generation limit of the original licensing agreement represents a structural boundary that internal R&D cannot bypass.⁸ As global x86 architectures continue to evolve, Hygon's frozen microarchitectural baseline steadily depreciates relative to global performance benchmarks, even as its legacy software compatibility retains immediate commercial value domestically.

The remaining elements of the framework yield mixed results. Scale economies operate standardly in fabless chip design, where fixed costs for mask sets and IP licensing amortize over expanding shipment volumes—a dynamic supported by Hygon's recent revenue growth. Conversely, process power and brand equity find minimal support in official disclosures; neither commands premium pricing over domestic peers, and ongoing gross-margin compression indicates limited pricing power.

Porter's five forces, applied honestly.

Supplier power: high, and rising. This force appears prominently in financial filings. As a fabless designer subject to Entity List restrictions, Hygon relies on a limited pool of foundry, packaging, and testing partners. Management explicitly acknowledges that certain supplier inputs are "scarce and proprietary" and that switching vendors would incur high costs.³ The contraction in consolidated gross margin from 2024 through the first half of 2026 illustrates how supplier pricing power absorbs gross profitability upstream.

Buyer power: high. Customer concentration remains severe, with the top five customers generating 90.28% of total revenue and a single related party accounting for over half of total sales.³ Furthermore, state tender committees control public-sector procurement with an explicit policy goal of sustaining multi-vendor competition, limiting individual supplier bargaining leverage.

Threat of substitutes: high and strengthening. Alternative architectures pose a structural threat. Huawei is not merely offering rival chips, but establishing an integrated domestic computing ecosystem—combining Kunpeng CPUs, Ascend AI accelerators, the CANN software framework, and verified supernode clusters.¹⁴

Threat of new entrants: moderate. While capital requirements and technical complexity present high entry barriers, state-backed investment funds continue to capitalize domestic chip ventures. However, the most critical barrier—access to an x86 architectural license—remains closed to new market entrants.

Rivalry: intense, and increasingly on price. Hygon's FY2025 disclosures highlight price competition and competitive tender bidding as active pressures on operating margins.¹³

The capital-markets lesson. The broader financial lesson of Hygon lies in the divergence between the capital required to acquire its underlying technology and the valuation assigned to it by public markets.

The core microarchitecture was acquired in 2016 for $293 million plus ongoing royalties.¹ Hygon's public listing on the STAR Market in August 2022 raised approximately $1.6 billion—ranking among the largest Chinese IPOs of that year—at an issue price of RMB 36 per share, with the stock closing its trading debut with significant gains.¹⁷[^18] By 2024, surging demand for AI hardware accelerated top-line revenue growth and attracted international market attention.[^19] By May 2026, equity markets valued the company at a multiple of its cumulative lifetime R&D expenditures.²

This outcome does not indicate that AMD mispriced the original joint-venture transaction; AMD monetized market access it could not directly exploit while securing needed capital during a period of corporate distress. Rather, it demonstrates that in markets where regulatory authorization forms the primary competitive constraint, the license holder captures the vast majority of commercial value, irrespective of original engineering origin. This dynamic recurs wherever regulatory barriers, rather than pure technical capability, define market access.

Which sets up the central disagreement about this company.

↑ Back to Top

VIII. Analysis: Bear vs. Bull Case & Current Risk Radar

Two intelligent investors can examine the same Hygon filing and reach opposite conclusions, as official disclosures provide evidence supporting both perspectives. What follows is the analytical debate surrounding the business, framed by the empirical record.

The bull case.

The x86 position inside Xinchuang remains uncontested. This represents the strongest leg of the thesis, independent of management rhetoric. For a core banking engine, a telecom billing system, or a power grid dispatch platform, Hygon remains the sole domestic supplier offering silicon that executes legacy x86 binaries without software modification. As national substitution deadlines approach, state enterprise buyers with complex, legacy software stacks are entering procurement cycles. Hygon's addressable share of that remaining enterprise demand is disproportionately large.

The AI accelerator business benefits from external demand tailwinds. Restrictions on high-end Nvidia hardware, combined with a rapidly expanding domestic artificial intelligence ecosystem, require Chinese enterprises to secure alternative accelerator silicon. Hygon's adaptation of 365 models and DTK's automated CUDA translation lower migration costs for enterprise buyers.³ The 66.5% year-over-year revenue growth recorded in the first half of 2026 reflects genuine market demand rather than administrative allocation alone.⁴

The fabless business model generates high operating margins when unconstrained. Bypassing capital-intensive foundry ownership allows chip designers to convert research expenditures into operating margin. Even after two years of compression, maintaining a gross margin above 55% on RMB 9.10 billion in first-half 2026 revenue represents a lucrative operating profile.¹⁵ Furthermore, annual operating cash flow more than doubled in 2025 to RMB 2.10 billion.³

Optionality to establish an interconnect standard. Launching Hygon System Link represents an effort to position the company's CPU as the foundational socket around which domestic accelerator vendors integrate their hardware.³ If adopted broadly, this architecture would allow Hygon to monetize third-party AI silicon deployments. While unproven commercially, it provides long-term strategic optionality that public markets currently discount.

The bear case.

Upstream supply constraints cap total growth. As a sanctioned fabless designer on the Entity List, Hygon must compete for domestic wafer allocation, advanced packaging capacity, and high-bandwidth memory against unsanctioned and state-prioritized peers. Balance sheet metrics indicate these constraints are active: by mid-2026, inventory reached RMB 7.52 billion and supplier prepayments rose to RMB 4.77 billion, while operating cash flow turned negative at RMB 428 million.¹⁵ Such extensive supplier prepayments signal defensive inventory hoarding and capacity reservation in a tight supply environment.

The single-generation architectural baseline limits long-term scaling. Iterating on a frozen 2017 microarchitecture meets traditional server needs, but encounters bottlenecks when workloads require newer instruction-set extensions. For example, low-precision data types that accelerate CPU-based AI inference were added to the broader x86 instruction set after technology transfers ceased in 2019.⁸ Adding CPU cores and enlarging die area can temporarily compensate for architectural gaps, but doing so increases power consumption and consumes scarce packaging capacity.

Concentrated related-party transactions drive top-line expansion. In 2025, sales to a single affiliated customer more than doubled, generating 56.68% of total revenue.³ This heavy reliance creates operational vulnerability: if procurement from this key affiliate normalizes, top-line growth could slow sharply. Because Hygon reports its CPU and DCU sales as a consolidated category, tracking segment-level underlying demand remains difficult for external investors. Furthermore, four years after Shanghai Stock Exchange regulators requested a plan to reduce related-party dependencies, revenue concentration with affiliates has expanded rather than contracted.⁹

The terminated merger reveals governance and capital allocation friction. The corporate integration process spanned six months and six progress filings before being abruptly cancelled under disclosures citing immature market conditions.³¹⁸ Rather than resolving related-party dependencies, the failed transaction left management offering post-hoc rationales favoring operating independence—contradicting the vertical integration argument presented at launch.⁷[^20]

Commercial channel concentration persists despite corporate independence. While abandoning the merger avoids direct channel conflict with competing server makers such as Inspur (浪潮信息), Lenovo (联想), and ZTE (中兴通讯), Hygon retained operating neutrality by default rather than strategic choice. Because a single affiliated server manufacturer generates over half of Hygon's sales, commercial exposure remains heavily concentrated in practice.

Current risk radar.

Upstream export-control escalation. While finished processor exports are already restricted, secondary risks center on upstream tools and foundry equipment, including electronic design automation (EDA) software updates, spare parts for fabrication equipment, and specialty materials. Hygon explicitly identifies EDA software vendors and IP licensors among concentrated suppliers that carry high replacement costs.³ Stricter enforcement in these supply-chain inputs would slow design iterations and chip tape-outs over time.

Ecosystem encroachment by domestic competitors. The primary competitive risk in AI hardware is that Huawei's integrated platform—combining Ascend accelerators, CANN software, and SuperPoD clusters—becomes the standard reference architecture for state-directed procurement. If public enterprise buyers standardize on full-stack platforms, hardware decisions will occur at the system level rather than the chip level. Huawei's deployment of more than 750 SuperPoD installations and its frontier model training benchmarks highlight this competitive pressure.¹⁴

Structural gross-margin compression. Consolidated gross margins contracted by nearly eleven percentage points over eighteen months, dropping to 57.78% in 2025 and 55.21% in the first half of 2026 as rising foundry and packaging costs outpaced revenue growth.³¹⁵ Inability to pass input-cost increases on to enterprise customers indicates limited pricing power within protected procurement channels, threatening underlying profitability if cost inflation persists.

Capital intensity of dual processor roadmaps. Hygon simultaneously funds R&D for both general-purpose x86 CPUs and GPGPU accelerators without access to external technology sharing, spending RMB 4.57 billion on research in 2025 alone.³ Global peers amortize equivalent design expenditures across far larger international sales volumes. While Hygon remains profitable, maintaining two independent silicon roadmaps within a domestic market limits long-term returns on invested capital.

Valuation volatility and market sentiment. Hygon's stock price nearly doubled between early 2026 and its intraday peak in May before retracing by August.² High equity market valuations expose shareholders to sentiment-driven drawdowns unrelated to quarterly operational performance, while share-price volatility can complicate corporate transactions and capital allocation.

Weighing the evidence. The bull thesis relies on verified policy demand and genuine software switching costs across legacy enterprise workloads. The bear thesis rests on supply-chain bottlenecks, customer concentration, and disclosure limits documented in official filings. On the verified record, Hygon maintains a defensible position in domestic enterprise x86 computing, an unproven stance in frontier AI training, and an earnings profile obscured by consolidated reporting and heavy affiliated sales.

Clarifying that financial profile requires greater reporting transparency across three key areas.

↑ Back to Top

IX. Strategic Horizon & The KPIs That Matter

Every company has a handful of numbers that frame its trajectory. For Hygon, three key metrics dictate the investment case — and two of them are not currently published.

1. Disclosed DCU revenue, and its customer mix.

This is the single most critical key performance indicator, yet it remains unpublished. Hygon reports all silicon revenue under a single consolidated line item for high-end processors.³ As a result, every market estimate regarding the company's AI accelerator business relies on ecosystem announcements, model-adaptation counts, and working-capital inferences.

Investors must watch for a dedicated DCU revenue breakdown in future filings — and, more critically, whether that revenue stems from commercial internet platforms making voluntary repeat orders without administrative mandates. A state-owned entity purchasing domestic accelerators to satisfy a compliance target demonstrates policy efficacy. Commercial hyperscalers like 阿里巴巴 Alibaba or 字节跳动 ByteDance placing follow-on orders demonstrate product performance. Those represent two distinct operational realities that warrant different market valuations.

A parallel qualitative metric is whether Hygon's accelerator deployments are limited to inference and fine-tuning workloads or extend to frontier model training. Based on public disclosures through mid-2026, Huawei's Ascend platform serves as the primary domestic reference architecture for frontier training.¹⁴ Verifiable evidence of a large-scale Hygon DCU cluster handling frontier training in production would fundamentally alter the investment thesis.

2. Related-party revenue share.

This metric provides the clearest measure of whether Hygon operates as an independent merchant chip vendor or a captive supplier. Related-party sales accounted for 100% of revenue in the earliest historical periods reviewed by regulators, dropped to the mid-50% range by 2020, hovered near 40% in 2024, and surged to 56.68% in 2025.⁹³

The trajectory of this metric will answer the core strategic question left unresolved by the cancelled merger. If related-party revenue share declines while total revenue expands, Hygon will have demonstrated genuine customer diversification, justifying a narrower customer-concentration discount. Conversely, if affiliated sales remain above half of total revenue, Hygon's top-line expansion and its primary shareholder's procurement decisions remain fundamentally the same variable — regardless of expanding ecosystem partnership metrics.

3. Gross margin, read alongside the R&D capitalization rate.

These two metrics must be analyzed in tandem due to their underlying accounting relationship.

Hygon's consolidated gross margin contracted from the mid-60% range in prior years to 57.78% in 2025, falling further to 55.21% in the first half of 2026.³¹⁵ Amortization of self-developed intangible assets flows into cost of sales once the underlying product generates revenue — an accounting treatment Hygon adopted following regulatory questioning at IPO.⁹ Concurrently, the R&D capitalization rate fell to 9.26% in 2025 from 17.04% the prior year.³

A critical signal for investors is whether gross margin stabilizes while the R&D capitalization rate increases. Such a pattern would indicate that reported margins are being defended by deferring expenses onto the balance sheet rather than through pricing power or cost management. A healthier signal would be gross margin stabilizing while capitalization rates remain low or decline further, reflecting genuine pricing power or relief from upstream component costs.

Closing.

A popular narrative frames Hygon as a triumph of industrial adaptation: a decade-old licensing deal that outlived the geopolitical conditions that permitted it, transformed into a national compute champion by the very trade restrictions intended to contain it.

That interpretation is not false, but it omits the operational details essential for investment analysis.

The verified record documents a company that licensed a specific architectural baseline during a window of counterparty vulnerability and constructed a capable engineering organization around it — deploying 2,766 R&D staff, securing 1,101 granted invention patents, and developing a software stack that meaningfully lowers migration cost.³ Yet that same record reveals a company paying a rising nine-figure annual IP royalty to an unnamed minority shareholder, selling over half its output to an affiliated buyer, reporting two distinct silicon lines as one number, failing to complete its proposed structural merger, and surrendering roughly six points of gross margin to its suppliers in eighteen months.

Both descriptions derive from the same audited disclosures. Hygon's upcoming reporting periods will determine which one defines its commercial trajectory.

Hygon Information Technology Co., Ltd.

Table of Contents

Hygon Information Technology: The Trojan Horse of Chinese Silicon

I. Introduction & Episode Roadmap

II. The x86 Imperative & China's Enterprise Compute Conundrum

III. The 2016 AMD Joint Venture Alchemy & IP Mechanics

IV. The 2019 Entity List Shockwave & The Forced R&D Pivot

V. Product Anatomy, Segment Economics: CPU vs. DCU

VI. Governance, Management Credibility, & The Sugon Mega-Merger

VII. Playbook: Business & Investing Lessons

VIII. Analysis: Bear vs. Bull Case & Current Risk Radar

IX. Strategic Horizon & The KPIs That Matter

References