Dawning Information Industry (中科曙光 Sugon): China's Supercomputer Pioneer and the Chip Stake Worth More Than the Company

I. Introduction & Episode Roadmap

On the afternoon of April 21, 2026, a peculiar arithmetic problem sat on the screens of every analyst covering China's computing sector.

海光信息 Hygon Information Technology, a chip designer listed on Shanghai's STAR Market, closed the day with a market capitalisation of roughly 587 billion yuan. Its single largest shareholder — holding 27.96% of the company — was a server manufacturer called 中科曙光 Dawning Information Industry, better known by its brand name, Sugon. That stake alone was worth about 164.1 billion yuan.

Sugon's entire market capitalisation that day was roughly 131 billion yuan.¹

Read that again. The parent's stake in its own spin-out was worth about 33 billion yuan more than the whole parent. Which means the market, in its collective wisdom, had assigned a negative value to everything else Sugon owned: a fifteen-billion-yuan revenue business, thirty years of supercomputing engineering, liquid-cooled data centres in more than fifty Chinese cities, and a customer list that reads like a directory of China's state apparatus.

There are only a few ways to explain a stub value like that, and none of them are boring. Either the market believes Sugon's core hardware business destroys capital. Or it believes the Hygon stake is somehow trapped — unsellable, untaxable-at-any-reasonable-rate, structurally locked. Or the market is simply wrong, and a large, liquid, heavily-traded A-share has been mispriced in plain sight for months.

This is the story of how a company gets into that position.

Dawning Information Industry Co., Ltd. (曙光信息产业股份有限公司) trades in Shanghai under the ticker 603019. It is not a startup. It is not, in the Western sense, really a private company either. It was born inside a Chinese Academy of Sciences laboratory in the early 1990s as a national answer to a national humiliation, and it has never fully left that orbit. Its customers are governments, state-owned enterprises, and national supercomputing centres. Its founder was an academician. Its shares were listed only in 2014, eighteen years after the company was formed.²

In the year to December 2025, Sugon generated 14.96 billion yuan of revenue and 2.18 billion yuan of net profit attributable to shareholders.³ Those are respectable numbers. They are also, by the standards of the AI infrastructure boom raging around it, curiously modest — Sugon is not the largest server vendor in China, not the second, and not the third.

Three threads run through what follows.

The first is the core hardware business: what Sugon actually builds, who buys it, and whether the company holds any durable advantage in a market where the leaders are several times its size.

The second is Hygon — the chip company that emerged from Sugon's own orbit, absorbed an AMD x86 licence at exactly the right moment in history, went public separately in 2022, and then grew into something worth multiples of its parent. In May 2025 the two announced they would merge, in a share swap valued at roughly 116 billion yuan. In December 2025 they called it off.⁴ That episode is the analytical heart of this story.

The third is people. In January 2025 the 82-year-old founder resigned. By May 2026 his long-serving deputy held both the chairman's seat and the general manager's title — and had, two months earlier, disclosed a plan to sell a meaningful slice of his own shares.⁵

A company at the centre of a geopolitical technology race, with a valuation that makes no obvious sense, a collapsed mega-merger, and a founder succession still settling. Let's start where it started: in a laboratory, with an embargo.

↑ Back to Top

II. Origins: From a State Research Lab to a Public Company

The story that Chinese computer scientists tell about the late 1980s is called the "glass house."

When China's oil industry needed serious computing power for seismic modelling, it bought a supercomputer from an American vendor. But the machine did not simply arrive and get plugged in. It was installed inside a sealed, transparent room. American engineers operated it. Chinese researchers submitted their jobs and waited outside, watching through the glass while foreign technicians ran the calculations and reviewed what came out.⁶

You could not design a more efficient device for creating a generation of engineers determined never to be in that position again.

The 863 Program and the Institute of Computing Technology

China's response was institutional. The 863计划 863 Program — named for its March 1986 launch — was a state high-technology development effort that funded, among other things, indigenous high-performance computing. The work landed at the 中国科学院计算技术研究所 Institute of Computing Technology of the Chinese Academy of Sciences, the country's oldest and most prestigious computer research institution.

The man who led it was 李国杰 Li Guojie, born in 1943, a computer architect who had trained abroad and returned to China to build machines.⁷ The project he took on in the early 1990s was, by the standards of the day, close to absurd: build a parallel computer using commercially available microprocessors and a standard UNIX operating system, and do it on a budget of two million yuan — a sum that would not have covered the cabling on a comparable American machine.

The team improvised. Six researchers relocated to Silicon Valley for the better part of a year, living austerely and working relentlessly, joking that they were "foreign volunteer workers." Actual development took under twelve months.⁶

In May 1993, 曙光一号 Dawning No. 1 ran. It executed 640 million operations per second — trivial by modern standards, transformative in context, because it was China's first general-purpose parallel computer built on mainstream commercial microprocessors and standard software rather than exotic bespoke hardware. It received national certification in October 1993 and a second-class National Science and Technology Progress Award.

Then something happened that tells you more about the machine's real significance than any benchmark. Within days of Dawning No. 1 being unveiled, the Paris-based multilateral export control committee relaxed its restrictions, permitting sales into China of the class of supercomputers it had previously banned.⁶

That is the founding lesson of this company, and it has never stopped being relevant: the value of building it yourself is often not the machine — it is the change in what others will sell you.

From project to company

Most state research triumphs in China at that time ended as papers and awards. Li Guojie's team chose differently. In 1996, the intellectual property from Dawning No. 1 was valued at 20 million yuan and contributed as capital into a new commercial entity: 曙光信息产业有限公司, Dawning Information Industry Co.⁶

The early market response was humbling. The first-generation system sold three units. Chinese customers in the mid-1990s did not know what to do with a domestic parallel computer, and the sales cycle for institutional buyers who had never bought Chinese hardware was measured in years, not quarters.

But the customer set that eventually emerged — government departments, state research institutes, national supercomputing centres, state-owned enterprises — was exactly the customer set that a Chinese Academy of Sciences spin-out was uniquely positioned to serve. By 2012 Sugon held roughly 35% of China's high-performance computing market and was the only Chinese firm in the global top ten HPC vendors by revenue.⁶

A 2006 corporate restructuring converted the business into a joint-stock company. And on November 6, 2014, after an offering of 75 million shares priced at 5.29 yuan that raised 397 million yuan, Sugon listed on the Shanghai Stock Exchange under 603019.² Note what this was not: not a reverse merger, not a backdoor listing into a shell, not a STAR Market growth listing. It was a conventional main-board IPO by an eighteen-year-old profitable company.

Why the DNA matters

It is tempting to skip the origin story and get to the interesting present. Don't. Almost every consequential fact about Sugon in 2026 traces directly back to this lineage.

State-research parentage produced the customer base — which means demand tied to government budgets and policy cycles rather than organic enterprise buying. It produced the political salience — which is why the company later became a bargaining chip in a trade war. It produced the institutional shareholder structure, with the Institute of Computing Technology's ecosystem in the background rather than a founder-controlled block. And it produced, eventually, the sanctions exposure, because a company that openly built exascale-class machines for Chinese national programmes was always going to attract attention in Washington.

It also produced something subtler: a habit of spinning technology out rather than holding it in. Li Guojie's career included leadership of the 龙芯 Loongson processor programme, another Institute of Computing Technology project that became a separate company.⁷ The pattern of "incubate inside the state research ecosystem, then let the child list separately" is not an accident at Sugon. It is the operating model. It is also, as we will see, the reason the parent trades where it does.

Before we get to that, we need to understand what the parent actually sells.

↑ Back to Top

III. What Sugon Actually Sells: The Core Business Today

Walk into one of Sugon's demonstration data centres and the first thing that strikes you is what you don't hear.

Conventional server halls are loud — a continuous roar of fans pushing air across hot silicon, with cooling typically consuming a third or more of total facility power. In Sugon's immersion-cooled rooms, racks of servers sit submerged in dielectric fluid, the coolant boiling gently against the processors and condensing above them, carrying heat away by phase change rather than by moving air. It is nearly silent.⁸

That is the physical expression of Sugon's actual product philosophy: this is a company that sells thermodynamics and systems integration as much as it sells computers.

The segment reality check

Strip away the narrative and Sugon is, unambiguously, a hardware company.

In FY2025, IT equipment — rack servers, high-density and blade systems, storage arrays, cluster hardware — generated 12.50 billion yuan of revenue, growing a modest 6.81%. The remaining segment, covering software development, system integration and technical services, generated 2.45 billion yuan.⁹ Together they make up essentially all of the 14.96 billion yuan top line.

The interesting part is the direction of travel. In FY2024, IT equipment accounted for roughly 89% of revenue.¹⁰ By FY2025 that had fallen to around 84% — not because hardware shrank, but because the software and services line grew 75.34% in a single year.⁹

That is the single most important operating datapoint in Sugon's recent history, and it deserves plain-English interpretation. A hardware business growing 7% while an attached services business grows 75% is a business whose mix is improving faster than its scale. If it continues, it changes the margin structure. If it was a one-year artefact of a handful of large integration contracts, it changes nothing. One year does not settle the question, and management has not disclosed the contract composition underneath it.

The products, in plain terms

Sugon's catalogue divides into three practical buckets.

Servers and storage. The bread and butter — rack-mounted, high-density and blade servers, plus storage systems. This is a commodity-adjacent business where the buyer specifies a configuration and vendors compete on price, delivery and service. Margins reflect that.

Liquid cooling and high-density infrastructure. Sugon began liquid-cooling research in 2011, released what it described as China's first mass-producible cold-plate liquid-cooled server in 2015, and by 2019 had deployed blade-format immersion phase-change cooling at scale, reaching rack power densities of 160 kilowatts.⁸ For context: a conventional air-cooled rack typically handles somewhere in the single-digit-to-low-teens kilowatt range. When AI accelerators started generating heat that air simply cannot remove economically, this stopped being an engineering curiosity and started being a requirement.

City cloud computing centres. Sugon builds and operates municipal cloud infrastructure under a model it describes as "enterprise invests and operates, government purchases services" — the company funds the data centre, the local government buys capacity. By 2024 it had built close to fifty such city-level cloud centres.¹¹ This is infrastructure-as-utility for local government, and it is a genuinely hard thing for a foreign or purely commercial vendor to replicate, because it requires municipal relationships accumulated over a decade.

Channel and margin structure

Sugon sells overwhelmingly direct. Distribution is a minority channel and carries visibly lower gross margins than the direct business.¹² That is not a distribution failure — it is a structural statement about the business. When your buyer is a provincial government IT bureau or a national supercomputing centre, the sale is made by an account team that has worked the relationship for years, not by a reseller.

The consequence cuts both ways. Direct, relationship-driven selling into state buyers is defensible — it is genuinely hard for a newcomer to displace an incumbent with ten years of installed base and procurement history. It is also expensive, cyclical, and gives buyers enormous leverage on price, because a provincial procurement office running a tender knows exactly how many qualified domestic vendors will bid.

Company-wide gross margin ran around 27–30% through 2024 and 2025.¹ To calibrate: Hygon, selling chips into the same ecosystem, reported gross margin of roughly 57.8% in 2025.¹ The parent captures roughly half the gross margin percentage of the child. That gap is the entire argument about where value accrues in China's compute stack, expressed in a single ratio.

The financial arc, and what it actually shows

The recent trajectory is not a smooth line, and reading it as one would be a mistake.

FY2024 was a down year on the top line: revenue fell 8.40% to 13.15 billion yuan, while net profit attributable to shareholders rose 4.10% to 1.91 billion yuan.¹⁰ Management's framing was mix discipline — walking away from low-margin projects. Cost of sales fell 12%, faster than revenue, which is consistent with that story.¹³

But the composition of that FY2024 profit deserves scrutiny. Investment income contributed 562 million yuan, up 51% year on year and equal to roughly 29% of net profit — the bulk of it flowing from Hygon. Government subsidies contributed a further 529 million yuan, about 28%. Together, non-operating and equity-method sources accounted for something on the order of 57% of the company's reported net profit.¹² Meanwhile the cloud computing subsidiary posted a net loss of 186 million yuan with net assets down 39%.¹²

A profit line that grows while revenue falls, supported majority-wise by associate income and state subsidies, is not evidence of operating strength. It is evidence of an accounting structure that transmits somebody else's operating strength.

FY2025 was materially better on the operating line. Revenue rose 13.81% to 14.96 billion yuan and attributable net profit rose 13.87% to 2.18 billion yuan — but the number that matters is non-recurring-adjusted profit, which grew 33.97% to 1.84 billion yuan.³ Adjusted profit growing at more than twice the pace of headline profit means the core business improved faster than the reported figure suggests, and that subsidies and one-offs were a smaller share of the whole. That is a genuine improvement in earnings quality, and it is the strongest single piece of evidence in the bull case.

The fourth quarter carried the year: revenue of 6.14 billion yuan, up 20.3%.¹⁴ The board declared a dividend of 3.80 yuan per ten shares.³

Then 2026 accelerated. First-quarter revenue rose 23.7% with adjusted net profit up 53.3%.¹ The first half brought revenue of 7.43 billion yuan, up 26.99%, attributable net profit of 978 million yuan, up 34.19%, and adjusted net profit of 853 million yuan, up 50.01%.¹⁵ Second-quarter net profit of 750 million yuan was more than triple the first quarter's 228 million.¹⁵

The counter-evidence

Now the part a management presentation would place in a footnote.

FY2025 operating cash flow fell 51.75% to 1.31 billion yuan — against 2.18 billion yuan of reported net profit, a cash conversion ratio of roughly 0.61.¹⁶ Management attributed this to large-scale commissioned development project outlays and placements into large certificates of deposit. Accounts receivable grew 21.47% to 2.76 billion yuan, outpacing revenue growth of 13.81%, with credit impairment losses of 171 million yuan.¹⁶

Receivables growing half again as fast as revenue is a pattern with a small number of explanations: longer government payment cycles, looser credit terms to win competitive bids, or revenue recognised ahead of collection. None of them are fatal in a business selling to state buyers, where payment delays are endemic and eventual collection is usually reliable. All of them are worth watching, because the same pattern in a company with less creditworthy customers would be a serious flag.

One more oddity: selling expenses fell 9.09% in a year when revenue rose 13.81%.¹⁶ Operating leverage in sales costs is normally a good sign. In a market where competitors are spending aggressively to capture AI infrastructure share, cutting sales investment is at minimum a choice worth explaining, and the company has not explained it in detail.

So: a hardware business with improving mix and genuinely accelerating adjusted earnings, funded partly by an associate it doesn't control, converting profit to cash at a deteriorating rate, in a market it does not lead. That last clause is the one we haven't tested yet.

↑ Back to Top

IV. The Industry: Who Sugon Actually Competes With

Here is a useful exercise. Take IDC's ranking of China's x86 server market for 2024, and read down the list of vendors by revenue share.

浪潮信息 Inspur Information: 30.8%, roughly 88.5 billion yuan. 超聚变 Xfusion — the server division 华为 Huawei sold off in 2021 to escape US sanctions on the parent — 13.3%, about 38.2 billion yuan. 新华三 H3C: 12.6%, about 36.3 billion yuan. 联想 Lenovo: 9.8%. 宁畅 Nettrix: 8.3%. 中兴通讯 ZTE: 6.5%.¹⁷

Keep reading. Sugon does not appear in the top six.

That is the single most important corrective to the enthusiasm surrounding this stock. In the largest and most visible market Sugon competes in, it is not a leader. It is a second-tier player, in a market where the leader is roughly six times its total company revenue in that segment alone.

The three-tier structure

Chinese industry analysts describe the server market as three tiers. The first contains Inspur, H3C and the Huawei ecosystem — vendors with both domestic scale and global relevance. The second contains Lenovo and Sugon — real businesses with real capability, operating at a fraction of the leaders' volume.¹⁷

Scale in server manufacturing is not a vanity metric. It determines component purchasing power, the amortisation base for engineering, the density of the service network, and the ability to bid aggressively on large tenders and still make money. A vendor with a third of the market buys memory and power supplies on different terms than a vendor with five percent.

Sugon's position in the faster-growing accelerated-computing segment is, if anything, harder to defend. In China's accelerated server market — the AI-driven, higher-value half of the industry — the top three vendors by revenue in the first half of 2025 were Inspur, H3C and Lenovo, together taking close to half the market. Sugon was not among them.¹⁷ Internet companies accounted for nearly 69% of accelerated server purchasing, a customer segment where Sugon's government and SOE relationships confer no advantage whatsoever.

Where Sugon is genuinely different

Having stated the uncomfortable version, the honest analysis requires stating the other side properly, because "second tier in unit share" is not the same as "undifferentiated."

Supercomputing pedigree. Sugon's relationships with China's national supercomputing centres go back further than most of its competitors have existed as server vendors. These are not commodity procurements. They are multi-year engineering engagements involving interconnect design, system software, cooling and facilities. That is a cornered resource in the strict sense — access to a small set of institutional customers that cannot be bid away on price.

Liquid cooling. As AI accelerator power draw climbs, cooling stops being a facilities line item and becomes a design constraint on the compute itself. Sugon has been shipping liquid-cooled systems at commercial scale since well before the current AI cycle. Whether that translates into pricing power depends on whether competitors can catch up quickly — and cold-plate liquid cooling, in particular, is a technology many vendors have now productised. Immersion phase-change at 160kW densities is harder to copy. The company has not disclosed what share of its revenue carries a cooling-related premium, which is precisely the disclosure that would settle the argument.

The Hygon tie-in. This is the differentiator that actually matters commercially. Sugon builds systems around Hygon CPUs and DCUs — domestic silicon that is not subject to US export licensing. For a Chinese government buyer operating under 信创 Xinchuang domestic-substitution procurement rules, a Sugon-Hygon system clears a policy requirement that an Intel- or Nvidia-based system from a larger competitor cannot. That is a real, structural advantage in a specific and growing slice of the market.

Cluster-level delivery. The most encouraging recent development is a shift up the value chain. Rather than selling accelerator cards or individual servers, Sugon has moved toward delivering complete 万卡级 ten-thousand-card AI clusters — integrated systems with interconnect, cooling, power and software stack. By February 2026 the company had delivered three such complete clusters.¹ Per-unit revenue at cluster level is orders of magnitude above card-level sales, and the integration work is where a systems company can earn margin that a box-shifter cannot.

The evidence test

So is "differentiated" a fact or a narrative?

The honest answer is: partially proven. The Hygon tie-in is a fact, verifiable in the procurement rules. The supercomputing relationships are a fact, visible in the installed base. The liquid-cooling lead is a fact in engineering terms but unquantified in financial terms. The cluster-level shift is real but early — three delivered clusters is a proof of concept, not a business line.

What is not proven is that any of this produces durable pricing power. Sugon's gross margin, at roughly 27–30%, does not look like the margin of a company with a technical monopoly. It looks like the margin of a competent systems integrator in a competitive market. Companies with genuine product advantage generally show it in the margin line before they show it in the press release.

There is also a quieter competitive threat. Huawei's Ascend ecosystem and its Xfusion offshoot are pursuing the same domestic-substitution demand with a full stack of their own — chips, systems, software — and considerably more scale. Inspur has its own accelerated systems and dominant channel reach. Sugon's structural advantage is not that it is the only company that can build a Xinchuang-compliant AI cluster. It is that it is one of a handful, with a preferred position on one specific silicon platform.

Which brings us to the moment that made that platform indispensable, and it did not come from a customer. It came from Washington.

↑ Back to Top

V. The 2019 Entity List: The First Inflection Point

On June 24, 2019, the US Department of Commerce's Bureau of Industry and Security published a Federal Register notice adding five Chinese entities to the Entity List.

The names were: Wuxi Jiangnan Institute of Computing Technology, Sugon, Higon, Chengdu Haiguang Integrated Circuit, and Chengdu Haiguang Microelectronics Technology.¹⁸

The stated justifications were specific. The Wuxi institute was cited for military ownership and involvement in exascale high-performance computing. Sugon was cited on the basis of what BIS described as the company's own public acknowledgement that its high-performance computers have a variety of military end uses and end users. The three Haiguang-related entities were designated because of their relationships with Sugon, and were assessed as posing significant risk of involvement in activities contrary to US national security and foreign policy interests.¹⁸

The practical effect: any supplier anywhere in the world needed a licence from BIS to export controlled US-origin items to these entities, and licence applications would be reviewed with a presumption of denial.¹⁸¹⁹

The part that hurt

The reflexive reading of an Entity List designation is that it blocks a company's exports. For Sugon, that was almost irrelevant — it sold overwhelmingly into China.

The designation hit somewhere far more sensitive. In 2019, Sugon built servers. Servers require processors. Sugon's processors came from Intel and AMD. Its accelerators came from Nvidia. The Entity List did not primarily threaten Sugon's ability to sell; it threatened Sugon's ability to build.

This is worth sitting with, because it is the mechanism by which export controls actually bite, and it is frequently misunderstood. A hardware assembler in a globalised supply chain has no independent existence if the chips stop arriving. Sugon in mid-2019 was a company whose entire product line depended on components that had just become licence-gated with a presumption of denial.

The response

Publicly, the company was defiant, as Chinese firms in that position invariably are. The substantive response was structural, and it was much more interesting.

Sugon accelerated, hard, toward domestic silicon. Every engineering resource that could be pointed at qualifying, integrating and optimising Chinese processors got pointed there. The systems software work, the compiler and library work, the thermal and interconnect design work that had been organised around x86 from Intel and accelerators from Nvidia had to be re-organised around what China could produce itself.

The company that could produce it was, conveniently, sitting right there in Sugon's own corporate family.

Here is where the 2019 designation becomes one of those events that, in retrospect, looks less like a wound than like a forcing function. Absent the Entity List, Sugon would plausibly have continued as a competent integrator of American silicon — profitable, mid-tier, and permanently exposed to whatever Intel and Nvidia decided their Chinese channel was worth. The sanctions eliminated that option and made the Hygon relationship existential rather than optional.

It also, over the following six years, converted a policy risk into a policy tailwind. As the Chinese state made domestic substitution an explicit procurement priority, being on the US Entity List became something closer to a credential in state tenders. A vendor that cannot legally use restricted American components is a vendor whose supply chain cannot be interrupted by a future export control announcement.

What it did not fix

Two things should temper that reading.

First, being sanctioned did not make Sugon's chips good. It made Sugon dependent on whether Hygon's chips would be good, which was, in 2019, a genuinely open question. Hygon's first-generation products were derived from an AMD architecture licence of a specific vintage, and that licence did not extend forward.

Second, the Entity List is not a static document. It expands. Every year since 2019 has brought additional Chinese technology entities and additional categories of controlled items, and the tooling, EDA software and manufacturing equipment upstream of Hygon's silicon remain exposed in ways Sugon cannot control. A company whose competitive position rests on being immune to export controls has an unusual dependency on the assumption that the controls will not reach a layer it depends on.

Which raises the question we have deferred long enough: what exactly is Hygon, where did it come from, and how did a US-restricted Chinese entity end up with a legal x86 architecture licence in the first place?

↑ Back to Top

VI. Hygon (海光信息): The Spin-Out That Became the Story

The answer begins in February 2016, in a deal that Silicon Valley has spent the decade since arguing about.

AMD was, at that point, in genuine distress. Its share price had spent 2015 in the low single digits, its server market share had collapsed to a rounding error against Intel, and it needed cash to fund the Zen architecture programme that would eventually save it. So it did something no American chip company had done before or has done since: it licensed x86 to China.

The THATIC structure

The mechanism was deliberately baroque. AMD entered into a joint venture with a Chinese consortium named Tianjin Haiguang Advanced Technology Investment Co. — THATIC. The arrangement was worth $293 million plus royalties.²⁰

Underneath THATIC sat two further joint ventures. Chengdu Haiguang Microelectronics Technology (CHMT), in which AMD held a majority stake, held the licensed designs. Haiguang IC Design — Hygon — in which THATIC held the majority, licensed those designs from CHMT.²¹ The structure was engineered so that AMD retained control of the entity holding the IP, which was the argument AMD made to US regulators: the crown jewels never left American control.

The licence covered a single generation. AMD's leadership was explicit, repeatedly, that Zen 2 and subsequent architectures would not be transferred.²⁰

In 2018 Hygon began production of the Dhyana processor — functionally near-identical to AMD's EPYC server CPUs, differing chiefly in vendor identification codes. Linux kernel patches submitted for Dhyana support made the resemblance publicly visible to anyone reading the mailing lists.²⁰

A year later, all of it was on the Entity List, and AMD's Chinese x86 experiment was over as an ongoing technology transfer.²¹ But the generation that had already been transferred did not un-transfer. Hygon had a legal, paid-for x86 architecture — and, critically, x86 software compatibility, meaning the enormous installed base of enterprise software written for Intel and AMD servers would run on Hygon silicon without recompilation.

For a layman's analogy: China did not just get a car engine. It got an engine that fits the existing chassis, uses the existing fuel, and works with every mechanic already trained on the old one. In processors, the instruction set is the ecosystem, and ecosystems take decades to build.

From CPU to DCU

Hygon's second act was more consequential than its first.

Alongside CPUs, Hygon developed DCUs — deep computing units, its term for AI accelerator cards. These are the domestic answer to Nvidia's data-centre GPUs. And Hygon made the same architectural bet with DCUs that it had inherited with CPUs: compatibility. The DCU line is built around a CUDA-like programming model, meaning AI code written for Nvidia hardware can be ported with substantially less effort than to an architecturally alien accelerator.²²

This is the whole game in AI silicon. Nvidia's durable advantage was never purely the transistors — it was CUDA, the software layer that a generation of machine-learning engineers learned first and never left. Every challenger faces the same problem: superior hardware that requires rewriting the software is, commercially, inferior hardware. Hygon's approach sidesteps the worst of that by minimising the porting distance.

Hygon describes the combination as a 双芯战略 dual-chip strategy — CPU and DCU as parallel product lines driving each other — and pairs it with an ecosystem programme, the 光合组织 Guanghe alliance, which it says has gathered more than 6,000 partners.²² Ecosystem partner counts are a soft metric and should be treated as such, but the strategic intent is unmistakable: build the software gravity that makes the silicon sticky.

The listing, and the mathematics of the stake

Hygon listed on the Shanghai STAR Market in 2022. Sugon retained the largest shareholding — 27.96% as of recent disclosure — and remains both Hygon's biggest shareholder and one of its most important customers, since Sugon builds servers and clusters around Hygon silicon.¹²³

Then Hygon grew, fast. FY2025 revenue rose 56.9% to 14.4 billion yuan; net profit rose 32% to 2.54 billion yuan.²⁴ Note the symmetry: Hygon's FY2025 revenue was, within a rounding error, identical to Sugon's. The child had caught the parent in one year, and was growing four times faster.

It did not stop. Hygon guided first-half 2026 revenue growth of 55.56% to 70.20% year on year, with attributable net profit growth of 41.50% to 52.32%.²²

The market responded accordingly. Hygon's capitalisation reached roughly 587 billion yuan by April 21, 2026 — making Sugon's 27.96% stake worth about 164.1 billion yuan against a Sugon market capitalisation of roughly 131 billion.¹ The gap moved around considerably through 2026: by June 25, Hygon's capitalisation had climbed to roughly 837 billion yuan, valuing Sugon's stake near 234 billion; by early August it had retreated, with the stake worth closer to 188 billion.²³

What this actually means for valuation

A sum-of-the-parts framing is unavoidable here, and it needs to be done carefully rather than triumphantly.

Take Sugon's market value. Subtract the market value of the Hygon stake. What remains is the market's implied valuation of everything else Sugon owns — and through much of 2026, that residual was negative.

The bull reading is straightforward: an investor buying Sugon is buying Hygon exposure at a discount, and receiving a 15-billion-yuan-revenue, 2-billion-yuan-profit hardware business for free.

The problem is that stub-value arguments are among the most seductive and most frequently wrong ideas in equity analysis, and there are specific reasons the discount may be rational rather than absurd.

The stake is not liquid in any practical sense. Sugon selling down Hygon would forfeit its position as the largest shareholder of its own critical supplier, would face lock-up and disclosure constraints, and would be a politically loaded act given both companies' strategic status. A holding that cannot be monetised without destroying the strategic rationale for holding it is worth less than its mark-to-market.

There is also the double-counting problem. Sugon's reported earnings already include equity-method income from Hygon. An investor who values Sugon's earnings stream and separately values the stake at market is counting the same asset twice.

And there is the reflexive problem, which the merger collapse illustrated brutally: the higher Hygon's share price goes, the more negative Sugon's implied core-business value becomes, and the harder it is to argue that the market is making a simple arithmetic error rather than expressing a considered judgement about the parent's standalone quality.

The right way to hold this is as a question, not a conclusion. Is the discount a mispricing of a good business, or an accurate assessment of a mediocre one attached to a great asset? The most direct test anyone could have devised for that question was attempted in 2025 — and it failed.

↑ Back to Top

VII. The Merger That Almost Happened — and Didn't

On Monday, May 26, 2025, trading in both 603019 and 688041 was suspended.²⁵

The announcement that followed was, in the context of Chinese capital markets, remarkable. Hygon would absorb Sugon through a share swap — the subsidiary swallowing its own largest shareholder — with newly issued Hygon shares distributed to Sugon's shareholders. The transaction was valued at approximately 116 billion yuan, or about $16.4 billion.⁴ Shares were suspended for up to ten trading days while the structure was worked out.²⁵

The strategic logic was clean enough to fit in a sentence: put the chip designer and the systems builder inside one company, and China would have a vertically integrated compute champion running from transistor to cluster — the domestic answer to the integrated stacks being assembled elsewhere.

Markets loved it. From the announcement through early December, Sugon's shares gained more than 45%, while Hygon's rose around 60%.²⁶

The problem with a rising tide

And that, precisely, was the problem.

A share-swap merger is fundamentally an exchange ratio: how many Hygon shares does each Sugon share convert into? That ratio has to be negotiated, documented, approved by both boards, cleared by regulators, and voted on by both shareholder bases. In China, for a transaction of this scale involving two strategically significant listed companies, that process takes many months.

During those months, both stocks were repricing violently on AI enthusiasm — and repricing at different rates. Every week that Hygon outran Sugon, or Sugon outran Hygon, the negotiated ratio drifted further from what either side's shareholders would accept as fair. Executives described a cumulative move of over 61% in the two companies' share prices between early June and early December.²⁷

Merger arbitrage funds noticed the gap. Retail investors noticed it. And a fixed exchange ratio in a market moving that fast becomes not a valuation, but a lottery ticket with a known expiry date.

December 9, 2025

The termination announcement came after the close on December 9, 2025.

The stated reasons were procedural on their face: the transaction's large scale, the number of parties involved, the length of the planning period, and significant changes in the market environment meant that conditions for completing the restructuring were not yet mature.²⁸

The next morning, Sugon opened limit-down — the maximum permitted 10% single-day decline in Shanghai — and stayed there through the close. Hygon fell more than 5% intraday before recovering to end down just 0.36%.²⁷

The asymmetry of that reaction is the most eloquent commentary anyone offered on the deal. The market had understood the merger as Sugon shareholders receiving Hygon shares — that is, as a mechanism for closing the discount. When the mechanism disappeared, the discount reappeared instantly.

What management said

The post-termination investor briefing is worth reading closely, because this is where management credibility gets tested.

沙超群 Sha Chaoqun, Hygon's general manager, was direct about the mechanism, saying the market environment had undergone substantial changes compared to when the transaction was first planned, and pointing to the cumulative 61%-plus move in both share prices as reflecting broader AI chip sector volatility rather than anything specific to the transaction.²⁷

历军 Li Jun, Sugon's general manager, offered a strategic reframing: keeping the companies separate, he argued, allows Hygon to function independently as a chip supplier to the entire domestic server industry, while Sugon can work with multiple chip manufacturers as a systems vendor.²⁶ He described the two companies as capable of serving as "dual cores" for China's computing infrastructure, reducing single-point-of-failure risk for the industry.²⁷

Assess that honestly. The share-price explanation is specific, mechanically coherent, and consistent with the observable facts — it is a real answer, not a deflection. That is more than many managements offer when a deal collapses.

The strategic reframing is weaker. If independence genuinely serves both companies better — Hygon selling to all server vendors, Sugon buying from multiple chip suppliers — that was equally true in May 2025 when they announced the merger. Nothing about the industrial logic changed between May and December. Only the share prices changed. Presenting a valuation-driven abandonment as a strategic insight is the kind of retrofitted narrative that investors should note and discount.

There is also a live tension in Li Jun's framing that nobody resolved on the call. If Hygon is to supply the whole domestic server industry, it will be supplying Sugon's competitors. If Sugon is to work with multiple chip vendors, it will be diluting the one advantage that most clearly differentiates it. The "dual cores" formulation sounds reassuring and is, on inspection, a description of two companies becoming somewhat less useful to each other.

Discipline or paralysis?

The generous interpretation: walking away from a merger whose exchange ratio had become indefensible was disciplined. Boards that complete bad deals because they announced them destroy more value than boards that abandon them.

The critical interpretation is more uncomfortable. The exchange ratio problem was entirely foreseeable. Share-swap mergers between volatile stocks in a hot sector are a known hazard, and the standard toolkit for managing it — collars, floating ratios, walk-away provisions with price triggers, shortened approval timelines — has existed for decades. Six months of executive attention, adviser fees and investor expectation went into a structure that a market move of predictable magnitude was sufficient to break. That is not evidence of discipline. It is evidence of a structure that was not engineered for the environment it was launched into.

The broader context matters too. The collapse landed in a period when several large Chinese semiconductor combinations were being reconsidered as valuations moved faster than deal processes could accommodate. Sugon-Hygon was the largest and most visible, but the pattern — announce during enthusiasm, unwind when relative prices diverge — was not unique to it.²⁹

Is Merger 2.0 live?

Both companies committed to continuing close collaboration on system-level products.²⁷ The industrial logic of chip-to-system integration has not been refuted by anything that happened; only the financial mechanics failed.

So a renewed attempt is a genuine scenario rather than wishful thinking. What would make it more likely: a period of relative share price stability, a narrowing of the valuation gap, or a deal structure with a floating ratio. What would make it less likely: Hygon continuing to outgrow Sugon, which widens the gap and makes any ratio harder to justify to Sugon's shareholders; or a regulatory preference for keeping a critical domestic chip supplier independent of any single systems vendor.

An investor treating a second merger attempt as a base case is making an assumption management has not endorsed. An investor treating it as impossible is ignoring that both boards thought it was correct fifteen months ago.

Whoever decides the answer will be a different set of people than the ones who decided it the first time — because while the deal was collapsing, the company was changing hands at the top.

↑ Back to Top

VIII. Leadership in Transition: Founder Succession and Governance Signals

On January 23, 2025, Sugon's board received a written resignation letter from its chairman.

Li Guojie was 82 years old. He had chaired the company since its founding in 1996 — twenty-nine years, spanning the entire arc from a three-unit product launch to a fifteen-billion-yuan revenue business. He resigned from all positions: chairman, director, and head of the board's strategy committee. The stated reason was personal age considerations.³⁰

By the standards of Chinese corporate governance events, this was almost startlingly clean. No regulatory investigation. No detention notice. No unexplained "personal reasons" that later turned out to mean something else. An academician in his eighties stepped down because he was in his eighties.

Who Li Guojie was

It is worth pausing on the man, because founder-CEOs of this type are becoming rare in China's technology sector and their departures mark genuine transitions.

Born in 1943, Li was a computer architect by training and an institution-builder by practice. He led the Dawning No. 1 project. He led the Loongson processor programme, China's most sustained attempt at an indigenous instruction-set architecture.⁷ He became an academician of the Chinese Academy of Engineering. He spent his career at the intersection of state research priority and commercial execution — a position with no real Western analogue, requiring simultaneous fluency in scientific credibility, bureaucratic navigation and industrial delivery.

His practical role at Sugon in later years was strategic rather than operational. Board announcements credited him with setting long-term direction and driving breakthroughs in high-end computing.³⁰ The day-to-day company had been run by someone else for a very long time.

The successor

That someone was Li Jun.

Born in 1968, Li Jun studied at Tsinghua University and joined the Beijing Dawning subsidiary in 1996, the year the company was founded. He rose to president in the early 2000s and held the general manager role for more than two decades. He is described in Chinese coverage as Li Guojie's student — not merely a subordinate but an intellectual heir within the Institute of Computing Technology lineage.⁵

Following the founder's resignation, a majority of the board designated Li Jun to serve as acting chairman.³⁰ The arrangement ran through 2025, across the announcement and collapse of the merger — meaning Li Jun negotiated, defended and unwound a 116-billion-yuan transaction while holding the chairman's authority on a temporary basis.

On May 6, 2026, the first meeting of Sugon's sixth board of directors formally elected him chairman for a three-year term, by a unanimous vote of seven in favour, none against, none abstaining. He was simultaneously reappointed as general manager.³¹ The new board included Xu Zhiwei and Yu Hualong as non-independent directors, Zheng Yongqin, Dai Shufen and Yin Xucheng as independent directors, and Cui Zijie as employee director, with Weng Qinan continuing as CFO and board secretary.³¹

The governance question

Combining chairman and general manager in one person is not a scandal. It is, however, a specific governance choice with specific consequences, and it is worth naming them precisely rather than gesturing at "concerns."

The chairman's function, in principle, includes supervising the general manager's execution. When the same individual holds both, that supervisory loop closes on itself. Board oversight then depends entirely on the independent directors' willingness to challenge a chairman who is also the executive they are meant to be evaluating.

In Sugon's specific circumstances, the case for combination is reasonable. The company had just been through eighteen months of founder succession and a failed mega-merger. Continuity of authority has real value in that situation, and Li Jun's thirty-year tenure means there is no realistic alternative candidate with comparable institutional knowledge.

The case against is that the combination arrived at the same moment as everything else — a new board, a post-merger strategic vacuum, and an 8-billion-yuan capital raise in progress. Concentration of authority is least costly when the decisions are routine. These are not routine decisions.

Li Jun's economic alignment is genuine and unusual for a state-lineage Chinese technology company. As of March 31, he held approximately 42.14 million shares — 2.88% of the company, making him the second-largest shareholder overall and the largest natural-person shareholder.⁵

His disclosed compensation has moved in the opposite direction from the share price. It peaked above 5 million yuan in 2022 and fell to roughly 2 million yuan in 2025.⁵ A chief executive of a company this size taking a pay cut of that magnitude while the business accelerates is a datapoint that cuts in management's favour, and it deserves saying.

Then there is the other datapoint.

In March 2026 — two months before formally taking the chairman's seat — Li Jun disclosed a plan to sell up to 6.32 million shares within three months, potentially realising roughly 570 million yuan. The stated reason was personal funding needs. The stock fell 8% on the day of the disclosure.⁵

Pressure-testing the insider sale

The sympathetic reading is straightforward and largely persuasive on the numbers. Li Jun has worked at this company for thirty years and taken most of his compensation in equity that he has largely not sold. The disclosed plan covers roughly 15% of his holding — leaving him with about 85%, still the largest individual shareholder. Chinese disclosure rules for this type of sale are strict and were followed. A 58-year-old executive diversifying a fraction of a position that constitutes nearly all of his net worth, after a share price run-up, is doing what any competent financial adviser would recommend.

The skeptical reading has force too. The timing is awkward in a way that a more careful adviser might have flagged. Disclosing a large personal sale eight weeks before your formal elevation to chairman, and shortly after the board approved an 8-billion-yuan convertible bond that will dilute existing shareholders, invites exactly the question it received. If management's position is that the market undervalues Sugon's core business — which is the implicit premise of the entire equity story — then the chief executive selling into strength is at minimum a discordant note.

The most useful framing is neither accusation nor dismissal. It is this: an insider sale is a weak signal in isolation and a meaningful one in aggregate. What matters is whether it repeats. A single disclosed, proportionate diversification by a long-tenured executive is normal. A pattern of sales at each price peak, particularly around capital raises, would be something else. Investors have one observation. They should note it and count the next one.

Because the capital raise it sat next to is the largest financing decision in the company's history.

↑ Back to Top

IX. Capital Allocation: Betting 8 Billion Yuan on AI Infrastructure

On the evening of February 9, 2026, Sugon's board approved a plan to raise up to 8 billion yuan through a public issue of convertible bonds.³²

To calibrate the ambition: 8 billion yuan is more than half the company's annual revenue, roughly four times its FY2025 net profit, and around twenty times what the 2014 IPO raised. This is not incremental capex. It is a bet-the-strategy financing.

Where the money goes

The proceeds split three ways.³²

3.5 billion yuan to an AI-oriented advanced computing cluster system project — developing what the company calls 超节点 super-node hardware systems, high-speed interconnect systems, and system-level foundational software stacks.

In plain language: as AI models grow, the constraint stops being how fast any single chip computes and becomes how fast thousands of chips can talk to each other. A "super-node" is an attempt to make a large collection of accelerators behave, from the software's point of view, more like one enormous processor — through very high bandwidth interconnect, shared memory addressing, and a software stack that hides the seams. This is the same problem Nvidia addresses with NVLink and rack-scale designs, and it is genuinely hard systems engineering. It is also, notably, exactly the kind of work where a company with thirty years of supercomputer interconnect experience should have an edge over a commodity server assembler.

2.5 billion yuan to next-generation high-performance AI training-and-inference all-in-one machines — a self-developed full-stack solution combining chip, algorithm and complete system.

2 billion yuan to domestically-produced advanced storage systems, aimed at massive multimodal data processing. Less glamorous, more defensible: AI training pipelines consume storage bandwidth voraciously, and storage is a segment where domestic-substitution procurement rules apply with the same force as they do to compute.

The projections, and how to read them

Management has published expected average annual sales for the two AI projects once mature: 5.69 billion yuan for the cluster system project and 3.28 billion yuan for the training-inference machine project.

Those figures deserve scrutiny rather than acceptance. Combined, they imply roughly 9 billion yuan of incremental annual revenue against a current base of 15 billion — a step-change of about 60% from two projects. For a company whose IT equipment segment grew 6.81% in FY2025,⁹ that is a very large claimed acceleration.

To be fair to the company, project-return projections in Chinese convertible bond prospectuses are a regulatory disclosure genre with well-understood conventions. They are prepared under specified methodologies, they represent a mature-state run rate rather than a near-term forecast, and management is not putting them forward as guidance. Treating them as a forecast would misread the document. Treating them as evidence of demand would be worse.

Sugon itself flagged the dilution mechanics with unusual candour in the plan, noting that new project benefits take time to materialise, and that if the convertible bonds convert to equity quickly before project returns are fully realised, per-share financial metrics including earnings per share could decline modestly in the short term.³² That is a real disclosure of a real risk, and it is more forthcoming than many issuers manage.

The credibility test

So is this raise sized to demand, or to narrative?

The evidence for demand is not trivial. Sugon's own first-half 2026 results showed revenue up 26.99% and adjusted profit up 50.01% — acceleration, not deceleration.¹⁵ Hygon's guided first-half revenue growth of 55–70% confirms that domestic AI silicon is genuinely selling in volume.²² The three delivered ten-thousand-card clusters demonstrate that Sugon can execute at cluster scale, not merely propose it.¹ The company cited a global AI computing market expected to reach $158.7 billion in 2025 as context.³²

The evidence for caution is equally real. Sugon has not disclosed order backlog, contracted pipeline, or customer commitments underpinning the projects — the disclosures that would convert a projection into a validated plan. The company is entering a segment where Huawei's ecosystem and Inspur both have more scale. And the cash flow picture behind FY2025's earnings — operating cash flow down 51.75%, receivables outgrowing revenue¹⁶ — sits uneasily beside a decision to deploy 8 billion yuan of new capital. A business whose profits are converting to cash at 0.61 times is a business that should arguably be tightening working capital before expanding capacity.

There is also the choice of instrument. Convertible bonds are cheaper than straight equity when a company's shares are expensive, and cheaper than straight debt when a company wants to avoid fixed interest burden. But the conversion feature means dilution arrives precisely if the strategy works and the shares rise — which is the scenario in which existing shareholders would least want it. Issuing convertibles when the market values your core business at or below zero, on the theory that the shares are undervalued, is an internally inconsistent position that no one has been asked to reconcile.

The most useful conclusion is a conditional one. If AI infrastructure demand in China sustains at anything like the 2026 first-half rate, this raise will look prescient and the projections merely optimistic rather than fanciful. If domestic AI capex normalises — and every capex cycle in history has eventually normalised — Sugon will have added capacity into a decelerating market, funded by instruments that dilute on the upside and burden the balance sheet on the downside.

That conditional is the crux of the investment debate, and it is time to set out both sides properly.

↑ Back to Top

X. Bull Case vs. Bear Case

The bull case

Start with the strongest version, stated without hedging.

The policy tailwind is structural, not cyclical. 信创 Xinchuang domestic substitution is not a marketing programme; it is embedded in Chinese government and SOE procurement requirements. A Sugon system built on Hygon silicon satisfies a mandate. Systems built on restricted foreign components do not. That is demand created by policy rather than won in competition, and it is largely insulated from further US export controls — indeed, tightening controls expand it. Being on the Entity List since 2019 has become, commercially, closer to an asset than a liability.

The Hygon stake is a real and possibly underappreciated asset. Whatever discount one applies for illiquidity and strategic entanglement, a 27.96% position in a company growing revenue 55–70% is not worth nothing, and through much of 2026 the market implied that Sugon's operating business was worth less than nothing. Even a partial normalisation of that stub value would be significant.

The core business is inflecting, and the inflection is in the right line. The FY2025 adjusted profit growth of 33.97% against headline growth of 13.87%, followed by first-half 2026 adjusted growth of 50.01%,³¹⁵ indicates operating improvement, not subsidy-driven improvement. The 75.34% growth in software, integration and services⁹ indicates mix moving toward higher-margin work. These are the specific metrics that would move first if the strategy were working.

The customer relationships are hard to replicate. Nearly fifty city cloud computing centres¹¹ and long-standing national supercomputing centre engagements represent installed base and institutional trust accumulated over more than a decade. Competitors can undercut Sugon on a server tender. They cannot easily displace it from a municipal cloud platform it built and operates.

Cluster-level delivery moves the business up the stack. Ten-thousand-card AI clusters carry revenue per engagement and integration complexity that server sales never will.¹ If Sugon converts its supercomputing engineering into a repeatable cluster business, its competitive comparison set changes from "server vendors" to "the handful of firms that can build national-scale AI infrastructure."

The bear case

Now the version a short seller would present, and much of it survives contact with the evidence.

Sugon is not a leader in its own market. It does not appear in the top six of China's x86 server market by revenue share, where Inspur holds 30.8% and even sixth-placed ZTE holds 6.5%.¹⁷ It is not among the top three in accelerated servers.¹⁷ Every strategic claim the company makes has to be evaluated against the fact that several competitors are executing the same strategy with more scale.

The differentiation is thinly evidenced beyond Hygon. Liquid cooling is a real engineering lead with no disclosed financial quantification. Cluster delivery is three units. Company gross margin around 27–30%¹ is not the margin profile of a firm with a defensible technical monopoly. If the primary advantage is privileged access to Hygon silicon, then the advantage belongs to Hygon and is rented by Sugon — and Hygon's own general manager has now publicly positioned it as a supplier to the entire domestic industry.²⁶

Earnings quality is a legitimate question. In FY2024, investment income and government subsidies together accounted for roughly 57% of net profit.¹² FY2025 improved that composition materially, but operating cash flow fell 51.75% while receivables grew 21.47% against revenue growth of 13.81%.¹⁶ Cash conversion at 0.61 in a growth year is a pattern that demands a second year of data before it can be dismissed as timing.

Customer concentration is a structural cyclicality. Government and SOE demand tracks state budgets and policy priorities. It is stable in aggregate and lumpy in timing, and it is subject to a risk that commercial demand is not: a change in policy emphasis. The internet sector, which drives nearly 69% of China's accelerated server purchasing,¹⁷ is not Sugon's customer base.

The merger collapse is an execution datapoint. Six months, substantial cost, and considerable shareholder expectation produced nothing, because a foreseeable market move broke a structure that was not built to absorb it.²⁷ The strategic rationale offered afterward — that separation was better all along²⁶ — does not survive the question of why it was not better in May.

Governance flags are yellow, not red, but they are yellow. Combined chairman and general manager roles,³¹ an insider sale plan disclosed eight weeks before the formal elevation,⁵ and an 8-billion-yuan raise with unvalidated demand projections,³² arriving together in a single eighteen-month window, is a lot of concentration of discretion at one moment.

The AI capex assumption is doing heavy lifting. The entire bull case on the core business rests on domestic AI infrastructure demand sustaining. If Chinese AI capex normalises — because model training economics shift, because compute efficiency improves faster than demand, or simply because the buildout front-loads — Sugon will have committed 8 billion yuan into a decelerating market.

Porter's Five Forces, applied

Barriers to entry: high. Capital intensity, government procurement qualification, security clearances, engineering depth in interconnect and cooling, and Xinchuang localisation requirements collectively make greenfield entry impractical. This protects Sugon — but it equally protects Inspur, H3C and Xfusion, all of whom are already inside the wall.

Rivalry: intense. Six or more credible vendors chase the same domestic tenders. Rivalry among established incumbents, not entry by newcomers, is what caps returns here — which is why margins sit in the high twenties despite the barriers.

Buyer power: high. Government and SOE buyers procure through competitive tender, know the vendor set precisely, and control payment timing. The receivables trend is buyer power showing up in the financial statements.

Supplier power: extreme, and this is the crux. Sugon's most important supplier is Hygon, which is also its largest asset and a company earning roughly double Sugon's gross margin on the same end-market demand. When a supplier captures twice the margin percentage of its customer, the economics of the value chain are telling you where the scarce resource sits. Sugon's 27.96% ownership stake partially converts supplier power into shareholder return — which is exactly why the stake matters strategically as well as financially, and exactly why the merger was proposed.

Substitutes: moderate and shifting. Within China, the substitute for a Sugon-Hygon system is a Huawei Ascend system, an Inspur system, or an imported platform where policy permits. Policy is what constrains the last option, and policy can change in either direction.

Hamilton Helmer's 7 Powers, applied

Run Sugon through Helmer's framework and the results are clarifying, mostly because of what is absent.

Scale economies: no. Sugon lacks the volume that would confer them, and its larger competitors have them instead.

Network economies: no. Servers do not become more valuable as more people buy them. Hygon's software ecosystem has something like this; Sugon does not.

Switching costs: partial. Real for city cloud platforms Sugon built and operates, and for supercomputing centres with deep integration. Weak for standalone server purchases.

Branding: minimal in economic terms. Sugon's name carries institutional credibility in China, but institutional buyers procure by specification and price, not brand premium.

Cornered resource: yes, and this is the strongest power on display. Sugon's privileged relationships with national supercomputing centres and its Chinese Academy of Sciences lineage constitute access that cannot be bought or competed away. Its ownership position in Hygon is a second cornered resource in the literal sense.

Process power: arguably present in liquid cooling and large-cluster integration — accumulated engineering practice that competitors would need years to replicate. Unproven in the margin line, which is where process power should eventually appear.

Counter-positioning: no, for Sugon. But emphatically yes for Hygon, which occupies a position no US chip designer can contest by regulation and no purely domestic architecture can contest on software compatibility.

The uncomfortable synthesis: the most powerful competitive position in this story is Hygon's, not Sugon's. Sugon's genuine power is a cornered resource in state-approved computing relationships — valuable, durable, and fundamentally limited in how far it can scale. That is a description of a good niche business with a great minority investment, which may well be a fine thing to own, but it is a different proposition from a compounding platform.

↑ Back to Top

XI. Durable Lessons & What to Watch

The business lesson

The most transferable insight from this story has nothing to do with servers.

In state-directed technology sectors, the spin-out can end up worth more than the parent — and the parent may have no practical mechanism to capture that value. Sugon incubated the relationship that produced Hygon, gave it a home, became its largest customer, and holds 27.96% of it. Hygon then grew into a company whose stake in Sugon's hands exceeds Sugon's own market value.

Why does this pattern recur in China specifically? Because the state research ecosystem generates technology faster than any single corporate vehicle can commercialise it, and because separate listing is the standard mechanism for funding commercialisation. The Institute of Computing Technology produced both Sugon and Loongson. Sugon's orbit produced Hygon. Each generation lists separately, and value accrues to whichever layer of the stack turns out to be scarce.

For investors, that produces a recurring, potentially tradeable structure: parent companies trading at or below the value of their listed stakes. The pattern is real. The trap is equally real — a discount can persist for years, and "cheap on sum-of-the-parts" is not a catalyst. What closes the gap is a corporate action, and the corporate action here was attempted and failed.

The investing lesson

Geopolitics can manufacture demand and valuation faster than management can execute against it.

The 2019 Entity List designation was intended to constrain Sugon. Within six years it had helped make Sugon's domestic-silicon positioning a procurement advantage. Then the AI cycle repriced both Sugon and Hygon so violently that a strategically sensible merger became financially unworkable in the space of six months.

The general form: when a policy tailwind arrives, share prices reprice in days, corporate structures reprice in quarters, and operating businesses reprice in years. Any transaction that requires those three clocks to agree is fragile. The Sugon-Hygon merger is a clean case study in how a deal that made sense on Monday can be impossible by Friday, not because the industrial logic failed, but because the arithmetic moved.

The KPIs that matter

Three metrics, and deliberately not more.

First: the growth rate of Sugon's software development, system integration and technical services segment, relative to IT equipment. This is the single cleanest test of whether Sugon is moving up the value stack or remaining a hardware assembler. FY2025 showed 75.34% versus 6.81%.⁹ If that divergence persists across FY2026 and FY2027, the mix-shift thesis is validated and margins should follow. If it reverts, Sugon is a box vendor with a good investment.

Second: Hygon's revenue growth and DCU adoption. This is the direct driver of the value of Sugon's largest asset and of a meaningful share of Sugon's reported earnings. Hygon's disclosure of its dual-chip product mix, and the trajectory of DCU revenue specifically, is the proxy for whether domestic AI accelerators are winning share against the alternatives.

Third: realised sales against the convertible bond projects, versus the disclosed projections. Management has published expected mature-state annual sales of 5.69 billion and 3.28 billion yuan for the two AI projects. Those numbers are now on the record. Tracking actual reported revenue from cluster systems and training-inference machines against them is the most direct available test of whether this management team's forward projections deserve credence — a test that will take years to complete and will be worth more than any single quarter's results.

A fourth item is worth watching without being a KPI: operating cash flow relative to net profit. One year of 0.61 conversion is a question. Two would be an answer.

The risk radar

Export control expansion. Further Entity List additions or tightening of controls on semiconductor manufacturing equipment, EDA software or advanced packaging could reach layers upstream of Hygon that neither company controls.

Execution risk on the capex programme. 8 billion yuan deployed into super-node systems, training-inference appliances and domestic storage, against competitors with more scale, with no disclosed order backlog.

Demand concentration. Government and SOE buyers with tender-driven procurement, lengthening payment cycles, and budget sensitivity to macro conditions.

Merger optionality, in both directions. A renewed Hygon combination would be a significant positive catalyst for the discount. Its continued absence leaves the stub-value question unresolved indefinitely.

Governance concentration. A single individual holding chairman and general manager authority, newly confirmed, having recently disclosed a personal share sale, presiding over the largest capital deployment in company history.

AI capex normalisation. The most underweighted risk, because it is the one the current numbers give no warning of. Infrastructure cycles turn after the orders stop, not before.

↑ Back to Top

XII. Epilogue

Where does this leave things in the late summer of 2026?

The operating business is doing better than it has in years. First-half revenue rose 26.99% and adjusted profit rose 50.01%,¹⁵ the fastest growth Sugon has posted in the AI era, with the mix shifting toward higher-margin integration work and the first ten-thousand-card clusters delivered. The 8-billion-yuan convertible bond programme is underway, aimed at super-node systems, training-inference appliances and domestic storage.³² A new chairman sits confirmed to a full three-year term, holding the general manager's title alongside it.³¹

And the central question remains exactly where it was on that April afternoon when the arithmetic first became absurd.

Hygon keeps growing faster than Sugon. Every quarter it does, the stake grows relative to the parent, and the implied value of Sugon's own operating business gets squeezed further toward — or below — zero. The merger that would have collapsed the two into one entity, and resolved the question by making it moot, was attempted and abandoned. Both managements say collaboration continues. Both have now publicly articulated reasons why independence might be preferable, reasons that were equally available before they proposed combining.

So an investor looking at 603019 is really being asked to choose between two readings of the same company.

The first: Sugon is a real business — thirty years of supercomputing engineering, a cooling technology lead, fifty city cloud platforms, national computing centre relationships, and an operating inflection that is finally showing up in adjusted earnings — that happens to also hold an extraordinarily valuable minority stake. On that reading, the market's negative stub value is an error, and the error will eventually correct as the core business demonstrates it can earn its keep independently.

The second: Sugon is a competent second-tier hardware company in a market dominated by larger rivals, generating gross margins half those of its own supplier, converting profit to cash at a deteriorating rate, whose most valuable characteristic is an ownership position it can neither sell nor consolidate. On that reading, the discount is not an error. It is the market pricing an operating business that has never proven it can generate returns above its cost of capital without help from the child it created.

The evidence, honestly assessed, does not yet settle it. FY2025 and the first half of 2026 lean toward the first reading. The market share data, margin structure and cash conversion lean toward the second. The next two years of segment disclosure — services growth versus hardware growth, cash conversion, and realised revenue from the projects an 8-billion-yuan bond issue is funding — will produce the evidence that resolves it.

Until then, Dawning Information Industry remains one of the more interesting puzzles in Chinese equities: a company that built China's first supercomputer, survived being sanctioned by the world's largest economy, and now finds its principal claim on investor attention resting not on what it makes, but on what it owns.

Dawning Information Industry Co., Ltd.

Table of Contents

Dawning Information Industry (中科曙光 Sugon): China's Supercomputer Pioneer and the Chip Stake Worth More Than the Company

I. Introduction & Episode Roadmap

II. Origins: From a State Research Lab to a Public Company

The 863 Program and the Institute of Computing Technology

From project to company

Why the DNA matters

III. What Sugon Actually Sells: The Core Business Today

The segment reality check

The products, in plain terms

Channel and margin structure

The financial arc, and what it actually shows

The counter-evidence

IV. The Industry: Who Sugon Actually Competes With

The three-tier structure

Where Sugon is genuinely different

The evidence test

V. The 2019 Entity List: The First Inflection Point

The part that hurt

The response

What it did not fix

VI. Hygon (海光信息): The Spin-Out That Became the Story

The THATIC structure

From CPU to DCU

The listing, and the mathematics of the stake

What this actually means for valuation

VII. The Merger That Almost Happened — and Didn't

The problem with a rising tide

December 9, 2025

What management said

Discipline or paralysis?

Is Merger 2.0 live?

VIII. Leadership in Transition: Founder Succession and Governance Signals

Who Li Guojie was

The successor

The governance question

Ownership, pay, and the share sale

Pressure-testing the insider sale

IX. Capital Allocation: Betting 8 Billion Yuan on AI Infrastructure

Where the money goes

The projections, and how to read them

The credibility test

X. Bull Case vs. Bear Case

The bull case

The bear case

Porter's Five Forces, applied

Hamilton Helmer's 7 Powers, applied

XI. Durable Lessons & What to Watch

The business lesson

The investing lesson

The KPIs that matter

The risk radar

XII. Epilogue

References