兆易创新 GigaDevice: The Chip Designer That Planted a Trillion-Yuan Seed

On the evening of July 9, 2026, 兆易创新 GigaDevice Semiconductor filed a routine-sounding document with the Shanghai Stock Exchange titled a half-year profit pre-increase announcement. Chinese listed companies file these when results will deviate sharply from the prior year. Most of them are dull. This one was not.

The company told investors that first-half revenue would come in at roughly RMB 11.5 billion, an increase of about 177% over the same period a year earlier, and that net profit attributable to shareholders would be approximately RMB 6.9 billion — up 1,099%.¹ For context, the comparable half-year profit in 2025 had been RMB 575.5 million.¹ A company that earned a little over half a billion yuan in six months was now earning nearly seven billion in six months. Strip out one-off gains and the adjusted figure was still about RMB 4.85 billion, up 791%.¹

The market reaction was immediate and euphoric. A-shares hit their daily limit. The Hong Kong line, which had only begun trading six months earlier, jumped more than 21%.⁴

Now fast-forward one month. On August 10, 2026, GigaDevice's Hong Kong shares closed at HK$501, down 5.1% on the day.² The stock's 52-week high was HK$1,248.² The Shanghai line closed at RMB 403.03, against a 52-week high of RMB 846.66.³ In roughly four weeks, a company that had just reported the best half-year in its twenty-one-year history had lost something close to half its market value.

That gap — between an operating result that is spectacular and a share price that is behaving as though something is badly wrong — is the whole story. It is not a contradiction. It is the memory semiconductor business doing exactly what the memory semiconductor business always does.

This article is about how a fabless chip designer from Beijing's 中关村 Zhongguancun district ended up at the center of the most violent commodity cycle in modern electronics, and about three decisions that got it there.

The first was a product decision. In 2013 GigaDevice launched a family of microcontrollers — the small, cheap, ubiquitous processors that run washing machines, electric meters, drones and car door handles — deliberately engineered to be drop-in compatible with the industry-standard parts made by STMicroelectronics. For years this was a respectable niche business. Then the global chip shortage of 2021 arrived, STMicroelectronics lead times blew out past a year, and thousands of Chinese engineers who had never seriously considered a domestic microcontroller suddenly had no other option. What began as a substitution of last resort became a franchise.

The second was a capital allocation decision that looks, in hindsight, like one of the highest-return-on-invested-capital moves in Chinese industrial history — and which, in a twist that generates most of the analytical confusion around this company, contributes almost nothing to GigaDevice's reported earnings. Between 2016 and 2019, GigaDevice and its founder helped seed a domestic DRAM manufacturer in the city of 合肥 Hefei. That company is now 长鑫存储 ChangXin Memory Technologies, listed on the STAR Market since July 27, 2026 with a debut-day market capitalisation above RMB 3.3 trillion.²⁶ Its chairman is GigaDevice's chairman. GigaDevice owns 1.80% of it.²⁷

The third is happening now. The AI buildout has consumed so much of the world's memory manufacturing capacity that ordinary, unglamorous, older-generation memory chips — the ones GigaDevice sells — have gone from commodity to scarce good. Prices for the products GigaDevice ships more than doubled in the first half of 2026.⁶ The company listed in Hong Kong on January 13, 2026, into the teeth of that boom.⁵

Three inflections, three very different kinds of luck-plus-preparation. What follows examines each, tests the durability claims management makes about them, and looks hard at the things a sceptical investor should be uncomfortable about: an income statement that has become partly a trading book, a chairman who sold RMB 4.4 billion of stock during the run, and a supply relationship with a related party that is scaling toward RMB 5.7 billion a year.⁴⁶

The company's fundamentals are genuinely better than they have ever been. Whether they are durably better is the open question, and the honest answer is that nobody — including management — knows yet.

↑ Back to Top

II. Origins: From Tsinghua to NOR Flash (1989–2016)

朱一明 Zhu Yiming was seventeen when he arrived at 清华大学 Tsinghua University in 1989 to study physics.⁸ He was from 盐城 Yancheng, a coastal city in northern 江苏 Jiangsu province that produced very few semiconductor entrepreneurs. To make money as a student he took programming work for companies in Zhongguancun, the Beijing technology district that was then a chaotic bazaar of computer shops and software startups.⁸

That job produced the observation that shaped everything after it. Working with Chinese firms, Zhu noticed that essentially every integrated circuit in the products he touched had been designed and manufactured somewhere else — overwhelmingly in the United States. China assembled electronics. China did not design the parts inside them.

He did what ambitious Chinese physics students did in the 1990s: he went abroad. After a master's at Tsinghua he continued to the State University of New York at Stony Brook, then moved into Silicon Valley industry, working at the network security company iPolicy Networks and, from 2001, at Monolithic System Technologies, where he led memory chip development projects.⁸

Memory is a peculiar corner of semiconductors, and understanding why matters for everything that follows. A processor is a differentiated product — Intel's chip is not AMD's chip. A memory chip is closer to a barrel of oil. Its function is standardised, its interfaces are standardised, and the buyer's only real questions are price, availability, and whether the part will behave identically to the one it replaces. That makes memory brutally cyclical: when supply is short, prices go vertical; when supply catches up, prices collapse below cash cost. Almost every company that has ever entered the memory business has eventually been destroyed by it. The survivors — Samsung 삼성전자, SK Hynix SK하이닉스, Micron — survived by being enormous.

At Monolithic System Technologies, Zhu had ideas about improving memory chip design. By late 2003 he had concluded that those ideas were not going to be adopted inside somebody else's company, and he began thinking about starting his own.⁸

In early April 2005 he incorporated GigaDevice in Silicon Valley, backed by Li Jun, a Tsinghua alumnus who was persuaded by Zhu's concept for a static random-access memory design.⁸ An angel investor, Hsun K Chou, put in $100,000.[^9] Then came the pivot that determined the company's nationality. Within weeks, Xue Jun of the Tsinghua Science Park venture arm argued that the American memory market was saturated and offered no room for a startup, and convinced Zhu to move the whole thing to Beijing.⁸ Another roughly $920,000 came from a group of Tsinghua alumni, and GigaDevice relocated in late April 2005, incubated inside the 启迪 Tus business incubator.[^9]⁹

So: a company founded in California in April 2005 and living in Beijing by May 2005, on under a million dollars, aiming at the most capital-intensive and cyclical segment in semiconductors. It should not have worked.

What made it work was a decision about where in memory to compete. GigaDevice did not go after DRAM, the high-volume main memory in every computer, where the incumbents' scale was insurmountable. It went after NOR flash.

NOR flash deserves a plain-English explanation, because it is still the core of this business. Think of a device's memory as a filing system. NAND flash — what is inside an SSD or a phone — is bulk storage: cheap per gigabyte, but you have to read it in big blocks, like fetching an entire filing cabinet drawer to read one page. NOR flash is different: the processor can read any individual byte directly, instantly, without loading a block first. That makes it useless for storing movies and essential for storing the code that boots a device. Every automotive electronic control unit, every router, every TV, every industrial controller needs a small amount of code that must be readable instantly and reliably for ten or fifteen years. That is NOR flash. It is small, it is old technology, and it is completely unavoidable.

In 2008, GigaDevice's team shipped what was reported as China's first domestically designed SPI NOR flash on a 180-nanometre process.[^9] By industry standards that node was already ancient — the leading edge was around 45 nanometres — which was precisely the point. Old process technology meant cheap, widely available foundry capacity and a design problem GigaDevice could actually solve.

The business model was fabless: GigaDevice designed chips and outsourced manufacturing, initially leaning on Chinese foundries including 中芯国际 SMIC, which counted GigaDevice as a significant NOR flash customer.¹⁰ No fabs meant no multi-billion-dollar capital bills, no depreciation cliff when the cycle turned, and — crucially for a small company — no obligation to fill a factory at any price. It also meant GigaDevice never controlled its own supply. That trade sits at the heart of both the bull and bear case two decades later.

The company restructured as a joint-stock company in December 2012 and listed on the Shanghai Stock Exchange on August 18, 2016, issuing 25 million shares at RMB 23.26 each after receiving China Securities Regulatory Commission approval the previous month.⁹ The shares limit-up almost immediately.⁹ It was a modest IPO by any global standard — roughly $88 million.⁹

But it gave Zhu Yiming something he had never had: an acquisition currency. And he already knew exactly what he wanted to buy.

↑ Back to Top

III. The First Inflection: A Failed Acquisition, and the DRAM Seed That Became CXMT (2016–2019)

The problem with owning the NOR flash niche is that the niche is small. Global NOR flash is a low-single-digit-billion-dollar market. DRAM, by contrast, is a market measured in the high tens of billions of dollars in ordinary years. If Zhu wanted to build what he had once described as a Chinese Samsung, NOR flash alone was never going to get him there.

There was one obvious shortcut. Integrated Silicon Solution, Inc. — ISSI — was an American memory company that had been taken private by a Chinese-backed consortium and parked inside a domestic holding vehicle, 北京矽成 Beijing Sigmastar. ISSI made niche DRAM: smaller-capacity, longer-lifecycle DRAM parts for industrial, automotive and networking customers, exactly the kind of DRAM the giants found insufficiently interesting. Buying it would have handed GigaDevice a DRAM franchise, a customer list, and two decades of design know-how in one transaction.

GigaDevice signed a letter of intent in December 2016 and disclosed a draft transaction report on April 18, 2017: it would acquire 100% of Beijing Sigmastar for RMB 6.5 billion, of which RMB 4.952 billion would be paid in newly issued shares and RMB 1.548 billion in cash.¹⁴ For a company that had raised under RMB 600 million at IPO months earlier, this was an audacious swing.

It fell apart in the most instructive way possible.

Three things went wrong at once. First, on February 17, 2017, the securities regulator tightened the rules on private placements by listed companies, capping new issuance at 20% of existing share capital and requiring at least eighteen months between raises — which made the share-based portion of the consideration far harder to execute.¹⁴ Second, the performance guarantees looked heroic: Beijing Sigmastar had earned RMB 15.6 million, RMB 57.2 million and RMB 75.5 million in the relevant recent periods, yet the deal's profit undertakings called for RMB 299 million, RMB 442 million and RMB 572 million across 2017 to 2019.¹⁴ A sceptical reader of that draft would have asked how a business quadruples profit on command.

Third, and fatally, one of ISSI's suppliers objected. GigaDevice said it only learned of the opposition in late July 2017.¹⁴ The logic was straightforward and, for China's semiconductor ambitions at that moment, humiliating: ISSI's main products were DRAM, and there was no foundry in mainland China capable of manufacturing DRAM on a contract basis.¹⁴ ISSI depended entirely on foreign manufacturing partners. If control passed to a mainland acquirer, the supplier reserved the right to stop supplying.

GigaDevice terminated the transaction in early August 2017.¹⁴

Here is where the story turns from a failed deal into the single most consequential decision in the company's history. Zhu Yiming's conclusion from the ISSI collapse was not "DRAM is not for us." It was: the reason we cannot buy a DRAM business is that China has no DRAM factory, therefore China needs a DRAM factory.

In October 2017 — roughly two months after the termination — GigaDevice announced a cooperation agreement with the Hefei municipal industrial investment arm to develop 19-nanometre-class DRAM on 12-inch wafers.¹⁵ The project carried a budget of approximately RMB 18 billion, with GigaDevice and the Hefei state investor funding it in a 1:4 ratio, implying a GigaDevice contribution of up to roughly RMB 3.6 billion.¹⁵ The operating entity was 合肥睿力集成 Hefei Ruili Integrated Circuit, which sat under the group that would become CXMT.¹⁵

Think about the asymmetry of that arrangement for a moment. A fabless designer with a market capitalisation in the low billions of yuan was nominally on the hook for a fifth of a project that would eventually require far more than RMB 18 billion — DRAM fabs always do. The Hefei government took the overwhelming majority of the risk, the land, and the balance sheet exposure. This was industrial policy in its purest Chinese form: a local government using its fiscal capacity to buy a seat at a strategic technology table, with an entrepreneur supplying credibility and technical leadership.

In July 2018, Zhu stepped away from day-to-day executive duties at GigaDevice to take the helm of the Hefei venture, then branded Innotron.¹⁶ He remained chairman of GigaDevice. From that moment, the founder of a listed company was simultaneously running an unlisted company that would become the listed company's largest supplier. That dual role has been the source of both GigaDevice's most important structural advantage and its most legitimate governance criticism, and it is worth flagging early rather than treating as a footnote.

GigaDevice's actual cash exposure to the venture ended up far smaller than the 2017 framework implied. In April 2019 it made a RMB 300 million convertible debt investment in the project, and stated at the time that it had no other investment in it and that the project entity operated independently.¹⁵ The economics of the funding arrangement had shifted decisively toward the state.

The factory worked. In September 2019, CXMT announced the start of production of a 10-nanometre-class first-generation 8Gb DDR4 part, with a phase-one design capacity of 120,000 wafers per month.¹⁵ For China, this was a genuine technological milestone: the country's first credible domestic DRAM manufacturing at scale, built on intellectual property foundations that traced back to the German memory maker Qimonda.

Seven years later, CXMT priced its Shanghai IPO at RMB 8.66 per share on July 27, 2026, selling 6.688 billion shares to raise RMB 57.92 billion — around $8.6 billion, the largest offering ever on the STAR Market and China's second-largest IPO since Agricultural Bank of China in 2010.²⁶ The shares opened at RMB 49.50 and traded as high as a 531% gain intraday, taking the company's market capitalisation past RMB 3.3 trillion and briefly making it the most valuable company listed on mainland exchanges — ahead of Industrial and Commercial Bank of China.²⁶ First-quarter 2026 revenue had been RMB 50.8 billion, up over 700% year on year on AI server memory demand.²⁶

GigaDevice's residual stake in all of that is 1.80%.²⁷

An investor's instinct on reading that sentence is to reach for a calculator. Resist it for now — Section IX explains in detail why that stake produces almost nothing on GigaDevice's income statement, and why the company itself went out of its way in May 2026 to tell shareholders to stop over-interpreting it.²⁷

The strategically important legacy of the Hefei bet is not the equity. It is the pipe. Having a friendly domestic DRAM fab, run by your own chairman, is what allows a fabless company to sell DRAM at all in a world where every wafer is spoken for. That relationship is the foundation of the third leg of the business, and it is why the failed ISSI acquisition may have been the most valuable deal GigaDevice never did.

While the DRAM saga unfolded in Hefei, however, a quieter product line back in Beijing was about to have a very good war.

↑ Back to Top

IV. The Second Inflection: GD32, STM32, and the 2021 Shortage Windfall

To understand what happened to GigaDevice in 2021, you first have to understand what a microcontroller actually is, and why substituting one is so hard.

A microcontroller — an MCU — is a complete tiny computer on a single chip: a processor core, a bit of memory, and the input/output circuitry to talk to motors, sensors and switches. It is the least glamorous product in semiconductors and one of the most profitable, because it is embedded in things that stay in production for a decade. Your car has dozens. Your air conditioner has one. Your electricity meter has one.

The critical commercial fact about MCUs is that customers almost never switch. An engineer who has written firmware for a particular MCU family has learned its quirks, bought its development tools, built test fixtures around it, and passed safety certifications on it. Changing MCU vendors means rewriting software, redesigning the circuit board, and re-certifying the product. It is closer to changing the foundations of a house than swapping a supplier of screws.

STMicroelectronics understood this better than anyone. Its STM32 family, built on Arm Cortex-M processor cores, became the default MCU for an entire generation of engineers worldwide — a de facto industry standard with an enormous ecosystem of code libraries, tutorials, and trained developers.

GigaDevice's move, beginning with its first GD32 parts in 2013, was to attack that lock-in from the only angle available: compatibility.¹¹ The GD32 devices were engineered so that the flagship parts were pin-to-pin equivalents of their STM32 counterparts — same package, same pin functions, same physical footprint — and largely software compatible. An engineer could, in principle, lift an STM32 off a board, drop a GD32 into the identical solder pads, and ship.

This was a deliberate act of judo. Rather than trying to build an ecosystem, GigaDevice borrowed one. Every hour ST spent training an engineer on STM32 was an hour that partially transferred to GD32.

For seven years, this earned GigaDevice a respectable but unremarkable position. The company kept broadening the line — Cortex-M3, M4, M23, M33 and M7 cores, plus, in 2019, what it claimed as the world's first general-purpose 32-bit MCU built on the open RISC-V architecture.¹¹ By late 2021 the portfolio spanned 28 series and roughly 400 part numbers.¹¹ But compatibility alone does not move a conservative engineering buyer. Why would you take the career risk of a second-source part when the original is available?

In 2021, the original stopped being available.

The global semiconductor shortage of 2020–2022 hit MCUs harder than almost any other category, because MCUs are made on mature process nodes in fabs that nobody had been building. Automotive demand snapped back faster than anyone forecast, consumer electronics demand exploded with remote work, and lead times for Western MCUs stretched past a year. Purchasing managers at Chinese appliance makers, industrial equipment firms and consumer electronics companies faced a binary choice: qualify a domestic alternative, or stop the production line.

The results were immediate. GigaDevice's MCU revenue in the first half of 2021 reached RMB 797 million, up 222.1% year on year, after already growing 70.1% in 2020 on a 75.9% increase in unit volume.¹¹ By October 2021 the company reported cumulative GD32 shipments of over 800 million units across more than 20,000 customers.¹¹

At group level, 2021 revenue rose to RMB 8.51 billion with net profit of RMB 2.34 billion at a gross margin of 46.0%.¹⁸ Against 2020, that was roughly a doubling of revenue and a tripling of profit.

The analytically important question is not what happened in 2021. It is what happened after, because that is what distinguishes a windfall from a franchise.

Here the evidence is genuinely supportive of the durability claim, with an important caveat. Once a Chinese engineer had qualified a GD32 into a design in 2021, the switching-cost logic that had protected ST began protecting GigaDevice instead. The design win persisted through the product's life. By 2024, Frost & Sullivan ranked GigaDevice eighth globally in MCUs and first in mainland China, and MCUs contributed 23% of company revenue.¹² In 2025 the MCU line generated about RMB 1.91 billion, roughly 20.8% of the group.⁶ That is a smaller share than 2024 only because memory revenue grew faster — the MCU business itself kept expanding.

The caveat is that GigaDevice's MCU position remains overwhelmingly domestic and overwhelmingly at the value end. Being first in China and eighth in the world simultaneously tells you the global MCU market is dominated by companies — STMicroelectronics, Infineon, NXP, Renesas, Microchip — whose automotive and industrial franchises GigaDevice has not yet meaningfully penetrated. And the compatibility strategy that opened the door also caps the ceiling: a company defined as the affordable equivalent of somebody else's standard does not easily command premium pricing.

GigaDevice has been working the obvious answer, which is to climb into applications where qualification barriers are highest. Its GD32A5xx series was developed against the AEC-Q100 Grade 1 automotive reliability standard, and the second-generation GD32A7 family uses a high-performance Cortex-M7 core aimed at body domain control, telematics, lighting and battery management.³⁷ Its automotive NOR flash has been certified to ISO 26262 ASIL D functional safety.³⁷ These are slow-burn markets — a design win in a car takes three to five years to reach volume — but they are the only markets where an MCU business becomes structurally defensible rather than merely price-competitive.

There is a live test of pricing power available right now. In the second quarter of 2026, GigaDevice raised MCU prices, attributing the increase to wafer and packaging cost inflation, and reported record first-half MCU shipments alongside expansion into newer applications including 3D printing, robotics, optical modules and server power supplies.²¹ Raising price while raising volume is the textbook signature of genuine demand rather than share-buying. Whether it survives the next downturn is the thing to watch.

The MCU business, then, is the closest thing GigaDevice has to a non-cyclical asset: real switching costs, a genuine ecosystem borrowed from a bigger rival, and a demonstrated ability to hold share after the emergency that created it passed. It also turned out to be exactly the wrong kind of business to be leveraged to when the cycle inverted.

↑ Back to Top

V. The Hangover: What 2022–2024 Revealed About the Business

If 2021 was the party, 2023 was the morning after, and it was ugly enough to be worth dwelling on — because it is the single best evidence base for what happens to this company when conditions normalise.

The mechanics of a semiconductor bust are simple and always the same. During the shortage, every customer double-ordered. Distributors stockpiled. Contract manufacturers built buffer inventory. When demand finally softened, all of that phantom demand vanished at once, and the industry discovered it had built a mountain of chips nobody needed at prices nobody would pay.

GigaDevice's numbers tell the story with unusual clarity. Revenue peaked at RMB 8.51 billion in 2021, held roughly flat at RMB 8.13 billion in 2022, then fell to RMB 5.76 billion in 2023 — a decline of about 29% in a single year.¹⁸ Net profit went from RMB 2.34 billion in 2021, to RMB 2.05 billion in 2022, to RMB 161 million in 2023.¹⁸ That is a 93% collapse in earnings on a 32% decline in revenue from the peak.

Gross margin tells you why. It ran at 46.0% in 2021, 45.5% in 2022, and 30.3% in 2023.¹⁸ Fifteen points of gross margin evaporated in twelve months.

This is the fundamental economic truth about GigaDevice that no amount of narrative about domestic substitution or AI demand should be allowed to obscure: the company's cost base is largely fixed in the short run — wafer purchase commitments, R&D headcount, mask sets — while its selling prices are set by a commodity market it does not control. That produces enormous operating leverage in both directions. When prices rise, almost every incremental yuan of revenue drops to the bottom line. When prices fall, the same thing happens in reverse.

Two things about the downturn deserve credit, though, and they matter for assessing management.

First, the company did not stop investing. R&D continued through the trough, which is why the product portfolio emerging in 2025 and 2026 was broader than the one that entered 2022. GigaDevice had used its 2019 acquisition of the sensor company 思立微 Silead, valued at RMB 1.7 billion and completed after regulatory approvals in 2019, to add touch and fingerprint sensing to the portfolio.³⁸ By 2024, sensors contributed 6.1% of revenue.¹² More recently it has pushed into analog and power management, including a joint laboratory with Navitas Semiconductor announced in October 2025 to work on high-efficiency digital power.³⁶ None of these are large businesses yet. Their strategic function is to give the company more shots on goal and to reduce the share of revenue that is pure commodity memory.

Second, the balance sheet stayed clean. Because GigaDevice owns no fabs, the downturn did not produce the depreciation death-spiral that flattens integrated memory makers. The company entered the 2026 upcycle with roughly RMB 15 billion of net cash and negligible debt risk.⁶ For a business in the most cyclical category in electronics, an unlevered balance sheet is not a luxury — it is the thing that determines whether you survive to see the next upturn.

The recovery began in 2024 and was, at first, modest: revenue of RMB 7.36 billion and net profit of RMB 1.10 billion, with gross margin recovering to 35.7%.¹⁸ Then 2025 delivered revenue of RMB 9.20 billion, up 25.1%, and net profit of RMB 1.65 billion, up 49.5%, at a 39.1% gross margin.¹⁸ The company declared a dividend of RMB 525 million, or RMB 7.50 per ten shares, representing 31.9% of net profit — a payout ratio that is respectable but not remarkable, and consistent with a company that intends to keep reinvesting.²⁸

The pattern across five years is worth stating plainly. GigaDevice has now been through a full memory cycle as a public company. It made very good money at the top, nearly no money at the bottom, and lost no ground competitively along the way. Its 2025 revenue exceeded its 2021 peak, which is more than several memory peers can say. That is the evidence for the claim that this is a share-gaining business inside a cyclical industry rather than merely a cyclical business.

What it is not evidence for is any claim that the cycle has been tamed. The 2023 trough happened with the domestic-substitution tailwind fully in force. Substitution moved GigaDevice's baseline up; it did not flatten the curve.

And then, in late 2025, the curve began pointing up more steeply than anyone in the industry had ever seen.

↑ Back to Top

VI. The Third Inflection: The AI Memory Supercycle and an Eleven-Fold Profit

The 2026 memory shortage was not caused by a shortage of memory factories. It was caused by artificial intelligence eating them.

The mechanism is worth explaining carefully because it is the single most important thing driving GigaDevice's current earnings, and it is not intuitive.

AI accelerators need extraordinary memory bandwidth — they must feed data to thousands of processing cores simultaneously. The solution the industry adopted is High Bandwidth Memory, or HBM, which stacks DRAM dies vertically and connects them with thousands of tiny vertical wires. HBM works, but it is enormously wasteful of manufacturing capacity: a stack consumes several dies' worth of silicon and yields poorly, so producing one unit of HBM consumes roughly two to three times the wafer capacity of an equivalent amount of conventional DRAM.

Because HBM sells at multiples of conventional DRAM pricing, every rational memory manufacturer redirected capacity toward it. HBM's share of global DRAM wafer capacity rose from about 2% in 2020 to roughly 25% in 2026.¹³ The share of a server's bill of materials accounted for by DRAM went from around 50% in 2023 to about 75% in 2026.¹³

The consequence was a supply vacuum in everything else. Conventional DRAM contract prices rose 90–95% quarter on quarter in the first quarter of 2026, with PC DRAM up 105–110%.²⁰ DDR5 chip-level pricing went from $6.84 per gigabyte in September 2025 to $27.20 per gigabyte by December 2025.²⁰ Older DDR4 parts, which the majors were actively discontinuing, in some configurations traded above newer DDR5 — an inversion of the normal price hierarchy that tells you how completely supply had broken.²⁰ Roughly 70–80% of the world's advanced memory capacity became locked into AI applications.²⁰

Now overlay GigaDevice's product mix on that picture. The company sells NOR flash, SLC NAND, niche DRAM and MCUs — every one of them a mature-node product that competes for exactly the capacity the giants were abandoning. TrendForce data showed NOR flash and SLC NAND contract prices rising more than 100% in the first half of 2026.⁶

The results followed mechanically. First-quarter 2026 revenue reached RMB 4.19 billion with net profit of RMB 1.46 billion.⁶ Then came the half-year pre-announcement, and doing the arithmetic on it is illuminating: if the first half produced roughly RMB 11.5 billion of revenue and the first quarter produced RMB 4.19 billion, the second quarter alone generated around RMB 7.3 billion — nearly 75% more than the quarter before it. Sequential acceleration of that magnitude in a hardware business is almost always price, not volume.

Management's own explanation, given in the profit announcement, was exactly that: storage chip products experienced simultaneous increases in both volume and price because of supply constraints, while microcontrollers benefited from industrial, consumer and automotive demand.¹

So what does this actually tell an investor?

It tells you that GigaDevice's earnings power is a leveraged call option on legacy memory pricing. That is not a criticism; it is a description. In a shortage this severe, a fabless designer with existing customer qualifications and existing wafer allocations is one of the most operationally geared entities in the entire supply chain, because its incremental cost of selling a chip at triple the price is close to zero.

But there is a second-order point that cuts the other way, and it is the one most commentary misses. GigaDevice does not own fabs. In a capacity shortage, the scarce asset is the wafer, and the owner of the wafer captures the rent. GigaDevice buys wafers into the same inflation. The company's own decision to raise MCU prices citing wafer and packaging cost inflation is direct evidence that input costs are climbing.²¹ Its blended gross margin will therefore be determined by whether its selling prices outrun its foundry prices — a question about relative bargaining power, not about the shortage itself.

The evidence so far says GigaDevice is winning that contest. Specialty memory gross margin ran at 42.8% in 2025 and expanded sequentially through the upcycle rather than compressing.⁶ Expanding margin during input-cost inflation is a meaningful signal: it suggests the embedded memory customer, facing firmware compatibility issues, long product lifecycles and requalification costs, absorbs price increases rather than shopping around.⁶ That is a real and specific mechanism, not a slogan.

On the July 2026 investor call, management laid out an aggressive forward view. It said it expects to become the world's largest NOR flash supplier within two to three years, citing demand from edge AI, automotive electronics and AI servers, with the strongest gains in mid-to-large capacity parts.²¹ The product evidence behind that ambition is at least visible: in November 2025 the company launched its GD25NX series of xSPI NOR flash with a dual-voltage design, aimed squarely at higher-capacity, higher-bandwidth applications rather than the small commodity parts that built the franchise.³⁵ On SLC NAND — the small, simple, highly reliable flash used in industrial and networking gear — management said supply is severely constrained because major producers have shifted capacity to 3D NAND for AI markets, and that more than half of suppliers have exited or plan to exit the category, pushing prices to historical highs.²¹ Looking to 2027, management said it anticipates limited capacity additions and sustained elevated pricing across DRAM and flash.²¹

That 2027 comment deserves scrutiny, because it is not obviously consistent with what the company said earlier. Commentary around the Hong Kong listing period framed the outlook as niche memory prices continuing to rise through 2026 but entering a period of high-level volatility from 2027 onward.¹³ "High-level volatility" and "sustained elevated pricing" are not the same forecast. The shift in tone happened across roughly six months during which the stock quadrupled. An investor tracking management credibility should note that the narrative got more confident as the share price got higher, which is the direction of drift one should be most sceptical about.

Meanwhile, the market has already begun voting the other way. In late July 2026 a global chip selloff wiped more than $1 trillion from the twenty most valuable semiconductor stocks in a matter of days, with SK Hynix, Samsung and Micron losing $176 billion, $173 billion and $113 billion of market value respectively.³¹ Memory names collectively entered a bear market on fears that AI infrastructure spending is peaking faster than expected.³² GigaDevice fell with them.

The stock is now telling you what the earnings cannot: that the market has moved on from asking how big 2026 will be, to asking what 2027 looks like on the other side.

Before that question can be answered, though, there is a piece of the 2026 story that has nothing to do with memory chips at all.

↑ Back to Top

VII. Capital Markets: The Hong Kong Listing and the Anatomy of a Frenzy

On January 13, 2026, GigaDevice's H-shares opened in Hong Kong at HK$235 against an offer price of HK$162 — a 45% first-trade gain.⁵ They closed the session at HK$222.80, up 37.5%, valuing the company at roughly HK$155.2 billion, or about $19.9 billion.⁵ The Shanghai line, meanwhile, barely moved, rising 0.6% to RMB 263.50.⁵

The offering itself was small: approximately 28.92 million H-shares raising HK$4.68 billion gross, about $600 million, with net proceeds of roughly HK$4.611 billion.¹²²³ Against a company that already carried a market value in the tens of billions of dollars, this was a sliver — under 5% of the share count.

The demand was not small. The Hong Kong public tranche was oversubscribed 542.22 times; the international tranche 18.52 times.²³ Retail investors committed HK$468 million of their own money against HK$193.7 billion of margin financing.⁵ In grey-market trading before official listing, the shares changed hands at up to a 40% premium on turnover exceeding HK$600 million.²³ Cornerstone investors included China Pinnacle Equity Management, Yunfeng Capital, 小米 Xiaomi, TCL Industries and 华勤技术 Huaqin Technology.⁵

Two of those names are worth pausing on. Xiaomi and Huaqin are customers or potential customers — Xiaomi makes phones and an expanding range of connected devices, Huaqin is one of the world's largest contract design manufacturers of smartphones and laptops. Cornerstone investments by customers are not merely capital; they are a public statement about supply relationships in a market where allocation is the scarcest commodity.

The obvious question is why a profitable, cash-rich, already-listed company with roughly RMB 15 billion of net cash needed to raise $600 million at all.

The stated use of proceeds was research and development plus strategic and industry investments, including potential acquisitions.⁵ That is a broad mandate, and the honest reading is that money was not the primary motivation.

Three other motivations are more plausible. First, offshore currency: an H-share listing gives a Chinese company Hong Kong dollars and a Hong Kong-listed acquisition currency, useful for buying assets or building operations outside the mainland — which matters for a company whose customers increasingly need to demonstrate supply chains that are not exclusively China-domiciled. Second, index and investor access: a Hong Kong line brings in international institutional money that cannot easily hold A-shares. Third, and least discussed, optics of internationalisation at a moment when the entire Chinese semiconductor complex was rushing to Hong Kong. The listing was executed with international legal advice and marketed as a step in a global growth strategy.²⁵²⁴

The subsequent trading history is where the capital-markets lesson lives. That HK$162 offer price now looks like a gift: the shares reached HK$1,248 at their peak, a rise of more than 670% from the offer in under seven months.² They then gave back most of it, trading at HK$501 on August 10, 2026 — still triple the IPO price, but roughly 60% below the high.² The 200-day moving average sits near HK$518, meaning the stock has round-tripped an entire year of gains in weeks.²

The A-share line behaved similarly but less violently, ranging from RMB 116.48 to RMB 846.66 over twelve months and trading at RMB 403.03 on August 10.³ The H-shares' wider swings are a structural artefact: with fewer than 29 million H-shares outstanding against roughly 700 million total shares, the Hong Kong float is tiny, and tiny floats amplify everything.

The valuation arithmetic on that August 10 price is instructive precisely because it is so unstable. The A-shares carried a trailing price-to-earnings ratio of about 93.5 against a market capitalisation of roughly RMB 284 billion.³ The Hong Kong line showed a trailing multiple near 101 and a forward multiple around 45.² At the early-July peak, the trailing multiple had been approximately 157 times, on an enterprise value of roughly 42 times sales.⁶

Those forward and trailing numbers differ by more than a factor of two, and the entire investment debate lives in that gap. A trailing multiple of 100 on peak-cycle earnings would be extraordinary. A forward multiple in the mid-40s on a company doubling revenue is defensible. Which one is right depends entirely on whether 2026's earnings represent a new base or a spike. Sell-side work published during the run reframed the stock at roughly 46 times 2026 estimated earnings and 25.5 times 2027 estimates.⁶ That reframing is only as good as the 2027 estimate, and the 2027 estimate is a forecast about memory pricing eighteen months out — a quantity that has never once been forecast accurately by anyone.

For a peer check: Winbond 華邦電子, the Taiwanese company closest to GigaDevice in product mix, traded around 55 times earnings and 7.7 times sales; Macronix 旺宏電子 at about 8.7 times sales.⁶ GigaDevice's multiple sat well above both even after the correction, which tells you the market is paying for something beyond the current product portfolio.

Some of that premium is the third leg of the business, which is finally starting to show up in the numbers.

↑ Back to Top

VIII. Niche DRAM and Customized Memory: The Third Leg, Still Building

The DRAM ambition never actually died after the Hefei venture was handed over to the state. It simply went quiet and underfunded for several years. In April 2020, GigaDevice received regulatory clearance for a private placement of up to roughly RMB 4.3 billion, with the bulk earmarked for developing and industrialising its own DRAM chips.¹⁷ Its first self-branded DRAM part followed — a modest, small-capacity device aimed at consumer and industrial applications rather than at the mainstream computing market where the giants live.¹⁹ For four years, it was a rounding error in the accounts.

Then, in July 2024, GigaDevice quietly established a wholly-owned subsidiary in the Hengqin free trade zone next to Macau, named 珠海横琴芯存半导体 Zhuhai Hengqin Xincun Semiconductor.²² Its scope was integrated circuit design and sales. Its purpose was DRAM.

The company has been feeding it steadily. RMB 800 million went in during September 2024. Another RMB 500 million followed in January 2026, with funds subsequently distributed to affiliated entities in Hefei and Xi'an. In July 2026 the board approved a third injection of RMB 500 million — RMB 50 million to registered capital and RMB 450 million to capital reserves — lifting the subsidiary's registered capital from RMB 150 million to RMB 200 million.²²

The unaudited first-quarter 2026 figures for that subsidiary are the reason it deserves a section of its own. Zhuhai Xincun generated revenue of approximately RMB 1.106 billion and net profit of approximately RMB 614 million in the three months to March 31, 2026, against total assets of RMB 3.237 billion and net assets of RMB 2.237 billion.²²

Hold those numbers against the group. GigaDevice's consolidated first-quarter revenue was RMB 4.19 billion and net profit RMB 1.46 billion.⁶ The DRAM subsidiary therefore accounted for roughly a quarter of group revenue and — strikingly — something on the order of two-fifths of group net profit in that quarter.

That is a remarkable result from a business that did not exist two years earlier. It also demands a health warning: those are subsidiary-level unaudited figures that may not be directly comparable to consolidated group accounting after intercompany eliminations, and a single quarter at the peak of the sharpest DRAM price spike in history is not a run rate. The honest framing is that niche DRAM has gone from a rounding error to a material contributor with startling speed, in conditions that flattered it enormously.

What is GigaDevice actually selling here? Two distinct things, and the distinction matters.

The first is conventional niche DRAM. The company sources DRAM wafers from CXMT and sells finished parts under its own brand into industrial, networking, consumer and automotive applications. Its historic strength is DDR3 — an old standard that the global giants have almost entirely abandoned but which remains embedded in enormous installed bases of equipment. Management has said the product mix is now shifting from DDR3 toward higher value-added categories: LPDDR4 and a new-process 48Gb-class product completed tape-out early in 2026, with mass production targeted for the second half of the year, and small-capacity DDR4 also entering production in the second half.²¹ DDR5 small-capacity development is underway, with automotive named as a target market.²¹ Frost & Sullivan ranked GigaDevice seventh globally in niche DRAM on 2024 sales.⁷

The second is customized memory, and this is the genuinely interesting option. Rather than selling a standard part, GigaDevice works with a customer to specify a memory device tailored to a particular system — different capacity points, different interfaces, different packaging, sometimes memory and logic combined. Management reported breakthrough progress with key customers and projects in AI smartphones, AI PCs and robotics, with revenue contributions expected in the second half of 2026 and the business framed as a medium-term growth driver.²¹

Why does customization matter strategically? Because it is the one path out of commodity economics. A standard DDR4 part is interchangeable by definition; a customized memory device co-designed with a customer's system is not. It carries a design win, a switching cost, and a margin structure closer to a specialty component than a commodity. It is, in effect, an attempt to apply the MCU business's stickiness to the memory business.

It is also, at this moment, mostly a promise. Revenue has not yet been reported. The customers have not been named. The claim that AI smartphones and robotics will need bespoke memory is plausible but unproven at scale.

The enabling asset behind both product lines is the CXMT supply relationship, and its trajectory is the single most important operating disclosure in this company's filings. Related-party procurement from CXMT was RMB 764 million in 2023 and RMB 1.18 billion in 2025; the estimate for 2026 is approximately RMB 5.7 billion.⁷⁶ Management has described this as providing robust capacity assurance.²¹

That is roughly a sevenfold increase in two years, and it is the mechanism by which GigaDevice converts its founder's Hefei bet into revenue. In a world where every DRAM wafer on Earth is allocated and the global majors are actively discontinuing the products GigaDevice sells, guaranteed access to a domestic fab is not a nice-to-have. It is the entire business.

It is also, viewed from a different angle, a concentration risk and a governance question of the first order. Which brings us to the parts of this story that the bull case tends to skip.

↑ Back to Top

IX. Myth vs Reality: Four Consensus Stories, Stress-Tested

Every hot stock accumulates a folklore. GigaDevice's is unusually rich, because the company sits at the intersection of three separate narratives — Chinese semiconductor self-sufficiency, the AI memory boom, and a founder who built a trillion-yuan company on the side. Four claims circulate widely enough to be worth testing directly.

Myth one: the CXMT listing made GigaDevice shareholders rich.

This is the most persistent and the most wrong. When CXMT's market value crossed RMB 3 trillion, a natural piece of arithmetic suggested GigaDevice's 1.80% stake was worth tens of billions of yuan against a carrying value of roughly RMB 2.86 billion at the end of 2025.⁶ Enormous, surely, relative to a company whose entire market capitalisation was a few hundred billion.

The accounting says otherwise, and GigaDevice itself said so first. In May 2026, responding to abnormal share price movements after a run of more than 20% over three consecutive trading days, the company issued a clarification: its association with CXMT is limited, its stake is 1.80%, it cannot significantly influence CXMT's operations, and — the critical sentence — the investment is classified as an "Other Equity Instruments Investment," so changes in CXMT's market value arising from the IPO will not affect GigaDevice's net profit.²⁷

Here is what that classification means in plain terms. Under Chinese accounting standards, as under IFRS, an entity may make an irrevocable election to designate a non-trading equity investment at fair value through other comprehensive income. When it does, the investment is remarked to fair value on the balance sheet — so shareholders' equity and book value rise — but the gains flow into a reserve within equity and never pass through the income statement. Unlike debt instruments, these gains are never recycled to profit and loss, not even on disposal. Under the alternative treatment, fair value through profit or loss, every price move would hit reported earnings.

GigaDevice chose the former. The practical consequences are threefold. Reported earnings per share are unaffected by CXMT's share price, in either direction. Book value per share rises, potentially by a very large amount. And no cash arrives unless the stake is sold — which, given that GigaDevice's chairman also chairs CXMT and that the relationship secures GigaDevice's DRAM supply, would be a strategically odd thing to do.

So the stake is real, it is valuable, and it is almost entirely economically inert from an earnings standpoint. A shareholder should think of it as a large, illiquid, non-cash-generating asset on the balance sheet whose main function is to cement a supply relationship.

Myth two: the eleven-fold profit surge is all operating leverage.

It is not. Roughly RMB 2 billion of the first-half profit came from non-recurring gains on securities — an amount that by itself exceeded GigaDevice's entire net profit for full-year 2025.⁴

The source was pre-IPO strategic placements in Chinese listings. GigaDevice put RMB 53.27 million into 650,600 shares of Lianxun Instruments at its IPO price of RMB 81.88; by June 30, 2026 those shares had appreciated more than twenty-sevenfold to RMB 2,322.33, producing a paper gain exceeding RMB 1.45 billion.⁴ A second position, in Zhenbao Technology, which listed on June 24 at RMB 44.56, appreciated more than fourteenfold by month-end for a paper gain over RMB 600 million.⁴

Three observations follow. First, these are marks, not cash — strategic placement shares carry lock-ups, and a twenty-seven-fold gain in a matter of weeks is precisely the kind of mark that can reverse violently. Second, unlike the CXMT holding, these positions clearly do run through the income statement, which is why the headline profit and the adjusted profit diverge so sharply. Third, and most importantly for anyone modelling this company: the underlying operating business still delivered adjusted profit growth of 791%.¹ The core story is real. The headline number simply overstates it by about 30%.

But it raises a question a sceptical investor should ask out loud: why is a semiconductor designer running what amounts to a proprietary equity book at all? Strategic investments in ecosystem companies are common and defensible. Booking a gain larger than your entire prior-year profit from marking two positions is a different thing, and it makes earnings quality materially harder to assess.

Myth three: as a fabless company, GigaDevice is a pure beneficiary of the shortage.

Partly true, partly backwards. GigaDevice is a wafer buyer competing for the same scarce capacity as everyone else. The company's own decision to lift MCU prices explicitly on wafer and packaging cost inflation confirms input costs are rising.²¹ What GigaDevice actually has is access — long foundry relationships and, in DRAM, a related-party fab — rather than immunity. Access is genuinely valuable in a shortage. It is not the same as owning the capacity, and in the next downturn it will not protect pricing.

Myth four: domestic substitution guarantees the franchise.

Chinese self-sufficiency policy has unquestionably been a tailwind, and the Hong Kong listing was received in that frame.⁵ But substitution cuts in more than one direction. The same policy that funds GigaDevice's customers to buy domestic funds its domestic competitors too — and China's NOR flash and MCU markets have attracted a long list of local entrants. More subtly, CXMT itself sells DRAM. A fab that supplies GigaDevice's niche DRAM business is also a company with its own brand, its own customers and, after raising RMB 57.92 billion, considerable ambition. The interests are aligned today. They are not structurally guaranteed to remain so.

There is also a live geopolitical overhang worth flagging rather than dramatising. The US Commerce Department has at various points considered adding CXMT to its Entity List, alongside subsidiaries of SMIC and 长江存储 YMTC.³³ Reporting has also indicated that CXMT and YMTC were at one stage removed from a draft restricted list amid trade negotiations.³⁴ The status is genuinely unsettled. If CXMT were restricted, the mechanism of harm to GigaDevice would be indirect but real: constrained equipment access at its key DRAM supplier, pressure on the carrying value of its stake, and possible spillover onto GigaDevice's own foundry relationships.

Which leads to the question of who is steering all of this, and how they have behaved when it counted.

↑ Back to Top

X. Management, Governance, and the Activist's Notebook

There is a scene that a short-seller's presentation would open with, and it is worth setting out precisely because the honest interpretation is more interesting than the cynical one.

On April 30, 2026, GigaDevice disclosed that its chairman and executive director Zhu Yiming planned to reduce his holding of A-shares by up to 11.21 million shares, about 1.60% of total share capital, in the window running to July 29, citing personal capital needs.²⁹ He executed almost the full amount: 11.1106 million shares, or 1.58% of the company, sold between May 6 and June 12, 2026, at prices ranging from RMB 339.44 to RMB 538.90, for total proceeds of approximately RMB 4.4 billion.³⁰ After the reduction his stake stood at roughly 6.80%.³⁰ This followed a prior reduction plan disclosed around the 2025 annual results.²⁸

The cynical reading writes itself: founder cashes out billions at the top of a cycle while telling the market about a multi-year franchise.

The evidence complicates that reading in a way that deserves acknowledgement. Zhu's selling was completed on June 12. The profit pre-announcement that sent the stock limit-up came on July 9.¹ He sold before the good news, not after — at prices between roughly RMB 339 and RMB 539, against a share price that subsequently reached RMB 846.66.³ Whatever else that was, it was not opportunistic front-running of a positive disclosure. If anything, he left a great deal of money on the table.

The fair criticisms are different and more structural.

The related-party pipe. GigaDevice's chairman also chairs its largest and fastest-growing supplier. Procurement from that supplier is scaling toward roughly RMB 5.7 billion in 2026.⁶ GigaDevice has disclosed the transactions and Zhu abstains from voting on them, which is the correct procedure.⁷ But procedure does not answer the underlying question a minority shareholder should ask: who sets the price? In a market where DRAM wafer allocation is the scarcest resource on Earth, the terms on which GigaDevice obtains that allocation are the single largest determinant of its gross margin — and those terms are negotiated with an entity its chairman leads. That could be a spectacular advantage for GigaDevice shareholders. It could equally be a mechanism by which value migrates toward the far larger, now-public CXMT. Disclosure lets you see the volume. It does not let you audit the pricing.

Two chairs, one person. Zhu stepped back from GigaDevice's CEO role in 2018 to run the Hefei venture.¹⁶ He is now chairman of a RMB 3.3 trillion company and chairman of a RMB 284 billion company that buys from it.²⁶³ Where does his attention go when the two companies' interests diverge — over capacity allocation in a shortage, or over customers in a downturn? There is no allegation of wrongdoing here. There is an unavoidable structural tension that any governance-focused investor should price.

Earnings quality. Two positions in newly listed Chinese companies produced marks that exceeded the entire prior year's profit.⁴ Combine that with the CXMT stake's balance-sheet-only treatment, and GigaDevice's reported financial statements now contain three quite different things: an operating chip business, a proprietary securities book that runs through profit, and a strategic equity stake that does not. Reading the headline net profit figure without decomposing it produces a materially wrong picture.

Capital allocation drift. The Hong Kong proceeds were earmarked for R&D plus strategic and industry investments including potential acquisitions.⁵ A company that has just demonstrated it can make RMB 2 billion from placement gains, holding RMB 15 billion of net cash, listing at a peak valuation, and telling investors it may buy things, is a company where an activist would want to see a specific and disciplined framework. None has been publicly articulated in detail.

Where does management earn credit?

The strategic record is genuinely strong. The ISSI collapse could have ended the DRAM ambition; instead it produced a durable domestic supply channel. The GD32 compatibility strategy was a correct long-term read of how MCU markets actually work. R&D continued through the 2023 trough rather than being cut to protect an earnings number. The balance sheet was never levered into a cyclical peak. Across a full cycle, the company gained share and emerged larger.

The communication record is more mixed. The optimism about 2027 hardened noticeably as the share price rose, from an earlier framing of high-level volatility to a later framing of limited capacity additions and sustained pricing.¹³²¹ Management also set a public target of becoming the world's largest NOR flash supplier within two to three years.²¹ That is a falsifiable, dateable claim, and it is exactly the kind of statement investors should file away and check. Winbond, the incumbent GigaDevice would have to pass, approved capital expenditure of NT$42.1 billion for 2026 specifically to expand NOR and SLC NAND capacity.⁶ The incumbent is not standing still.

Which sets up the competitive question properly.

↑ Back to Top

XI. War Game: Five Forces, Seven Powers, and the Map of the Battlefield

Imagine you are running strategy at Winbond in Taiwan in mid-2026. You have watched a Chinese competitor take 18.5% of global NOR flash and second place worldwide, and now announce it intends to take first within three years.¹²²¹ You have just committed the largest capital expenditure programme in your recent history to defend the category.⁶ What do you conclude about how this fight goes?

Start with Porter's five forces, because they explain the structure that governs everyone in it.

Supplier power is the dominant force, and it is high. In semiconductors generally, and in 2026 acutely, manufacturing capacity is the binding constraint. GigaDevice buys wafers from foundries including SMIC and 华虹半导体 Hua Hong, and DRAM from CXMT. In a shortage, the entity that owns the fab captures the rent. This is the structural reason a fabless model that looks capital-light in good times is strategically fragile in scarce times, and it is why the CXMT relationship — whatever its governance complications — is the most important asset GigaDevice does not own.

Buyer power is genuinely low in the core, high at the edges. An automotive tier-one that has qualified a NOR flash part to ISO 26262 ASIL D standards for a ten-year vehicle programme has no practical ability to renegotiate on price.³⁷ The requalification cost exceeds any plausible saving. An engineer with firmware written against a specific memory device faces the same trap. This is why GigaDevice's specialty memory gross margins expanded rather than compressed during input-cost inflation.⁶ At the consumer end — set-top boxes, low-end appliances — buyers are far more price-sensitive and switching is easier.

The threat of new entrants is moderate to high, and specifically domestic. The fabless model has low barriers by semiconductor standards. China's policy environment has funded a long list of NOR flash and MCU design houses. The barrier that actually protects GigaDevice is not capital but qualification: the years it takes to accumulate design wins in automotive and industrial. That barrier is real but it erodes at the low end continuously.

Substitution is the quietest and most underrated threat. Two forms matter. In NOR flash, the long-run risk is integration — as MCUs and system-on-chips embed more flash on-die, the need for a separate external NOR chip diminishes. In DRAM, the current price spike is itself driving customers toward memory compression, flash-based tiering and smarter memory management specifically to reduce DRAM consumption.¹³ High prices are always the cure for high prices; here they are actively funding the engineering effort to design around the product.

Rivalry is intense and about to intensify. Winbond and Macronix are the direct Taiwanese comparables in NOR and SLC NAND.⁶ Micron, through its ISSI unit, and Nanya Technology 南亞科技 compete in niche DRAM. Domestically, a growing set of designers including Puya Semiconductor 普冉股份 attacks the low end of NOR. And looming over all of it: Samsung, SK Hynix and Micron are not absent from these categories by choice — they left because HBM paid better. If AI memory demand cools, that capacity comes back, and it comes back into GigaDevice's markets.

Now Hamilton Helmer's 7 Powers, which asks a sharper question: which of GigaDevice's advantages actually persist against a competitor who understands them?

Counter-positioning is the power GigaDevice has exercised most successfully. Its entire history is a series of moves into segments where incumbents could not profitably follow: 180-nanometre NOR flash when everyone chased the leading edge; STM32-compatible MCUs that ST could not respond to without cannibalising its own pricing; DDR3 and small-capacity DRAM that the majors were actively exiting. Counter-positioning is powerful precisely because the incumbent's rational response is to do nothing. It is also, by construction, temporary — it lasts exactly as long as the incumbent's alternative remains more attractive.

Switching costs are the most durable power in the portfolio, and they are strongest in MCUs and automotive-grade memory. Firmware, board layouts, safety certifications and ten-year product lifecycles create genuine, measurable customer captivity. This is the mechanism that converted a 2021 emergency into a persistent franchise.

Network economies exist in an unusual, borrowed form. The GD32 ecosystem — 400-plus part numbers, third-party libraries, tens of thousands of customers and a large population of engineers trained on a compatible architecture — makes each additional GD32 design win slightly more likely.¹¹ It is a weaker version of what ST built, but it is real, and it compounds.

Cornered resource is where the analysis gets uncomfortable. GigaDevice's privileged access to CXMT's DRAM capacity, in a world where DRAM capacity cannot be bought at any price, is close to a textbook cornered resource. It is also uniquely fragile, because it rests substantially on one individual holding two chairmanships. Cornered resources are supposed to be structural. This one is personal.

Scale economies are modest. GigaDevice amortises R&D across roughly 4.3 billion chips a year.¹² That is meaningful for a design house. It is trivial next to the scale economics of a fab owner.

Branding and process power are, realistically, not present in any investable sense. Memory chips are not bought on brand, and GigaDevice's manufacturing processes are its foundries'.

The war game's conclusion is uncomfortable for both bulls and bears. GigaDevice has two durable powers — switching costs and a borrowed ecosystem — that support a good business at mid-cycle economics. It has one spectacular but conditional power in DRAM allocation. And it has a counter-positioning history that has worked repeatedly but is definitionally impermanent. What it does not have is anything that protects it from the price of memory falling.

↑ Back to Top

XII. Bull, Bear, and the Three Numbers That Matter

Set out the two cases as an investor would actually argue them.

The bull case rests on four legs.

The first is that this is a share-gaining company, not merely a cyclical one. GigaDevice entered the 2023 trough as a roughly RMB 8.5 billion revenue business and emerged from it larger than before, with a broader portfolio spanning memory, microcontrollers, sensors and analog.¹⁸ It is the only integrated circuit design company ranked in the global top ten across all four of NOR flash, SLC NAND, niche DRAM and MCUs.¹² Diversification across four cyclical categories is not the same as being non-cyclical, but it does mean the company is not a single-product bet.

The second is that the structural supply picture in legacy memory has genuinely changed. When more than half of SLC NAND suppliers exit a category and the global majors discontinue DDR4, the surviving suppliers face a permanently better competitive structure — fewer competitors, older equipment already depreciated, customers with no alternatives.²¹ That is a different situation from an ordinary cyclical upswing.

The third is customized memory and DRAM as an option not yet in the numbers. If the customized memory business converts its AI smartphone, AI PC and robotics engagements into revenue in the second half of 2026, GigaDevice would be adding a genuinely differentiated, design-win-based product line to a commodity portfolio.²¹

The fourth is the balance sheet and the optionality on it: substantial net cash, no meaningful debt, a stake in the largest Chinese memory manufacturer carried well below market, and a listing structure that gives access to both onshore and offshore capital.⁶

The bear case is shorter and does not require anything to go wrong.

Memory is a commodity, and commodity prices mean-revert. GigaDevice's own history documents what happens: gross margin fell fifteen points and net profit fell 93% in a single year the last time the cycle turned.¹⁸ The current earnings base is built on legacy memory prices that roughly doubled in six months.⁶²⁰ There is no version of the bull case in which those prices are permanent.

The specific bear mechanism to watch is capacity. Every high-margin period in semiconductors funds the capacity that ends it. Winbond's NT$42.1 billion programme is already committed.⁶ Macronix reactivated capital expenditure.⁶ CXMT raised RMB 57.92 billion, of which RMB 29.5 billion is earmarked for production-line upgrades and DRAM technology.²⁶ Every one of those investments lands in GigaDevice's markets in 2027 and 2028. Meanwhile, the demand side is already engineering around high memory prices through compression and tiering.¹³

Layer on the second-order risks. Foundry concentration means GigaDevice's cost of goods is negotiated, not controlled. Export control uncertainty around CXMT could impair the DRAM channel.³³³⁴ Earnings quality is complicated by a securities book capable of producing multi-billion-yuan swings.⁴ Related-party procurement is scaling faster than any other line in the business.⁶ And the valuation — even after a roughly 50% drawdown in the A-shares and 60% in the H-shares — still embeds an expectation that 2026 is closer to a base than a peak.²³

The global memory complex has already begun repricing that expectation, with the sector entering a bear market in late July 2026 on concerns that AI infrastructure spending is peaking.³² Whether that concern proves right is unknowable today. What is knowable is that GigaDevice's share price is now more sensitive to the 2027 memory price forecast than to anything management does operationally.

So what should an investor actually track?

Three numbers, and deliberately not more.

The first is specialty memory gross margin. This is the cleanest available read on whether GigaDevice's pricing power is structural or merely cyclical rent. It ran at 42.8% in 2025 and expanded through the upcycle.⁶ The critical test comes when memory prices stop rising. If margin holds because embedded customers cannot easily requalify, the switching-cost thesis is validated. If it collapses in line with spot pricing, GigaDevice is a price-taker with good marketing, and the correct multiple is much lower.

The second is DRAM revenue as a share of group revenue, together with the CXMT procurement run rate. These two move together and tell you whether the third leg is becoming a business or remains a cyclical windfall. Procurement scaling toward RMB 5.7 billion in 2026 is the input; DRAM revenue and the Zhuhai subsidiary's contribution are the output.⁶²² Watch whether the mix genuinely shifts from DDR3 toward LPDDR4, DDR4 and customized parts as management has said it will, or whether the growth turns out to have been DDR3 price inflation wearing a new label.²¹

The third is inventory. In memory, inventory days are the earliest and most reliable indicator of a cycle turn — not revenue, not orders, not management commentary. When a fabless memory company's inventory starts building while average selling prices flatten, the top is usually already behind.

Everything else — the CXMT stake's mark, the securities gains, the Hong Kong premium to the A-shares, the analyst target prices — is noise layered on top of those three signals.

The deeper question this company poses is one that recurs throughout industrial history. GigaDevice's founder made a bet in 2017, after a humiliating deal failure, that the way to solve a supply problem was to help build the supply. Nine years later that bet produced a separately listed company worth roughly twelve times GigaDevice itself, and GigaDevice kept 1.80% of it plus a phone line to the fab.²⁷²⁶³

Was that brilliant capital allocation or a giant value transfer away from GigaDevice's own shareholders? The answer depends entirely on how much the phone line is worth — and that, in turn, depends on how long the world stays short of memory.

Gigadevice

Table of Contents

兆易创新 GigaDevice: The Chip Designer That Planted a Trillion-Yuan Seed

II. Origins: From Tsinghua to NOR Flash (1989–2016)

III. The First Inflection: A Failed Acquisition, and the DRAM Seed That Became CXMT (2016–2019)

IV. The Second Inflection: GD32, STM32, and the 2021 Shortage Windfall

V. The Hangover: What 2022–2024 Revealed About the Business

VI. The Third Inflection: The AI Memory Supercycle and an Eleven-Fold Profit

VII. Capital Markets: The Hong Kong Listing and the Anatomy of a Frenzy

VIII. Niche DRAM and Customized Memory: The Third Leg, Still Building

IX. Myth vs Reality: Four Consensus Stories, Stress-Tested

X. Management, Governance, and the Activist's Notebook

XI. War Game: Five Forces, Seven Powers, and the Map of the Battlefield

XII. Bull, Bear, and the Three Numbers That Matter

References

Gigadevice

Table of Contents

兆易创新 GigaDevice: The Chip Designer That Planted a Trillion-Yuan Seed

I. Introduction: An Eleven-Fold Profit and a Halved Share Price

II. Origins: From Tsinghua to NOR Flash (1989–2016)

III. The First Inflection: A Failed Acquisition, and the DRAM Seed That Became CXMT (2016–2019)

IV. The Second Inflection: GD32, STM32, and the 2021 Shortage Windfall

V. The Hangover: What 2022–2024 Revealed About the Business

VI. The Third Inflection: The AI Memory Supercycle and an Eleven-Fold Profit

VII. Capital Markets: The Hong Kong Listing and the Anatomy of a Frenzy

VIII. Niche DRAM and Customized Memory: The Third Leg, Still Building

IX. Myth vs Reality: Four Consensus Stories, Stress-Tested

X. Management, Governance, and the Activist's Notebook

XI. War Game: Five Forces, Seven Powers, and the Map of the Battlefield

XII. Bull, Bear, and the Three Numbers That Matter

References