From its listing on the ChiNext board on April 16 to the announcement of planning an H-share issuance on the evening of August 24, only 130 days passed. The market's initial reaction was almost reflexive: Is it short of cash?
But if you know that developing a complete enterprise SSD controller chip + firmware + product validation, from project initiation to mass production, takes over 5 years—you might change your perspective.
What they are rushing for is not money, but time.
130 Days: The Market's Initial Reaction Might Be Misguided
Let's clarify the facts first.
On April 16, 2026, Shenzhen DapuStor Corporation (DapuStor) listed on the ChiNext board of the Shenzhen Stock Exchange, with the stock code 301666, an issue price of CNY 46.08 per share, and net proceeds of CNY 1.879 billion. It is the first company to successfully list while unprofitable after the ChiNext board adopted its third set of listing standards—an identity that inherently carries the implication of an "institutional breakthrough."
About four months after the listing, the company simultaneously disclosed its first semi-annual report since going public and an announcement regarding the planning of an H-share issuance. On September 9, an extraordinary general meeting reviewed and approved the H-share issuance and related proposals, which were passed with an overwhelming majority of votes.
The pace is indeed fast. The phrase "seeking another IPO just 5 months after listing" captures exactly this speed.
And in the context of the capital market, "fast" is easily interpreted as "anxious," and "anxious" is readily read as "short of cash."
The semi-annual report seems to provide corroborating evidence: the net cash flow from operating activities was CNY -871 million.
But the same semi-annual report also presents another set of figures: cash and cash equivalents increased from CNY 253 million at the beginning of the year to CNY 1.849 billion, while the asset-liability ratio dropped from 85.65% to 51.68%.
Do these two sets of figures contradict each other? They do not. They represent two sides of the same coin:
Profits have not vanished; they have turned into raw materials in the warehouse and accounts receivable in the hands of customers.
When a company's half-year revenue surges from over CNY 700 million to CNY 4.7 billion—a year-on-year growth of 531%—it must stock up inventory in advance and front the money for production. Money turns into goods, and the goods have not yet turned back into cash. This is a typical pattern for high-growth manufacturing; it does not mean there is a problem with operations, but it does indicate that: growth at this speed requires a continuous supply of external funding.
This is the first layer. What is truly interesting lies in the next.
In the AI Era, Why Have Drives Suddenly Become So Important?
To understand DapuStor, one must first understand a term—the storage wall.
Let's use an analogy. Imagine a top-tier restaurant that has hired ten Michelin-starred chefs (GPUs) with extremely fast knife skills and precise heat control. However, the kitchen has only a narrow pass for serving dishes, allowing ingredients to be handed in only a little at a time.
What is the result? The chefs spend half their time waiting for ingredients.
This is the real dilemma faced by today's AI computing clusters. The computing power of GPUs is growing extremely fast, but the speed of feeding data cannot keep up. As models become larger, context windows grow longer, and the number of concurrent users increases, "moving data" is becoming the bottleneck that slows down the entire system.
This bottleneck specifically manifests in two scenarios, each with different requirements:
During model training, the requirement is to "read a massive amount of data in one go"—sustained high bandwidth to feed in massive datasets.
When the model is deployed for service (i.e., inference), the requirement is to be "on-demand and fast every time"—when a user asks a question, the system must compute the answer in milliseconds without a single stutter. This tests the stability of latency, not the peak speed.
Why is inference so demanding? Because training is batch processing; an occasional slowdown can be tolerated. Inference, however, is real-time and online. A single long-tail stutter will collapse the user experience across the entire request chain.
Consequently, something originally considered an "accessory" has undergone a fundamental shift in its position:
SSDs (Solid State Drives) are no longer just data repositories; they are an integral part of the computing power chain.
Data corroborates this shift: according to Counterpoint Research, the proportion of enterprise SSDs in global NAND flash shipments surged from 26% in the second quarter of 2025 to 48% in the second quarter of 2026, nearly doubling in a year, and is projected to exceed half by the end of 2026.
When drives begin to determine whether GPUs can operate at full load, their value proposition changes entirely.
What DapuStor produces are precisely the enterprise SSDs used in data centers—not the kind in your personal computer, but those installed in cloud servers, operating with non-stop read/write cycles 24/7, where a failure would lead to catastrophic consequences.
A 5-Year Clock
Now we can answer the core question: Why the rush?
Because this business operates on a very unique clock.
According to DapuStor's prospectus, the timeline for a complete enterprise SSD technology system, from project initiation to mass production, is roughly as follows: Controller chip R&D: 2–3 years. Firmware algorithm R&D: 1.5–2 years. Finished product validation: In total, completing a full technology system typically takes over 5 years.
5 years. This is the key to understanding the entire matter.
What's more critical is that this is not a "do it once and you're done" situation. This industry has two non-stop engines pushing everyone forward:
The first engine: Interface generations. PCIe evolves from 4.0 to 5.0 to 6.0, with each generation doubling the bandwidth. Server platforms are updating, and customers are expanding; if you cannot keep up, you are eliminated in this round of procurement.
The second engine: Flash memory media. NAND dies evolve from TLC to QLC and beyond, with a structural upgrade approximately every two generations. The controller chip must be compatible in advance with dies that have not yet been shipped at scale—you have to design today's chips for the flash memory of two years from now.
Therefore, an enterprise SSD manufacturer always operates simultaneously on three timelines: selling the current generation, validating the next generation, and conducting preliminary research on the generation after that.
And the rules here are extremely ruthless: falling behind by one generation does not mean falling behind by a little bit; it means missing an entire product cycle.
The reason is that customer adoption for enterprise SSDs is extremely heavy—from sampling to final adoption, a product typically undergoes 6 to 18 months of multiple rounds of validation. Conversely, once adopted, the stickiness is extremely strong, and customers will not easily switch suppliers.
What does this mean? Once a window is missed, it is not just a delay of half a year; it means waiting for the next expansion cycle.
Looking back at DapuStor's position now, the picture becomes clear.
Currently shipping are its two generations of self-developed controllers, the DP600 (PCIe 4.0) and DP800 (PCIe 5.0). The DP800 is among the world's first mass-produced enterprise PCIe 5.0 products and has been delivered in volume. Meanwhile, the DP900, targeting the next generation—specifically the PCIe 6.0 controller—is still in the R&D phase, advancing according to the fundraising investment plan.
On the capacity front, the 122TB QLC ultra-high-capacity drive is already commercialized, and the 245TB product has been launched and sent for testing.
Overlaying these two timelines, the conclusion is clear: While its current generation products are selling best, the ticket to the next generation must be paid for by investing heavily while the current generation is still hot-selling.
This is not aggressive; it is dictated by physical time.
Therefore, the possible truth behind launching the H-share issuance after 130 days is: This is not being pushed forward by a funding gap, but by the R&D milestones of the DP900.
The pace of technological R&D investment is never determined by current profits. From 2023 to 2025, the company's total R&D expenses were approximately CNY 928 million; in the first half of 2026, R&D expenses reached CNY 178 million, a year-on-year increase of 34.15%, with R&D personnel accounting for as much as 67.61%.
Two-thirds of the company's workforce is engaged in R&D. This in itself is the answer: it is a company driven by the technology clock.
Where Exactly Does the Difficulty of "Full-Stack In-House R&D" Lie? Breaking It Down into Three Layers
The label most frequently associated with DapuStor is "full-stack in-house R&D of controller chips + firmware algorithms + modules." This sounds like a slogan, but when broken down, it actually describes a very difficult undertaking.
Layer 1: Controller Chip—The Brain of the SSD.
It determines how flash memory is managed, how data is scheduled, and how errors are corrected. You can purchase a generic controller externally, like buying a standard engine; or you can design it yourself, at the cost of several years and massive investment.
The value of doing it in-house is not just about saving money, but about speed—when a customer raises a specific requirement, you can directly modify the chip to respond, rather than waiting for the upstream manufacturer's schedule.
Layer 2: Firmware Algorithms—The Operating System of the SSD.
It sounds abstract, but it directly determines whether the drive is stable, fast, and long-lasting in use. Take the concept of "write amplification" for example: a characteristic of flash memory writing is that the actual amount of data written is often greater than the intended amount, and the excess part wastes lifespan for nothing. Good algorithms can suppress this loss.
DapuStor has a highly representative product feature—transparent compression for KV Cache. This is a targeted optimization specifically for large model inference scenarios.
A drive manufacturer engaging in "large model inference optimization" is, in itself, the best footnote to the storage wall.
Layer 3: Reliability—The most inconspicuous, yet the hardest to copy.
The divide between enterprise and consumer SSDs essentially lies in this layer:
Consumer drives are designed for 8 hours of daily use; enterprise drives require 24/7 non-stop operation, working continuously for over 5 years; the enterprise requirement for bit error rate is typically at the 10⁻¹⁸ magnitude—meaning only one error correction failure might occur per 10^18 data reads.
Achieving this level does not rely on some "black technology," but on a whole set of hard, meticulous work: how to protect data during a sudden power loss, how to control temperature, how to manage bad blocks, and how to ensure consistency across every batch of shipments.
This is also the key to understanding the use of raised funds. About 30% of this H-share fundraising is allocated to "advanced packaging and testing, and testing production lines." Many interpret this as "expanding capacity," but in the enterprise SSD business, the testing production line itself is a technical capability—flash memory dies from different batches and brands require parameter re-adjustment and re-validation; high-temperature aging, power-off testing, and long-term consistency validation can only be done with one's own production lines.
Without this capability, you cannot guarantee that every delivered drive is equally stable in a multi-brand supply environment.
It Is Betting on Three Futures of AI
If controllers, firmware, and reliability are the basic skills, then DapuStor's more cutting-edge layout reveals even more about what it aspires to become.
Three routes perfectly correspond to three needs in the AI era:
Ultra-high capacity → Solving "storing enough." The 122TB QLC drive is commercialized, and the 245TB has been sent for testing. Large model corpora, vector databases, and the like are expanding at an astonishing rate. The larger the capacity of a single drive, the more data a cabinet can hold, and the lower the cost.
Storage Class Memory (SCM) → Solving "fast enough." Its positioning lies between memory and flash memory, used to bridge the speed gap between the two—ultra-low latency and ultra-long lifespan. This is a capability possessed by only a handful of manufacturers globally; DapuStor released a new generation of Gen5 products in 2025 and has shipped them in volume.
Computational Storage → Solving "helping the CPU do the work." Integrating processing capabilities inside the drive allows data filtering, encryption, and compression to be completed within the drive, without occupying the host CPU. In today's environment where CPU resources in AI servers are increasingly tight, this is an architectural approach to alleviating the storage wall.
The common premise for these three routes is: they all require a foundation of in-house controllers and in-house firmware. The lifespan of QLC relies on algorithms to back it up, the low latency of SCM relies on the controller architecture for support, and computational storage requires stuffing processing units directly into the chip.
Therefore, "full-stack in-house R&D" is not just a slogan—it is the common foundation for whether these three routes can be successfully navigated.
How Will the Funds Be Spent?
The company disclosed three sets of figures for the allocation of H-share raised funds.
40% → Next-generation controller chips (PCIe 6.0) and enterprise SSD R&D. This is the entry fee for the generational race. The necessity of spending this money is determined by the speed of interface standard evolution, and has nothing to do with whether the company is making money in the current period.
30% → Advanced packaging and testing, and testing production lines at the Shenzhen headquarters and the Greater Bay Area. This is the infrastructure for reliability; as mentioned earlier, it determines whether the company can deliver stably in a multi-brand supply environment.
20% → Overseas market expansion and sales network. On a technical level, this has another implication: the demands for next-generation AI storage are often defined in advance by global top cloud providers and AI companies when designing data centers. To participate in this "joint definition," you must walk into the customers' technical planning meetings and gain entry into the circles of international standards and specifications. The company disclosed that it has participated in the formulation of multiple industry storage standards and specifications—this 20% is buying exactly this right to participate.
The company's own explanation for going to Hong Kong is: to enhance international credit and meet overseas customers' requirements for supplier compliance and transparency in information disclosure.
In this context, this statement is very practical: A company that wants to participate in defining next-generation AI storage must first convince customers that it has the ability to continuously invest heavily in R&D over the next five years. And this trust is largely built on financial transparency and international auditing.
Incidentally, here is a piece of data: in the first half of 2026, the company's overseas revenue reached CNY 2.706 billion, accounting for 57.30% of total revenue.
More than half of its business is already on an "international scale," but its capital identity remains entirely a domestic face. This mismatch is the deepest motivation behind this trip to Hong Kong.
Three Questions Remain Unanswered
First, whether the DP900 can be mass-produced on schedule. According to industry norms, it usually takes more than 2 to 3 years for a generation of controllers to go from R&D to mass production, and it also requires the upstream and downstream ecosystems—such as server platforms, operating systems, and flash memory supply—to mature together. There is a considerable distance between smooth R&D and the realization of mass production.
Second, whether ultra-high-capacity products like the 245TB can be delivered at scale. From "initiating sampling" to volume shipment, there are several hurdles in between, including customer validation, yield ramp-up, and multi-brand adaptation. Sampling is a positive signal, but it is not yet delivery capability.
Third, whether cutting-edge products can be truly embraced by the ecosystem. For technologies like SCM and computational storage, being able to develop them is merely a prerequisite; the real test is the software ecosystem—operating systems, databases, and the application layer must coordinate to release the value. From "I can do it" to "being used on a large scale," there lies the maturity of an entire ecosystem, which cannot be driven by a single company alone.
As for the stock price in the secondary market—the company's market capitalization once surged after listing, followed by a noticeable pullback from its May high—this is the result of the market repricing between valuation and prosperity. It belongs to public market behavior and is a different level of issue from operational fundamentals; this article will not comment on it.
Two Clocks
DapuStor has a clear timeline: founded in 2016, it launched its first PCIe SSD in 2019, and since then has progressed from PCIe 3.0 to becoming among the world's first to mass-produce PCIe 5.0 enterprise SSDs, while simultaneously laying out SCM and computational storage.
It took the company about ten years to complete the journey from chip design to the realization of full-stack in-house R&D.
And the next generational window is only two to three years away.
Therefore, the phrase "seeking another IPO just 5 months after listing" might be focusing on the wrong thing. What is truly worth asking is not how short 130 days is, but:
Can this company, while its flagship products are still hot-selling, simultaneously secure the R&D funding for the next-generation controller, the packaging and testing capabilities, and the technical synergy with overseas customers?
The time of the capital market is measured in quarters, while the time of technology is measured in generations. When the two are inconsistent, the rational approach is to follow the technology.
What an H-share listing can provide is a funding channel less constrained by a single market cycle, and a credential of credit that can be scrutinized by global customers.
And whether this credential can ultimately be cashed in for technological leadership, the answer is not in the prospectus, but in the shipping orders of the next few generations of products.
This article is written based on the publicly disclosed prospectuses, periodic reports, and temporary announcements of the listed company, as well as third-party public research data. The article aims to analyze corporate strategy and industrial logic, and does not constitute any investment advice, nor does it constitute a recommendation to buy or sell related securities. The market has risks, and decisions should be made with caution.