The supply of key raw materials for memory is cut off, and Changxin has unexpectedly become the biggest winner?

Have you ever heard of tungsten hexafluoride?

It is a key material in the production of 3D NAND and HBM. In semiconductors, it is mainly used for CVD deposition of tungsten films. In DRAM, tungsten is used in structures such as contact holes, interconnections, and word lines.

Tungsten hexafluoride is itself toxic, highly corrosive, and will generate HF when it comes into contact with water. It has high requirements for purity, packaging, cylinders, transportation, and on - site supply systems.

Without it, the tungsten metal interconnection layer inside the chip cannot be deposited, and the circuits of advanced processes cannot be connected.

However, starting from July 1st, two Japanese companies, Kanto Denka and Central Glass, will permanently stop producing tungsten hexafluoride. Approximately 25% of the global production capacity will disappear in July.

The direct reason for the production halt is that China's exports of high - purity tungsten powder to Japan will basically drop to zero starting from January 2026.

The production of tungsten hexafluoride mainly comes from tungsten powder, and approximately 80% of the world's tungsten resources and refining come from China.

It's okay that the fire of AI has spread to memory, but now even raw materials are affected.

So who is the beneficiary?

A. NVIDIA, B. SK Hynix, C. Obtuse angle.

The answer is none of the above. It is ChangXin.

NVIDIA of course hopes for a stable memory supply. However, the supply disruption of tungsten hexafluoride will inevitably drive up the costs of HBM and server memory, which will ultimately be passed on to its procurement price.

SK Hynix, the main customer of Japanese tungsten hexafluoride, now has to turn to South Korean local suppliers SK Specialty and Foosung. However, these two companies have notified customers that the contract price in 2026 will increase by 70% to 90%.

While the three giants are busy dealing with the raw material crisis and the pressure of HBM capacity expansion, ChangXin has just seized this God - given opportunity.

ChangXin has entered Corsair's AVL

Corsair is one of the world's largest consumer - grade memory brands. In the past, it only sourced memory chips from Micron, Samsung, and SK Hynix.

Now, ChangXin chips have appeared in its Vengeance DDR5 - 6000 kit, with the model number CMK5X16G3E60C36A2 - CN. The suffix CN indicates that it is exclusive to the Chinese market.

Screenshots from CPU - Z and HWiNFO64 clearly show that the manufacturer is listed as ChangXin Technologies. This product supports Intel XMP and AMD EXPO overclocking profiles, with a timing of CL36 and an operating voltage of 1.35V. Its specifications fully meet the mainstream standards.

After the news that Corsair is using ChangXin chips spread, many people's first reaction was the price.

After all, the memory price has skyrocketed in the past year or so. DDR5 modules have risen from $60 - 70 at the beginning of 2025 to $200 - 300, and some high - capacity kits have even exceeded $400.

However, after Wccftech interviewed multiple memory manufacturers during Computex 2026, the feedback was that the price of ChangXin's DDR5 is close to that of Samsung, SK Hynix, and Micron. A memory brand manufacturer clearly stated that the procurement price of ChangXin chips is "almost the same" as that of the three giants, with a single - digit percentage difference.

The outside world has always thought that ChangXin would enter the market with low prices. After all, when it comes to "domestic products", people think of "affordable alternatives".

Currently, the shortage of DDR5 is a consensus in the industry. Therefore, the price is no longer determined only by cost, but also by deliverability.

For brand manufacturers like Corsair, if they cannot obtain chips from Samsung, Hynix, and Micron, or if the delivery time, quota, and price are unstable, then even if ChangXin's chips are only a little cheaper, or even close in price to the three giants, they are still worth purchasing.

This is the real meaning of ChangXin's pricing being close to that of the three giants. It is not proving that it has completely caught up with the three giants, but rather that the memory market has shifted from "who is cheaper" to "who can still supply goods".

The problem with the three giants now is not that the price is too high, but that they simply have no goods. Their production capacity has been absorbed by HBM, LPDDR5X, and server DDR5, leaving a serious shortage of chips for consumer - grade PCs.

PC manufacturers and memory brands need not cheaper chips, but more available chips.

ChangXin just fills this gap. It has no order pressure from AI data centers and no long - term contracts with NVIDIA, AMD, or hyperscale cloud providers. Therefore, its production capacity can be fully dedicated to the consumer - grade market.

However, the significance of the Corsair incident is not entirely about the price itself.

It really shows that ChangXin has passed the verification system of international brands.

In the PC industry, there is a term called AVL, which stands for Approved Vendor List. It can be understood as a "list of qualified suppliers".

For PC manufacturers, memory brands, and server manufacturers, buying memory chips is not about randomly choosing a supplier in the market.

Suppliers need to be verified, including yield rate, stability, compatibility, batch consistency, after - sales risk, and long - term supply capacity. Only after passing the verification can they enter the customer's AVL.

A mature large - scale manufacturer like Corsair will definitely not randomly put others' memory chips into its product line.

It needs to test compatibility, verify the yield rate, confirm batch stability, evaluate after - sales risk, and ensure that the supplier can provide long - term supply.

ChangXin's entry into Corsair's AVL indicates that it has taken the first step into the international market.

ChangXin also has an advantage that the three giants can hardly replicate now, which is the certainty of raw material supply.

As mentioned earlier, Japanese tungsten hexafluoride manufacturers stopped production on July 1st. Samsung and Hynix had to turn to South Korean local suppliers, resulting in increased production costs and reduced stability of raw material supply.

However, ChangXin does not have this problem. Because China has an abundance of tungsten resources and a solid foundation in fluorine chemical industry.

In terms of tungsten, China is the world's largest supplier; in terms of fluorine chemical industry, China also has a very complete industrial chain.

The basic synthesis route of tungsten hexafluoride is not a mystery. Usually, tungsten powder reacts with fluorine gas, followed by rectification and purification.

Therefore, ChangXin has the confidence to price its memory chips close to those of Samsung and SK Hynix.

AI PCs provide a larger stage for ChangXin

Furthermore, if the sales volume of the AI PC market increases, ChangXin's entry into Corsair's AVL may bring more benefits to ChangXin.

An AI PC is not simply adding an NPU. It will change the basic configuration of a PC.

Microsoft's Copilot + PC requirements clearly state that the memory should start at 16GB. In the past, the minimum memory configuration was generally 8GB.

In the first half of 2026, the AI hardware community generally found that local models with 7B to 13B parameters require 16GB to 32GB of memory to run smoothly. A 70B model requires more than 64GB.

Although most consumers will not directly run a 70B model, the selling point of an AI PC is its edge - side capabilities.

If an AI PC only has 8GB of memory, its capabilities will be very limited. Open a few browser tabs, run a few applications in the background, and add a local AI assistant, and the memory will basically be full.

Therefore, 16GB is changing from "sufficient" to "barely enough", and 32GB is changing from "high - end configuration" to "standard configuration".

This poses a huge cost pressure on PC manufacturers.

The BOM of an AI PC is expanding. The CPU needs to support an NPU, the screen needs a higher resolution, the battery needs a larger capacity, and the heat dissipation needs a stronger design.

Each item costs money.

If the memory price continues to rise, it will be difficult for AI PCs to move from high - end prototype machines to the mass market. Therefore, OEM manufacturers such as Lenovo, HP, Dell, Asus, and Acer, who are preparing to enter the AI PC market, must have a stable supply source of DDR5 and LPDDR5X.

The fact is that the three giants cannot provide enough goods now.

NVIDIA's Grace CPU Superchip uses 960GB of LPDDR5X per unit, which is equivalent to the memory usage of 60 high - end mobile phones. The price of server - grade DDR5 is expected to double by the end of 2026 compared to the beginning of 2025.

Moreover, HBM consumes about three times the wafer area of DDR5.

PC manufacturers want to promote AI PCs, but the memory supply cannot keep up.

This is ChangXin's opportunity, and its value will be magnified by AI PCs.

Specifically, as of the first quarter of 2026, ChangXin's three 12 - inch wafer fabs in Hefei and Beijing have all reached full production, with a stable monthly production capacity of 290,000 - 300,000 wafers.

During the same period, SK Hynix's monthly DRAM production capacity is about 550,000 wafers, and Samsung's is about 720,000.

Although ChangXin still lags behind in production capacity at present, its growth rate is very fast.

In March 2026, ChangXin's Shanghai Super Wafer Fab was officially launched, with a planned total production capacity of 400,000 - 600,000 wafers per month. It is planned to move in equipment in the second half of 2026, start formal production in 2027, and reach full production in 2028. In contrast, Micron's newly announced factory will not start production until 2028.

Moreover, ChangXin's capacity expansion speed for DDR5 is faster than that of the three giants because it does not need to divert production capacity to HBM.

More importantly, ChangXin's customer structure is completely different from that of the three giants.

As AI PCs move towards the mainstream market, PC manufacturers need memory suppliers other than the three giants.

ChangXin has caught up with the cycle when the PC industry is raising the memory configuration threshold again. This cycle may last for several years.

There will be more and more AI applications, local models will become larger, and edge - side inference will become more complex.

16GB will soon become "outdated at entry - level", 32GB will become the "standard configuration", and even 64GB may become the "standard configuration".

During this process, ChangXin's position in the AVL will become more and more stable.

NVIDIA locks in Hynix, and the DDR5 market will continue to expand

At the beginning of June 2026, NVIDIA and SK Hynix signed a multi - year memory cooperation agreement in South Korea.

NVIDIA promised to purchase next - generation memory products from Hynix, and Hynix promised to provide customized memory solutions for NVIDIA's Vera Rubin AI system, RTX Spark PC, and Jetson Thor robot platform. The agreement covers multiple years and involves roadmap coordination, capital expenditure planning, advanced packaging technology, and production capacity scheduling.

AI data centers offer larger orders, higher profits, and deeper binding. Therefore, within the three giants, the priority of DDR5 memory chips will naturally decline.

Hynix's HBM production capacity in 2026 is expected to increase by 50% compared to 2025. Samsung plans to increase the monthly production capacity by about 60,000 wafers in the first half of 2026, mainly for HBM4 production. Micron announced in December 2025 that it would exit the consumer - grade memory and storage market and focus on serving AI data center customers.

These actions all indicate the same thing. The three giants, who have already reduced the production of memory chips, will further cut production from now on.

After all, the profit margin of HBM is much higher than that of DDR5. Who doesn't want to earn more money?

The LPDDR5X usage per unit of NVIDIA's Vera Rubin platform is expected to exceed 1TB, more than twice that of the Grace platform. AMD's MI450 processor will also use a large amount of HBM4.

Do you think that's all? Obviously not.

According to foreign media reports, hyperscale cloud providers are all hoarding AI servers, and their memory demand will remain high in the next three to five years. Even if the three giants expand production, the new production capacity may not be sufficient.

This leaves more room for ChangXin to grow.

Even without considering AI PCs, ordinary PCs, consumer memory, edge devices, and some servers all require a large amount of DDR5 and LPDDR5X.

On June 9th, Jensen Huang himself said in an interview that the memory crisis will last for "quite a few years".

The current DDR5 supply is so tight that it exceeds the inventory mismatch of one or two quarters. In other words, the memory you buy today is actually the goods that the manufacturer waited for half a year to obtain.

If PC manufacturers want to continue promoting AI PCs, they cannot wait for Samsung, Hynix, and Micron to expand production. They must bring in new manufacturers to the AVL.

There is also a very interesting thing. On the same day that NVIDIA announced the multi - year memory cooperation with SK Hynix, South Korea's KOSPI index closed down more than 8%. SK Hynix closed down 7.7%, and it once dropped 10.3% during the session. Samsung Electronics plummeted 10.2% at the opening.

South Korean financial media believes that this round of decline does not mean that the market suddenly loses confidence in AI. Instead, it is because South Korean chip stocks have risen too fast before.

Hynix and Samsung have been traded as HBM leaders, and a lot of expectations for AI orders have been priced into their stock prices in advance. As long as the expectations loosen slightly, some funds will withdraw first.

However, a drop in the stock price does not mean a change in the industry direction. NVIDIA is still locking in Hynix, cloud providers are still scrambling for AI servers, and HBM production capacity is still the tightest and most profitable part. For the three giants, their production capacity will definitely continue to shift towards AI memory, leaving more and more room for ChangXin.

The supply shortage of DDR5 will be a long - term situation.

Multiple supply chain insiders interviewed by Tom's Hardware during Computex 2026 all said that the shortage of DRAM and NAND will continue throughout 2027.

Mainboard manufacturers have even started to restart the DDR4 mainboard production line. AMD has re - released the Ryzen 7 5800X3D 10th Anniversary Edition, and Intel said it will continue to ensure products that support "old - fashioned memory technology".

This article is from the WeChat official account "Zimu AI", author: Miao Zheng. Republished with permission from 36Kr.