DRAM has been the graveyard of  semiconductor empires. Competing  

in this market is not for the faint of heart.

After decades, the DRAM industry consolidated  

to just three big companies:  Samsung, SK Hynix, and Micron.

But now they are being challenged by  a new kid on the block. A Chinese DRAM  

monster emerging out of Hefei: ChangXin  Memory Technologies, or CXMT (長鑫存儲).

In today's video, we profile the third of the  

People's Republic of China's  semiconductor titans: CXMT.

## Beginnings

The Chinese have tried to  do DRAM many times before.

In 1975, a team at Peking University  working at Factory 109 announced  

that they had finished a silicon-gate NMOS  technology that can produce a 1K DRAM chip,  

about five years after the  Japanese and Americans did it.

And over the years, the Chinese Academy of  Sciences announced various DRAM milestones:  

4K in 1978, 16K in 1981, and 64K in  1985. Chinese announcements lagged  

the leading edge Japanese or Korean  announcements by maybe 1-3 years.

Despite this heritage of reaching  technical milestones, the Chinese  

struggled to achieve commercial  success in these days. Perhaps  

in part due to financial struggles as well  as insufficient transfers from lab to fab.

One of China's most well respected semiconductor  makers was Wuxi Huajing - formerly known as  

Factory 742. After a few catchups, they  fell hard at the 256K DRAM level. A  

product was not released until 1993,  about 7 years behind leader Samsung.

As I covered in a prior video, Huajing eventually  fell into a crisis thanks to excess bureaucracy,  

a few bad technical choices in importing foreign  technology, and the Asian Financial Crisis.

Then in the 1990s, the Chinese-Japanese joint  venture Shougang-NEC produced 64 megabit DRAM,  

based on an NEC DRAM design. However,  an ambitious fab expansion was halted  

due to bad market conditions in 2001,  and the company faded from relevance.

The formidable Chinese semiconductor foundry  SMIC did produce DRAM at their 12-inch wafer fab,  

Fab 4. But DRAM prices plunged again, and  the company got sued by TSMC for IP theft.  

Facing financial strain in 2007, they  started slowly backing away from DRAM.

In early 2008, they agreed to sell their DRAM  fab facility to foundry customer Qimonda,  

exiting the space altogether at a  loss of several millions of dollars.

So the TLDR of this long but sad history? Mainland  China is the world's largest semiconductor market,  

consuming an estimated 40%  share of industry production.

Yet the top DRAM suppliers are headquartered  in South Korea or the United States.

I do want to note that there are foreign-run  DRAM factories in China. For instance,  

SK Hynix has a very large one in Wuxi, though the  

United States government has prevented  the company from keeping it up to date.

## DRAM Markets

The issue with competing in DRAM - as  you might have noticed from the brief  

historical overview - is that DRAM is a  commodity subject to wild price swings.

Memory buyers are finicky and panicky.  Their demand for DRAM is tied to larger  

consumer market trends like 5G smartphones  or the ChatGPT AI boom. When times are good,  

the big buyers tend to over-order,  driving up prices. But if things  

turn - and they can turn very fast  - then those orders can evaporate.

The issue is that researching, developing,  and mass-producing new DRAM products is  

capital intensive. Building a modern fab  of any size takes far more time than the  

market allots you. So if things turn and  orders are canceled, then suppliers are  

left in the lurch - potentially on  the hook for billions of dollars.

Therefore, to compete you not only need  technical prowess and operational excellence,  

but also stable financing. So it  makes sense then that the Korean  

memory-makers are part of larger conglomerates.

Samsung with Samsung Electronics and  Samsung Group. And SK Hynix to SK Group,  

a conglomerate with telecom,  energy and chemical arms.

## Hefei

Hefei (合肥) is the capital of Anhui province and  boasts a population of about 9 million people.

The population is comparable to that  of Austria and Switzerland. While the  

physical size is only being about that of Jamaica.

It is an ancient city. And situated quite close  to metropolises like Shanghai or Hangzhou,  

but remains less well known. The  published photos look quite nice.

Like a few other Chinese cities, Hefei  has a state-owned investment fund,  

doling out funds for industrial  development in local areas. In 2008,  

it invested in the OLED company BOE Technology  (京东方). I profiled them in a prior video.

BOE went public and grew into a global giant,  

motivating another move by Hefei into  semiconductor memory, DRAM in particular.

Seeing China’s dependence on foreign  DRAM-makers as a national security risk,  

Hefei investment officials  reached out to a Chinese  

memory company called GigaDevice about  founding a domestic DRAM manufacturer.

Launched in 2016, the effort was internally  called Project 506 and led to the creation of  

Innotron Memory (合肥睿力集成电路有限公司). The company  later changed its name to CXMT in 2019.

Hefei pledged land and a reported $7.2 billion  into the effort. Additional financial support  

came from China's Integrated Circuit  Industry Investment Fund, or Big Fund,  

which the State Council set up in 2014  to grow its semiconductor industry.

## GigaDevice

What is GigaDevice? Or as they were  once called, Zhaoyi Innovation (兆易创新)?

GigaDevice's founder Zhu Yiming is a relatively  young founder. After graduating from Tsinghua  

University, he received an education in  the United States at Stony Brook college.

After working in Silicon Valley for a few  years, he returned to Beijing to found  

GigaDevice. They started producing SRAMs, small  microcontrollers, and then NOR flash memory.

NOR flash is a predecessor to  NAND. It is not a large market,  

but an easier one to enter. NOR flash chips  do not hold a lot of data, but vendors often  

use them for storing the firmware and codecs  for smaller internet-connected devices like  

earbuds. This is because these chips are  fast, quite reliable, and long-lasting.

GigaDevice competed against Taiwanese  producers like Winbond and Macronix,  

becoming a respected industry leader. And in 2020,  

Apple chose to use their NOR flash chips for  the AirPods product, displacing Windbond.

Zhu Yiming led both Gigadevice and CXMT  (though today he serves only as chairman,  

not CEO). Gigadevice provided design  services for the memory and purchased  

the supply. CXMT would do the manufacturing.

## Technology Transfer

How did CXMT learn the complicated  art and science of making DRAM?

In a 2019 presentation, Zhu Yiming noted  that CXMT's technology base came from two  

sources. The first is the bankrupt  German DRAM manufacturer Qimonda.

Qimonda had once been part of Infineon, which in  turn was once part of the German giant Siemens.  

As part of a restructuring, Infineon spun  them off as a separate company in 2006.

Qimonda had been a technical R&D  pioneer especially in technologies  

relating to stack capacitor cells, where  the capacitors go upwards. Though most  

of their actual production were of "trench  capacitors", which go down into the substrate.

At one point, they had been the second  largest DRAM vendor. But in 2008 they fell  

into the red due to unfavorable memory prices.  Despite a 280 million euro rescue package from  

the German government, Qimonda fails to survive  the global financial crisis and goes insolvent.

Their patents were sold to a Canadian  "IP company" called Wi-LAN, which then  

licensed those patents to CXMT, maybe in  2019. Zhu mentioned that they acquired  

over 10 million technical documents  and 2.8 terabytes of Qimonda's data.

Like I said, Qimonda had been a technical  pioneer, but their technology will have been  

at least 8 years old when it passed to CXMT.  It might not have been super useful in terms  

of cutting edge production, but I reckon it  provided formidable air cover from litigation.

The second source - and probably the  more relevant one - came through the  

Japanese company Elpida. Elpida is  the result of a merger between NEC  

and Hitachi's DRAM operations.  In 2013, Micron bought Elpida.

In 2016, Elpida's former president Yukio  Sakamoto set up a consulting company called  

Sino King. Sino King hired employees from  both Elpida and Powerchip Semiconductor  

Manufacturing Corporation, a Taiwanese memory  company. They then joined the Hefei project.

However, technology acquisition  can only be a temporary thing.  

It is critical that a transfer is complete, with  CXMT producing their own innovations and patents.  

And if you measure innovation by patents, they did  that. By the end of 2017, a year after founding,  

CXMT had applied for 354 patents and  planned to apply for 1,509 more in 2018.

## Fab 1 and DDR4

The speed with which Innotron  or CXMT took off is impressive  

even by semiconductor manufacturing standards.

Right off the bat, the company started on  their fab's first building, or phase, in  

March 2017. The roof was sealed in June 2017 and  equipment started moving in the following quarter.

Equipment move-in was completed  in early 2018 and they started  

working on producing engineering  samples of their first product,  

a 8-gigabyte LPDDR4 DRAM produced  using a 19 nanometer process node.

Is that good? As you probably know,  process node naming for semiconductor  

logic chips like CPUs and GPUs is  just all marketing. But with memory,  

the naming still has some ground in reality.  It still tells you the lithography half-pitch.

But companies in the industry are not excited  about flagging this sort of information. So  

at 20 nanometers, they started calling the  nodes 1x, 1y and 1z. Yes, the names are weird.

Anyway, DDR4 DRAM is an older product. Samsung  first demonstrated prototypes in 2012. And they,  

SK Hynix and Micron all started shipping  it in volume in 2014. So when CXMT started  

shipping DDR4 products in 2019, they were  about 5-7 years behind the leading edge.

## Fujian Jinhua Yes, they were behind. But in shipping a product,

CXMT evaded the sad fate of another Chinese  DRAM startup, Fujian Jinhua Integrated Circuit.

Fujian Jinhua was founded about  the same time as CXMT, in 2016,  

and was funded by the Fujian provincial  government. So a similar model to CXMT.

They recruited aggressively from Taiwan,  opening a design center there to cater to  

designers who did not want to move across  the Strait, which is not an unreasonable  

request. But in doing so, they eventually  ran afoul of the sharp-elbowed Micron.

In 2017, Micron sued them and  the Taiwanese foundry UMC for  

stealing trade secrets. UMC had  apparently acquired and passed  

over trade secrets as part of a DRAM  project they accepted from Fujian.

This litigation got very complicated.  And overall, it was a strange legal case,  

but the whole thing was eventually settled in  the end with UMC agreeing to pay $60 million  

in fines. The case also brought the wrong  type of attention onto Fujian's practices.

In October 2018, the first Trump  administration placed Fujian Jinhua  

on an "entity list", locking out their access  to American-made semiconductor technologies.

This crippled their ability to compete with  CXMT - which seemed to have taken greater pains  

to show the legitimate origins of their IP.  CXMT has since sprinted ahead in the market.

## DDR4 to DDR5

You shouldn't feel too comfortable with  

CXMT's five year DRAM gap because  CXMT very quickly made ground.

Throughout 2020, the Hefei fab ramped up  production despite COVID restrictions,  

eventually making about 40,000 wafers  per month by the end of the year.

Sounds like a lot, but that is just  3% of the market. Samsung put out an  

estimated half million wafers per month  in 2020. SK Hynix did 400-450,000 wafers  

per month. And those guys were already moving  on to the next generation of DRAM products.

So CXMT needed more capacity and to catch up.  In 2020, they raised another $2.2 billion from  

China's semiconductor Big Fund to help set  up the next phase of their fab in Hefei.

In 2021, they started working on  the next generation of products,  

DDR5. The memory standards group JEDEC published  

the standard in 2017 and Samsung produced  the first prototypes the following year.

Both Samsung and SK Hynix started shipping DDR5  in volume in mid-2021 using a 10-nanometer class,  

or D1a, process node. Their nodes incorporate  EUV, which is not available to the Chinese.

Despite these shortcomings, CXMT started ramping  its DDR5 products to the market in late 2024,  

closing the gap with the Big Three from 5-7  years to about 3-4. Depending on how you mark it.

Moreover, it is a good product. Despite being  produced with a 1z process node marked somewhere  

between 15-16 nanometers, it has a bit density,  or unit of storage per unit area, of about 0.239  

gigabits per square millimeter. Which is  denser than what Samsung and SK Hynix have.

It was rumored that CXMT DRAM units  were qualified and adopted by major  

Chinese smartphone vendors  like Xiaomi and Transsion.

## Catching Up

How is CXMT catching up so fast?  Several reasons. Some related to  

the company and some related to the product.

First, the company is spending  money like there is no tomorrow,  

investing billions of dollars  into new equipment and R&D.

Second, they hired a huge number of R&D staff  from abroad, mostly South Korea and Taiwan,  

to transfer technology know-how  to their native Chinese staff.

Both CXMT and Gigadevice relentlessly  poached talent from Taiwan and Korea.  

Nikkei Asia reported in 2018 how  they like to target key regional  

managers at existing Taiwanese  memory companies like Nanya,  

offering them a salary package 3-5 times  higher to come and bring over their whole team.

In 2024, the Korean Economic Daily looked up  300 CXMT employees on LinkedIn and found that  

35% of them were Korean - former semiconductor  veterans with many years at Samsung or SK Hynix.

Third, the technical nature of  the DRAM product itself makes  

it hard for the Big Three to build  a lasting moat. DRAMs are produced  

per industry standards like DDR4  and DDR5 to ensure compatibility.  

The specs are decided by industry-wide  organizations like the aforementioned JEDEC.

There is some room to add a special  spin within these specs. For example,  

producers can work out a better internal design,  

add unique features on top of the spec, or  produce the product with a more advanced node.

But in the end, it's still a  DDR4 or DDR5 memory module.  

It is not possible to achieve the same  differentiation as with logic chips,  

which can be built with proprietary  architectures and IP like x86, Arm or even CUDA.

## Going 4F2? Finally and most significantly,

you don't need the most advanced node or  equipment to produce a leading product.

Leading edge DRAMs are produced with nodes in  the low teens to high single-digit nanometers.  

A 10-nanometer pitch is quite achievable  

with older 193-nanometer immersion  technology and quadruple patterning.

So you can buy trailing-edge equipment from  suppliers or second-hand resellers. And then  

push your big, young and hungry R&D staff to  run more experiments across the whole problem  

space to slowly grind out tiny optimizations that  add up. In other words, the DeepSeek approach.

Case in point. The last major architectural change  

came nearly 20 years ago in 2007 when  the industry switched from the 8F2 cell  

to the smaller 6F2 cell to achieve further  cost reductions and grow memory capacity.

For next generation DRAM, the industry is  deciding whether to adopt the 4F2 cell design,  

which is a 33% shrink of the  6F2, and is thus quite similar  

to it. But it does bring a bunch of  serious manufacturing challenges.

In 2023, CXMT researchers presented a fascinating  

paper discussing their work in overcoming  scaling challenges relating to 4F2 cells.

They presented an 8-gigabyte DRAM test chip  fabricated with an access transistor they call  

a Vertical Channel Transistor, but looks  to me like a Gate-all-around transistor.

This is new stuff and more than a few people  noticed. Dylan accused them of breaking sanctions.

CXMT gave a statement to South China  Morning Post saying that this was pure  

research and had nothing to do with current  production. DDR5 is still ramping up,  

but I wonder how soon before we  start seeing some 4F2 leapfrogging.

## Conclusion

CXMT is no doubt an R&D leader, but in terms of  

sheer production capacity they  still have ground to make up.

The market analysis firm TechInsights  estimated that in Q3 2024, CXMT had about  

8% market share. Meanwhile, Samsung will  have 38%, SK Hynix 27%, and Micron 19%.

CXMT is projected to grow that 8%  market share to about 12.5% by 2029,  

thanks to a new Beijing fab they are working  on, but still remain in fourth place.  

They must keep on spending in order to take  ground, but will have plenty of money to spend.

In April 2023, they thought about going public on  

the Shanghai STAR stock exchange at  a valuation of about $14.5 billion.

But in December of that year, they canceled  that due to bad market conditions. So they  

instead raised about $1.5 billion in March  2024 from existing investors GigaDevice,  

Hefei and others at a $19.5 billion valuation.

The Chinese DRAM beast continues to  see its star rise, driven forward  

by relentless R&D and state funds. The Big  Three can't really rely on export controls  

for this one. They need to push harder  and stock up cash for the fight ahead.

And it can now be said without a doubt  that China has a domestic leading edge  

DRAM supplier of their very own. A  fifty year dream, finally fulfilled.