news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

for the korean semiconductor industry, this year is much better than last year.

on the one hand, the prices of dram and nand have recovered. on the other hand, the ai ​​dividend has not yet bottomed out, and hbm is still in short supply.

ict import and export statistics for august released by south korea's ministry of science, technology, information and communications on the 13th showed that semiconductor exports in august increased by 37.6% compared with the same month last year. as demand for semiconductors expands due to the recovery of the artificial intelligence market and it equipment market, total semiconductor exports have grown by double digits for 10 consecutive months. in terms of memory semiconductors, semiconductor export growth increased significantly compared with the same month last year due to increased demand for high value-added products such as high-bandwidth memory (hbm).

but behind this thriving scene, there are always people who are not so optimistic.

damo who reports sorrow but not good news

morgan stanley released a semiconductor industry report titled "winter looms" on the 15th of this month, which triggered an earthquake in the korean storage industry.

in the report, morgan stanley expects sk hynix's dram average selling price (asp) to fall 7.7% next year and 25.0% by 2026. as smartphone and computer sales continue to be sluggish, coupled with unclear demand for artificial intelligence (ai), dram prices are expected to fall starting early next year and lead to an industry downturn.

morgan stanley also predicts that hbm will experience oversupply. the fundamental basis for this view lies in the weakness of the it market. in addition to ai servers, long-term stagnation in demand for smartphones and computers has dragged down the performance of the semiconductor industry. according to market research firm trendforce, global smartphone production is expected to decrease by 5% in the fourth quarter of this year compared with last year, and apple's iphone 16 is also facing poor sales.

in the report, it also significantly lowered the target share price of sk hynix, cutting it to half of the previous target price to 120,000 won, and lowered the investment rating by two levels. after the release of this report, sk hynix closed at 152,800 won on the 19th, down 6.14% from the previous trading day. foreign investors sold a net 361.7 billion won of sk hynix shares.

in fact, this is not the first time that morgan stanley has badmouthed korean semiconductors. the morgan stanley research team has always held a radical view on south korea’s semiconductor sector. the semiconductor industry downward warning reports they issued all have one thing in common, that is, south korea semiconductors have hit a "high point."

in november 2017, morgan stanley shocked investors by publishing "thanks for the memory, time for pause." at that time, even though global it companies entered the "super cycle" and received widespread praise, morgan stanley still warned of the downside risks to the industry. half a year later, morgan stanley once again pointed out the deterioration of the storage semiconductor industry in a message to investors, and continued to warn the next year that "the worst is yet to come."

in 2021, morgan stanley released another report titled "memory-winter is coming". although the dram demand and inventory management of korean semiconductor companies were going smoothly at that time, four months later, morgan stanley issued pessimistic forecasts, which directly caused the collapse of the korean storage industry, and the stock prices of samsung and hynix fell in response.

some korean investors pointed out that the content released by morgan stanley this time is very similar to the 2021 report. the wording and sentence structure in the report are almost the same as those in past reports. questions about the reliability of the report are endless. some critics believe that, once sean king determines that certain conditions for a deterioration in the semiconductor industry are met, he will issue a negative outlook report.

it is worth mentioning that only three months ago, morgan stanley's analyst team also predicted that the stock price of the memory industry will reach a new high this year, and that in the short term, there will be no problems in the industry. at that time, morgan stanley the top stock selected is sk hynix.

the initial push quickly turned into a quick sale. on september 13 this year, morgan stanley’s seoul branch sold 1,011,719 shares of sk hynix. based on the closing price, it was approximately 164.7 billion won, approximately three times the amount sold (351,228 shares) the previous day (12th). that was more than double the amount sold on the day by other major foreign securities firms such as jpmorgan (500,462 shares) and macquarie (209,411 shares).

after completing the large-scale sell-off, morgan stanley immediately lowered the target price and investment view of sk hynix from "overweight" to "underweight".

three months ago, hynix was still being touted, and three months later, it secretly sold its shares and then badmouthed it. this is indeed hard for koreans to accept.

optimistic hynix and micron

although morgan stanley has badmouthed sk hynix, hynix itself is very optimistic.

sk hynix stated in its second-quarter results release conference call in july this year that hbm, which is currently making large-scale investments, has a market structure and mass production characteristics that are different from ordinary dram and cannot be simply compared with past cases.

kim kyuhyun, head of dram marketing at sk hynix, said: "based on past experience, increased investment may trigger concerns about oversupply, but the current investment and production increase are mainly focused on hbm, which is completely different from ordinary dram. therefore, it cannot simply be used ' "increased investment = oversupply' logic." he also pointed out that considering hbm's chip size penalty and relatively low production efficiency, even if investment increases, the increase in bit output will be limited. as hbm generations advance, production constraints will further intensify.

from a demand perspective, sk hynix's prospects are also quite optimistic. kim kyuhyun said: "as competition in the ai ​​industry intensifies, the demand for hbm is increasing sharply, and even if suppliers expand production capabilities, supply shortages continue." in addition, as the demand for ai in various applications continues to emerge, memory requirements will also increase further.

sk hynix expects hbm revenue this year to expand by more than 300% compared to last year. it also predicts that since most of its production capacity is invested in hbm, the supply shortage of ordinary dram may continue.

kim woo-hyun, chief financial officer of sk hynix, said: "if the tight supply continues and the demand for ordinary dram accelerates to recover, the profitability of ordinary dram may surpass hbm." he also pointed out that large-scale investment from 2017 to 2018 cloud data centers are about to enter a refresh cycle. it is expected that the growth rate of server dram except hbm this year and next year will be in the mid-20% range.

this is in sharp contrast to the report issued by morgan stanley. investment banks such as morgan stanley previously believed that as major memory manufacturers such as sk hynix and samsung electronics increase investment, production is bound to increase. morgan stanley predicts that memory manufacturers' capital expenditures on semiconductor manufacturing equipment will exceed us$100 billion next year, a record high, ultimately leading to a situation of "increased semiconductor investment = oversupply".

korean media believe that investment banks such as morgan stanley failed to correctly understand the market structure, especially in the memory market, and increasing capital investment does not mean that output will increase proportionally. for example, hbm's fifth-generation hbm3e product needs to improve its performance every year according to customer requirements, such as increasing the number of chip stacking layers from 8 to 12 layers. starting in 2026, the competition for the sixth-generation hbm4 will also officially begin. with the evolution of products, the difficulty of the process has also increased sharply, which has limited the increase in output.

it should be noted that sk hynix recently announced the mass production of 12-layer high-bandwidth memory 3e (hbm3e), becoming the world's first supplier to mass-produce this memory. it plans to start delivering hbm3e to nvidia later this year. the news quickly drove up sk hynix's stock price.

as the first memory company to emerge from the previous industry downturn, hynix really has nothing to worry about. hbm is still in short supply. the mass production of 12-layer hbm3e is also ahead of other manufacturers. as for dram demand, it is also quite strong. it is a leader in both technology and market, and morgan stanley's badmouthing does not seem to have much impact.

next door, micron is as optimistic as hynix. its latest financial guidance exceeded market expectations and it is expected that revenue in the first quarter of fiscal 2024 will reach us$8.7 billion, higher than the average analyst estimate of us$8.32 billion.

at the same time, micron expects gross margin to rise significantly to about 39.5% and adjusted earnings per share of $1.74, exceeding analysts' expectations of $1.65.

micron's growth momentum also relies on the surge in hbm demand driven by ai. it mentioned in june that its hbm chips have been fully booked in 2024 and 2025.

in terms of the overall hbm market outlook, micron's view is obviously contrary to morgan stanley, which expects the hbm total available market (tam) to grow from approximately us$4 billion in 2023 to more than us$25 billion in 2025.

the company is also making progress on hbm next year. according to its press release, micron expects its hbm, high-capacity d5 and lp5 solutions, and data center ssd products to bring it billions of dollars in revenue in fiscal 2025.

micron also expects its hbm market share to be equal to the company's overall dram market share in 2025.

according to trendforce, micron ranked third in dram revenue in the second quarter of 2024, with a market share of 19.6%, behind samsung's 42.9% and sk hynix's 34.5%.

regarding the latest developments from hbm, after its 8-hi hbm3e entered volume production in february, micron confirmed that it has begun shipping production-capacity hbm3e 12-hi 36gb units to key industry partners, with 12-hi expected to be ramped up in early 2025 hbm3e production proportion and increase the proportion of 12-hi products in overall shipments throughout the year.

micron obviously didn't take damo's words to heart.

is ai a bubble?

although morgan stanley's main argument is that the year-on-year growth of south korea's semiconductor exports has slowed down, the core focus is still on hbm. the weakness of the it market has led to a decline in ai demand, which in turn has affected the hbm market, which is in line with the ai ​​bubble in recent months. the theory is very similar.

in fact, many people believe that although technology giants such as microsoft, google, and meta are competing to invest in ai infrastructure, it is doubtful whether such huge expenditures can be sustained if they cannot launch ai services for the public. some people believe that since the launch of chatgpt in 2022, the ai ​​industry has experienced a rapid peak of attention, but now, just like the "internet bubble" in the past, attention has gradually cooled down.

this view began during google's second-quarter earnings conference last month. although google invests up to 12 billion u.s. dollars (about 16 trillion won) in ai every quarter, the answer to the timing of profitability is still vague, and microsoft's ai cloud business revenue has not met market expectations. how does ai make money and how much can it make? , is still a difficult question to answer.

although profits are far away, the capital expenditures of the four major technology giants (microsoft, google, tesla, and apple) have reached a new high this year, with a total expected to reach 206 billion u.s. dollars. nvidia’s next-generation ai chip “blackwell b200” has design flaws this resulted in production delays and further contributed to the spread of ai bubble theory.

correspondingly, hbm, which is closely related to ai, has also come to the forefront. if the ai ​​bubble theory is true, the currently prosperous hbm market will soon follow in the footsteps of the standard dram market.

needless to say, hynix needs to say more. nvidia ceo huang jensen also refuted the ai ​​bubble theory. he said at the goldman sachs technology conference held in the united states this month: "there are too many demands. everyone wants to be the first. we are doing our best to meet them." supply."

the korean industry believes that the "blackwell" design defect incident is actually good news. due to technical problems with blackwell, which is expected to be released at the end of the year, nvidia made some changes to its design. it is reported that the company changed the original hbm3e (fifth generation) 8-layer stack product to the hbm3e 12-layer stack product. as the number of hbm stacking layers increases, the technical requirements also increase accordingly, which is a big plus for sk hynix, samsung electronics and micron, which can manufacture hbm.

a korean securities researcher said: "during the internet bubble period, the five-year average price-to-earnings ratio of the leading company cisco was 37 times, with the highest price-to-earnings ratio of 132 times. in comparison, nvidia's five-year average price-to-earnings ratio from 2019 to 2023 was about 40 times , the current price-to-earnings ratio has dropped to just over 30 times due to recent adjustments.”

another researcher said: "unlike the internet bubble period, most ai-related companies have established healthy cash flows based on profit models and are making new ai investments on this basis. the current ai industry is not going from 0 to 100, but the process from 100 to 130, so it is different from the internet bubble.”

it is not clear whether ai is a bubble, but for the memory industry, it is indeed a bit dependent on the future. as long as the giants are willing to spend money, hbm will not have to worry about selling. once the giants reduce investment, korean manufacturers will suffer immediately.

a korean factory that can’t help itself

for samsung and hynix, hbm is a rare ray of light after experiencing multiple dram ups and downs cycles. it has not plunged like dram prices in the past two years, but is in short supply like nvidia's gpus. although the overall market the proportion of hbm in china is not large, but it can already contribute considerable profits.

this is why hynix and samsung have frequently mentioned customized hbm this year. so far, memory is still a general market that is supplied according to strict standards. customers such as nvidia have been manufacturing ai in accordance with the hbm specifications released by sk hynix and samsung electronics. accelerator, in essence, hbm is still a general product similar to dram.

however, starting from hbm4, the situation is different. previously, the logic layer under the stacked dram only served to connect the gpu and dram. but starting from hbm4, it can be equipped with system semiconductors such as low-power functions required by customers. function.

this is a major benefit for korean manufacturers. customization is obviously more profitable than generalization. the general market often means price wars, but the right to speak on customization is in the hands of the memory manufacturers themselves, and this this is also one of the reasons why micron was strongly opposed before. this kind of customization is bound to further strengthen the voice of market leaders, and the lower-ranked manufacturers will only become more marginalized.

fortunately, the good news released by hynix and micron has cleared up the haze of morgan stanley's report, but the problem still exists: customized hbm can certainly make a lot of money, but if the orders of the ai ​​giants are reduced, the extra production will be who can pay for the product?