As generative AI models like OpenAI sweep across the globe, the demand for high-performance GPU has been on the rise. For instance, NVIDIA’s GPU products are in tight supply and often reportedly out of stock. Industry analysis reveals that NVIDIA’s high-performance GPU is manufactured by TSMC and packaged using CoWoS technology. However, TSMC’s advanced CoWoS packaging capacity is insufficient to meet the AI demands. Later, TSMC actively expanded CoWoS production, and NVIDIA’s GPU shipments gradually stabilized as capacity increased.

Notably, new challenges present following the ease of GPU shortage problem.

Recently, Mark Zuckerberg, CEO of Meta, stated publicly that the GPU shortage in AI data centers is being alleviated, while the future bottleneck will be power supply.

As GPU supply become less tight, companies kick-start big investment in building data centers and other infrastructure facilities to be in tune with the AI development trends.

Zuckerberg believes that the next crux for AI development will be power supply. He noted that many new data centers could consume 50-100MW, and large data centers could reach up to 150MW. As the scale of data center power consumption continues to grow, AI industry may hit a power supply bottleneck.

The energy industry, unlike the AI sector, does not find building new power plants a simple task. Given factors such as regulations (Especially nuclear energy), power transmission planning, and construction, it could take several years from planning to the actual integration of new power into the grid. Therefore, capital investments fail to yield results in a short time, and the delivery of additional power supply lags far behind the construction of data centers.

To address the potential future power crisis, media reports indicate that Meta is currently collaborating with Silicon Ranch, a solar energy company in Georgia, to supply power to its data centers.

Likewise, Tesla CEO Elon Musk also issued a warning regarding AI development in April 2024, stating that the next shortage facing AI will be power, and there may not be enough electricity to run all the chips next year.

Besides, OpenAI CEO Sam Altman has also stated that AI will consume more power than expectation, and future AI development will require breakthroughs in clean energy.

Thanks to the surge in AI demand, Samsung and SK Hynix, the world’s leading memory chip manufacturers, anticipate the prices of DRAM and high-bandwidth memory (HBM) to remain strong this year due to increasing demand for high-performance chips. According to the Korea Economic Daily, Samsung and SK Hynix have converted over 20% of their DRAM production lines to produce HBM.

Last week, SK Hynix CEO Kwak Noh-jung announced at a press conference that their HBM chips have already been sold out for 2024 and are nearly sold out for 2025.

During an investor relations meeting hosted by Samsung Securities on May 9th and 10^th, a SK Hynix official stated that the company’s HBM chips are supplied through binding annual contracts, which detail supply volumes, payment methods, and deadlines.

The price drop in HBM3 will be offset by the price increase in HBM3e, so the gross profit margin is expected to remain stable in 2024, according to the Korea Economic Daily, citing the SK Hynix official. It is also reported that NVIDIA is the major customer of SK Hynix’s eight-layer HBM3e.

According to the report, a Samsung IR executive also stated that the company’s HBM production has been sold out. Based on the current supply and demand situation, Samsung has predicted that HBM will not be in oversupply in 2025.

Samsung noted that the gap between its eight-layer stacked HBM3e and SK Hynix’s is closing, claiming it has taken the lead with 12-layer HBM3e.

In late March, South Korean media Alphabiz reported that Samsung may exclusively supply 12-layer HBM3e to NVIDIA, indicating NVIDIA is set to commence large-scale purchases of Samsung’s 12-layer HBM3e as early as September.

In terms of the price outlook for traditional DRAM and solid-state drives (SSDs), both SK Hynix and Samsung remain positive.

According to the latest forecast by Trendforce, the HBM market is poised for robust growth, driven by significant pricing premiums and increased capacity needs for AI chips. HBM prices are expected to Increase by 5–10% in 2025.

In terms of market value, HBM is projected to account for more than 20% of the total DRAM market value starting in 2024, potentially exceeding 30% by 2025.

(Photo credit: SK Hynix)

Samsung’s planned launch of its budget model version for the upcoming Galaxy Z Fold 6 foldable phone may be cancelled, according to information obtained by South Korea’s media outlet TheElec.

According to sources cited by TheElec, component suppliers to Samsung, which were gearing up for mass production of parts for these foldable devices, indicated that they have not received orders for the budget model Fold 6.

Initially, the South Korean tech giant had planned to unveil three foldable phones this summer: the Galaxy Z Fold 6, the Galaxy Z Flip 6, and an affordable version of the Fold 6.

The Fold 6 features a digitizer layer on the screen to support the S-Pen stylus, but the budget model was intended to be thinner and omit this feature.

It is reported that Samsung, after conducting a durability test of 200,000 folds for the budget model, with water and dust resistance functions added, has founded out that the product turned out not to be thinner than its Chinese rival Huawei’s equivalent, thus lacking significant differentiating points from these rival products.

Huawei’s Mate X3 and X5, launched last year, were both around 11.08mm thick, according to information obtained on Huawei’s website. Samsung’s Galaxy Z Fold 5 was 13.4mm thick, while Xiaomi’s Mix Fold 3 was 10.9mm thick.

According to analysis from TrendForce, global shipments of foldable phones reached 15.9 million units in 2023, marking a 25% YoY increase and accounting for approximately 1.4% of the overall smartphone market. In 2024, shipments are expected to rise to about 17.7 million units, growing by 11% and slightly increasing the market share to 1.5%. However, this growth rate remains below market expectations, with the segment’s share predicted to exceed 2% only by 2025.

Trendforce points out that the slowdown in the growth of foldable phones could be attributed to two main factors: Firstly, consumer retention is low due to frequent maintenance issues faced by first-time foldable phone users, leading to a lack of confidence in the product. As a result, users may opt for high-end flagship smartphones when considering replacements. Secondly, the current price points of foldable phones have yet to reach the sweet spot for consumers, making it challenging to meet sales targets based solely on pricing.

(Photo credit: Samsung)

SK Hynix has updated its latest timeline regarding HBM4E, the company’s 7th generation high bandwidth memory, on Monday. According to reports from Wccftech and TheElec, the memory heavyweight’s generation changes of two-year gap has been shortened to one, and is now on track to launch HBM4E by 2026.

The news was revealed by Kim Gwi-wook, head of SK Hynix’s HBM advanced technology, during the International Memory Week 2024 held in Seoul yesterday. Earlier in February, Chun-hwan Kim, Vice President of SK Hynix, stated that the company plans to commence large-scale production of HBM4, their sixth generation of the HBM family, in 2026, as well.

According to TheElec, HBM4E will be the first chip from SK Hynix to be made through its 10-nanometer (nm) class Gen 6 (1c) DRAM. It is reportedly to be made of 32Gb DRAM and use 1c DRAM as the core die.

According to Wccftech, HBM4E’s feature bandwidth is said to be 1.4 times higher than that of its predecessor, HBM4. This indicates a significant enhancement in power efficiency, providing a preview of the advancements we can anticipate in next-gen AI accelerators.

Earlier in April, SK Hynix announced it has recently signed a memorandum of understanding with TSMC for collaboration to produce next-generation HBM and enhance logic and HBM integration through advanced packaging technology. The company plans to proceed with the development of HBM4, or the sixth generation of the HBM family, slated to be mass produced from 2026, through this initiative.

Samsung, SK Hynix’s major competitor on memory, also schedules to start mass production of HBM4 in 2026.

(Photo credit: SK Hynix)

TrendForce’s latest findings revealed that as of 2024, Taiwan is expected to lead the global semiconductor foundry capacity in advanced manufacturing processes (including 16/14nm and more advanced technologies) with a 66% market share, followed by Korea (11%), US (10%), and China (9%). However, the semiconductor production capacities of advanced nodes in Taiwan and Korea are projected to decrease to 55% and 8%, respectively, by 2027.

It is worth noting that though semiconductor heavyweights including TSMC, SK Hynix and Samsung keep raising the amount of investments,  Taiwan and Korea, as the two countries holding the highest market share in advanced nodes, are expected to fall in their market shares.

On the other hand, in the US, where the government has been pushing incentives and subsidies more aggressively, its global capacity share in advanced manufacturing processes is expected to jump from 10% in 2024 to 22% in 2027.

Per the overall foundry capacity, Taiwan is expected to hold approximately 44% of global market share as of 2024, followed by China (28%), South Korea (12%), the US (6%), and Japan (2%). The overall trend is expected to be in line with advanced nodes. In contrast, the overall semiconductor production capacities of Taiwan and South Korea are projected to decrease to 40% and 10%, respectively, by 2027.

China, where foundries focus more on expanding mature process capacities and are backed by government subsidies, is projected to perform relatively strong in the overall global market share, growing from 28% in 2024 to 31% in 2027. Its market share of the matured process (including ≥28nm nodes) capacity is expected to rise from 33% in 2024 to 45% in 2027.

According to a report from THE CHOSUN Daily on May 10th, citing forecast from The Semiconductor Industry Association (SIA) and The Boston Consulting Group (BCG), in 2022, Taiwan and Korea held 69% and 31% shares of the production of the most advanced semiconductors below 10 nanometers, while their market share on the advanced nodes may fall to 47% and 9% in 2032, respectively.

The report as mentioned earlier stated that the dramatic decline in South Korea’s semiconductor production share is primarily attributed to key players like Samsung Electronics and SK Hynix, who currently dominate the global advanced semiconductor market alongside TSMC. However, instead of investing in South Korea, they have opted to establish their latest factories in the United States.

The US government announced earlier in April that it would provide up to USD 6.4 billion in subsidies to Samsung for expanding advanced chip production capacity at its Texas plant. In addition, SK Hynix plans to spend $3.87 billion building an advanced packaging plant and research center for artificial intelligence products in Indiana.

(Photo credit: Samsung)