
AI-generated summary
The demand for AI servers continues to expand, causing HBM to compete with traditional DRAM for limited advanced process and wafer production capacity. It will still take time to improve the yield and output of new-generation products.
Memory supply is expected to remain tight in 2027. (Diagram, Reuters)
[Reporter Zhuo Yijun/Taipei Report] According to the latest research from TrendForce, the demand for AI servers continues to expand, and HBM and conventional DRAM is still competing for limited advanced process and wafer production capacity. In addition, it will still take time to improve the yield and output of new-generation products. It is expected that memory supply will continue to be tight in 2027. With the overall memory price maintaining an upward trend, suppliers are optimistic about HBM in 2027. The attitude towards price increases has also become clearer. In order to reflect the continued tightness of overall supply, coupled with the increase in the proportion of high-priced HBM4 and the gradual increase in volume of HBM4e in the second half of 2027, TrendForce has revised its HBM price outlook for 2027, and is expected to blend. The annual growth rate of ASP will reach 121%.
TrendForce pointed out that recently, GPU and ASIC manufacturers have been discussing reducing HBM mounting capacity. The main consideration is not the weakening demand for terminal computing, but the limited supply of HBM and the rapid increase in overall system costs. Compared with the previously planned 12hi HBM products, 8hi requires a smaller number of DRAM dies, which helps reduce the cost of a single stack and the overall material cost of the GPU. It can also support more GPU shipments with the same HBM bit output. Therefore, as HBM supply continues to be tight, 8hi is expected to become a priority evaluation solution for many GPU and ASIC manufacturers in 2027, as a compromise between performance, supply and cost for GPU manufacturers.
Please read on...
However, the HBM capacity reduction does not mean that the price per unit capacity (per Gb) will be simultaneously reduced. Since each HBM needs to be equipped with a base die, the related costs will not decrease proportionally as the number of stacking layers decreases; the DRAM capacity of the 8hi product that can amortize the cost of the base die is lower than that of 12hi, so the converted per-Gb cost is higher. TrendForce predicts that the per Gb price of 8hi products in 2027 may be about 10-20% higher than that of 12hi.
Overall, TrendForce believes that the HBM market will continue to be in supply shortage in 2027. Under the dual considerations of tight supply and rising costs, the penetration rate of 8hi specifications will further increase, becoming a priority solution for more GPU and ASIC manufacturers to evaluate; however, its per Gb cost structure is at a premium compared to 12hi, which is still an important variable that customers need to include in their evaluation when negotiating specifications and supply chain layout. Therefore, although HBM capacity adjustment in 2027 can partially alleviate GPU cost pressure, it will be difficult to reverse the trend of substantial increase in HBM overall price.
Grasp the economic pulse with one hand. Click here to subscribe to Free Finance Youtube channel
AI outlook — possibilities, not facts
The annual growth rate of HBM blended ASP will reach 121% in 2027
Likely · Within years
8hi HBM will become a priority evaluation solution for many GPU and ASIC manufacturers in 2027
Likely · Within years
The per Gb price of 8hi products in 2027 may be about 10-20% higher than that of 12hi
Possible · Within years

On September 29, the State Council Information Office held a press conference to introduce the progress of building a scientific and technological power during the "15th Five-Year Plan" period. Yin Hejun, Minister of Science and Technology, said that my country's scientific and technological innovation has leapt to a new level in the past five years. In 2025, the entire society's R&D investment will exceed 3.92 trillion yuan, accounting for 2.8% of GDP.

The China Meteorological Administration released the "Meteorological Data Management Measures" today, which will be officially implemented on November 1. This measure establishes rules and regulations around data supply, circulation and utilization, safety assurance, supervision and management, etc., and fills the gaps in the system of meteorological data confirmation and authorized operation.

Experts pointed out that the current specific progress of AI in the field of cancer is mainly reflected in assisting early detection and diagnosis, such as discovering subtle clues of pancreatic cancer from CT images that were judged to be normal in the past. However, they emphasized that cancer is complex and diverse, and AI cannot solve biomedical problems by relying on computing power alone. Experimental verification and clinical trials are still needed.
Yang Jun, deputy director of the Guangdong Provincial Department of Science and Technology, announced that the Guangdong Provincial People's Government has decided to hold the International Humanoid Robot Basketball Challenge (RBA) in Guangzhou in mid-to-early April 2027, aiming to promote the integrated development of science and technology and industry in cutting-edge fields such as artificial intelligence and humanoid robots. The event consists of competitive competitions and exhibition competitions, which are open to the world. Domestic and foreign companies, universities, etc. can participate.

Nvidia announced a record US$150 billion share buyback and launched the Open Agent Safety Platform, with CEO Jensen Huang urging US-China collaboration on AI safety despite political tensions, emphasizing that competitors can still work together to prevent AI harm.

When Jen-Hsun Huang attended the annual dinner of the South Korea-US friendship non-profit organization "Korea Association" in New York and received the Van Fleet Award, he once again emphasized that AI is not a living organism, but just a software. He opposed giving it too many human characteristics and called for an alignment mechanism to ensure the safe and responsible development of AI.