TrendForce: H200 expected to become NVIDIA's main supplier after Q3 2024
TrendForce predicts that due to the increasing demand for the Hopper GPU, NVIDIA's H200 GPU will become the main supply product after the third quarter of 2024. It is expected that NVIDIA's overall revenue will reach $30 billion in the second quarter of the 2025 fiscal year, with a year-on-year growth rate of 154%. Nearly 90% of GPU products will belong to the Hopper platform, with the H100 maintaining a non-price reduction strategy to drive market share for the H200. The demand for AI servers equipped with H20 in the Chinese market is expected to significantly increase starting from the second quarter of 2024
According to TrendForce's latest research, due to the increased demand for NVIDIA's core product Hopper GPU, NVIDIA's data center business is expected to drive the company's overall revenue to more than double in the second quarter of the 2025 fiscal year, reaching $30 billion. Supply chain investigations show that CSPs (cloud service providers) and OEM customers will increase their demand for H200 in the near future, with this GPU expected to become NVIDIA's main supply after the third quarter of 2024.
NVIDIA's financial report shows that the data center business revenue in FY2Q25 increased by 154% year-on-year, outperforming other businesses, leading to an increase in the overall revenue contribution to nearly 88%. TrendForce predicts that by 2024, nearly 90% of NVIDIA's GPU product line will belong to the Hopper platform, including H100, H200, special edition H20, and the GH200 solution integrated with their own Grace CPU, mainly targeting specific applications in the HPC (high-performance computing) and AI markets. It is expected that starting from the third quarter of this year, NVIDIA will not reduce prices for H100, and after fulfilling old orders, H200 will become the main market supply. In the Chinese market, NVIDIA's demand for AI servers equipped with H20 is expected to significantly increase from the second quarter of 2024 due to cloud customers deploying large-scale language models (LLMs), search engines, and chatbots locally.
TrendForce expects that as the demand for AI servers equipped with H200 increases, it will fill the gap left by the possible delay in shipping of the Blackwell new platform due to the need for supply chain readiness. It is estimated that the revenue of NVIDIA's data center business in the second half of this year will not be greatly affected. Additionally, it is projected that NVIDIA's Blackwell platform will be officially mass-produced in 2025, with a die size twice that of the existing Hopper platform, driving the growth of CoWoS demand after becoming mainstream next year.
TrendForce stated that TSMC, the main supplier of CoWoS, recently raised its monthly production capacity plan to the end of 2025 to nearly 70-80K, close to doubling the capacity in 2024, with NVIDIA expected to account for more than half of the capacity.
In NVIDIA's product roadmap for this year, H200 is the first GPU to adopt HBM3e (fifth-generation high-bandwidth memory) 8Hi (8-layer stack), and the subsequent Blackwell series chips will also be fully upgraded to HBM3e. Micron and SK hynix completed HBM3e verification at the end of the first quarter of 2024 and started mass production in the second quarter. Micron's products are mainly used for H200, while SK hynix supplies both H200 and B100 series. Although Samsung introduced HBM3e later, it has completed verification and officially started shipping HBM3e 8Hi, mainly for H200, with verification work on the Blackwell series steadily progressing