Edge AI Penetration And Enterprise Investment To Accelerate Adoption of GenAI
Generative AI’s development is approaching a new stage, with investors now focusing on investment from enterprise and development of edge AI given concerns of slowing cloud investment from 2025. However, we believe cloud investment growth will remain strong as NVIDIA’s accelerated product cycle, coupled with the introduction of GenAI’s application from edge AI, may accelerate the development and investment for data centre hardware. Maintain OVERWEIGHT on the sector.
WHAT’S NEW
- Generative AI’s development is approaching a new stage in 2024. In 2023, Generative AI’s was almost entirely led by cloud and training applications in 2023 as most AI developers were building their own state-of-the-art LLM from the ground up. With the basic framework and foundation models more readily available, we are now approaching the start of the implementation stage where key users like enterprises and end-device OEMs are starting to integrate these existing foundation models into their operations and products. As such, we are expecting investments from enterprise to pick up significantly to bolster growth of the AI server market in 2024. At the same time, AI-enabled devices will drive a significant replacement cycle from 2H24 onwards, as well as spreading the reach of AI applications to a broader audience.
- Cloud investment growth will remain elevated. With investors turning their attention to enterprise and edge AI, the Greater China cloud hardware plays have registered significant price corrections from the peak seen during mid-23. However, it is unreasonable to discount the growth opportunities from cloud, as we expect cloud investment to remain elevated on the back of NVIDIA’s expedited product launches and shipment ramp-up. In Oct 23, NVIDIA released its new product launch roadmap in which its data centre business’ product cycle was shortened from two years to one year. Given that these new products offer a much better performance (2-3x faster vs previous generations), we believe this will speed up the development rate for AI and investment cycles pertaining to cloud. This, coupled with the robust production ramp-up (around 2x by 2024) for high-end chips, and an accelerating adoption of GenAI thanks to the penetration of AI-enabled devices, could keep cloud AI investment elevated in the foreseeable future.
- Data centre GPU supply chain to benefit from robust shipment ramp-up. As mentioned above, NVIDIA and its suppliers are building up the production volume of data centre GPUs. The global capacity of high bandwidth memory (HBM) and Chip-on-Wafer-on-Substrate (CoWoS) – two key bottlenecks in the supply chain – is expected to grow 105% and 185% respectively from 2023-24. This development, coupled with the market’s robust order backlog and NVIDIA’s accelerated product launches, could see NVIDIA’s high-end data centre GPU server shipment surge 82% and 53% yoy in 2024 and 2025 respectively. On the other hand, we also expect inference servers’ shipment (L40/L40s/B40) to accelerate from 2024 onwards with a growth of 73.1% and 77.8% yoy in 2024 and 2025 respectively, driven by demand from enterprises as well as inference applications at cloud. This will benefit NVIDIA’s GPU supply chain (eg GPU module and baseboard assembly, HBM suppliers and foundries), its key ODM partners (mostly supplies to AI startups and CSPs) as well as OEM players (mostly supplies to enterprise).
- GP servers to recover in 2024. GP (general purpose) server investment fell in 2023 due to: a) both CSPs and enterprises reducing spending on GP infrastructures amid a tough macro economy, and b) CSPs’ investment leaning heavily on AI servers. As such, GP server shipment should have registered a 9.8% yoy decline in 2023. However, with destocking largely done, and a stronger-than-expected application of GP servers in GenAI, we are now more positive on the recovery of GP servers from 2024 onwards. Enterprise clients will start to resume spending on cloud infrastructures, while the major CSPs from the US and China have all indicated limited further downsides in capex for GP servers. As such, we now expect GP servers to register a positive 4.6% and 2.6% recovery in spending in 2024 and 2025 respectively (previous estimates: -1%/-0.5% yoy), benefitting both server ODMs (Inventec, Wistron) and branded server names (HP, Dell, Lenovo).
ACTION
- Maintain OVERWEIGHT. We expect the GPU server market to grow 159.7%/85.2%/56.8% yoy in 2023-25 respectively, on the back of a 101.9%/70.0%/55.7% yoy growth in shipment volume. We believe the GPU server market will be driven by both training and inference applications going forward, compared with the largely training-driven market in 2023.
- Recommend BUY on Lenovo to capture the edge/enterprise opportunities. As the largest PC brand globally with 60% exposure to commercial segment, Lenovo will be one of the earliest beneficiaries of the spread of AI-enabled devices. Despite its strong share price performance in the past six months, Lenovo’s current valuation of 8.6x FY25F PE is still undemanding given its exposure to multiple AI investment trends. We raise our target price to HK$12.10, based on a higher multiple of 9.9x FY25F PE, still on par with peers’ average valuation.