AI 硬件：第二波浪潮（2025-2027 年）.pdf

1、For Reg.AC certification and important disclosures,see the last page(s)of this report.AI Hardware:The Second Wave(2025-2027)11 September 2024Brett Simpson,AnalystArete Research Services LLP+44(0)20 7959 1320Janco Venter,AnalystArete Research Services LLPJ+27 71216 3220Nam Hyung Kim,AnalystArete Rese

2、arch,LLC +1 424 228 79142Key Summary TakeawaysWe see a second wave of AI growth coming in the 2025-2027 period.Training clusters are scaling up,while we see AI inference demand layering in as new architectures drive material reduction in cost per token.See slides 3-12.The speed of technology change

3、is frightening in the 2025-2027 period.Back-side power,102T Ethernet,serdes to 400G,Next generation COWOS,UALink,wafer scale technology,HBM4,use of photonics,etc.See slide 4.We review industry capacity plans through 2027 for both foundry and HBM,and see a fresh round of LTAs ahead.HBM4 pricing likel

4、y to jump 50%on a per GB basis.COWOS could reach 100k wpm by YE27.See slides 5-7.We look at the challenges around monetization and the opportunities ahead from enterprise,consumer,SaaS co-pilots,and government.Hyperscaler capex this year is broadly capacity add next year.See slides 9-11.With smartph

5、ones and PCs embracing EDGE AI starting in 25,we look at the opportunities and challenges in scaling this up over the next few years.See slide 12.Who can fund the buildout of 1m accelerator clusters?We believe all major CSPs are planning to build GW sites to realise this hardware scaling capability

6、by decade-end.3The Exponential R&D Arms Race Continues2025-20272028-20301M XPUs?100K B200s$50-100bncapex?Cluster Sizes continue to scale exponentially the next five years to support next gen foundation models.Leading-edge model makers need access to this capability at each generation.We are heading