云加速器研究-Blackwell 业务扩张，价格保持稳定-Cloud Accelerator Study_ Blackwell Broadens Out, Pricing Holds Up

Summary of Key Points from the Conference Call Industry Overview - The report focuses on the GPU cloud pricing and availability within the semiconductor industry, particularly regarding AI demand and cloud service providers like AWS, Google Cloud, and Azure [2][4]. Core Insights and Arguments - AI Demand Environment: There are ongoing investor concerns about the durability of AI demand, prompting a revisit of GPU cloud pricing and availability [2]. - Availability of Accelerators: - The B200 accelerator is now more widely available, with spot instances appearing at AWS and GCP for the first time in November 2025 [4]. - The B300 has also been spotted at AWS, indicating faster market penetration compared to the B200 [4]. - Pricing Trends: - Pricing for older NVDA generation GPUs has seen a 1.8% month-over-month decline, while prices for newer models like the H100 and H200 have increased by 3.3% and 1.2% respectively [4]. - The pricing for older accelerators remains stable, suggesting that cloud vendors still find economic value in these legacy chips [2][4]. - AMD's Market Position: There is limited traction for AMD's offerings, with no instances available across the covered clouds, although some manual checks indicate availability at Oracle [4]. Additional Important Information - Legacy GPU Availability: Older generation GPUs, including Ampere and Hopper, continue to be widely available, with no significant decline in their location counts [4]. - Google and Amazon ASICs: Google’s TPU and Amazon’s Trainium are available at stable prices, although Trainium2 pricing is noted to be volatile [4]. - Competitive Landscape: The report highlights the competitive dynamics between NVIDIA and AMD, with NVIDIA maintaining a strong position in the market despite AMD's potential for growth in cloud and AI sectors [54][55]. Data Highlights - Spot and On-Demand Pricing: The report provides detailed pricing comparisons for various accelerators, indicating significant premiums for on-demand pricing over spot pricing, with some instances showing premiums as high as 6.84x for H100 [7][11][32]. - Performance Metrics: The theoretical performance and price/performance ratios for key accelerators are analyzed, showing NVIDIA's GPUs generally outperforming AMD's offerings in terms of price efficiency [37][44]. This summary encapsulates the critical insights from the conference call, focusing on the semiconductor industry's current state, particularly in the context of AI demand and cloud services.