This big AI bubble argument is wrong

Concerns surrounding the potential collapse of the AI market have surged recently, with many investors fearing an impending bubble. A significant point of contention involves the value of GPUs—critical components for training and executing AI models. As new models are released, some worry that older GPUs will rapidly depreciate, leading to substantial financial losses for cloud companies. This view, however, is misleading. In March, I highlighted the depreciation risks faced by some AI firms, including CoreWeave. By August, investor Jim Chanos echoed similar fears, particularly regarding the lifespan of GPUs. The consensus suggests that while cloud providers account for GPU depreciation over five to six years, the actual useful life may only be one to two years. Such a mismatch could indeed spell trouble for the AI sector's earnings in the future. Contrary to this prevailing belief, experts argue that GPUs remain viable for much longer than anticipated. Stacy Rasgon, a prominent chip analyst at Bernstein, asserts that GPUs can effectively operate for about six years. He emphasizes that the depreciation practices of major hyperscalers align well with this reality. Operating costs for GPUs in AI data centers are notably low, especially when compared to the market rates for renting GPUs through cloud services. As a result, the profit margins from utilizing older GPUs are substantial, even amid continuous advancements in GPU technology. According to Rasgon and his colleagues, vendors can maintain healthy profits on five-year-old A100 models, suggesting that a five to six-year depreciation period is indeed justifiable. Delving deeper into why these GPUs retain their value, Matt Rowe, senior director at AI cloud provider Lambda, explains that the effective lifespan of GPUs can extend to seven or even eight years. While many companies adhere to a six-year depreciation schedule, strategies such as warranty extensions and redeployment are effectively prolonging the lifecycle of these assets. Rowe points out that warranty contracts—which often last five years—ensure that failing GPUs are replaced, thereby enhancing the longevity of the entire GPU fleet. He notes that older models like Nvidia's K80s, P100s, and V100s typically outlast their expected lifespan, demonstrating the resilience of these technologies. Additionally, Erwan Menard, senior VP at Crusoe, elaborates on the lifecycle of GPUs within their operations. He explains that GPUs are not only used for cutting-edge AI model training but also for less demanding inference tasks, allowing older models to remain productive for years. This adaptability highlights the extensive range of workloads that can still be efficiently handled by older GPUs. AI cloud providers often assess user expectations and budgets when selecting GPUs. Menard illustrates this with an example of an AI service offering a free tier that utilizes older, less powerful hardware, ensuring an initial user experience that can later be upgraded as customers convert to paid services, which demand more advanced GPUs. As the industry increasingly turns to cost-effective open-source models, older GPUs might see even more utilization. While energy efficiency is a legitimate concern, Menard argues that the overall cost-effectiveness of older GPUs often makes them the preferred choice, as they remain cheaper to operate despite their higher energy consumption. In summary, the notion that GPUs are rapidly losing value may not hold water. With thoughtful strategies in place, the AI sector can continue to leverage older GPU technologies effectively, mitigating fears of a looming bubble.