During a recent interview, Mark Zuckerberg highlighted the enduring value of AI chips, noting that even used H100 GPUs have increased in worth over the past year. This insight contrasts with the common belief that AI hardware rapidly loses value, a notion that has fueled much of the current market speculation. Zuckerberg's comments, made while discussing data center operations, suggest that owning such hardware offers long-term financial advantages, akin to holding real estate rather than a depreciating asset.
Investors often assume AI chips are like perishable goods, losing most of their value within a few years. However, Zuckerberg's perspective implies that these chips may have a longer useful life, especially as demand for compute power remains high. His remarks align with historical precedents where infrastructure investments, like canals and railroads, initially appeared speculative but proved durable over time. This analogy suggests that AI hardware could similarly outlast initial expectations, providing ongoing value to those who own it.
Meta's accounting practices further support this view. The company extended the estimated useful life of its servers and network equipment to 5.5 years, significantly impacting its depreciation costs and net income. This change highlights the financial implications of viewing AI hardware as a long-term asset rather than a short-lived expense. In contrast, cloud renters face ongoing costs without the flexibility to repurpose or retain the hardware, underscoring the advantages of ownership.
The debate over chip depreciation is not without dissent. Amazon recently shortened the useful life of its servers, citing rapid technological advancements. This divergence in perspectives highlights the uncertainty surrounding AI hardware's lifecycle. While scarcity may currently justify higher valuations, the introduction of more efficient chips could shift the balance. The market's cautious stance reflects this uncertainty, with investors closely watching utilization rates and used chip prices for signals.
Ultimately, the industry is still determining how to classify AI hardware. If older chips maintain value and find new applications, they could bridge the adoption gap by making AI more accessible to smaller businesses. This scenario would accelerate AI integration across various sectors, driven by the durability and reusability of these powerful chips.