The growing longevity of graphics processing units has emerged as a key trend in the AI hardware market, with industry leaders highlighting shifts in depreciation patterns. NVIDIA CEO Jensen Huang stated at the GTC Berlin conference that modern GPUs now maintain productive utility for eight years or longer, a significant extension from earlier four-year expectations. This durability stems from software programmability and continuous performance enhancements throughout a unit's lifecycle, enabling sustained relevance across diverse applications.

Elon Musk reinforced this perspective during SpaceX's Q3 2026 earnings discussion, noting that advancing AI model capabilities increase 'useful intelligence per watt' over time. He observed that older GPU generations, such as H100 units, now command higher rental rates than when initially deployed, reflecting their enduring value in compute-intensive workflows. The SpaceX chief emphasized that compute value scales proportionally with AI efficiency gains, stating: 'The value of that compute is proportionate to the useful intelligence per watt.'

Analyst Gavin Baker, recognized for semiconductor market insights, pointed to pricing dynamics where GPU rental costs have risen sharply despite declining token prices. Baker noted that increased token efficiency per GPU hour has driven up the value of compute time, with rental rates climbing 'vertical' as AI workloads become more optimized. This trend aligns with broader market observations about residual value retention.

Market data indicates NVIDIA's older A100 and H100 models currently trade at premiums exceeding standard depreciation schedules, with B200 units fetching 158 percent of their original cost. This depreciation pattern contrasts with traditional linear scheduling models, as retired GPUs transition through secondary markets rather than becoming obsolete. The sustained demand for compute resources is further underscored by financial metrics showing AI infrastructure investments yield payback periods under one year, with $40 billion annual revenues from $50 billion in dedicated compute capacity.

These developments reflect a structural shift in how computing hardware maintains economic value, driven by software-hardware synergy and evolving AI workload requirements rather than hardware obsolescence cycles.