Nvidia CEO claims AI chips are the lowest risk asset for lenders

Redefining hardware lifecycles

Nvidia CEO Jensen Huang is challenging the standard timeline for how long computing hardware remains useful. In the past, graphics processors typically followed a four-year lifecycle before becoming obsolete. Huang argues that AI accelerators do not follow this same pattern because of their unique performance and versatility.

During a conversation with Yahoo Finance's Deirdre Bosa at the GTC conference in Berlin, Huang described the company's chips as 'productive,' 'fungible,' and 'durable.' He suggests that these systems can stay in active use for eight years or even longer. This longevity is driven by continuous software improvements that keep the hardware relevant as technology evolves.

By extending the expected useful life of these chips, Nvidia is attempting to change how the market perceives the depreciation of its products. Instead of seeing them as rapidly aging tools, Huang wants the industry to view them as long-term infrastructure. This perspective is intended to influence how investors and lenders evaluate the risks associated with AI spending.

A new class of collateral

Huang makes a bold claim regarding the financial security of these chips, stating that the risk of lending against Nvidia hardware is 'the lowest of any asset.' This statement suggests that the hardware maintains significant residual value across many different customers and types of workloads. Because the chips can be repurposed for various tasks, they do not lose their utility as quickly as other types of hardware.

The CEO believes that financiers have not yet fully grasped the flexibility of these AI accelerators. Because the chips can be moved between different applications, they could serve as highly attractive collateral for loans. This would allow companies to borrow more money against their existing hardware to fund further expansion.

If lenders begin to accept this view, it could create a more stable environment for funding massive data centers. It shifts the conversation from the immediate cost of the hardware to its long-term value as a tangible asset. This could potentially lower the barrier for companies looking to build out their AI capabilities without liquidating other types of capital.

The performance-per-watt metric

The argument for long-lasting hardware value is supported by other industry figures, including Elon Musk. Musk recently suggested that older AI accelerators might actually be more profitable for inference workloads than newer models. He bases this on the idea of performance per watt, which measures how much intelligence a chip can generate for every unit of electricity it consumes.

Musk argues that companies should prioritize this efficiency metric over raw processing power when evaluating their hardware needs. Under this logic, an older chip that is highly efficient could be more valuable than a newer chip that consumes significantly more power to achieve similar results. This approach highlights a growing trend where the cost of electricity becomes a primary factor in hardware ROI.

This shift in thinking is bolstered by the fact that some older chips are already showing surprising resilience in the market. For instance, Nvidia's H100 chips have reportedly been worth more today than they were when they were first released. This suggests that as AI models and inference techniques improve, the underlying hardware can actually become more effective over time.

Market dynamics and resale value

The second-hand market for AI accelerators provides a striking example of these trends in action. Reports indicate that the B200 chip has been rented out or traded at prices equivalent to 158% of its original purchase price. This high resale value underscores the intense demand for high-quality AI computing power across the global economy.

However, industry insiders warn that the value of older accelerators is not determined by price alone. The worth of a used chip depends on several specific factors, including its physical condition and the specific system configuration it belongs to. It also depends heavily on the specific workloads the hardware is expected to handle in its next life.

While high resale prices are a strong indicator of demand, the diversity of AI applications means that some chips may be more valuable than others. A chip optimized for one type of training might be less desirable for another type of inference. Therefore, the 'durability' Huang speaks of is tied to the fact that these chips can be adapted to many different roles as the AI landscape matures.