Lending Against Obsolescence

Credit has always relied on one fundamental principle: lenders need confidence that if something goes wrong, something valuable remains. For centuries, that meant tangibles like land, factories, and infrastructure. Today, however, the AI boom has introduced a new and strange asset: the GPU.

Billions of dollars are invested in these chips that power artificial intelligence, and creditors have begun originating and securitising AI debt with this infrastructure, but unlike more traditional assets, their value can erode as rapidly as technology advances. The question credit markets face is therefore unprecedented: can investors lend against an asset that may become obsolete before the loan even matures?

AI models require enormous amounts of physical infrastructure, such as data centres, electricity, and cooling systems. So, whilst AI looks like software, it behaves a lot more like infrastructure, requiring a lot of upfront capex. Every technological revolution has relied on a physical foundation. The industrial revolution depended on railways, the internet on fibre-optic cables, and today, it is artificial intelligence with GPUs.

Traditionally, secured lending has relied on collateral with established secondary markets. Land retains its value for decades, factories can be repurposed, and machinery can be resold. In contrast, GPUs are expensive, rapidly evolving, with depreciation uncertain and the value depending on AI demand. Their value is not only derived from its physical condition but also whether demand for that generation of compute still exists. That leaves investors facing three vital questions with GPU-backed debt:

• If the AI startup defaults, who buys the GPUs?
• What would their liquidation value be once newer generations have entered the market?
• Given this, how do lenders price this risk when collateral is determined by technological relevance?

The most direct answer for now is that markets do not fully know. Instead, lenders appear to rely on something far more predictable than the chips themselves: the cash flows they generate.

GPUs do not behave in the same way that normal equipment would behave on a balance sheet. Rather than calculating depreciation based on terminal value and time, GPUs depreciate through what could be described as technological entropy. The chip may continue to function perfectly, but its ability to generate money declines over time. Every few years the world keeps improving so yesterday's cutting-edge technology becomes today's ordinary technology. A GPU doesn't essentially deteriorate by sitting in a warehouse, but it decays because AI models get bigger, algorithms become more demanding, new chips become more efficient, and electricity costs matter.

A recent transaction illustrates this shift particularly well. Nebius, an AI cloud company, recently raised $775 million in its first-ever secured debt financing. This senior-secured debt is backed by GPU infrastructure and contracted cash flows, maturing on the 31st of October 2030 with SOFR plus a 250bps spread.

At first glance the structure appears questionable, with a maturity of 2030 when most GPUs have a half-life of two to three years, highlighting the mismatch in the deal’s structure. However, what this deal does show is investor sentiment in AI infrastructure.

The debt is not simply a bet on the residual value of the GPU’s but also on the cash flows generated by these GPUs. For the lender, the collateral is the downside protection and the cashflows are the repayment source attached to the infrastructure. Reportedly, this deal covers more than 100% of the capex required to deploy the underlying GPU infrastructure, demonstrating that lenders are underwriting a going concern, not just the hardware.

Historically, securitised assets produced stable cash flows, land generates rent, shipping generates freight, and factories generate production. Here, GPUs produce compute. The productive asset has changed but the underlying credit question has not. The question is not whether GPUs depreciate, which they undoubtedly do. Rather, it is whether the cash flows generated during their useful life are attractive enough to support long-term lending.

The Future Outlook

The emergence of GPU-backed lending signals that the market is betting AI compute demand will remain strong enough to sustain long term lending despite technological change. As AI infrastructure spending accelerates and equity financing becomes more expensive, more companies are likely to finance through secured debt. Simultaneously, lenders will increasingly rely on the contracted cash flows these chips generate.

Overall, the sustainability of this market and the trajectory that it will take relies on the continued demand for AI compute. If AI investment continues to expand, GPU lending could become a permanent feature of credit markets. However, if that demand weakens, lenders may discover that technological obsolescence depreciates collateral faster than anticipated. The question is no longer whether a GPU can hold its value, but whether it can earn enough before it doesn’t.