How can GPU assets become a new type of investment?

Generative AI has pushed 'computing power' from being a technical resource to a tradable economic good. In the past, when companies discussed cloud services, they mostly cared about flexibility and paying for what they used; now, the bottleneck for many AI projects has become GPU supply, data center electricity, and delivery timelines. As a result, computing power has started to be 'reserved, contracted long-term, and secured' like real estate. When computing power can be measured, divided, and sublet, it takes the first step towards assetization: transforming from a service into a priced 'usage right.' This real-estate-like concept is rewriting the financial structures and investment narratives of cloud providers.

The core of GPU leasing is turning expensive hardware into continuous cash flow. You don't necessarily need to own H100/H200 GPUs; you just need to rent the right cards, network, and environment when needed. This has given rise to two paths: one is the 'new cloud (neocloud),' which integrates large-scale procurement and data centers to offer businesses pay-as-you-go computing power; the other is the 'GPU marketplace,' listing idle computing power scattered across different regions, creating a price mechanism akin to a spot market. Prices have thus become comparable: research compiling H100 rental rates around 2025 shows a significant gap between the lower-end rates on market-type platforms and the higher quotes from large public clouds, indicating that computing power is moving towards transparent competition and narrowing profit margins.

GPU = 'Mortgageable Machine Room Property'

When computing power is locked in by long-term contracts, it starts resembling real estate rental income: the contract itself becomes a financeable instrument, with GPUs and servers acting like 'movable house structures.' For example, CoreWeave had already seen financing cases using Nvidia chips as collateral as early as 2023. Subsequently, it announced the completion of a large 'collateral-based' debt financing and expanded its data centers to boost supply. The essence of this model is leveraging signed demand and existing equipment to secure more funding, purchase more GPUs, and create a rolling expansion.

However, the real estate-like transformation also brings real estate-style risks: First, asset residual value. The rapid iteration of GPUs means that if new cards emerge and render old cards' cost per unit of computing power uncompetitive, the 'resale value' of collateral may be downgraded; Second, rental volatility. When market prices drop rapidly and customers’ bargaining power increases, cash flow will be squeezed; Third, financing terms and compliance controls. What the financial markets care about is not just the demand narrative but also factors like the location of the collateral, contractual restrictions, and internal control quality — all of which have become market focal points in related companies’ disclosures.

More noteworthy is that the real estate-ization of computing power turns 'electricity' into a prime location factor. Whether a data center can secure a stable power supply and grid connection timelines often determines whether a 'computing power building' can be delivered on schedule. Thus, investors no longer focus solely on GPU specifications but also on power contracts, data center locations, the quality of long-term customer agreements, and the cost of capital. As cloud computing power becomes a new type of infrastructure asset, the market will evaluate it using the language of infrastructure/real estate: occupancy rate (utilization), rent (price per GPU hour), capitalization rate (cost of capital), and the hardest to estimate — the depreciation curve (technological obsolescence). Given the strong long-term demand for AI, the winners may not be determined by 'who buys the most GPUs' but by 'who can turn computing power into more stable, predictable, and finance-market-friendly cash flows.'

(Chip and Computing Power Series No. 37)

$Advanced Micro Devices (AMD.US)$$Intel (INTC.US)$$NVIDIA (NVDA.US)$$Alphabet-C (GOOG.US)$

33K Views