The AI compute arms race enters a new phase as credit markets struggle to keep up.

Author: Kerman Kohli (Founder of Fastbreak Capital)

Compiled by Deep潮 TechFlow

Shenchao Overview: As the global AI arms race enters its second half, the bottleneck in computing power infrastructure is expanding from a simple shortage of chips to the trillions-of-dollars credit financing market. Traditional credit institutions commonly apply conservative models used to evaluate office buildings and conventional data centers when assessing GPU computing power, holding three prevailing misconceptions: that AI is a bubble, that startup tenants lack sufficient creditworthiness, and that GPU hardware will depreciate to zero within a few years like ordinary servers.

Kerman Kohli, founder of Fastbreak Capital, exposes the blind spots in traditional credit risk assessment. He points out that the potential demand for computing power in quantitative finance, biopharmaceuticals, cutting-edge scientific research, and defense is limitless, and the Jevons Paradox ensures that reduced costs will only multiply consumption. Even more ironically, credit institutions, in their pursuit of so-called "investment-grade safety," compel startups to sign 2- to 3-year contracts with steep discounts of 30% to 50%. Amid today’s annual computing power rental increases of 100%, this practice inadvertently traps funders in a de facto "shorting of computing power" position.

This article deeply analyzes the underlying logic of compute power financialization, revealing how credit-backed compute assets are ushering in a generational transformation surpassing traditional real estate, as demand for inference memory intensifies.

The current hash rate market is facing numerous bottlenecks, but one of the most severe is access to capital. More specifically, there is a lack of credit capital available to fund hash rate infrastructure. We now confront a construction gap amounting to trillions of dollars, yet capital struggles to flow smoothly into this sector.

Traditional credit institutions are accustomed to crudely equating computing power with real estate. They often hold three core beliefs that appear rational on the surface but are in fact contradictory:

• Firmly believe that AI is an asset bubble that will eventually burst;

• Believes the hash power off-taker is not a high-rated company and has a very high probability of default;

• Believes the underlying collateral (GPU hardware) has no value retention and will depreciate to zero within a few years, like ordinary data center servers.

These viewpoints may seem logical to outsiders, but they are financially incompatible. To unlock the trillions in capital liquidity across the entire ecosystem, we must thoroughly correct each misconception.

The core argument of AI bubble skeptics is that compute demand is limited and supply will eventually catch up; furthermore, improvements in large model design or algorithmic efficiency will reduce demand for hardware.

This assertion collapses under the weight of real-world data. The annual demand growth for computing power from agentic AI has reached a staggering 2.5x to 5x, while global advanced wafer and packaging supply can at most double by 2030—widening the supply-demand gap each year. Global high-bandwidth memory (HBM) capacity for 2027 has already been fully pre-ordered by major players, and leading computing providers are carrying tens of billions of dollars in unfilled backlogs.

More importantly, there are numerous core industries in the real economy with insatiable demand for computing power:

• Quantitative High-Frequency Finance: Deep modeling of complex derivatives and holographic risk (e.g., the billions of dollars in dedicated computing power committed by top market maker Jane Street);

• Biopharmaceutical giant: Invests unlimited GPU computing hours to simulate molecular docking, target screening, and new drug development;

• Leading scientific researchers: Leverage massive computing power to connect inspirations across unknown disciplines and perform complex simulations of the physical world;

• Defense and military sectors: Computing power scale directly determines absolute dominance in drone swarms and battlefield simulations;

• Frontline of Cyber Defense: Hackers and security teams continuously use brute-force computing power to uncover deep zero-day vulnerabilities in systems.

These entities simply do not obsess over “token budgets per employee” like traditional businesses. As long as human society continues to need better vaccines, more precise financial risk control, stronger national defense, and deeper scientific breakthroughs, computing power remains the purest form of productivity, and its demand is inherently limitless. Moreover, Jevons Paradox is accelerating in the AI domain: when the cost of a single inference drops by 90%, societal usage frequency and application scenarios will surge more than tenfold.

Figure 1: Capital expenditure and contract backlog trends among major global cloud providers and AI giants. Cumulative procurement commitments have reached trillions of dollars, with the supply gap widening year over year.

Equating computing power with real estate is an extremely dangerous conceptual misconception. Potential tenants of real estate are absolutely constrained by geography; whereas computing power is a highly homogeneous, globally delivered, universal digital commodity via fiber optics in seconds.

Many lending institutions are overly fixated on tenants’ “investment-grade creditworthiness,” routinely demanding leases of two to three years and requiring substantial upfront payments of 20% to 30%, or even as high as 50%. Lenders mistakenly believe that long-term contracts secure a safety cushion, but in practice, they place themselves in an extremely disadvantageous position.

In a world over the next 5 to 10 years where computing power is severely undersupplied, if a startup client defaults or violates terms, at least ten backup clients are queued to take over that computing power allocation. Flagship chips like the B300 are nearly impossible to obtain on the spot market. Even more absurdly, against the backdrop of a 100% surge in spot rental rates for computing power over the past year, lenders are forcing borrowers to sign three-year contracts at steep discounts of 30% to 50% in exchange for so-called certainty—effectively turning lending institutions into “short sellers” of computing power in financial substance.

This distorted desire for certainty causes lenders to unnecessarily sacrifice significant floating returns and extend their capital recovery periods. This conservatism undermines their own return on capital.

The most popular argument among AI pessimists is that compute hardware depreciates rapidly and will become e-waste within a few years. However, objective facts have proven this assumption absurd: NVIDIA’s A100 GPU, released in 2020, is still operating at full capacity in data centers today, with rental orders booked through 2029—establishing that GPUs can maintain a stable commercial service life of at least nine years. Based on this, the effective economic lifespan of the Blackwell architecture and its subsequent models could very well exceed a decade.

More disruptively, cutting-edge computing hardware is exhibiting a "appreciation" characteristic rarely seen in traditional industrial goods: since its release two years ago, the real secondary market value of the B200 has surged by 50% compared to its initial price; even a six-year-old A100 chip still maintains a robust residual value of around 25% of its original price. The six-year straight-line depreciation model previously used by most Wall Street analysts has now been proven to be an excessively pessimistic and overly conservative assumption.

Figure 2: Actual second-hand market value and rental trends for mainstream GPUs. Even six years after its release, the A100 retains a residual value of 25%, while the B200 commands a market premium of over 50%, completely overturning the traditional assumption of full depreciation within three years.

Some may question: As more advanced, efficient chips continue to emerge, won’t they accelerate the obsolescence of older cards? This counterargument overlooks three fundamental realities:

1. The core bottleneck in modern AI inference is highly memory-dependent, requiring high-bandwidth memory;

2. VRAM is a hard currency asset in a state of severe shortage;

3. Global semiconductor manufacturing capacity simply cannot produce sufficient chips and memory components in the short term.

Based on these irreversible physical and industrial principles, the entire financial capital market must reevaluate the lifetime value of GPU assets with a fresh perspective.

My core assessment is that, over the coming decades, computing power will surpass real estate to become the largest and most liquid asset class in human history. However, harnessing this superwave requires capital providers to possess both deep technical insight into the underlying technology and sophisticated modern financial engineering tools. The era of computing power assetization, anchored by credit and structured finance, has just begun its grand unfolding.