If you only look at tech news from 2025, you would think the world is great: AI investment continues, North American data center construction accelerates, and crypto miners have finally 'emerged from the cycle,' successfully transforming the originally high-volatility mining business into stable AI computing power services.
Article author, source: Anita, MarsBit
But in the credit department on Wall Street, the atmosphere is completely different
Credit investors are not discussing model performance and do not care which generation of GPU is stronger. They are focused on the core assumptions in the Excel spreadsheets, starting to feel a chill: it seems we are using a 10-year real estate financing model to purchase a perishable product with a shelf life of only 18 months.
Reuters and Bloomberg's continuous reports in December revealed the tip of the iceberg: AI infrastructure is rapidly becoming a 'debt-intensive industry'. But this is just the surface; the real crisis lies in the deep structural financial mismatches—when high-depreciation computing assets and highly volatile miner collateral are forcibly bundled with rigid infrastructure debt, a hidden default transmission chain has already formed.
One, the deflation on the asset side: the brutal revenge of 'Moore's Law'.
The core logic of debt is the debt service coverage ratio (DSCR). Over the past 18 months, the market assumed that AI computing rental rates would be as stable as housing rents, or even resistant to inflation like oil.
The data is ruthlessly shattering this assumption.
According to tracking data from SemiAnalysis and Epoch AI for the fourth quarter of 2025, the unit AI inference cost has decreased by 20-40% year-on-year over the past year.
The popularization of model quantization, distillation technology, and the efficiency improvements of ASICs have led to exponential growth in computing supply efficiency.
This means the so-called 'computing rental' has inherent deflationary properties.
This constitutes the first duration mismatch: the bond issuer purchases GPUs at the peak price in 2024 (CapEx), but locks in a rental yield curve that is destined to plummet after 2025.
If you are an equity investor, this is called technological advancement; if you are a creditor, this is called collateral depreciation.
Two, the alienation of the financing side: packaging venture risks into infrastructure returns.
If the returns on the asset side are thinning, the rational liability side should be more conservative.
But the reality is exactly the opposite.
According to the latest statistics from The Economic Times and Reuters, the total debt financing for AI data centers and related infrastructure in 2025 surged by 112% to reach $25 billion. This surge is primarily driven by 'Neo-Cloud' vendors like CoreWeave and Crusoe, as well as transitioning mining companies, which are adopting asset-backed lending (ABL) and project finance on a large scale.
The fundamental change in this financing structure is extremely dangerous.
In the past: AI was a game for tech VCs, and failure meant equity going to zero.
Now: AI has become a game of infrastructure, where failure means debt default.
The market is mistakenly placing high-risk, high-depreciation tech assets (Venture-grade Assets) into a low-risk financing model (Utility-grade Leverage) that should belong to highways and hydropower plants.
Three, the miners' 'fake transformation' and 'real leverage'.
The most vulnerable link appears among cryptocurrency miners. The media likes to praise the miners' transformation to AI as 'de-risking', but from a balance sheet perspective, this is risk stacking.
Data from VanEck and TheMinerMag reveals an counterintuitive fact: the net debt ratio of leading publicly listed mining companies in 2025 has not substantially reduced compared to the peak in 2021. In fact, the debt scale of some aggressive mining companies has surged by 500%.
How did they achieve this?
Left hand (asset side): Still holding highly volatile BTC/ETH, or using future computing revenue as implicit collateral.
Right hand (liability side): Issuing convertible notes or high-yield debt to borrow USD for purchasing H100/H200.
This is not deleveraging; this is rollover.
This means miners are playing a 'double leverage' game: using the volatility of Crypto as collateral to bet on the cash flow of GPUs. In favorable conditions, this provides double profits, but once the macro environment tightens, 'price drops' and 'rental rates for computing power decline' will happen simultaneously. In credit models, this is referred to as correlation convergence, which is a nightmare for all structured products.
Four, the missing repo market.
What keeps credit managers awake at night is not the default itself, but the liquidation that follows.
In the real estate subprime crisis, banks could at least auction off houses. But in AI computing financing, if a miner defaults, what can creditors do with the ten thousand H100 graphics cards they reclaim?
This is a severely overestimated liquidity secondary market.
Physical dependency: High-end GPUs cannot just be plugged into personal computers; they rely heavily on specific liquid cooling cabinets and power density (30-50kW/rack).
Hardware Obsolescence: With the release of NVIDIA Blackwell and even Rubin architectures, the older cards are facing nonlinear depreciation.
Buy-side vacuum: When systemic sell-offs occur, there are no 'lenders of last resort' willing to take on outdated electronic waste.
We must be cautious of this 'collateral illusion'—the LTV on paper looks safe, but the secondary repo market that can absorb billions in selling pressure does not exist in reality.
This is not just an AI bubble; this is a failure of credit pricing.
It needs to be clarified that this article does not deny the technological prospects of AI, nor does it deny the real demand for computing power. What we question is the flawed financial structure.
When deflationary assets driven by Moore's Law (GPUs) are priced as inflation-resistant real estate, and when miners who have not truly deleveraged are financed as quality infrastructure operators—the market is actually conducting a credit experiment that has not been fully priced.
Historical experience has repeatedly proven: credit cycles often peak earlier than technology cycles. For macro strategists and credit traders, the primary task before 2026 may not be predicting which large model will win, but rather re-examining the true credit spread of those 'AI Infra + Crypto Miners' combinations.
