The code does not lie, but the narrative does. On paper, Anthropic secured a $0.9 billion loan from Eagle Point Infrastructure to fund a $16 billion data center project in Texas. That is a leverage ratio of 17.7x on equity. In crypto markets, a 94% loan-to-value on a single asset class is called a liquidation event waiting to happen. In the world of AI infrastructure, it is called a growth strategy. I have spent the last four years modeling cross-border payment rails and liquidity flows. The same structural patterns that govern DeFi leverage now govern the AI compute race. The only difference is the collateral: instead of a volatile token, Anthropic is pledging future compute capacity against a debt instrument. The question is not whether they can build the facility. The question is whether the market will validate the value of that compute before the loan matures.
This is the first signal that the AI industry is entering a phase of capital intensity that mirrors the late-stage crypto bull runs of 2021 and 2024. The same euphoria, the same reliance on cheap debt, and the same assumption that demand will grow exponentially to justify the upfront cost. But in macro, the only constant is reversion to the mean. The code does not lie, but the narrative does. The data center will be built. The real story is what happens when the loan comes due.
Context: The Infrastructure Arms Race
Anthropic, the company behind Claude, has been a tenant on Google Cloud since its inception. That arrangement was part of a multi-billion dollar investment deal with Google, which provided both capital and compute credits. But tenants eventually want to own the building. The Texas project, funded by a $0.9 billion loan from Eagle Point, is the clearest signal yet that Anthropic is pivoting from a capital-light, cloud-dependent model to a capital-intensive, infrastructure-owning model.
Eagle Point is not a technology venture capital firm. It is an infrastructure debt specialist that typically funds toll roads, energy grids, and data centers for hyperscalers like Amazon and Microsoft. Their willingness to lend to an AI company that has not yet achieved profitability is a bet on the secular growth of AI compute demand. But it is also a bet on the market's ability to price that compute correctly. In crypto, we call this a 'basis trade' — borrowing against an asset that you believe will appreciate faster than the cost of debt. The risk is that the basis flips, and the asset depreciates.
From a macro perspective, this project is part of a broader trend. The global AI data center capex is projected to exceed $500 billion by 2027, according to Goldman Sachs. The majority of that spending is debt-financed. The interest rate environment, with the Fed holding rates at 5.25-5.5%, makes this debt expensive. The only way to justify the cost is if the output — AI model inference and training — generates a return on invested capital above the cost of debt. That is an assumption that has not been tested in a bear market for AI.
Core: The Technical and Economic Mechanics
Let us break down the numbers. The $16 billion total project cost is spread over multiple phases, likely 5-10 years. The initial $0.9 billion loan covers the first phase. Standard data center construction costs are roughly $10-15 million per megawatt of power capacity. A $16 billion facility could support 1,000-1,500 megawatts of IT load. That is enough to power a small city. For comparison, OpenAI's existing data center capacity is estimated at around 500 megawatts. Anthropic is building a facility three times larger.
What does that compute buy? At current prices, a single NVIDIA H100 GPU costs around $30,000. If 40% of the project budget is allocated to GPUs, that is $6.4 billion, or roughly 213,000 H100s. That is a cluster larger than any single deployment publicly known outside of Meta and Google. But the real constraint is not the number of GPUs; it is the interconnect bandwidth. Training a model like Claude 4 requires high-speed NVLink or InfiniBand networking. The cost of networking hardware can be 20-30% of the total GPU budget. This is analogous to the high gas fees on Ethereum during peak congestion — the bottleneck is not the compute, but the communication.
From a crypto perspective, I see a direct parallel to the DeFi liquidity traps of 2021. Projects borrowed heavily against their governance tokens, assuming that the value of those tokens would rise indefinitely. When the market turned, the debt became toxic. Anthropic's debt is secured against future revenue from API calls. That revenue is a function of model demand and competitive pricing. If another model — say, an open-source variant from Meta or a cheaper alternative from Mistral — captures market share, Anthropic's revenue may not grow fast enough to service the debt. The loan terms are not public, but typical infrastructure debt has a maturity of 5-7 years and an interest rate of 8-12%. Annual interest on $0.9 billion at 10% is $90 million. That is a fixed cost that must be paid before any profit.
Contrarian: The Decoupling Thesis and the Liquidity Trap
The prevailing narrative is that this project positions Anthropic to compete with OpenAI and Google by owning its compute. I will take the contrarian view: this project may actually increase Anthropic's vulnerability to a market downturn, and it signals a decoupling of AI companies from the capital markets that will eventually lead to a consolidation.
First, the decoupling thesis. In 2023 and 2024, AI companies were valued primarily on their model capabilities. Compute was rented, not owned. That meant that if demand fell, companies could scale down their cloud usage. By owning a data center, Anthropic is converting a variable cost into a fixed cost. This is a classic mistake in capital-intensive industries: the belief that owning assets is always cheaper than renting. In practice, owning only makes sense if utilization is consistently high. If demand drops, the fixed cost of the data center becomes a drag on profitability. This is the same trap that crypto miners fell into during the 2022 bear market. They locked in long-term power contracts and bought ASICs at high prices, only to see the spot price of Bitcoin fall below their breakeven cost.
Second, the liquidity trap. The loan from Eagle Point is likely structured as a priority claim on the data center's cash flows. In a default scenario, Eagle Point seizes the facility. Anthropic's equity holders are wiped out. This is a classic debt overhang problem. The company's ability to raise additional equity is constrained because new investors will demand a higher return to compensate for the existing debt. This is exactly what happened to Celsius Network in 2022: a large debt load made it impossible to recapitalize when the market turned.
From a macro perspective, the timing is critical. The Fed is expected to cut rates in late 2025 or 2026. If Anthropic's loan is floating-rate, the interest cost will decrease. But if the loan is fixed-rate, they are locked in at a high rate. The project completion date is likely 2027-2028. By then, the AI market may have matured. The question is whether the market will reward Anthropic's bet with sufficient revenue to cover the debt.
Takeaway: The Cycle Position
I have been in this industry long enough to recognize the pattern. Every bull market creates a narrative that justifies outsized capital expenditure. In 2021, it was 'DeFi is the future of finance.' In 2024, it is 'AI infrastructure is the new oil.' Both narratives contain a kernel of truth, but both ignore the cyclical nature of capital flows. The code does not lie, but the narrative does.
When the liquidity squeeze comes — and it will come, as it always does — the companies with the most leverage will be the first to fall. Anthropic's $16 billion bet is a leveraged bet on the future of AI. It may pay off, or it may become a cautionary tale. The signal I am watching is not the construction progress, but the API revenue growth rate. If Anthropic's API revenue does not grow at least 50% year-over-year for the next three years, the debt will become a burden that no amount of model accuracy can fix.
In a bull market, everyone is a genius. In a bear market, only the infrastructure survives. The real winners here may be Eagle Point and the lenders who collect interest regardless of the outcome. Anthropic is playing a high-stakes game. The rest of us are watching from the sidelines, waiting for the next data point to confirm or refute the thesis.
If you can't measure it, you can't manage it. The same applies to AI infrastructure. The market will eventually price in the risk. When it does, the narrative will shift. The question is whether Anthropic will have enough time to adjust before the loan comes due.
