The market is a liar. It tells you that a $16 billion data center is a vote of confidence in Anthropic’s future. But smart contracts execute truth, not intent. I audited the void and found a backdoor — the real story is not about AI progress, but about capital structure, leverage, and the hidden liabilities embedded in this mega-project.
Hook
Eagle Point, a little-known infrastructure lender, is providing $1.3 billion in loans to fund a $16 billion data center in Texas for Anthropic. The ratio is 1:12.3. That means for every dollar of debt, there is $12.3 of total project cost. This is not a typical venture debt round. It is a structured finance play that mirrors the 2017 ICO arbitrage I ran: find the mispricing in the capital stack, execute before the crowd understands.
Floor sweeps are just data points in motion. In this case, the floor is the debt-to-equity ratio. The data point: $1.3B loan, $16B total. That implies Anthropic is putting in $14.7B of equity or other sources. But where does that equity come from? The company’s last reported valuation was ~$30B (post-money 2024 round). If they are committing ~50% of the total project from their own balance sheet, that is a massive bet on future cash flows. Either they expect a 10x return on capital, or they are hiding something.
Context
Anthropic, the AI lab behind Claude, has been building its own infrastructure for years. Previously, they relied on Google Cloud credits (part of a $2B investment from Google). Now they are going independent — building a dedicated data center in Texas, likely to train Claude 4 and beyond. The project is described as a "mega-project" that could reshape the tech landscape. But the details are sparse: no chip contract announced, no timeline, no power purchase agreement.
Eagle Point is not a tech VC. It is a specialized infrastructure lender that typically funds toll roads, pipelines, and energy projects. Their involvement signals that this data center is being treated as a hard asset with predictable cash flows — like a toll road for AI compute. But AI compute is not a toll road. The demand is volatile, the hardware depreciates fast, and the competitive landscape shifts every quarter.
Core
Let’s break down the numbers.
A $16B data center, by industry standards, allocates ~40-50% to compute hardware (GPUs, network, storage). That’s $6.4B to $8B for chips. At current NVIDIA H100 pricing (~$30K each), that buys 213,000 to 266,000 GPUs. For reference, OpenAI’s training cluster for GPT-4 was estimated at ~25,000 H100s. So this is a 10x scale-up. But Blackwell B200 is now shipping at ~$40K, which would reduce the count to 160,000-200,000 units. Still massive.
The remaining 50-60% ($8B-$9.6B) goes to land, building, power infrastructure, cooling, and networking. Texas electricity costs ~3-5 cents per kWh, half of California’s. That’s a smart cost play. But the grid (ERCOT) is fragile. The 2021 freeze showed that. If the data center draws >1 GW, it could strain the local grid. No backup power plan has been disclosed.
Now the loan: $1.3B from Eagle Point. That is a small fraction of the total. Why so little? Because the loan is likely secured against the first phase — maybe $3B-$4B of the project. The rest is equity from Anthropic (which they raised from previous rounds) and possibly tax incentives from Texas. The loan structure is classic: low leverage now, high leverage later. As the project generates cash flow, they can borrow more. But AI inference margins are unpredictable. If Claude’s API revenue doesn’t hit $10B/year by 2027, the debt service will eat into equity.
Contrarian Angle
Retail analysts celebrate this as a sign of AI dominance. Smart money sees a red flag: the capital intensity is spiraling out of control. Anthropic is essentially betting that the market for AI compute will grow exponentially, and that they will capture a large share. But the history of tech infrastructure is littered with overbuilds. Remember the 2000 telecom bubble? Fiber optic networks were laid at massive cost, only to be sold for pennies on the dollar. AI data centers could follow the same path.
Moreover, the loan from Eagle Point is not a vote of confidence in Anthropic’s technology. It is a vote of confidence in the asset as collateral. If Anthropic defaults, Eagle Point can seize the data center and lease it to someone else. This is a classic "asset finance" play, not a strategic partnership. The real risk is that Anthropic’s model quality fails to keep pace with OpenAI or Google, and the compute becomes a stranded asset.
Another hidden factor: chip supply. NVIDIA’s lead time for Blackwell is 12-18 months. If Anthropic wants to launch Claude 4 in 2026, they need to order chips now. But no deal has been announced. This could mean they are waiting for a competitor (AMD, Intel, or even self-designed chips). Or they may be overconfident about supply. Either way, execution risk is high.
Takeaway
I am not saying this project will fail. I am saying the math does not guarantee success. The 1:12.3 leverage ratio is a signal of a capital structure that prioritizes asset control over cash flow efficiency. Smart contracts execute truth, not intent — and the truth is that $16B in infrastructure requires $16B in revenue to justify the risk. Anthropic’s current revenue is likely under $1B. The gap is a void. I audited that void. It has a backdoor called debt.
Watch for: (1) any chip supply announcement, (2) the actual power purchase agreement (PPA), and (3) Eagle Point’s syndication of the loan to other institutions. If the loan gets sold to pension funds, the risk is distributed. If not, the risk is concentrated. That is the data point to follow.
Floor sweeps are just data points in motion. This one is moving fast. Do not be the one holding the bag when the market reprices the risk.