The market is wrong about Anthropic's $965 billion valuation. Not because the number is too high or too low, but because investors keep treating it as a balance sheet. It isn't. A valuation is a multiple of future promise. A 1.6-gigawatt data center is a physical pile of steel, copper, and combustible gas. The gap between those two numbers just got priced, and the terms reveal more about Google's leverage than Anthropic's ambition.
The deal, first reported as a $5 billion project and now scaled to $15 billion, plants Anthropic's first hyperscale campus in Hubbard, Texas: 2,800 acres, a behind-the-meter natural gas power plant, and a compute buildout designed around custom TPUs co-developed by Google and Broadcom. That last detail is the most important. Anthropic has stopped renting NVIDIA GPUs and started borrowing Google's entire hardware stack.
This is project finance, but it reads like a collateralized loan. Google — already Anthropic's largest shareholder at roughly 14% — is guaranteeing tens of billions in leases and power purchase agreements. In exchange, it receives a 20% equity stake in the physical project. Morgan Stanley simultaneously leads the financing syndicate and will likely bookrun the October 2026 IPO. Every layer of the capital stack is intermediated by the same web of relationships.
Let me be clear about what is actually happening, because I have audited enough structured products to recognize the pattern. Anthropic is splitting the buildout into two distinct layers: the physical facility (land, buildings, power generation) and the compute chips. Each layer is financed separately, through supplier financing agreements and lease structures. The entire purpose is to isolate the most volatile capital expenditures from the consolidated income statement ahead of the IPO. In financial engineering, this is called asset isolation. In plain English, it is off-balance-sheet leverage.
Note: Valuation is not a physical asset.
Run the numbers. A $965 billion valuation against a $15 billion single-project buildout leaves a capital-expenditure-to-valuation ratio near 1.5%. Compare that to ExxonMobil, a company that genuinely earns its $500 billion market cap with $20 billion to $25 billion in annual capex. If Anthropic scales to oil-major status, it needs $200 billion or more in infrastructure, not $15 billion. Where does that come from? Not from organic revenue. The only answer is more debt, more equity dilution, or more Google guarantees. Each option carries a price.
The clever part of this structure is the decoupling. Custom TPUs refresh every 12-18 months; buildings and gas turbines last 25-30 years. By walling off the physical layer, Anthropic can upgrade chips without being trapped by real-estate depreciation cycles. That is genuinely elegant. But elegance has a cost. Supplier financing still requires minimum purchase commitments. Anthropic has likely signed take-or-pay agreements on TPU volumes that will become contingent liabilities on the IPO balance sheet. Investors will not see these obligations in the headline numbers. They will appear in footnotes, and footnotes are where hidden leverage lives.
I have watched this movie before. In 2020, I audited a dYdX beta and walked away convinced that order-book centralization was the only path for institutional capital, not because AMMs were mathematically broken but because liquidity fragmentation kills settlement confidence. The same liquidity-first logic applies here. Anthropic's entire strategy is a bet that it can keep the capital flowing at a cost below its unit-economics threshold. That is not a technology bet. It is a credit bet.
Underwrite this the way I would underwrite a DeFi protocol. A DAO treasury full of illiquid governance tokens looks rich until the banks demand physical collateral. The moment an AI lab's promised revenue is discounted back at a risk-free rate, the gap between narrative and economics becomes visible. This is exactly the dynamic we saw in Terra's algorithmic stablecoin: the market lent against an assumption, not an asset. Anthropic is in a healthier position, but the principle is identical.
The industry-level signal is even bigger. The Meta-BlackRock partnership already pushed AI infrastructure into asset-finance territory. Brookfield's involvement in the Department of Energy's Paducah site confirms that sovereign land is being used as an AI subsidy mechanism. Anthropic is the third leg of a new asset class: institutional capital buys the balance sheet, tech companies buy the compute, and everyone hopes the demand curve keeps rising. This is the financialization of AI compute, and it will reshape who controls the data center supply chain.
The timing of this financialization matters. The Federal Reserve's credit cycle is already tightening, and AI infrastructure debt is exactly the kind of long-duration, asset-heavy exposure that gets compressed when the risk-free rate moves. Every basis point in credit spreads translates into millions of dollars in annual interest expense on a $15 billion project. The same macro logic that broke crypto liquidations in 2022 is now embedded in the AI capital stack.
Note: Off-balance-sheet is just a lease with extra steps.
Now for the contrarian angle. The narrative in the market says Anthropic has secured a decisive compute advantage. That is the wrong way to read the term sheet. Anthropic has not bought independence. It has borrowed control from a counterparty with five overlapping identities. Google is simultaneously Anthropic's largest shareholder, its key chip designer, its cloud landlord, its debt guarantor, and the operator of its most direct competitor, Gemini. Any one of these roles would create conflicts of interest. Combined, they make a mockery of the independence narrative that Anthropic has carefully cultivated.
There is also the matter of execution risk. The developer, Nexus Data Centers, has a thin public track record in hyperscale projects. A 1.6-gigawatt campus with an on-site power plant is one of the most complicated engineering tasks in the world. Natural gas plants take 24 to 36 months to build. Data center shells take 12 to 18 months. Google and Broadcom's custom TPU production is competing with Google's own TPU v7 volumes for the same TSMC capacity. Any mismatch in timing produces idle chips or idle buildings, both of which burn cash at a staggering rate.
The honest question is not whether Anthropic gets the compute. It will, eventually. The honest question is at what price and with what delay. A 12-month delay in a 24-month chip generation cycle is not a small risk; it is a structural risk. The company's model training roadmap is now hostage to a supply chain that has never successfully executed at this scale.
Note: The first casualty of any project overrun is the unproven developer, not the model lab.
I also want to flag the governance trap. Morgan Stanley will negotiate the financing terms on one side and manage the IPO on the other. That is a known pattern in modern finance, but when the deal size crosses the hundred-billion-dollar threshold, the conflict becomes material. The loan pricing and the IPO pricing are negotiated by the same institution. Investors will need to demand audited valuations of the real estate, the power purchase agreements, and the guaranteed minimum TPU volumes before they trust the pre-IPO balance sheet. I would also expect the Federal Trade Commission to examine whether Google's role has crossed from investment into control. Microsoft's relationship with OpenAI already triggered regulatory scrutiny. This arrangement is structurally more entangled.
Retail is still asking whether Claude beats GPT-5. That is the wrong question. The takeaway for this cycle is not about model benchmarks or token predictions. It is about credit. The AI capital stack now resembles the CDO era: an ambitious borrower, a AAA-rated guarantor, and a financing structure so clever that no one fully understands its downside. The next inflection point will not come from an Anthropic model release. It will come from the first major project delay, the first TPU delivery miss, or the first margin-call trigger embedded in a supplier financing agreement.
In the coming quarters, the data points that matter are not benchmark scores. They are the timing of the final investment decision, the selection of the EPC contractor, the first TPU tape-out, and the credit rating assigned to the project debt. Any one of these misses will move the narrative more than a model release.
Watch the EPC contractors. Watch the Fed's impact on credit spreads. Watch the FTC's docket. This is why I have been quietly telling our readers to watch AI credit the way we watched stablecoin collateral quality in May 2022. Same error, bigger scale. The narrative has shifted from who trains the best model to who can service the most debt. That shift, not the next benchmark, is the bull case and the bear case for AI infrastructure.