The numbers land like a punch. $15 billion. 1.6 gigawatts. 2,800 acres. A behind-the-meter natural gas plant. And a valuation of $965 billion that cannot, by itself, buy a single megawatt of physical capacity.
Ledger update: Capital is fleeing. Not out of AI — into the cracks between balance sheets. Anthropic's Texas data center deal is not a procurement contract. It is a financial engineering artifact. The kind that makes IPO underwriters sweat and antitrust regulators sharpen their pencils.
Here is the structural truth: Anthropic just borrowed Google's AAA credit rating to build infrastructure it does not own, on land it does not control, with chips it does not yet have, from a developer with no proven track record at this scale. And Wall Street is calling it innovation.
Context: In October 2026, Anthropic plans to go public. Its annualized revenue is estimated at roughly $1 billion — against a $965 billion valuation. That is not a growth multiple. That is a leap of faith. The company's answer to the gap is not more equity. It is a multi-layer financing structure that pushes the most volatile capital expenditures off its future IPO balance sheet entirely.
The deal, first reported at $5 billion and 612MW in March, has ballooned to $15 billion and 1.6GW on a 2,800-acre site outside Hubbard, Texas. The structure splits compute infrastructure into two discrete layers. Layer one: physical facilities and power. Layer two: chips. Each layer is financed separately, by different consortia, under different terms. Anthropic's design is deliberate. Separate the depreciating concrete from the fast-iterating silicon. Keep the volatility off the books.
The core insight, buried in the fine print: Anthropic is not buying computing power. It is buying an option on computing power, collateralized by Google's balance sheet and priced into a future IPO story. Google provides billions in lease and power purchase agreement guarantees. In exchange, it takes 20% equity in the project — on top of its existing ~14% stake in Anthropic itself.
Let me be precise about what this means, because my audit experience tells me the accounting matters more than the press release. This is project finance. Classic SPV logic. You carve out the risky asset, ring-fence it, and use limited recourse debt to keep the parent's balance sheet clean. The innovation is not the structure — it is the scale and the counterparty. Google is simultaneously Anthropic's largest shareholder, its chip supplier, its guarantor, its landlord, and its most direct competitive threat. Five roles. One counterparty.
The strategic calculus is brutal and rational. Anthropic chose custom Tensor Processing Units co-designed by Google and Broadcom over NVIDIA GPUs. This is the loudest "de-NVIDIA-ification" signal since Microsoft launched its Maia chip. For Anthropic, the custom TPU path means lower inference costs, theoretically, and deeper optimization of the training stack. It also means technical path dependency: once your model architecture, tooling, and inference stack are tuned to TPUs, switching costs become existential. Anthropic has effectively mortgaged its next five years of technical roadmaps to Google's silicon roadmap.
Compare the competitive landscape. OpenAI runs on Azure and is building Maia. xAI built Colossus — 200,000 GPUs in Memphis, self-operated. Meta constructs its own 240,000-H100 clusters. Google DeepMind eats its own dogfood on TPUs. Every frontier lab except Anthropic has chosen vertical integration. Anthropic has chosen a "semi-internalized" path: it does not own the infrastructure, but it controls it long-term — via contracts, guarantees, and financial engineering. Capital efficiency is maximal. Apparent asset lightness at IPO is maximal. Strategic independence is minimal.
Now the uncomfortable math. At $965 billion valuation and a $15 billion single-project investment, Anthropic's capex-to-valuation ratio sits around 1.5%. Compare ExxonMobil: $500 billion market cap, $200-250 billion annual capex — roughly 4-5%. If Anthropic's long-term infrastructure needs scale to meet its valuation, we are talking $200 billion-plus in cumulative investment. No operating cash flow can bridge that gap. No VC round can. The only answer is repeated forays into this exact kind of leveraged, off-balance-sheet architecture — which means repeated dependence on Google's willingness to guarantee.
Morgan Stanley is the tell. The same bank leads the Nexus financing syndicate and has been tapped for the IPO. This dual mandate — lender and underwriter — is not a conflict in legal terms. It is a convergence of incentives that demands scrutiny. When the lead arranger of the debt is also the lead banker on the equity story, priced by the same market, at the same moment, the question is not whether bias exists. It is whether the disclosure captures it.
Here is the contrarian angle nobody is talking about. The deal's weakest link is not Anthropic, not Google, not even the money. It is Nexus Data Centers. The developer's public record in hyperscale projects is thin. 1.6GW is not an incremental step. It is an order-of-magnitude leap. An 18-to-24-month delay in this project does not just dent Anthropic's training timeline. It triggers Google's guarantees, exposes the banking syndicate's credit, and lands Anthropic with chips it cannot rack and models it cannot train. The entire edifice rests on the delivery schedule of an unproven developer, a custom chip tape-out cycle of 18-36 months, and a gas plant that takes longer to build than the data center it powers.
There is a second blind spot: minimum purchase commitments. Anthropic likely signed take-or-pay provisions with Broadcom and Google for the TPUs. Those commitments become contingent liabilities on the IPO balance sheet. If the market notices before October 2026, the $965 billion valuation gets a haircut. If it notices after, the IPO investors eat the loss. This is not hypothetical. Enron taught the market that off-balance-sheet structures can beautify a story without changing its substance. The lesson did not expire.
Add the antitrust vector. Google holds 14% of Anthropic, guarantees billions in project debt, owns 20% of the data center SPV, supplies the chips, and rounds it out as landlord. The FTC has already scrutinized Microsoft-OpenAI. Google-Anthropic is structurally farther down the "control via capital" path. If regulators see the five-role tangle as a merger in disguise, the IPO timeline and financing terms both face forced adjustment.
Alpha dropped: Follow the money. The money says this is the year AI infrastructure became a financial engineering battleground. Meta-BlackRock's $14 billion El Paso co-financing. Brookfield-NextEra's $10 billion Paducah project with DOE backing. And now Anthropic's $15 billion SPV stack. The pattern is clear: big tech no longer builds with retained earnings. It builds with other people's capital, structured by banks, guaranteed by incumbents, and delivered by developers who may or may not be ready.
The risks, ranked: project delay and delivery failure, first. Regulatory review, second. The AI capex cycle turning, third. All three are live. All three are priced nowhere.
Anthropic's $965 billion valuation is a claim about the future. The $15 billion deal is a claim about the present. The gap between them is not an engineering problem. It is a question of whether financial innovation can outrun physical reality. Watch the FID announcement. Watch the EPC contractor selection. Watch whether Google's guarantee terms include a revocation clause if Anthropic stumbles. And watch the FTP filings, where the truth about those take-or-pay obligations will finally hit daylight.
The structure works — until it does not. And when it stops working, the first to know will be the gas plant's carbon counters, the consortium's credit committee, and the IPO market's pricing engine. In that order.