The $965 Billion Balance Sheet Anthropic Does Not Own

A $965 billion valuation does not purchase a single megawatt of power. That is the first fact to hold on to. Anthropic is now one of the most valuable private companies in the world, with a rumored October 2026 IPO designed to convert that mark into liquid equity. But the company cannot finance its own compute infrastructure. The proof sits near Hubbard, Texas, where a $15 billion data center project has been assembled for a tenant who owns none of it. Anthropic does not own the building. It does not own the turbines. It does not own the chips. The structure requires Google to guarantee billions of dollars in leases and power-purchase agreements, a debt syndicate led by Morgan Stanley to hold construction risk, a developer named Nexus Data Centers with a thin public record in hyperscale delivery, and a 1.6-gigawatt on-site natural gas plant that turns Anthropic into something closer to a municipality than a software company. The code compiles, but the reality bankrupts. The distance between a $965 billion mark and the physical liabilities standing behind it is the entire story. Anthropic is the frontier laboratory behind the Claude model family and its API. Its product is intelligence sold as tokens. Its cost function is compute. The old operating model was simple: rent GPUs from hyperscalers, pay the invoice, train the model, repeat. That model is gone. The new model is a stack of special purpose vehicles, supplier financing agreements, power contracts, and at least one AAA credit backstop. This is not a press release. It is a financing manual. THE DEAL The structure, as reported: the project has expanded from $5 billion at 612 megawatts, per an FT report from March 2026, to $15 billion at 1.6 gigawatts across 2,800 acres near Hubbard, Texas. The site includes behind-the-meter natural gas generation. Electricity is produced on site, bypassing grid interconnection queues that commonly run three to five years. The hardware is not NVIDIA. It consists of custom tensor processing units co-designed by Google and Broadcom, acquired through a vendor financing structure. The physical plant and power infrastructure sit in a separate financing vehicle in which Google holds approximately 20 percent equity, after providing billions in lease and power-purchase guarantees. Google is already a 14 percent shareholder in Anthropic. Morgan Stanley runs the debt syndicate for the project and is simultaneously the lead banker on the IPO. For crypto-native readers, this should trigger recognition. This is the DeFi treasury playbook applied to artificial intelligence. The tokens are called leases, guarantees, and sponsor equity. The economic shape is identical: a high-valuation operating company moves its capital-intensive assets off its balance sheet and rents them back, converting hidden liabilities into recurring expenses. THE BALANCE SHEET The entire structure exists to keep $15 billion of fixed capital off Anthropic's IPO balance sheet. This is project finance in textbook form: create a special purpose vehicle, transfer land, buildings, and generation assets into it, raise non-recourse debt at the project level, and sign a long-term lease so the operating company can train models without a massive depreciation line. 'Textbook' should never be confused with 'true.' Off-balance-sheet means the fixed assets will not appear on the income statement. It does not mean the costs disappear. Lease payments become operating expenses. Take-or-pay provisions become contractual obligations in the IPO prospectus. Minimum purchase commitments to Broadcom and Google for TPUs become obligations with delivery deadlines and penalty clauses. The historical precedent for this class of engineering is Enron. Two consequences follow for public-market investors. First, the apparent asset-light quality of Anthropic's pre-IPO financials will be a function of accounting classification, not economic substance. A lease is a lease, and a guarantee is a guarantee, whether labeled operating or financing. Second, the equity market has demonstrated, repeatedly, that it can price hidden liabilities once they are disclosed; the problem is the window before disclosure. The financial history of the last two decades is a sequence of exactly this mechanism. The lesson is not to avoid complexity. The lesson is to demand the footnotes. In my practice as a due diligence analyst, the first question in any financing structure is always the same: what is on the balance sheet, and what has merely not been accounted for yet? In 2017, I audited the vesting contract of a prominent Asian utility-token ICO and found an integer overflow that would have allowed early investors to drain 40 percent of the total token supply. I published the mathematical flaw as a GitHub issue. The project devalued quickly. The lesson was not about Solidity. The lesson was that the whitepaper promised one thing while the code omitted the failure case. The same discipline applies here. The real balance sheet of Anthropic consists not of the $15 billion facility but of every obligation hidden inside the word 'separate.' THE FIVE ROLES Now Google. The counterparty analysis is where this deal becomes structurally uncomfortable. Google is a 14 percent shareholder in Anthropic. Google is the architect of the TPU hardware family Anthropic will use for training and inference. Google Cloud is a primary distribution channel for Claude. Google is the guarantor of billions in lease and power-purchase obligations. And Google is a 20 percent equity partner in the project's real estate and energy assets. Five roles. One counterparty. The related-party problem is not whether Google intends to harm Anthropic. It is whether prices for power, rent, and chips can ever be independently verified as market-rate. When landlord, guarantor, chip supplier, shareholder, and cloud distributor converge on one entity, transfer pricing becomes untestable from the outside. I have seen this pattern in crypto. It resembles a single market maker serving as exchange, custodian, oracle, and lender for one token. The audit committee approves the arrangement; the conflict remains structural. Compare this with OpenAI-Microsoft. Microsoft holds a substantial equity stake in OpenAI, services its compute needs, and integrates its models. But Microsoft does not simultaneously design the chips, own the real estate, and carry the guarantees in the same manner. The Anthropic-Google arrangement reaches one layer deeper into the capital structure. The phrase 'strategic partner' no longer describes it. The word 'control' is closer, even if no one at either company is willing to say it aloud. Microsoft and OpenAI are already under regulatory scrutiny for a similar entanglement. A $965 billion company whose largest shareholder, primary guarantor, and landlord is also its most direct competitor is a dossier waiting for a regulator to open it. THE PHYSICS The engineering numbers deserve their own examination. 1.6 gigawatts is enough to power roughly 1.2 to 1.6 million American homes. It can support between 150,000 and 300,000 of the latest accelerator cards, depending on the per-card power envelope of one to two kilowatts. This is a first-tier installation by any international standard. The behind-the-meter gas strategy is rational in the short run. Grid interconnection is the bottleneck of American AI expansion; on-site generation removes it. Gas-fired generation costs roughly $40 to $60 per megawatt-hour, against an industrial tariff range of $50 to $80 in constrained regions. The unit economics are defensible. The unit physics are less comfortable. At a 50 percent capacity factor, this one plant emits roughly 3 to 4 million tonnes of carbon dioxide per year. That is Scope 1 emissions, owned directly by the project. No carbon-offset accounting can remove it. The 'sustainable AI' narrative does not survive contact with a 1.6-gigawatt gas plant. The cooling problem also remains unstated in the deal summary: at a power usage effectiveness of 1.2, the site needs roughly 2,000 megawatts of heat-exchange capacity. Near Hubbard, where water is not abundant, liquid cooling at this scale is an engineering question with real cost and real permitting risk. Beyond construction, there is the operating problem. A behind-the-meter natural gas plant requires fuel procurement, emissions compliance, and 24/7 operations staff. A 1.6-gigawatt facility has the same operational profile as a mid-sized utility. Anthropic is an AI company; its founders are not utility operators. Someone will run this plant. The contracts will specify who, but the management distraction and the compliance overhead will be real regardless. In my experience, the most common cause of project underperformance is not bad engineering. It is management attention divided across too many physical assets. Then there is time. Custom TPU silicon takes 18 to 36 months from design to volume production. Data centers take 12 to 18 months to construct. Natural gas plants take 24 to 36 months. The sequence must be orchestrated precisely. If power arrives first, the building arrives second, and chips arrive third, the tenant is paying for idle capacity through a web of leases and guarantees. There is also the unspoken supply-chain constraint. Broadcom's advanced-node capacity, fabricated at TSMC, is shared between Google's TPU roadmap and the new Anthropic custom line. When two large customers of the same foundry compete for the same 3-nanometer and 4-nanometer wafer starts, delivery priority is not a technical detail. It is a business decision. The source material does not say how priority will be assigned. It also does not say whether Anthropic has made minimum purchase commitments to lock in allocation. If it has, those commitments are contingent liabilities. If it has not, the chips may not arrive on schedule. The timeline mismatch is not a detail. It is the project. THE DEVELOPER Nexus Data Centers is the weakest link in this chain. The public record of hyperscale delivery from this developer is thin. A 1.6-gigawatt campus with on-site generation, cooling loops, substations, network backhaul, and accelerator clusters is one of the most complex construction programs in the American economy. If Nexus slips by 12 to 24 months, which is hardly an outlier for this asset class, Anthropic faces a worse outcome than spending more money. It faces the arrival of chips that are two generations old. Silicon in this market advances every 12 to 18 months. Compute that arrives late is not compute. It is depreciating inventory. THE RISK WATERFALL Let me map the default order. First loss sits with Anthropic's equity and future cash flows. Second loss sits with the project special purpose vehicle and its syndicated creditors. Third loss sits with Google, which issued the guarantees and will pay if the project cannot. Fourth loss sits with the bank group, which is underwriting 40-year infrastructure risk with roughly 36 months of gas-price visibility. Fifth loss sits with the developer, which is executing a project that may exceed its actual capability. There is also the matter of Morgan Stanley wearing two hats. The same institution that arranged the debt syndicate is leading the IPO. Syndicated lenders are expected to push for favorable terms for themselves; IPO underwriters are expected to push for a favorable price for the issuer. When both functions sit in one firm, the conflict is not hypothetical. It is a structural variable that will affect the timing of the float, the terms of the covenants, and the quality of the disclosures. The market will not price this waterfall at announcement. It will price it when the first milestone slips. The transaction is permanent; the mistake is not. THE COMPETITIVE SET The strategy becomes coherent in context. OpenAI runs on Microsoft Azure, is developing its own Maia silicon, and operates under a contractual envelope running into the hundreds of billions. xAI built Colossus in Memphis, exceeding 100,000 GPUs on a self-owned cluster. Meta operates self-owned AI fleets in the hundreds of thousands of accelerators. DeepMind trains Gemini on in-house TPUs. Anthropic has chosen a different path: it owns no warehouse, no turbine, and no wafer. It contracts. The formal term is asset-light. The practical term is semi-internalization. The asset-light approach improves the optics of an IPO. It converts a $15 billion capital problem into a lease problem. But every asset-light structure embeds a premium to the capital provider. Someone must absorb construction risk, power-price risk, and chip-delivery risk. In this structure, Google bears part of the guarantee, the syndicate bears part, and Nexus bears the execution portion. That premium is already in the terms. The question is whether the long-run marginal cost of leased compute stays competitive against the vertical integrators. That answer depends entirely on the hidden provisions: the take-or-pay volumes, the gas-price indexation, and the TPU refresh rights. The market should also watch NVIDIA's response. Anthropic has been one of the largest buyers of NVIDIA GPUs for frontier-scale training. This deal shifts a meaningful portion of that demand toward custom Google-Broadcom silicon. NVIDIA has responded to such defections before by adjusting allocation, pricing, and the strength of the CUDA lock-in. The effect on the broader AI chip market is not neutral. Every frontier lab that builds or buys custom silicon reduces the unit volume over which NVIDIA amortizes its own roadmap. The broader the custom-silicon trend spreads, the more pressure NVIDIA faces to push its next-generation architecture on price. THE ARITHMETIC The valuation gap is the core problem. ExxonMobil, at roughly $500 billion in market capitalization, spends $20 to $25 billion per year in capital expenditures, a ratio near 4 to 5 percent. If Anthropic is worth $965 billion and needs comparable capital intensity, the implied annual infrastructure spend is $40 to $50 billion, not $15 billion once. The revenue base, estimated near $1 billion annualized, cannot fund that. This is not a criticism. It is arithmetic. The IPO is not a strategic option. It is the mandatory exit ramp from a funding obligation that outruns private capital. The $15 billion project equals roughly 1.5 percent of the current valuation. A full build-out of Anthropic's stated compute ambitions implies $300 to $500 billion of infrastructure exposure across the next several years. There is no version of this plan in which the balance sheet stays as clean as the headline suggests. The supply of bank credit is cyclical. When AI return-on-investment expectations compress, the same way they compressed for technology in 2022, the credit standard for AI infrastructure tightens, and the cost of the next $50 billion tranche rises. Once the equity market prices the total liability stack, the adjustment will not be smooth. It will happen at the moment of maximum liquidity. That moment is the IPO. PATTERN RECOGNITION I spent two months in 2022 reverse-engineering the UST algorithmic stablecoin before the collapse. I calculated that the seigniorage loop required geometrically increasing demand for LUNA, an impossibility that the market ignored until the end. What I have learned from that exercise and from analyzing Uniswap v2 liquidity dynamics is that every finance structure has the same feature: it postpones the question of whether the underlying assets actually generate cash. AI infrastructure is real, in shortage, and useful. The assets are not fake. The risk is that the financing makes them look smaller than they are, and the counterparties look more independent than they are. WHAT THE BULLS GOT RIGHT Let me present the other side, because it is substantial. First, Google's guarantee is a priced signal. An AAA-rated counterparty does not deploy guarantee capacity for charity. Google's credit committee, with its own internal models, has assigned a low probability of default to Anthropic's obligations. In crypto, I have reviewed protocols where the audit says the code is sound. I do not trust the audit; I trust the exploit. But a guarantee from Google is, effectively, an exploit with the sign flipped. It is real backing. Second, the two-layer financing structure is genuinely efficient. By decoupling the chip layer from the physical layer, Anthropic can refresh silicon every 18 to 24 months without renegotiating a 30-year building contract. That is not an accounting trick. It is a designed fit between the lifecycles of assets and the lifecycles of technology. For an industry still learning how to fund gigawatt-scale compute, this may be the most practical structure available. Third, the asset class is being created in real time. Meta-BlackRock's $14 billion El Paso project and the Brookfield-NextEra $10 billion Paducah project are not footnotes. They are the first deployments of long-duration specialist capital into AI infrastructure. The entry of infrastructure investors lowers the cost of compute for the entire industry. That is a structural positive that exists independently of Anthropic's specific outcome. The bull case is not that this deal is clean. The bull case is that this deal is the first honest recognition that AI's next phase is a balance-sheet problem, and that the industry is inventing the instruments to solve it. TAKEAWAY Valuation does not move electrons. Guarantees do. Track the construction milestones, the TPU deliveries, and the IPO disclosures on minimum purchase obligations and related-party terms. Everything else, including the $965 billion number, is narrative. The transaction is permanent; the mistake is not. Illusion has a price tag; truth has none. The open question is whether the IPO prospectus will be written for clarity or for compliance. Read it as if the counterparties wrote it. Because they did.