Predictability is a myth; only volatility is real. The most volatile number in the AI stack is not a model benchmark, a token price, or a chip allocation. It is the distance between a private valuation and a physical power purchase agreement.
Anthropic is valued at $965 billion. The same week, reporting confirmed a $15 billion, 1.6-gigawatt data center project in Hubbard, Texas: 2,800 acres, a behind-the-meter natural gas plant, custom TPUs co-designed by Google and Broadcom, and a financing architecture engineered to keep the entire asset base off Anthropic's balance sheet. The developer, Nexus Data Centers, carries a conspicuously thin public record in hyperscale delivery.
$965 billion is a story about future earnings. $15 billion, 1.6 gigawatts, 2,800 acres is a story about concrete, steel, gas turbines, semiconductor lead times, and permitting schedules. The gap between those two stories is where this analysis begins.
I have spent eighteen years reading the gap between narrative and infrastructure. In 2017, I audited the Parity multisig wallet and published a technical pre-mortem three days before the exploit, forecasting a $30 million loss while the market celebrated the contract's growth. In 2022, I published a mathematical breakdown of UST's seigniorage death spiral six hours before the peg dissolved. The recurring lesson: when capital structures and physical mechanisms diverge, assume the mechanism wins. Sentiment is delayed settlement.
The AI infrastructure cycle is now producing the same pattern at a scale the crypto market can only envy. Anthropic's Texas project is not merely a data center deal. It is a financial instrument disguised as infrastructure, a potential template for the next wave of AI capital formation, and a live stress test of how far a $965 billion valuation can stretch before the physics underneath it push back.
Context: The Financing Frontier
Anthropic plans an IPO in October 2026, with Morgan Stanley as reported lead banker, targeting a $965 billion valuation. That would make it one of the largest public companies on the planet before meaningfully disclosing audited financials.
The infrastructure deal behind that valuation has escalated in real time. The Financial Times first reported a $5 billion project at 612 megawatts in March 2026. The current structure stands at $15 billion and 1.6 gigawatts — nearly a threefold expansion in both capital and power within months.
That scope escalation is a double-edged signal. Either genuine compute demand breakthroughs are driving the expansion, or the momentum dynamics of a financing market where bigger is always read as better are driving it. The two possibilities carry very different risk profiles, and current reporting does not yet distinguish between them.
This financing cycle is not happening in a vacuum. The bull market in AI-related equities has created an appetite for infrastructure-yield products that did not exist eighteen months ago. Private credit desks, pension funds, and infrastructure investors are rotating into AI compute assets because offtake contracts — signed by companies like Anthropic with AAA-rated guarantors like Google — resemble the stable cash flow profiles of traditional infrastructure. That analogy is only valid if the offtaker remains solvent and the technology remains relevant. Both assumptions are untested at this scale.
The physical design is equally instructive. The project occupies 2,800 acres in Hubbard, Texas, and includes a behind-the-meter natural gas plant. The logic is defensive: grid interconnection queues in many US regions run three to five years, and transmission instability is existential risk for AI training clusters. A behind-the-meter plant functions as a micro-utility, exchanging regulatory complexity for fuel procurement risk, emissions exposure, and single-source energy fragility.
The compute layer is custom. Anthropic is acquiring tensor processing units co-designed by Google and Broadcom under separate supplier financing agreements. This is a deliberately engineered split: physical infrastructure and power are financed in one vehicle; chips are financed in another. The stated goal: isolate the most volatile capital expenditures from the IPO balance sheet.
The counterparty web is dense. Google is providing billions in lease and power purchase agreement guarantees, receiving a 20 percent equity stake in the project entity in return. It already holds approximately 14 percent of Anthropic. Read that list carefully. Google is simultaneously Anthropic's largest shareholder, its primary compute supplier, its principal chip co-designer, its lease guarantor, its power purchaser of last resort, its landlord — and its direct competitor in frontier models. That is not a partnership. That is a five-layer dependency with competition embedded at every layer.
Core: The Anatomy of the Off-Balance-Sheet Engine
Project finance is a century-old discipline built for oil fields, pipelines, and liquefied natural gas terminals. The core logic: isolate a capital-intensive asset in a special purpose vehicle, fund it with limited-recourse debt secured by the project's own cash flows, and keep the asset and its debt off the sponsor's balance sheet. Anthropic and Google have now applied that logic to frontier AI compute.
The architecture splits the project into at least two distinct layers. The first holds the long-lived physical asset: land, buildings, cooling systems, the gas plant. The second covers chip acquisition through supplier financing agreements. Each layer carries different capital providers, risk profiles, and claim structures.
The sophistication is real. A data center has a useful life of 25 to 30 years. A TPU generation is obsolete in 18 to 24 months. These are different asset classes with different physical dynamics, and financing them separately is sound engineering — matching long-dated capital to long-lived assets and shorter-dated supplier credit to short-lived chips.
The less discussed component is the IPO optics. When Anthropic files to go public, its balance sheet will show a comparatively light fixed-asset base. The heavy machinery, the gas turbines, the 1.6 gigawatts of negotiated power — those will sit in vehicles owned by other parties, leased or contracted. This is the classic off-balance-sheet playbook, and it is not inherently fraudulent. But the question for investors is not whether the structure is legal. The question is whether the risks have been priced, or merely relocated.
The relocation is consequential. Anthropic's annualized revenue is estimated in the single-digit billions. A $15 billion project is a capital commitment several multiples of that revenue. The long-run capital expenditure intensity of frontier AI is not yet reflected in any public valuation. When the market reprices that, the adjustment will be binary.
Quantifying the Gap
Let me quantify the scale. 1.6 gigawatts is roughly the electricity consumption of 1.2 to 1.6 million American homes. The OpenAI/Microsoft project in Mount Pleasant, Wisconsin, runs approximately 2.4 gigawatts. xAI's Colossus in Memphis has scaled to roughly 300 megawatts. This single project places Anthropic in the first tier of US AI data centers.
The benchmark comparison is stark. ExxonMobil, at roughly $500 billion market capitalization, spends between $20 and $25 billion annually on capital expenditures — a ratio near 4 to 5 percent. Anthropic at $965 billion with one $15 billion project sits near 1.5 percent, and this is one project in one state. The multi-year trajectory implies hundreds of billions in additional infrastructure commitments. The capital intensity will not stay at 1.5 percent. Either the valuation declines or the capital expenditures explode.
The carbon ledger is equally substantial. A 1.6-gigawatt gas plant at a 50 percent capacity factor produces roughly 3 to 4 million tons of carbon dioxide equivalent annually. That is a Scope 1 emissions footprint that contradicts any credible carbon neutrality commitment and will become a regulatory and reputational liability as these projects multiply.
Water and cooling are the unmentioned constraint. A 1.6-gigawatt facility at a power usage effectiveness of 1.2 requires roughly 2,000 megawatts of heat exchange capacity. Direct liquid cooling at this scale demands either enormous water supplies or closed-loop systems with massive cooling infrastructure. In Hubbard, Texas, water availability is a genuine engineering constraint that has not been addressed in any public reporting. This is the kind of detail that delays projects by quarters.
The financial transformation is worth stating plainly. Off-balance-sheet treatment converts capital expenditures into operating expenses. Anthropic will pay rents, tolling fees, and compute service charges for the life of these contracts. On an income statement, that looks disciplined. On a lifetime cost basis, Anthropic is paying the cost of capital of every entity in the chain, including Google's expected equity return on its 20 percent stake. There is no free lunch in project finance. The lease is a loan wearing a costume.
The engineering economics of the gas plant deserve scrutiny. Natural gas generation at $40 to $60 per megawatt-hour compares favorably to US industrial electricity prices of $50 to $80 per megawatt-hour. But that spread is a hedge against grid volatility, not a guarantee of savings. Gas prices are volatile, and the plant's carbon exposure is permanent. More importantly, a behind-the-meter plant sacrifices the grid's statistical multiplexing: when the plant trips, there is no alternative source, and the entire training cluster goes dark. At the scale of a hundreds-of-thousands-of-GPU cluster, a ten-minute outage is not a rounding error. It is a synchronization catastrophe.
The Execution Chain
The execution chain is where this structure is most fragile. Nexus Data Centers is the developer, with a limited public record at this scale. A 1.6-gigawatt campus with an on-site power plant is not a routine delivery. It requires simultaneous execution of substation work, cooling loops, network backhaul, gas supply contracts, emissions permitting, and chip rack-and-stack across hundreds of thousands of square feet.
The typical failure mode in hyperscale delivery is not one catastrophic event; it is the accumulation of 18 to 24 months of delays across interdependent trades. In an industry where AI chips move on a 12-to-18-month generation cycle, a 24-month delay means the data center comes online with obsolete silicon. That is a sunk-cost event at a scale this industry has not yet priced.
There is also a temporal mismatch risk in the chip layer. Custom silicon from architecture to mass production runs 18 to 36 months. The Anthropic TPU order will compete with Google's own TPU allocations within Broadcom's finite production capacity at TSMC. If the chip design cycle and the construction cycle drift out of phase, you get either a data center waiting for chips or chips waiting for a dark data center. Both are expensive; only one is visible in the reporting.
The engineering timeline deserves forensic reconstruction. The gas plant alone requires 24 to 36 months from notice to proceed to commissioning — typically longer than the 12-to-18-month shell construction for the data center itself. That implies either the power plant is already sanctioned or the entire schedule carries a hidden variance of a year or more. A phased delivery model is the only credible reading: the campus will come online in stages, and Anthropic's compute capacity arrives in tranches, not as a single 1.6-gigawatt event.
The expansion from 612 megawatts to 1.6 gigawatts supports that reading. The original scope was likely Phase 1; the full project carries a multi-tranche, multi-year delivery profile. That is not bearish by itself. But the market is being sold a headline number, not a schedule. The difference between the two is where underperformance is born.
The Precedent Pattern
This deal does not exist in isolation. Meta and BlackRock have co-financed a $14 billion AI data center project in El Paso. Brookfield and NextEra structured a $10 billion project with the Department of Energy at Paducah, deploying what is effectively a sovereign-land model. Three structures in parallel, all within the same financing cycle, point to the same conclusion: big tech is no longer willing to absorb the full capital intensity of AI infrastructure on its own balance sheet. The sector is shifting from corporate capital expenditure to third-party capital plus long-term offtake agreements.
This is the financialization of AI infrastructure, and it changes the risk topology of the industry. Equity risk migrates from tech balance sheets to the capital markets, but operational risk — construction, power, chip delivery — remains in the real economy. Markets price the former; reality settles the latter. The divergence between those two pricing timelines is the source of the next systemic shock.
The banking industry is adapting. Morgan Stanley's simultaneous roles as financing syndicate lead and IPO lead banker suggest investment banks are building a combined product line: infrastructure finance plus equity capital markets. This mirrors the rise of telecom infrastructure REITs in the early 2000s — a new asset class built on the back of a technological boom, with the same potential for over-construction and valuation disappointment.
The Governance Layer
Morgan Stanley's dual role deserves the scrutiny it is not receiving. Banks routinely hold lending and advisory positions with the same client. But when the loan book is in the billions and the IPO valuation approaches a trillion dollars, the conflict surface expands. Underwriting discipline and lending discipline do not always pull in the same direction. If the project underperforms, the bank's incentives as lender and as underwriter are not perfectly aligned with the investors brought into the equity story.
The disclosure burden on Anthropic's IPO filing will be substantial. Related-party transactions with Google — leases, guarantees, power purchase agreements, TPU supply terms — must be described and quantified. The quality of that disclosure will determine whether the market treats $965 billion as a floor or a ceiling. Investors will need to evaluate whether the project's pricing reflects arm's-length terms or a strategic discount extended by a shareholder with motives beyond financial return.
The Competitive Positioning
The strategy comparison sharpens the picture. OpenAI runs on Azure with a parallel self-developed Maia chip effort, seeking diversification while carrying a multi-hundred-billion-dollar commitment to Microsoft. xAI built its Colossus cluster at record speed and operates it directly. Meta self-built one of the world's largest compute fleets. Google DeepMind trains on its own TPUs at internal cost.
Anthropic has chosen the most contract-intensive, least-vertically-integrated path among the frontier labs. It does not own the infrastructure. It owns contracts, guarantees, and promises — a portfolio of claims on compute, backed by Google's balance sheet. The strategy maximizes capital efficiency and keeps the IPO balance sheet light. But it embeds a structural cost: the long-term marginal cost of compute includes the risk premium of every capital layer beneath it. In a sustained frontier-model price war, higher lifetime cost per token is the slowest and most certain way to lose.
The path dependency is the deeper issue. If Anthropic trains its models, optimizes its inference stack, and tunes its toolchains around Google-designed silicon for three years, switching costs to alternative hardware become prohibitive. The company will tell investors it retains optionality. The physics of software-hardware co-optimization will say otherwise.
And NVIDIA is the quiet loser in this arrangement. Anthropic has been among the largest purchasers of NVIDIA GPUs. The pivot to custom TPUs signals that frontier labs can escape the NVIDIA tax through co-design — a strategic threat NVIDIA will answer with tighter CUDA lock-in and bespoke offerings. The supply-chain map of frontier AI is being redrawn, and Broadcom emerges as the arms dealer of record.
The structural irony deserves naming: the more successful Anthropic's models become, the more compute they consume, and the deeper the dependency on Google grows. In that feedback loop, Google's five roles create a compounding claim on Anthropic's future. Every victory tightens the leash. The optimal outcome for Anthropic's technology is not aligned with the optimal outcome for Anthropic's independence.
Contrarian: The Composed Risk Nobody Is Pricing
History does not repeat, but it rhymes in binary. And this structure rhymes almost perfectly with the composability failures I have modeled for years in decentralized finance.
The canonical DeFi collapse is not a single protocol breaking. It is a cascade: one collateral asset drops 20 percent, liquidation cascades through lending pools, oracle lag compounds the pressure, and the crash was not forecast because nobody modeled the dependency graph. The June 2020 liquidity shock in Aave and Compound was precisely this — not one bug, but a lattice of interlocking assumptions about price stability, oracle integrity, and liquidity depth, failing in sequence.
The Google-Anthropic dependency graph has the same topology. Google is shareholder, chip designer, guarantor, landlord, and competitor. Each role is a separate contract with a separate risk profile. But they are not independent. A shock to Google — antitrust action, earnings pressure, a strategic AI pivot, a regulatory cap on cross-investment — propagates simultaneously through every layer of Anthropic's compute stack. This is not a hedge. It is correlated risk concentration dressed as a diversified partnership.
The antitrust vector is the least reported dimension of this deal. Google holds 14 percent of Anthropic, provides billions in guarantees, and will own 20 percent of the project entity. The FTC's scrutiny of the Microsoft-OpenAI relationship has already established that capital-plus-contract control of frontier AI labs is a regulatory concern. The Google-Anthropic arrangement is structurally similar, with tighter bonds: real estate, energy, and chip design are bundled into the dependency, not merely compute rental. Regulatory remedies — forced divestiture, guarantee unwinding, governance changes — could arrive on the same timeline as the IPO. October 2026 is not simply an execution deadline. It is a regulatory exposure window.
Then there is the tail-risk question no one in the syndicate appears to be asking: who absorbs the shock if AI sentiment turns and the $965 billion valuation proves mid-cycle rather than early-cycle? The banks hold construction and term loans. Google holds guarantees and equity. Nexus holds delivery risk. Anthropic holds offtake commitments. If API revenue decelerates, minimum purchase commitments and lease payments become fixed charges against a shrinking top line. The guarantee structure exposes Google. The syndicate is exposed. And the IPO investor, presented with a clean balance sheet, absorbs the least visible layer of liabilities.
Stability is an illusion maintained by ignoring latency. The latency here is the delay between the capital markets underwriting the valuation story and the capital markets recognizing that the story is secured by an off-balance-sheet chain of promises terminating at a single counterparty.
I want to be precise about what this analysis does not claim. Project finance is not fraud. Google's role is not a conspiracy. Custom TPUs are not irrational. Each layer, examined in isolation, is defensible. The problem is the system. And the lesson of systemic interdependence — learned from 2017, learned from 2020, learned from 2022 — is that individually rational layers produce collectively fragile systems.
There is a specific resonance for the crypto industry. We ran this experiment at smaller scale for years. Terra/Luna failed not because seigniorage theory was wrong, but because reserve adequacy was an illusion — the gap between a protocol's promise and the collateral behind it. FTX failed not because exchange technology broke, but because the distance between its balance sheet and the claims on it was unbridgeable. The AI infrastructure market is entering the same phase. The $965 billion valuation is the promise. The Texas project, the Google guarantees, the Nexus delivery schedule, the TSMC production slots — that is the reserve adequacy. Until the infrastructure ships and throughput is proven, the valuation is unsecured, uncollateralized, and dependent on the continued goodwill of a direct competitor.
My own risk modeling from 2020 taught me that the most accurate models include counterparty concentration as an explicit variable rather than assuming independent defaults. Applied here: Google's credit rating is effectively the reserve currency of this entire structure. One downgrade event ripples through every syndicate participant and raises the cost of future AI infrastructure capital for every lab in the market. The contagion would not be contained to Anthropic. It would repriced the entire frontier compute complex.
Takeaway: The Next Watch
Watch three indicators. First, the final investment decision and the EPC contractor announcement for the Texas project. If the developer's role changes or the schedule slips more than one quarter, treat it as the first crack. Second, any FTC or DOJ inquiry into the Google-Anthropic capital structure. The Microsoft-OpenAI precedent makes this a when, not an if. Third, the next Anthropic disclosure: minimum purchase commitments, lease guarantees, and related-party transaction terms will reveal whether the off-balance-sheet structure is a genuine efficiency or a delayed charge against the IPO story.
The deeper question is whether this financing template becomes the industry standard. If it does, we are watching the birth of a genuine asset class: AI infrastructure as a securitized, tradable, off-balance-sheet claim on future compute. That asset class will accelerate the buildout and create a new systemic risk — one that lives on no single balance sheet and therefore on no one's risk dashboard.
Predictability is a myth; only volatility is real. The volatility here is not in model benchmarks. It is in the distance between a $965 billion valuation and a 2,800-acre construction site in Hubbard, Texas.
The next 24 months will determine whether Anthropic's off-balance-sheet empire is the most sophisticated capital structure in AI history — or the most elaborate pre-IPO lease in modern finance.