The AI Crowding-Out Effect
Anthropic's S-1 reveals over four hundred billion dollars in compute commitments and complex off-balance-sheet vehicles. The true barrier for AI is no longer the algorithm: it is capital crowding out the real economy and an electrical grid that cannot be compiled overnight.
When Anthropic filed its initial Form S-1 registration statement, public commentary focused on voting trusts, revenue run-rates, and the governance mechanics of its Long-Term Benefit Trust. Almost nobody analyzed the capital structure disclosures.
Buried in the commitments note is a figure that belongs to a sovereign state rather than a software startup: over 400 billion dollars in aggregate future compute and cloud obligations. Alongside it sits a 63 billion dollar special purpose vehicle arranged with Wall Street syndicates to purchase custom Broadcom processors off the balance sheet, a 15 billion dollar revolving credit facility that explicitly carves compute leases out of its debt caps, and twenty-year take-or-pay leases on powered data center shells.
Software was supposed to be the ultimate asset-light business. You wrote the code once, ran it on shared servers, and enjoyed near-zero marginal cost at scale. Frontier AI has inverted that equation entirely. The industry now operates like an unholy hybrid of an aluminum smelter, an aircraft leasing desk, and a sovereign wealth fund. The primary constraint on artificial intelligence is no longer the model architecture. It is a collision between financial gravity, physical lead times, and the cost of capital for everyone else.
The balance sheet tells the story before any code even runs. When you tally the disclosed contractual liabilities, the sheer scale of modern frontier commitments dwarfs conventional corporate debt.
Seeing these obligations stacked together highlights an unsettling asymmetry. The actual corporate debt on the balance sheet is negligible, while off-balance-sheet compute guarantees stretch outward through 2045. That staggering volume of pledged future cash does not exist in an economic vacuum. It alters the gravitational field of corporate finance.
The Great Capital Vacuum
Capital is finite, and institutional liquidity has boundaries. Every hundred billion dollars locked into cloud capacity backstops, hardware leasing syndicates, and private credit facilities is capital extracted from the wider economic pool. When hyperscalers and frontier labs demand unprecedented sums to underwrite physical compute clusters, they fundamentally reshape the price of money for everyone else.
Sovereign borrowers, municipal utility districts, auto-parts suppliers, and conventional software startups all drink from the same reservoir of institutional debt. When mega-cap technology sponsors issue billions in high-yielding, investment-grade paper to build data centers, they reset the baseline hurdle rate across the credit market. An asset manager overseeing pension capital faces an easy decision: why finance a municipal water treatment facility or a domestic tooling plant at six percent when tech-backed compute debt yields eight percent with fortress sponsorship?
This dynamic illustrates textbook macroeconomic crowding out. As institutional pools reorient their mandates toward AI infrastructure, the cost of borrowing for conventional borrowers climbs while their available credit shrinks.
Capital does not multiply simply because engineers invent larger neural networks; it merely changes hands. When hundreds of billions flow toward specialized silicon clusters, traditional industries find themselves competing against an insatiable bidder with access to subsidized liquidity.
The Emergence of an AI-Only Economy
For an ordinary commercial enterprise, borrowing money in this climate has become prohibitive. A regional trucking carrier looking to modernize its fleet or a precision machine shop attempting to finance five-axis CNC equipment faces commercial interest rates between 9 and 11 percent. Private credit desks that spent decades underwriting middle-market manufacturing loans have systematically shifted their loan books toward hyperscaler shells, energy development leases, and GPU collateral packages. A credit committee evaluating risk rarely chooses an industrial fabrication plant in Ohio when it can purchase syndicated debt paper backstopped by multi-year cloud contracts from trillion-dollar tech giants.
The consequence is a deeply bifurcated economy operating under two separate sets of financial physics. In the first tier, AI initiatives enjoy abundant liquidity, sovereign wealth allocations, and subsidized cost of capital. In the second tier, productive physical enterprises and standard software businesses are squeezed by tight bank covenants and steep hurdle rates. When the bar for funding rises this high, only projects promising astronomical intelligence windfalls can justify the carrying cost of debt. Everything else gets shelved.
This credit divide splits the market down the middle, separating those who can tap sovereign-backed tech syndicates from those dependent on traditional bank lending.
As capital flows disproportionately into data centers and custom silicon, the rest of the productive base pays the penalty in elevated borrowing costs and deferred capital expenditure. Yet this financial distortion is only the first layer of the problem. An even sharper constraint appears when financial contracts meet physical reality.
The Speed of Light Meets the Speed of Copper
The financial distortion is only half the dilemma. The more dangerous friction is the violent mismatch between software time and physical time. In modern software engineering, rapid iteration is the default expectation: a training run completes on Tuesday morning, model weights propagate across clusters by afternoon, and container fleets automatically scale up and down to match incoming packet traffic in fractions of a second.
The electrical grid, however, does not possess an autoscaler. Bringing a gigawatt of fresh compute capacity online requires navigating physical supply chains that move at an agonizingly deliberate pace. High-voltage step-up transformers carry procurement backlogs of three to four years. Interconnection study queues at regional transmission operators like PJM or ERCOT regularly stretch four to six years before construction crews even break ground. If a planned data center campus requires dedicated baseload generation, whether from a combined-cycle natural gas turbine or the recommissioning of a dormant nuclear reactor, the permitting and construction cycle stretches well past eight to ten years.
Comparing the cycle times of software releases against physical infrastructure exposes the chasm between digital expectations and heavy industrial manufacturing.
Software moves at the speed of light, and financial contracts execute at the speed of wire transfers. But copper, turbines, switchgear, and concrete obey the immutable timelines of metallurgy, civil engineering, and regulatory review. Signing power purchase agreements or announcing reactor restarts makes for compelling headlines, but it does not compress the calendar. You cannot hot-reload a 500-kilovolt substation.
This multi-year physical bottleneck creates an existential accounting crisis. If you take delivery of tens of billions of dollars in advanced silicon while waiting four years for utility substations to be energized, that hardware sits dark in warehouse crates. Because advanced accelerators lose their competitive edge within three to four years, your assets are actively depreciating into obsolescence before they ever draw a single watt of power. If those servers sat directly on a corporate balance sheet, standard GAAP depreciation would obliterate quarterly operating income and breach debt covenants. This exact crunch is what forced tech finance to invent off-balance-sheet engineering.
Off-Balance-Sheet Financial Engineering
This is the unpublicized reality behind Anthropic's 63 billion dollar special purpose vehicle. If a frontier lab or cloud provider carries hundreds of billions of dollars in fast-depreciating silicon on its balance sheet, its return on invested capital collapses before the first cluster reaches full utilization. A new generation of accelerators arrives every eighteen months, cutting the economic half-life of prior hardware in half. If an enterprise issues corporate bonds to buy processors that spend their first two years waiting in staging facilities for utility power lines, straight-line depreciation charges and financing costs will destroy its reported earnings.
To bypass this trap, corporate treasuries borrowed a playbook from commercial aviation and offshore infrastructure. Instead of buying the silicon directly, the company establishes an off-balance-sheet special purpose vehicle backed by Wall Street bank syndicates and private credit funds. The SPV purchases the chips from hardware suppliers like Broadcom, retains legal title to the hardware, and secures residual-value guarantees to cushion against steep market declines. The frontier lab then contracts to lease compute capacity from the vehicle as an operating expense, while carving the underlying debt obligations out of its primary corporate credit covenants.
The corporate structure isolates the hardware liabilities from the operating entity, routing capital and debt through bankruptcy-remote entities.
The architecture is undeniably sophisticated, but financial engineering cannot alter physical reality. Moving assets into an off-balance-sheet vehicle shields headline earnings from immediate depreciation, but the cash obligations remain binding. The bank syndicates expect their interest distributions on schedule, and the hardware continues its relentless march toward obsolescence whether the cooling towers are humming or silent.
When Gravity Asserts Itself
The tech sector has experienced this brand of infrastructure overreach before. During the telecom boom of the late 1990s, telecommunications carriers laid millions of route miles of long-haul fiber-optic cable across oceans and continents. Equity investors and bondholders convinced themselves that exponential internet traffic growth would quickly monetize the debt required to lay the glass. The technology itself was authentic, and over the ensuing two decades, society used every single strand of that dark fiber. Yet the capital structure was premature. The companies that financed the initial trenching collapsed into bankruptcy, and secondary operators later acquired the physical assets for pennies on the dollar.
Today's AI infrastructure cycle mirrors that trajectory, but with two compounding vulnerabilities. The first is scale: the aggregate commitments disclosed in S-1 filings and private credit syndicates measure in hundreds of billions of dollars per lab, far outstripping the capital expenditures of late-nineties telecom. The second vulnerability is physical durability. Glass buried beneath the earth does not decay; it can sit dark in a conduit for fifteen years and remain fully functional when new optical lasers light it up. Advanced semiconductor chips, by contrast, lose twenty-five to thirty-five percent of their economic value every single year due to semiconductor manufacturing advancements and architectural shifts.
Comparing the economic lifespan of dark fiber against advanced AI processors reveals why the modern compute buildout cannot afford idle years.
When fiber-optic carriers went bankrupt, the physical networks remained intact, waiting patiently for software and consumer demand to catch up. Stranded silicon enjoys no such luxury. If a gigawatt data center cluster sits idle for three years waiting for electrical grid interconnects, the underlying hardware arrives at the finish line economically obsolete. The equity and debt backing those processors cannot be recovered by flipping a switch years later.
Until capital expenditure rhythms synchronize with the physical realities of the electrical grid, the crowding-out effect will extract a steep toll from the wider economy. Underwriting frontier intelligence is a monumental engineering feat, but allocating hundreds of billions of dollars into stranded, fast-depreciating silicon while raising the price of capital for every other productive industry is an unsustainable trade. Financial alchemy can disguise idle depreciation for a few quarters, but it cannot outrun physics.
← Essays & Opinions