Alphabet raises eighty-five billion dollars to extract rents from a data-center bubble.
The credit markets are behaving as they always do, chasing double-digit returns without regard for the bezzle at the end of the pipe. Tech equities have climbed thirty-one percent in a quarter, and the financial press has spent the week celebrating Anthropic’s quarterly coding-tool revenue as though a narrow enterprise subscription can validate $670 billion in capital expenditure. It cannot. The profitability is real for Anthropic; the $670 billion is a speculative overbuild that will either find a buyer or evaporate into stranded assets, and the bond market is being asked to finance the difference.
While the $85 billion figure landed as a sudden headline, it follows months of reporting on the sheer scale of capital required to keep the training clusters running, and it sits alongside the blockbuster IPO preparations for other AI firms that could define the next cycle of equity dilution. Alphabet has also issued a hundred-year bond in British pounds, a trade that makes sense only if you believe AI will still be a growth industry in 2126. The railroad executives of 1873, flush with cash from the new Northern Pacific, would have recognized the confidence — and would have recognized, too, the duration mismatch that follows. A century bond locks in an obligation for a hundred years; the chips and models it finances will be obsolete in five. The railroad bonds of the 1870s defaulted not because the railroads were useless but because the cash flows never caught up to the claims made on them at issuance.
Here it is worth being precise about what “an AI infrastructure build-out” actually means, because the public discourse and the quarterly earnings-call vocabulary have the misleading habit of treating it as a vague digital cloud rather than as physical and logistical extraction. The $670 billion is not going to code. It is going to silicon fabrication nodes, cooling systems for racks of GPUs that will dissipate heat at densities no conventional HVAC can handle, water consumption for direct-to-chip liquid cooling, and the transmission-layer infrastructure required to move petabytes of training data between data centres. The power draw for this stack is measured in gigawatts, and the only variable in the equation is whether the power comes from a long-term purchase agreement with a nuclear operator or from the regional grid that residential ratepayers subsidize. As Kate Crawford documented, every inference call requires a physical machine drawing power and consuming water, and the only difference between a functional AI model and a stranded asset is whether the hyperscaler can lock in a corporate buyer to cover the power-purchase agreement before the chips depreciate.
The pattern here precisely fits what Doctorow calls the bezzle — John Kenneth Galbraith’s term for the interval between the commission of a fraud and its discovery, when it registers as wealth for the perpetrator and no loss for the victim. Doctorow extended the term to tech booms, and it applies with precision. The $159 billion in bonds that the hyperscalers have issued globally this year, up from $17 billion in 2024 according to Dealogic, is not a reflection of underlying cash-flow generation, because the cash flows from AI inference are currently orders of magnitude lower than the training and infrastructure costs they require. The bonds are a reflection of investor appetite for the next iteration of the tech-cycle narrative.
The Anthropic profit number — widely reported this week — is a genuine data point, and it deserves the concession that it provides a floor to the bezzle. If a subscription-based coding tool can cover its operational costs at scale, the enterprise market is willing to pay some threshold for AI augmentation. But the achievement should be treated as what Doctorow, after Lee Vinsel, calls criti-hype: a narrow capability gain being sold as evidence for an unbounded shift. Anthropic’s $10.9 billion revenue target against $670 billion in capex from the four hyperscalers is one dollar of revenue for every sixty-one dollars of infrastructure spending. That is not a business; it is a bet placed with other people’s capital. Repeating the claim that “Anthropic is profitable” as proof that “the entire AI infrastructure build-out is sustainable” is a compositional fallacy. Anthropic’s profitability on a narrow coding tool subscription says nothing about the economics of running general-purpose foundation models that consume millions in GPU-hours per day to generate text and images for free or near-free tiers. The subscription revenue covers the inference cost for a specific segment; it does not cover the $670 billion.
The capital markets know this, which is why they are pricing the risk exactly where the evidence puts it. The spread on 10-year Alphabet and Microsoft bonds over U.S. Treasurys sits below the average investment-grade corporate bond, meaning the market is treating Alphabet’s $85 billion equity raise and the $159 billion in bond sales as if the AI infrastructure were as safe as a regulated utility. But Oracle tells a different story. Oracle has issued $43 billion in bonds since September while projecting tens of billions of dollars in annual burn as it attempts to pivot from traditional enterprise software into cloud computing, and its bonds trade near the speculative-grade tier — a pricing signal that the underwriters are reading the overbuild risk accurately for the second-tier players who lack Alphabet’s legacy ad-tech cash-gusher to backstop the debt. CoreWeave, the former bitcoin miner turned AI-cloud darling, saw its bonds hammered last year when construction delays surfaced; this year its stock has rallied forty-three percent and it has raised another $20 billion. The rebound was taken as a vote of confidence. It is more usefully understood as a demonstration that, during a bezzle, anyone with a chart and a promise can refill the tank.
As Jordan Chalfin of CreditSights put it, “Overbuild risk isn’t going away.” That is the statement of a credit analyst who knows the difference between a spread that compensates for risk and a spread that only acknowledges it while the music is still playing.
The overbuild dynamic has a precise analog in Canadian telecom, where the incumbent carriers engaged in a fibre-to-the-home arms race in the early 2020s, each laying parallel fibre tracts to the same street in anticipation of a residential broadband demand that never materialized at the projected price point. Each was left with billions of dollars in stranded fibre while the Competition Bureau noted that the infrastructure duplication had done nothing to lower consumer prices because the market was a triopoly, not a competitive market. The hyperscalers are doing exactly this on a global scale, borrowing in Canadian dollars, yen, euros, and Swiss francs — as Alphabet and Amazon have done this year — to fund data-centre capacity that will sit idle if the enterprise-software buyers decide that the marginal cost of AI inference exceeds the marginal increase in employee productivity.
The reason the stack has swollen so rapidly is not that the returns are in view but that the Wall Street professionals buying the bonds are, as MSI noted last week, largely protecting their careers. If the bubble pops, no one gets fired for having bought Alphabet paper; if it doesn’t, you are a genius. The career-risk calculus extends to the equity side: Alphabet’s $85 billion raise, which will dilute existing shareholders, went through with barely a wobble because the fund managers who approved it are playing the same game.
The mechanism to prevent the overbuild from becoming a systemic credit event lies in the structural remedies that antitrust has been avoiding for forty years: mandating open interoperability at the API layer so that the hyperscalers cannot use model lock-in to force enterprise customers into continuous API subscriptions, and applying utility-style rate-of-return regulation to the cloud-compute monopolies that provide the physical substrate for the entire AI stack, capping their extraction at the weighted average cost of capital and forcing them to bear the depreciation risk of the chips they buy, rather than socializing the bond risk and privatizing the inference revenue. Requiring the Federal Energy Regulatory Commission to treat large-scale GPU clusters as rate-base assets subject to cost-of-service regulation, as the U.S. did with wholesale electricity in Order 888, would cap the infrastructure’s return at the weighted average cost of capital and force the hyperscalers to absorb chip-depreciation risk directly.
The $85-billion raise will close. The $159 billion in bonds will settle. The data centres will rise out of the agricultural land in Oregon and the desert scrub in Arizona, and the power-purchase agreements will lock in the grid capacity for the next twenty years. The question for the public investor is not whether the AI models will eventually generate enough value to justify the spend, but who will own the debt when the chips reach the end of their three-year depreciation cycle and the inference margins compress.
There is a public filing with the SEC on the registration statement for Alphabet’s offering. The prospectus runs to hundreds of pages, and it deploys the standard boilerplate that summarizes the trade: capital raised to fund working capital and general corporate purposes, including infrastructure investments. The capital is already being spent. The retail investor will never see the depreciation schedule, and the institutional underwriters will price the equity to reflect a risk the prospectus buries in the footnotes. The century bond will mature on a date none of its underwriters will live to see.