A two-trillion-dollar company that needs a cosigner

On November 4, the day before the vote, a company called TeraWulf filed a disclosure about a Kentucky data-center campus carrying roughly $19 billion in contracted Anthropic lease commitments, spread over a twenty-year lease. The debt behind those buildings comes with contractual credit support from Google — a cosigner, in effect, because the tenant's own name wasn't enough to get the financing done. TeraWulf said so itself.

We structured our Anthropic leases from the outset to address counterparty credit, financing costs and capital requirements. — TeraWulf

The sector's flagship IPO candidate needed its own customer to sign for its buildings ^[1]. The next day brought the election sweep, and the week brought the paradox. METR certified that Anthropic is now improving its models at roughly twice its prior-year rate, with three advanced models trained in six months ^[2]. Over those same weeks, the listing walked from a roughly $2 trillion pitch ^[3] to a marketed range of $1.5 to $2 trillion ^[4], and HSBC reported large Anthropic investors questioning the firm, with the valuation possibly downsized to $1 trillion to $1.5 trillion ^[5]. Dan Niles put the other side of the inversion in numbers: sequential revenue growth fell from roughly 145% in the second quarter to 43% in the third, with the fourth projected near 20% ^[6]. Faster models, slower receipts. Where did the doubled speed actually go? Shelf by shelf. The most capable shelf is the one no customer can buy. The three advanced models METR counted, trained in six months, remain unreleased ^[2]. The danger of shipping is on the record: DeepSeek and Moonshot AI used the outputs of OpenAI's GPT-6 Astra to train their own latest models, which means releasing the frontier hands it to competitors by distillation ^[7]. OpenAI reached the same conclusion: GPT-6 Cyber, a model that found 588 previously unknown Linux kernel flaws and 81 Firefox vulnerabilities, shipped only to a small set of approved customers through a gated program called Daybreak Red ^[8]. The pattern stretches back at least to May, when Anthropic's most powerful model, Mythos, was reportedly being sought by the NSA and Treasury rather than offered publicly ^[9]. One shelf down is what actually ships. On November 5, the same day as the METR report, Anthropic's public launch was a personal-finance assistant for Claude, with a limitation the company stated up front.

With Personal Finance in Claude, you can securely connect your bank and card accounts and ask in your own words. — Anthropic

A model reported capable of hacking major systems stays behind a gate; the model the public gets cannot move money. And what does ship for is being sold into a price war, as high-quality open-weight models from Meta, Nvidia, and Asian developers compress token prices ^[6]. The third shelf is where the acceleration actually lands: the cost side. Anthropic's balance sheet carries roughly $518 billion in decade-long, largely non-cancelable compute commitments under a take-or-pay clause — a contract where the full amount is owed whether the capacity is used or not. The company's own filing states the obligation plainly.

If our actual spend falls short, we must pay Google the difference. — Anthropic

Against that sits a $42 billion net loss for 2025 ^[3]. Training three models in six months does not shrink the commitment. It enlarges the invoice. The timing is worth setting down in dates rather than drama. On October 8, at an industry event, three attendees said they would buy Anthropic shares at a $3 trillion valuation ^[10]. On October 15, the company cut its range to $1.5 to $2 trillion, with investors citing revenue-growth deceleration ^[4]. On October 16, Uttam Dey was already redeploying into high-quality AI names on panic about slowing lab revenue and Chinese competition ^[11]. Then came the sweep. The public pitch walked, in about three weeks, back toward where private money had already stood since May ^[9], and the only material news about the company's capabilities in that window was that they were improving twice as fast ^[2]. The counter-evidence doesn't rescue the labs; it confirms the divide. Google's revenue rose 24% on enterprise AI demand, with a backlog grown to nearly four times trailing revenue ^[12]: the acceleration monetizes, but on the customer's side of the ledger. When David Katz named eight AI stocks for a rebound, the list ran Applied Materials, Generac, Qualcomm, TE Connectivity, Alphabet, Meta, Microsoft, Amazon — a basket of chip, power, and component suppliers plus the hyperscalers themselves ^[13]. No frontier lab, no capital-dependent upstart. So where did the money go? Into the one entity that sits on every side of the lab. Google is Anthropic's customer and holds roughly 14% of the company; it is the compute landlord behind the take-or-pay book; it is the expected backstop of the listing, alongside Amazon ^[3]. Investors rotated into exactly those names — Amazon, Microsoft, and Alphabet — while selling the second-tier infrastructure firms that depend on capital raising ^[14]. And now Google is the guarantor of the debt building the lab's capacity ^[1]. Buyer, shareholder, landlord, backstop, cosigner: one counterparty on every side of the deal. Hold the date on that TeraWulf filing. November 4, the day before the vote. The market's re-ranking of the frontier labs was signed on paper before it was ever read as an election result. What changed this week is not that AI became a liability. It is that the companies building the fastest models in the world stopped being priced as the investment-grade tenants of their own customers, and started being priced as what the paperwork already said they were: a promise that needs a cosigner.