The AI Industry Is Financing Its Own Buildout

The board of TSMC, the world's largest contract chipmaker, approved $500 million of common stock in Anthropic at the IPO price and another $500 million of preferred stock in OpenAI. The company that fabricates the chips has begun buying stock in the labs whose models run on them. The same meeting cleared a $40.93 billion capital budget and $265 billion for Arizona facilities ^[1]. The maker of the hardware is now a strategic shareholder in the labs whose demand justifies building it. This was not a new move. In June, memory maker Micron took a stake in Anthropic's $65 billion Series H round, which valued the lab at $965 billion, as part of a memory-supply agreement, and Claude models are now deployed inside Micron's own factories ^[2]. At Anthropic's coming listing, the anchor money is the vendor base: Nvidia and Alphabet are taking $10 billion each inside a $26 billion private placement, with AMD, Broadcom and Microsoft taking smaller slices ^[1]. The suppliers are underwriting the customer's IPO. Amazon sits at the top of the chain. It raised the largest corporate bond sale in history, $53.8 billion with roughly $126 billion of peak demand, partly to fund up to $50 billion into OpenAI ^[3]. The group's largest borrower is funding a rival lab. Beneath the equity sits a wall of guarantees. Nvidia has capped residual-value guarantees at $108.5 billion; Meta carries $41.5 billion of guarantees on its Hyperion campus ^[4]. Separate headlines, one web. The AI buildout is being financed by the industry selling itself its own capital: the same names that build, buy and rent the equipment are also the shareholders and guarantors funding it. The web exists because no single lab balance sheet can carry the buildout. Anthropic lost $42 billion in 2025 and has committed to $518 billion of compute spending over the next decade, much of it non-cancelable, meaning it cannot walk away even if revenue disappoints ^[5]. Some of that is take-or-pay: if Anthropic's own spending falls short, it must pay Alphabet the difference anyway. A residual-value guarantee runs the other way up the chain: Nvidia promises to buy equipment back at an agreed price, so the risk that the chips lose value sits on the guarantor, not on the data-center operator. Either way, the vendor becomes the backstop. A week ago the Democratic sweep was supposed to test whether any of this could survive politics. The week of the election delivered a miniature version of that test: a leveraged-tech unwind erased $3.2 trillion of market value, with the selling blamed on "concerns over circular financing in the AI sector," and within days dip-buying and Nvidia's recovery brought the whole structure back ^[6]. The sweep transmits through balance-sheet thickness. CoreWeave, carrying roughly $50 billion of debt, watched its stock and bonds swing on a single pledge from Texas's incoming governor to halt new data-center construction, framed as protecting the grid and water from a corporate free-for-all ^[7]. Alphabet, by contrast, launched its second euro-denominated bond sale of the year on election day itself ^[8]. Same shock, two balance sheets, opposite receptions. And CoreWeave's answer is not retreat. Days after the election it raised its 2026 forecast to 168% year-over-year revenue growth, held its $37–39 billion capital-spending target and 1.85 gigawatts of active power, and its chief executive dismissed the regulatory shift outright ^[9].

We do not anticipate any impact from the seismic regulatory shift that's occurred. — Mike Intrator

It is out raising at least $4 billion in an investment-grade loan led by Morgan Stanley and Mitsubishi UFJ ^[10]. The mid-tier's answer to political risk is more borrowing, not less. The banks have now put numbers on how much of the buildout the new politics touches. Goldman Sachs estimates state moratoria could affect about 45% of planned US capacity; Barclays finds 80% of a 300-gigawatt pipeline sits in states led by Democratic governors ^[11]. But the timeline is a cliff, not a cliff-edge. Susquehanna expects 2027 to remain the peak year for installations because most permits are already secured, while up to 60 of the 148 gigawatts planned through 2030 could face delay and 12 gigawatts of 2028–2030 projects could be canceled in a worst case ^[12]. The flows for next year are locked. The risk concentrates further out. When strain shows, the instinct runs up the ladder to the one balance sheet outside the web: the government. Palantir's chief executive and OpenAI's chief financial officer have both floated federal backstops, even nationalization, for strategic AI infrastructure. Treasury Secretary Scott Bessent has answered that AI developers must remain responsible for their own financial obligations ^[4]. So far the taxpayer's ledger is the one nobody in the web has been able to reach. Which is why the markets have spent the cycle pricing the same names in opposite directions. Bank of America's strategists call it an "AI credit-equity disagreement": the bond spreads and default-insurance costs of hyperscaler borrowers sit at their widest in at least five years, while equity multiples and volatility barely move ^[13]. The equity desks, meanwhile, name Alphabet, Amazon, Microsoft and Meta as the stable trade, precisely because of their diversified revenue and cheap capital ^[14]. Principal Asset Management's Seema Shah tells clients to shift from broad tech exposure to company-by-company balance-sheet work and calls diversification itself the safety ^[15]. The buildout's risk has migrated onto the balance sheets the stock market calls safe. And so far, the only desk charging for it is the one lending them the money.