The AI Industry Now Sells the Factory, Not the Model

On May 23, DeepSeek made its 75 percent price cut permanent, taking its V4-Pro model to between 12 and 19 times cheaper than the American labs' flagship models for comparable work ^[1]. The framing mattered as much as the number. One analyst explained what the cut actually was.

This is why the price cut is permanent rather than promotional. — Sanchit Vir Gogia

A cut offered that way is not a dip to ride out; it is a new floor for what intelligence costs. The same five months produced a different number. CoreWeave, the fastest-growing name in the data-center buildout, now pays 15 to 16 percent on its unsecured debt — borrowed money with no assets pledged behind it — while the swap market, where traders bet on default, puts the odds of one within five years at roughly 60 percent ^[2]. The two prices have moved in opposite directions: the thing the industry sells keeps getting cheaper, and the thing it runs on keeps getting dearer. Up and down the stack, every tier has found its own way to say the same thing. Jensen Huang has stopped pitching Nvidia as a chipmaker and now casts it as a factory that converts electricity into tokens ^[3]. He keeps sharpening the line.

Now, compute is revenue. — Jensen Huang

Sam Altman, whose company defined the business of selling models, has started pitching OpenAI as something else entirely ^[4].

We are taking a forward bet that it will continue to grow, and that not only will ChatGPT keep growing, but we will be able to become one of the important AI clouds, that our consumer device business will be a significant and important thing, that AI that can automate science will create huge value. — Sam Altman

Amazon's AWS lifted its 2026 spending guidance to $200 billion and tied the money to demand for AI agents ^[5]. Taiwan's MIC, the tech-industry forecaster, stated the industry's direction in one sentence ^[6].

AI applications are no longer confined to the cloud and are beginning to enter edge computing, end devices, and manufacturing sites. — Chris Hung

The pattern runs the length of the chain: nobody is selling intelligence anymore. Everybody is selling the plant. The conversion has a price, and it shows up in the debt. CoreWeave's valuation has slid from a peak around $90 billion to roughly $34 billion, and the interest on its roughly $50 billion of debt now consumes about a quarter of revenue ^[2]. Huang has said in public that CoreWeave will not get a backstop from Nvidia — the sector's key supplier, managing the optics of a credit event out loud ^[2]. OpenAI engineered a different door to capital: preferred equity, a class of investment paid out ahead of common shareholders, carrying a guaranteed minimum return of 17.5 percent for firms including TPG and Bain Capital, a structure built to keep customization costs off its main balance sheet ahead of a possible IPO. Thoma Bravo declined ^[7]. This week, Nvidia committed nearly $10 billion of its own money to European customers, citing a scarcity of capital ^[8]. And the neoclouds now sell themselves on the strength of their books. Nebius's chief revenue officer, asked about rivals, put the contrast bluntly ^[9].

We're seeing continued upward mobility in terms of the kind of price per megawatt. — Marc Boroditsky

When growth companies start competing on the quality of their balance sheets, the audience has changed. It is the creditor now. The buildout is not a passive borrower. Columbia Threadneedle, the asset manager, finds the buildout is driving a historic leveraging cycle in mega-cap tech bonds and data-center debt, and that the borrowing itself is pushing long-term yields higher across developed markets ^[10]. The expansion is bidding up the price of the very capital it must keep raising. August, an investment firm, puts the financing need at $5 trillion to $10 trillion and warns that borrowing costs rise as investors refuse to over-concentrate in one correlated trade, with roughly $150 billion of triple-C debt — the riskiest tier of corporate bonds — pressed against a maturity wall in 2027 and 2028 ^[11]. The buildout is making its own money more expensive. None of this settles whether the demand is real, and on that question the numbers do not agree. Crusoe's chief executive reports that older chips now rent for more than they did when first deployed, with 2027–28 capacity selling out ^[9]. Anthropic throttled Claude's peak-hour limits in March because compute could not keep up ^[12]. OpenAI has built a real advertising business, running at a $1 billion annual pace ^[13]. Goldman's credit strategist reads the widening spreads in AI debt as dispersion rather than widespread disruption, with $4.5 trillion in dry powder still available ^[14]. Every one of those counters contests the demand picture. None of them touches the pitch-and-pricing shift itself — the pitch every tier is now making, and the price being charged for it. And the sector's answer to the credit anxiety has not been to shrink the bet. It has been to enlarge it and wrap it in the macroeconomy. The White House's AI czar, David Sacks, has put the stakes in macroeconomic terms ^[4].

A reversal would risk recession. — David Sacks

JPMorgan Asset Management calculated that AI spending drove roughly two-thirds of U.S. GDP growth in the first half of 2025 ^[4]. That is the closing argument now: not that the numbers work, but that nobody can afford to see them stop. The calendar has already printed the next chapter. The $150 billion of riskiest AI-linked debt comes due in 2027 and 2028 ^[11], the same window in which the swap market's five-year verdict on CoreWeave matures ^[2]. The two prices have to reconcile somewhere, and the date is already on file.