The AI moat is now a balance sheet, so price the financing and not the model

Fortune has put a name to something the capex disclosures have been implying for two years: as large language models become interchangeable, competitive advantage shifts from the model to the balance sheet behind it, and to whoever can finance, build and run the infrastructure most cheaply ^[20]. That matters because it inverts the economics investors spent two decades rewarding, in which software was asset-light and margins were fat ^[21].

The scale is not subtle. Amazon, Microsoft, Alphabet and Meta have together committed $1.1 trillion to AI infrastructure since the boom began in 2023, and plan another $745 billion this year alone ^[1][2]. That single year is equal to roughly 68% of everything the four spent in the preceding stretch, and takes the combined figure to about $1.85 trillion ^[4][3].

The substitution thesis is coming from the buyers of those chips, not the sceptics. Fortune quotes Microsoft's Satya Nadella saying "every model is substitutable" and Amazon's Andy Jassy predicting "at least half a dozen" comparably good models ^[5][6]. The behaviour follows the statement: rather than betting the firm on one winning model, the hyperscalers are building capacity that can serve many of them ^[22].

Which pushes the contest into finance. Nvidia is working with Apollo, Blackstone, Goldman Sachs and other Wall Street firms to mobilise more than $500 billion of additional capital for AI infrastructure, and Google has assembled a $200 billion structure with Broadcom, Apollo, Blackstone and Morgan Stanley to fund Anthropic's chips and data centres ^[8][9]. That is more than $700 billion of build-out being arranged through structures rather than paid for out of operating cash ^[10]. The counterparties being financed are still lossmaking: both OpenAI and Anthropic lose money today ^[7].

If cost of capital is the moat, the incumbents start ahead. Microsoft, Amazon and Google have the balance sheets, the cheapest capital, and revenue from the same data centres used for training ^[11]. The credible challengers Fortune names are not software firms but capital pools: SpaceX, plus sovereign funds such as Saudi Arabia's PIF and Abu Dhabi's MGX, which combine cheap money, abundant power and the flexibility to work with both western and Chinese AI companies ^[12].

The bill is landing on enterprise software, which now competes with AI infrastructure for the same corporate budgets ^[13]. IBM's second quarter showed the mechanism, with customers postponing software purchases to secure AI capacity ahead of expected price rises, and the shares falling 25% in a single day in mid-July ^[14]. The buyers are paying too: Alphabet's capex has pushed free cash flow negative for the first time since its IPO, and Meta's fell sharply in the latest quarter ^[15][16].

Against that, the return is showing up where the assets are. Microsoft's cloud business grew 32% to $39.3 billion in the latest quarter, helping lift total revenue 18%, meaning cloud is growing at close to twice the company rate and added roughly $9.5 billion of quarterly revenue year on year ^[17][19][24]. AWS grew 37% ^[18].

What to watch is the gap between the two clocks. Cloud revenue is compounding now; the depreciation and the financing costs on $745 billion of this year's commitments arrive later, and the open question is whether the models generate enough value to justify them ^[2][20]. Fortune's read is that overcapacity may eventually appear but is still a long way off ^[23]. Treat capital structure, not model benchmarks, as the disclosure worth reading closely.