The Shadow Borrowing Behind Artificial Intelligence (1/N)

When financial engineering makes billions in GPU debt completely disappear.

Disclaimer. I am new to finance, and this is my first post on the topic. I am writing this series to understand the finance side of LLMs — how capital, debt, and financial engineering shape the AI industry — not as an expert. Treat what follows as notes from a learner, not financial advice. Corrections and pointers are welcome.

The Paradox

To someone researching large language models, the AI boom can look primarily like a technological race: larger models, longer training runs, more capable agents and ever-larger GPU clusters. Moonshot AI’s Kimi K3, for example, is a 2.8-trillion-parameter mixture-of-experts model, although only a fraction of those parameters are activated for each token. Models at this scale depend on enormous quantities of compute, advanced chips and the data centres required to house and power them. (ADD CITE)

But every model also has a balance sheet behind it.

The physical infrastructure supporting AI must be purchased, constructed and financed. Companies must pay not only for GPUs, but also for land, data-centen construction, electricity, cooling, water and connections to the power grid.

Initially, the largest technology companies could fund much of this expansion from their own cash flows. As the scale of investment grows, however, AI infrastructure is increasingly being financed through corporate bonds, private credit, long-term leases, special-purpose vehicles and project-level debt. The borrowing may sit outside the formal balance sheet of a hyperscaler (ADD SIDE NOTE), but repayment often still depends on that company, or one of its AI partners, continuing to pay for the capacity.

This creates a deeper tension within the AI race. And a race many companies are not willing to drop out of. For example, Anthropic has been clear that it is not willing to slow down its development of increasingly capable models, even as it warns about the risks of doing so. (ADD CITE)

Whether the present excitement eventually proves justified or excessive, the infrastructure race is accelerating. So is the financing required to sustain it.

The important question is therefore no longer only who can build the most capable model, but also:

Question. Who is financing the AI boom, where does the debt ultimately sit, and who bears the loss if future AI revenues fail to justify today’s investment?