Watch the video with chapters & transcript →
AI IN FINANCE
In short: Microsoft's AI annualized revenue run rate hit $37 billion in April 2025, up 123% year on year, from roughly $16.5 billion a year earlier.
In a new explainer for her AI for Finance channel, Ani Björkström, a Stockholm-based tech consultant working in financial services, breaks down the chart that has been circulating online: arrows looping between Nvidia, Microsoft and OpenAI, read by most viewers as proof that the AI boom is propped up on fake, self-dealing money. The video does not defend or debunk that reading outright. Instead, it walks through what the underlying deals actually are, names the accounting mechanism involved, and lays out the real numbers behind Microsoft and OpenAI's relationship — concluding that the arrangement is legal, decades-old, and genuinely risky at the same time.
Björkström's first point is definitional: "circular financing" is not an accounting term. It is a label journalists reached for to describe the shape of a chart. The Financial Times' Lex column, she notes, calls these arrangements "back-to-back" deals rather than circular ones, because real goods move in both directions — chips ship, servers run workloads, and Nvidia books real revenue while still holding the equity stake it paid for.
There is a genuinely illegal version of this pattern, and it has a name: round-tripping. That is when two companies push money back and forth purely on paper to inflate reported revenue, with no real goods changing hands — the practice Enron used, which resulted in criminal convictions. What Nvidia, Microsoft, Amazon, Google and the AI labs are doing has a different name: vendor financing, an arrangement in which a seller helps fund its own customer's purchase, through a loan, installments, or an equity stake, and in return keeps a claim on that customer's future success.
The video uses two escalating analogies. First, a car loan: the seller effectively hands over both the product and the financing to buy it, and the arrangement is not fraudulent because real value moves both ways — though real risk exists if the buyer stops paying. Second, a pharmaceutical company taking an equity stake in a small research lab: the lab gets funding it could never have borrowed conventionally, and the pharma company secures a future manufacturing customer plus a share of the upside if a drug succeeds. Their fates become tied together.
Swap in the real names and the AI story follows the same shape: hyperscalers (Microsoft, Google, Amazon) play the role of big pharma, AI labs (OpenAI, Anthropic) play the small lab, and Nvidia supplies the compute, cloud services and chips. Björkström's summary line: the hyperscalers are investing in the exact companies that generate most of their AI revenue.
Using Microsoft and OpenAI as the clearest example, the video cites Microsoft's disclosed AI annualized revenue run rate of $37 billion in April 2025, up 123% year on year from around $16.5 billion (roughly $4 billion a quarter) the year before. Citing documents reported by journalist Ed Zitron, it notes OpenAI spent about $2.9 billion in a quarter running its models on Microsoft's servers — implying an estimated 70% of Microsoft's AI revenue traces back to OpenAI alone, a company Microsoft has invested more than $13 billion in. Because OpenAI is private, Björkström flags these as reconstructions from leaks and reporting, not audited figures, and stresses they should be read as order-of-magnitude estimates. The same pattern repeats elsewhere: Amazon and Google have both invested billions in Anthropic, which has agreed to use Amazon's cloud for training and Google's chips to run on.
If AI pays off, Microsoft is paid twice on every ChatGPT query — a share of OpenAI's profit as an equity holder, plus a hosting fee. If confidence collapses, every vendor in the web loses twice at once: the investment is written off, and the customer paying the hosting bills disappears. The video draws one key distinction from the dot-com era: that bubble was inflated by debt, which spread losses into banks, pensions and insurers, whereas this round is largely funded from big tech's own cash. The caveat to that caveat is that the data-center buildout surrounding these deals is debt-funded, which the video flags as a separate, real risk. It also notes two reasons this isn't purely a big-tech problem: these firms now account for a large share of US GDP growth, and until OpenAI goes public, nobody outside the company can verify the books — a state Björkström calls concentration of risk.
| Pairing | Investment flow | Return flow |
|---|---|---|
| Microsoft → OpenAI | $13bn+ invested (equity) | OpenAI spends ~$2.9bn/quarter on Microsoft servers |
| Amazon & Google → Anthropic | Billions invested | Anthropic trains on Amazon's cloud, runs on Google's chips |
| Nvidia → AI companies | Equity stakes across the AI economy | Those companies buy Nvidia chips |
| Enron (illegal comparison) | Fake paper transactions, round-tripping | No real goods moved; ended in criminal conviction |
No. Round-tripping is illegal — money moves back and forth on paper with no real goods exchanged, as Enron did. The Nvidia-Microsoft-OpenAI arrangements involve real chips, real servers and real booked revenue, which the video identifies as vendor financing instead.
Any arrangement where a seller helps pay for its own customer's purchase — via a loan, installments, or an equity stake — so the buyer gets the product without upfront cash, and the seller books the sale while keeping a claim on the buyer's future success. It is decades old and legal.
Dot-com era vendor financing was largely debt-funded, which spread losses into banks, pensions and insurers. Today's AI deals are mostly funded from big tech's own cash — though the video notes the surrounding data-center buildout is debt-funded, calling that a separate risk.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?