The Accounting Trick Inflating the AI Bubble

In Q3 2025, Alphabet reported over $10 billion in gains on equity securities. Amazon reported $9.5 billion pre-tax gain in the same period. Microsoft disclosed nearly $6 billion in net gains over nine months.

Stock prices went up. Index funds bought more shares. Your 401(k) got a little more concentrated in big tech.

The thing is, none of that money came from selling products or winning new customers. It came from paper markups on AI startup investments. Specifically, stakes in Anthropic and OpenAI.

OpenAI isn't even public yet. But it's already moving the earnings reports of three of the largest companies on earth.

On a recent episode of The Accounting Podcast, David Leary and I got into the mechanics of what's actually happening here.

It turns out, there's an accounting angle.

The Round-Trip Accounting Problem

To understand why these valuations look the way they do, you need to understand “round-tripping.”

Say you have two fictional companies: a large tech incumbent and a well-funded AI startup.

The tech company wants in on AI, but it doesn't invest cash in the startup. Instead, it contributes $1 billion worth of cloud compute credits (essentially a promise to provide server time).

In exchange, it receives an equity stake that implicitly values the startup at a massive premium.

As the startup burns through those compute credits, the tech company gets to record that consumption as revenue. That’s $1 billion in cloud revenue, even though no cash changed hands.

It's a barter transaction dressed up as a sale.

Then a new funding round comes in at an even higher valuation. Under ASC 321, public companies holding stakes in private companies must mark up the value of those investments whenever there's an observable price change, like a new funding round.

That markup flows directly through to the income statement as a gain.

No new customers. No products sold. Just paper wealth recorded as profit.

Smart analysts can see through this. The passive investing machine can't.

The Index Fund Accelerant

More and more of the money flowing into the market is passively managed.

Index funds don't ask whether a company's profits came from real customers or from paper markups on AI startup equity. They see the profit number. A higher profit number pushes up the stock price. That increases the company's weighting in the index. The fund automatically buys more shares to stay aligned. Repeat.

Big tech already accounts for roughly 30% of the total U.S. stock market. And by some recent estimates, nearly all of the market's gains over the last twelve months have been AI-related.

Index funds don’t ask where the profit comes from.

Your 401(k) is along for this ride whether you opted in or not.

The Defensibility Question

OpenAI and Anthropic haven’t built anything that's truly hard to replicate.

Their core product is the large language model. Yes, these are expensive to build initially. But costs have been falling fast. DeepSeek proved that a well-resourced team can produce a competitive model for a fraction of what OpenAI spent. Open-source alternatives are proliferating.

The LLM itself is becoming a commodity. It’s more like a utility than a defensible business.

Compare that to the last tech boom's real winners. Facebook locked everyone onto one platform because your friends and family were there.

Google became the default search engine, and advertisers followed the eyeballs.

Microsoft Office is so embedded in enterprise workflows that switching costs companies years of institutional friction.

What does OpenAI have that prevents a competitor from undercutting them on price? I'm not sure the answer is anything durable.

The bull case is that the agent and workflow layer creates that stickiness. Once you build your business operations on top of Claude or ChatGPT, switching is painful.

That's the play Anthropic is making, and it's not a bad strategy.

But it's not the same as owning the underlying model. And these valuations are priced as if the LLM is the lock-in. It isn't.

What Happens If the Music Stops

If OpenAI doesn't IPO, or if the IPO disappoints, those paper gains start reversing. The index funds that bought in at inflated valuations start selling. The feedback loop runs in reverse.

That's how bubbles pop. Not always in a dramatic crash, but in a slow correction as the gap between paper profits and real cash flows becomes impossible to ignore.

I'm not saying AI isn't real or that the technology won't reshape the economy. It already is. But there's a difference between transformative technology and a correctly priced investment. The internet was both transformative and wildly overpriced in 2000. It took fifteen years for the actual business value to catch up to where the market assumed it would be in 2001.

We may be in a similar moment. And if you want to understand why, the answer is sitting in the footnotes of Microsoft's, Amazon's, and Google's quarterly filings.

This is what accountants are actually trained to see. Maybe we should be saying it louder.

Want to hear the full breakdown of how GAAP might be inflating the AI bubble? David and I map out the whole loop on Episode 488 of The Accounting Podcast.

You can listen and earn free CPE at earmarkcpe.com.