The Accounting Trick at the Heart of the AI Boom

Michael Burry doesn't short things because he thinks they're bad businesses. He shorts them when he thinks the market has fundamentally misunderstood what it's paying for.

In 2008, it wasn't the housing market itself he was betting against — it was the financial instruments built on top of it, priced on assumptions that didn't hold. When those assumptions broke, the gap between price and reality was catastrophic.

His current critique of OpenAI and Palantir follows the same logic. He's not saying AI doesn't work. He's saying the way these companies count their revenue — and how the market values it as a result — is built on accounting conventions that tell a much better story than the underlying economics support.

Here's exactly how it works.

The Gross Margin Illusion: Where You Put Salaries Changes Everything

This is the core of the gimmick, and it comes down to a single accounting decision: where do you classify your engineers' salaries on the income statement?

In a genuine software business, engineers build a product once and it gets sold a million times. Their salaries sit in Research & Development (R&D) — an operating expense that sits below the gross margin line. This is why pure SaaS companies report gross margins of 70–80%. The cost of delivering the product to each new customer is essentially just server capacity.

In a services business, engineers spend their time building custom solutions for specific clients. Their labour is directly required to deliver the revenue. That means their salaries should be classified as Cost of Goods Sold (COGS) — which sits above the gross margin line and reduces it directly.

Forward Deployed Engineers fall squarely into the second category. They're not building a general product. They're embedded at a specific client, doing bespoke work to make the software function in that client's environment. That's a service.

The gimmick: by classifying FDEs as R&D or "product development" rather than COGS, a company can report software-like gross margins of 80%+ while quietly running what is operationally a consulting business.

The numbers make this stark. If OpenAI hires 1,000 FDEs at $300,000 each — a conservative salary for this calibre of engineer — that's $300 million in annual costs. Buried in operating expenses below the gross margin line, it's invisible to the headline margin figure. Correctly classified as COGS, it would drop gross margins from "software company" territory into "professional services" territory — somewhere between 30% and 50%.

That difference isn't cosmetic. It's the difference between a company that deserves a 20x revenue multiple and one that deserves a 2–5x multiple. At OpenAI's scale, that's a valuation swing worth hundreds of billions of dollars.

Revenue Recognition: "Recurring" Isn't Always What It Seems

The second layer is how revenue gets recognised and described — specifically the distinction between genuine recurring revenue and contract revenue that only recurs because humans keep showing up.

Burry's critique of Palantir — which he laid out before Palantir became a market darling — centred on what he called revenue smoothing. FDE-heavy businesses tend to have "lumpy" revenue: a client pays for a large implementation, FDEs do the heavy lifting, and the contract looks like Annual Recurring Revenue (ARR) on the books.

But here's the honest version of that story. A significant portion of what looks like sticky software revenue is actually contract revenue dependent on ongoing human intervention. Remove the FDEs and many clients would churn — not because the software is bad, but because the software alone isn't enough to deliver the outcomes the client is paying for.

This matters enormously for how you value the business. ARR from self-serve software compounds beautifully — each new customer adds to a growing base with minimal ongoing cost. ARR that secretly requires a team of engineers at every account doesn't compound the same way. Every new dollar of revenue requires a roughly proportional increase in expensive human headcount.

The revenue line goes up. The unit economics quietly collapse.

The Anthropic Argument: If the Model Worked, You Wouldn't Need the Army

There's a third dimension to Burry's critique that's particularly pointed for OpenAI specifically.

The premise of large language models as a transformative technology is that they're reasoning engines — powerful enough to understand context, follow instructions, and integrate into business workflows without requiring a team of specialists to operate them. That's the pitch. That's what justifies the valuation.

If OpenAI genuinely needs to deploy armies of FDEs to win and retain enterprise contracts, it implies something uncomfortable: the API is not a self-serve utility. It's a complex raw material that requires significant human expertise to turn into something useful.

Burry's framing is direct. Just as he argued Palantir was "consulting in a software vest," the concern is that OpenAI is becoming an expensive AI consultancy — trading the high-margin scalability that makes software businesses extraordinary for high-touch market share that carries the economics of a services firm.

Competitors building models that require less handholding to deploy compound this problem. If the moat is the FDE relationship rather than the model itself, that's a moat that erodes as the underlying technology matures.

Valuation Arbitrage: The Number That Holds It All Together

All of these accounting choices — where salaries sit, how revenue is described, what gets called "recurring" — feed into a single number that determines what the company is worth: the Price-to-Sales multiple.

The market's current view is roughly: OpenAI is the next Microsoft, price accordingly. Software infrastructure companies with dominant positions trade at 20–40x revenue. Apply that to OpenAI's growth trajectory and you get a valuation north of $1 trillion.

Burry's view: OpenAI is becoming the next Accenture, price accordingly. Professional services firms trade at 2–5x revenue. Apply that to the same revenue figures and the valuation looks radically different.

The mechanism for correction is Free Cash Flow. Revenue can grow impressively while FDE headcount scales alongside it. But Free Cash Flow per share — the measure that eventually anchors what a business is actually worth — tells the honest story. When revenue growth requires proportional growth in expensive human capital, FCF doesn't compound the way a software business's does.

At some point, the market notices. When it does, the reclassification from "software multiple" to "services multiple" isn't a small adjustment. It's the kind of correction that ends a bull narrative.

---

The Question Worth Sitting With

Is OpenAI's FDE push a temporary bridge — a pragmatic way to help enterprises onboard to AI while the technology matures — or is it a permanent admission that the models aren't yet capable enough for businesses to use without hand-holding?

The bull case is that FDEs get you in the door, create deep integrations that are expensive to rip out, and eventually the software runs itself. Palantir has been promising this outcome for fifteen years.

The bear case — Burry's case — is that "eventually" is doing an enormous amount of work in that sentence, and the market is pricing it as if it's already happened.

The accounting choices being made right now aren't neutral. They determine whether these companies look like the next generation of software giants or expensive consulting firms with a compelling story. Burry is betting the story eventually meets the numbers.

History suggests he's worth taking seriously.