In a characteristically incisive intervention in the ongoing AI debate, Dr. Michael Burry has identified what may be the most overlooked vulnerability in the current market enthusiasm: the accounting sleight of hand that makes AI companies appear far more profitable and productive than they actually are. Writing in a collaborative Substack discussion with AI researchers and podcasters, Burry draws attention to stock-based compensation (SBC) as a massive hidden cost that distorts our understanding of both productivity gains and long-term returns.

This isn’t just accounting pedantry. Burry’s analysis strikes at the heart of the AI investment thesis and challenges the narrative that we’re witnessing unprecedented productivity gains. If he’s right, the emperor has clothes—but they’re being paid for by shareholders in ways that don’t show up in the metrics everyone is celebrating.

The SBC Shell Game

Stock-based compensation has become the tech industry’s favorite accounting tool for a simple reason: it allows companies to show higher profits while transferring enormous value from shareholders to employees. The mechanics are straightforward, but the implications are profound.

When a company pays an employee $1 million in cash, that’s an expense that reduces reported profit by $1 million. When a company pays an employee $1 million in stock, accounting rules allow much of this to be excluded from certain profit calculations, or at minimum, treated differently than cash compensation. The employee receives the same value, the shareholder experiences the same dilution, but the reported financials look healthier.

Consider what this means for Nvidia, the poster child of the AI revolution. If Burry’s calculation is correct—and given his track record, dismissing it would be foolish—then half of the profit that drives Nvidia’s valuation is an accounting mirage. The value is being created, but it’s being captured by employees through stock grants, not retained for shareholders through earnings.

The Productivity Paradox

Burry’s insight about productivity is particularly sharp. The narrative around AI companies emphasizes unprecedented productivity—look at how much value each employee creates! But this metric becomes meaningless when you account for how much each employee is actually being paid in total compensation.

Traditional productivity measures revenue or profit per employee. An employee generating $5 million in revenue looks highly productive. But if that employee is receiving $2.5 million in total compensation (salary plus the market value of stock grants), the productivity gain is far less impressive than it appears.

At companies like Nvidia, where senior engineers and researchers are receiving packages that can reach into eight figures when including stock appreciation, the productivity calculation becomes absurd. You can’t claim your workforce is extraordinarily productive when you’re paying them as if they’re extraordinarily productive. The productivity gain has simply been transferred to the employee as compensation.

The Million-Dollar Question

Burry poses an uncomfortable question: if your employees are now worth $25 million, what is the actual productivity gain? The implication is clear. If you’ve created technology that makes workers 10x more productive, but you’re paying them 10x more through stock grants, where’s the shareholder value? The productivity is real, but it’s being captured as employee compensation rather than shareholder returns.

This isn’t an argument against paying talented people well. It’s an argument that investors need to look at fully-loaded costs, not accounting-optimized profit figures, when evaluating whether these companies are as profitable as they appear.

The Software-to-Hardware Shift: ROIC’s Silent Killer

Here Burry identifies the second major vulnerability: the transformation of high-ROIC software businesses into lower-ROIC hardware businesses. This shift is not subtle, and its implications for long-term returns are severe.

Why Software Had Extraordinary ROIC

Software companies achieved remarkable returns on invested capital because the economics were fundamentally different from traditional businesses. Once you’ve written the code, the marginal cost of serving additional customers approaches zero. You don’t need to build new factories, purchase raw materials, or invest in physical infrastructure for each new customer.

Companies like Microsoft, Google, and Meta achieved ROIC figures that would be impossible in capital-intensive industries. They could generate enormous profits without tying up vast amounts of capital in physical assets. This is why software companies commanded premium valuations—they were capital-light profit machines.

The AI Hardware Reality

The AI revolution has fundamentally changed this equation. The companies leading the AI race are now spending unprecedented amounts on physical infrastructure:

These aren’t software companies anymore, at least not in the pure sense. They’re becoming infrastructure companies, utilities companies, with all the capital intensity that entails. Data centers, chips, power generation, cooling systems—these require massive upfront investment and ongoing capital expenditure.

The ROIC Death Spiral

Burry’s prediction is stark: ROIC is heading down fast and will continue falling through 2035. This isn’t speculation—it’s mathematical inevitability given the capital requirements of AI infrastructure. The question isn’t whether ROIC will fall, but how much and how quickly.

When ROIC falls, valuations must eventually follow. High price-to-earnings multiples are justified when companies can reinvest profits at high rates of return. When ROIC declines, those multiples become indefensible. A company earning $10 billion might deserve a $500 billion valuation if it can reinvest at 40% returns. The same $10 billion earning 12% returns might warrant only $150 billion.

The Historical Pattern

We’ve seen this movie before. Industries that start as high-return businesses inevitably attract competition and capital expenditure, driving returns toward the cost of capital. It happened to railroads, utilities, airlines, and telecommunications. The pattern is always the same:

1. Phase 1: New technology enables extraordinary returns
2. Phase 2: Competition and scale require massive capital investment
3. Phase 3: Returns normalize as capital intensity increases
4. Phase 4: Valuations adjust to reflect lower returns

AI companies are transitioning from Phase 1 to Phase 2. Burry is arguing they’ll reach Phase 3 faster than the market expects.

Nadella’s Hope vs. Burry’s Math

Microsoft’s CEO Satya Nadella represents the optimistic case: perhaps AI software will be so valuable that companies can maintain high ROIC despite massive hardware investments. Burry’s response is withering—he can’t see how this works mathematically, and suspects Nadella can’t either.

The Optimistic Scenario

Nadella’s hope rests on AI software being sufficiently valuable and defensible to generate returns that justify the infrastructure investment. If AI becomes truly transformative, perhaps companies can charge premium prices that offset the capital costs. The moat would come from the trained models, the data, the integrated ecosystems—software advantages that persist despite hardware commoditization.

Burry’s Skepticism

Burry sees fundamental problems with this narrative. First, hardware costs don’t disappear just because software is valuable. You still need to invest billions in infrastructure, and that capital still has a cost. Second, competition in AI is fierce—there’s no clear winner emerging who can charge monopoly prices. Third, the infrastructure itself may become commoditized, turning AI compute into a low-margin utility business.

Most tellingly, Burry notes this sounds like “only a hope” even to Nadella. Hope is not a strategy, and it’s certainly not a basis for current valuations that price in decades of sustained high returns.

Market Implications: When the Music Stops

If Burry’s analysis is correct, we’re witnessing a profound disconnect between current valuations and future returns. Markets are pricing AI companies as if they’ll maintain software-level ROIC indefinitely while scaling to unprecedented size. The math doesn’t work.

The Valuation Problem

Current market capitalizations assume these companies will continue generating extraordinary returns on capital. Microsoft trades at a valuation that implies it can reinvest massive sums at very high returns. But if ROIC is declining rapidly, these valuations become difficult to justify.

This doesn’t mean these companies won’t be profitable or important. It means they’ll be profitable in the way utilities are profitable—decent returns on enormous capital bases—rather than in the way pure software companies were profitable. That’s still a valuable business, but it’s not a $3 trillion business.

The Timeline: Why 2035?

Burry specifically mentions 2035 as the timeframe for declining ROIC. This isn’t arbitrary. It reflects the long buildout cycle for AI infrastructure and the time required for these trends to fully manifest. Infrastructure investments made today will impact returns for a decade or more.

This timeline also suggests that near-term stock performance might not reflect these fundamentals. Markets can ignore deteriorating ROIC for years, especially in a momentum-driven environment. But eventually, returns tell the truth. Companies with declining ROIC will see their valuations compress, regardless of how exciting the technology seems.

The SBC and ROIC Connection

Burry’s two critiques—SBC distortion and ROIC decline—are intimately connected. When ROIC falls but companies want to maintain the appearance of software-like profitability, SBC becomes an even more attractive tool. You can show impressive profit margins by paying people in stock rather than cash, even as the underlying returns on invested capital deteriorate.

This creates a vicious cycle. Lower ROIC requires more capital investment. More capital investment means more dilution or higher cash compensation needs. Companies respond with more SBC. SBC makes reported profits look better than they are. Better-looking profits support higher valuations. Higher valuations mean more valuable stock grants. More valuable stock grants mean more dilution.

Eventually, this cycle breaks. Shareholders notice the dilution. Or ROIC falls far enough that no amount of accounting creativity can hide the decline. Or both happen simultaneously, producing exactly the kind of market dislocation Burry has built his career on predicting.

What the Bulls Are Missing

The counter-argument to Burry’s analysis is straightforward: AI is genuinely transformative, productivity gains are real, and the companies capturing this value deserve premium valuations. This isn’t wrong, exactly—it’s incomplete.

Burry isn’t arguing AI isn’t transformative. He’s arguing that transformation doesn’t equal shareholder returns if those returns are being captured by employees through SBC or consumed by capital requirements that drive down ROIC. You can have revolutionary technology and mediocre investment returns at the same time.

The Productivity vs. Profitability Disconnect

The fundamental insight is that productivity gains don’t automatically translate into profits when:

1. Competition forces companies to pass gains to customers through lower prices
2. Employees capture gains through higher compensation (especially SBC)
3. Capital requirements consume profits through infrastructure investment
4. All three happen simultaneously in a competitive AI race

This is precisely the scenario Burry describes. Nvidia might be enabling extraordinary productivity, but if half that value transfers to employees and the rest requires massive ongoing investment, where’s the sustainable shareholder profit?

The Bear Case Gets Concrete

Burry has provided skeptics with two specific metrics to track: SBC as a percentage of reported profit, and ROIC trends over time. These aren’t abstract concerns about “bubbles” or “valuations”—they’re concrete financial metrics that either support or undermine current prices.

If these metrics deteriorate as Burry predicts, the investment thesis unravels regardless of how impressive the technology becomes. You can’t maintain a 30x P/E multiple on declining ROIC and margins that are half what they appear.

Why This Analysis Matters Now

Burry’s commentary comes at a crucial moment. AI enthusiasm is at its peak, capital expenditures are accelerating, and valuations reflect extremely optimistic assumptions about future returns. This is exactly when skeptical analysis is most valuable—and most ignored.

The pattern from his previous calls is evident. In the dot-com bubble, the technology was real but the profits were imaginary. In the housing bubble, the demand was real but the lending was unsustainable. Now, the AI transformation is real but the shareholder returns may be illusory—consumed by SBC and capital requirements.

The Contrarian Position

Betting against AI companies in early 2026 feels like betting against the internet in 1999. The technology is obviously important, the market momentum is overwhelming, and the consensus is that skeptics don’t understand the magnitude of the transformation. Burry isn’t denying the transformation—he’s questioning who captures the value.

This distinction is critical. You can be right about technology and wrong about investment returns. The companies building the internet were indeed transforming the world in 1999. Most of them still went bankrupt. The survivors took years to grow into their valuations. Shareholders who bought at the peak waited decades to break even.

The Long Game

Burry’s projection extending to 2035 is notable for its patience. This isn’t a call for an imminent crash. It’s a structural argument about deteriorating fundamentals that will eventually force valuation adjustments. This timeline is both the strength and weakness of his analysis.

The strength: It’s difficult to disprove because ROIC trends play out over years, not quarters. If he’s right, there’s plenty of time for the thesis to unfold.

The weakness: Markets can ignore deteriorating fundamentals for extraordinary periods, especially in transformative technologies. Even if ROIC is declining, stocks can keep rising on momentum, narrative, and liquidity.

This is the familiar Burry predicament: being analytically correct and market-timing wrong. The question isn’t whether he’s identified a real problem—the SBC and ROIC issues are demonstrably real. The question is whether and when markets will care.