Michael Burry vs NVIDIA: The Bear Case Hidden in the 10-K

In 2005, Michael Burry started buying credit default swaps against subprime mortgage bonds. Wall Street laughed. His investors sued him. He was right — it just took two years for the world to notice.

Now Burry is making another big call. This time the target is NVIDIA — the company that just posted $68 billion in quarterly revenue. The company whose chips power every major AI lab on Earth. The company that, by every traditional metric, is performing better than any semiconductor business in history.

So what does Burry see that the rest of the market doesn’t?

The Three Layers of the Bear Case

Most people think Burry’s NVIDIA thesis is “AI is a bubble.” It’s not. Or at least, it’s not just that. Over the past two months, Burry has built his case in three distinct layers, each more specific — and more grounded in data — than the last.

Let’s walk through each one.

Layer 1: The Shovel Seller

February 2026

This is the argument everyone knows. During the California Gold Rush, the people who got rich weren’t the miners — they were the ones selling shovels. Levi Strauss sold jeans. Samuel Brannan sold pickaxes. They made fortunes regardless of whether anyone actually found gold.

Burry’s claim: NVIDIA is the modern shovel seller. The company is profiting enormously from AI infrastructure spending, but the actual end users of AI — the companies deploying agents, writing code, running inference — may never generate enough value to justify what they’re paying for chips.

This is a framing argument, not a data argument. It’s powerful because it reframes NVIDIA’s strength (everyone buys their chips) as a weakness (everyone buys their chips whether or not the economics work). But it’s also easy to dismiss — NVIDIA has real revenue, real margins, and real customers with real workloads. Shovel sellers during an actual gold rush still made real money. The question is whether the rush continues.

Layer 2: The Circular Flow

February 2026

This is where Burry gets more specific — and more uncomfortable for bulls.

NVIDIA invested $30 billion in OpenAI. OpenAI then spends over $20 billion buying NVIDIA chips.

Read that again. NVIDIA is funding its own customer, who then uses that funding to buy NVIDIA’s products.

Burry’s point: some portion of NVIDIA’s revenue growth is self-financed through a circular capital flow. The revenue is real, but it’s partially propped up by NVIDIA’s own investment dollars flowing back to NVIDIA as chip purchases.

The bull counter is immediate: Microsoft did the same thing. Microsoft invested $13 billion in OpenAI, and OpenAI runs on Azure. Every major cloud provider has similar dynamics — Amazon, Google, and Microsoft all invest in AI companies that then spend on their cloud platforms. This isn’t a red flag. It’s how ecosystems work.

Layer 3: The 10-K Deep Dive

March 2026

This is the layer that matters most — and it’s the one getting the least attention.

Burry shifted from narrative to forensic accounting. He pulled NVIDIA’s Form 10-K (the annual SEC filing that companies can’t spin) and found three specific metrics that concerned him.

Finding 1: The Cash Conversion Cycle Is Extending — Permanently

Every company has a cash conversion cycle (CCC) — the number of days between paying for raw materials and collecting cash from customers. A shorter CCC means the business turns inventory into cash quickly. A longer CCC means capital is tied up longer.

NVIDIA’s CCC and Days Inventory Outstanding (DIO) are both extending. Not temporarily. Permanently.

The 10-K shows NVIDIA has $117 billion in total supply-related commitments — prepayments and purchase obligations with TSMC, memory suppliers, and packaging partners. In Q4 alone, supply commitments to TSMC rose by $44.9 billion in a single quarter to $95.2 billion.

Burry’s interpretation: NVIDIA is deliberately locking up supply chain capacity to a greater extent than ever before. If demand stays strong, this looks like genius — they’ve secured the entire wafer production pipeline. If demand softens even slightly, NVIDIA is stuck with obligations it can’t unwind.

Think of it like a restaurant that signed 10-year leases on 50 locations based on pandemic-era takeout volume. If dine-in trends change, you’re stuck paying rent on empty kitchens.

Finding 2: Hidden Stock Compensation Costs

NVIDIA’s reported earnings use GAAP accounting, which includes a line item for Stock-Based Compensation (SBC). This represents the cost of paying employees with stock (RSUs, options) instead of cash.

Burry’s finding: the actual economic cost of NVIDIA’s stock compensation is significantly higher than what GAAP reports.

Why? RSU grants are priced at the time of grant, but NVIDIA’s stock has risen roughly 10x since 2023. When those RSUs vest and employees sell, the dilution to existing shareholders is far greater than what the original GAAP expense captured. The difference isn’t because shares appreciated after vesting — it’s because the grants were made with the expectation of appreciation.

The bull counter: every major tech company has this exact same dynamic. Apple, Google, Microsoft, Meta — all of them have SBC that understates true economic dilution when stock prices rise. It’s a feature of high-growth tech accounting, not a NVIDIA-specific red flag. And SBC doesn’t consume operating cash flow — it’s a non-cash expense.

Finding 3: Market Structure Fragility

Stepping back from NVIDIA specifically, Burry warns of “increasing fragility and coiled tension” in US market structure broadly. The Magnificent 7 are all red year-to-date — Microsoft down 16.5%, Tesla down 10.8%, NVIDIA down 4.5%. Reuters published analysis warning that if AI labs like OpenAI or Anthropic fail, the $650 billion capex ecosystem built around them could rapidly collapse.

This is Burry’s systemic argument: it’s not just about NVIDIA’s 10-K. It’s about an entire $650 billion investment wave that depends on a handful of AI companies converting research breakthroughs into revenue.

The Stress Test: What the Bulls See

Here’s the thing about Burry’s thesis. Every single data point he cites is real. The $117B in commitments, the extending CCC, the SBC gap — all verifiable in SEC filings.

But the bull case isn’t based on ignoring these numbers. It’s based on context:

NVIDIA just guided $78 billion for the April quarter — beating consensus by 7%. Enterprise revenue doubled. SemiAnalysis benchmarks show the GB300 NVL72 delivers the lowest inference cost of any platform. Jensen Huang says enterprise AI is “by far larger” than hyperscaler as a long-term market.

The CCC is extending because demand is so strong that NVIDIA is prepaying to secure every available TSMC wafer. The supply commitments are growing because customers are begging for capacity. The SBC is high because NVIDIA’s engineers are the most valuable humans in the semiconductor industry and every competitor is trying to poach them.

The 2005 Question

This is the question that matters more than any individual data point.

In 2005, Burry was right about the housing market. Subprime mortgages were structured to fail. The math was clear. But the trade took two years to pay off, and along the way, he nearly lost his fund, his investors, and his reputation.

Is this the “2005 of AI”?

The structural parallels exist: massive capital flows into a single thesis, circular financing between participants, and metrics that look brilliant until they don’t. The supply commitments, the investment loops, the accounting nuances — these are the kinds of details that only matter when the music stops.

But there’s a critical difference. Subprime mortgages had no real underlying value. The loans were made to people who couldn’t pay them back. AI infrastructure spending, by contrast, is producing real products with real users. Claude writes code. GPT-5.4 does knowledge work. Gemini processes multimodal data at scale. These aren’t empty houses in Nevada.

The honest assessment: Burry’s 10-K analysis is methodologically sound. The data is real. The risks he identifies — overcommitment, circular flows, accounting gaps — are genuine. But NVIDIA’s fundamentals are also genuinely extraordinary. The question isn’t who’s right. It’s when.

If demand stays strong for 2-3 more years, the bears get crushed and the $117B in commitments converts to $300B+ in revenue. If demand plateaus or the AI labs can’t convert research into sustainable businesses, those commitments become anchors.

Burry’s track record says he’s usually right about the what. His challenge has always been the when.

Claim references: [de37f25e], [7f7c5098], [c4eb94ff], [fbeec218], [0c2adc91], [2ad6f7c4], [05990f42], [3b9fb4e5], [4bf934ca], [fc99e88b], [4770af7a], [ad1bf5d1], [ad45b563], [a4633f05], [1b29c734]