The trade appeared on a 13F filing like a cold diagnostic. Michael Burry's Scion Asset Management disclosed a put position against NVIDIA, then quietly bought call options at a mid-range strike, paying single-digit premiums per contract. The market read it as a bearish declaration. The code reads differently.
Let me be precise about what this position actually is. A short equity position hedged with long calls is a collar strategy, not a conviction short. It caps downside while preserving upside participation. Burry is not betting against the GPU. He is betting against the narrative that the GPU monopoly is permanent. Those are different positions with different risk profiles. The market conflates them. The market is often wrong.
I spent the last decade auditing DeFi protocols and building predictive models around liquidity stress. I learned one thing that applies directly to this trade: the bottleneck is never the infrastructure. The bottleneck is always the assumption that infrastructure remains scarce. NVIDIA's dominance is real. Its permanence is a hypothesis, not a theorem.
The CUDA moat is the real asset. Over 4 million developers are locked into the software ecosystem. Hardware parity is achievable. Software migration is not. AMD's MI300 series approaches H100 performance on raw benchmarks, but ROCm remains a decade behind CUDA in maturity, tooling, and mindshare. Google's TPU v5p competes on training efficiency but primarily serves internal workloads. Amazon's Trainium and Microsoft's Maia are early-stage experiments in vertical integration, not immediate threats. The code doesn't lie: CUDA's lock-in effect is the deepest moat in the industry, deeper than any single chip architecture.
But the code also shows cracks. Blackwell architecture, released in 2024, delivers 2-3x improvement over Hopper in training and inference efficiency. Rubin, expected in 2026, maintains the cadence. Yet the threat is not coming from the top. It is coming from the edge. ASIC chips designed specifically for transformer inference, like Groq and Cerebras, achieve higher energy efficiency in narrow vertical scenarios. The AI chip market is bifurcating from general-purpose GPUs to specialized accelerators. NVIDIA's general architecture may lose the efficiency war in inference-heavy workloads, even if it wins the training war.
Burry's core thesis is not valuation. It is capital expenditure cycles. He argues that NVIDIA's capital spending will "enter and pass through the top of the bubble," compressing future profitability. This is a systems-level argument, not a single-quarter earnings call. The company's gross margins sit above 70%, with net margins near 50% and return on equity approaching 100%. Those numbers are extraordinary. They are also unsustainable by definition. No company maintains 70% gross margins forever. The question is not whether margins compress. The question is when, and how fast.
The contrarian angle is this: Burry may be early. He has been early before, and it cost him. His track record with this exact strategy is mixed, by his own admission. The protective put structure suggests he knows this. The calls he bought are not a hedge against his own position. They are a hedge against the possibility that the market continues to reprice NVIDIA higher despite the thesis. This is not a bet on failure. It is a bet on reversion to the mean, with a defined risk ceiling.
Resilience isn't audited in the winter. It is audited during the build-out, when everyone assumes the demand curve is linear. NVIDIA's customers, the top five of which contribute roughly 40% of revenue, are building their own chips. Microsoft has Maia. Amazon has Trainium. Google has TPU. Meta has MTIA. These are not experiments. They are strategic imperatives. When your largest customers become your competitors, pricing power erodes. Not overnight. Not even in two years. But the trajectory is clear.
The "AI democratization" narrative is real but double-edged. NVIDIA's CUDA ecosystem lowers the barrier to entry for AI development, which expands the market. But it also enables the proliferation of competitors who will eventually build their own hardware. The moat is deep, but the water is being pumped out from both ends. The bottleneck isn't the infrastructure. It is the assumption that infrastructure demand remains elastic forever. AI capex cycles are historically violent. The 2024-2025 build-out resembles every prior infrastructure bubble, from fiber optics in 2000 to data centers in 2015. The same pattern: overbuilding, followed by consolidation, followed by margin compression.
The geopolitical layer adds another variable. Export controls on China have limited NVIDIA's addressable market. Huawei's Ascend and Cambricon are filling the gap. This is not a near-term revenue issue. It is a long-term market share issue. When you cede a geography, you cede the ecosystem that grows around it. The Chinese AI chip ecosystem will mature in isolation, then compete globally. This is the same playbook China executed with solar panels and EVs. The code doesn't lie: export controls create scarcity, and scarcity creates substitutes.
So what is the takeaway? Burry's trade is not a prediction of NVIDIA's collapse. It is a measured bet that the market has overpriced the permanence of a monopoly. The CUDA moat is real, the system-level solutions are sticky, and the software subscription model is smart. But the clock is ticking. By 2026-2027, custom silicon from hyperscalers will be production-scale. AMD's MI400 series will close the hardware gap. ASIC inference chips will erode the efficiency advantage. The question is not whether NVIDIA's dominance erodes. The question is whether the market has priced in that erosion.
The market is pricing NVIDIA as if the moat is permanent. Burry is pricing it as if the moat is temporary. The code doesn't lie. It just doesn't tell you the timeline. Resilience isn't audited in the winter. It is audited when the spring of competition finally arrives. That spring is coming. The only unknown is the date.
