Nvidia and Marvell Earnings: The CoWoS Bottleneck Is the Real Signal

Signal detected. Two earnings reports, one infrastructure bottleneck. Nvidia reports Wednesday. Marvell follows Thursday. The market will parse revenue beats and guidance like tea leaves. I'm reading the supply chain disclosures instead. That's where the truth lives. Here's the context most traders miss: both companies are fabless. They design the chips but don't fabricate them. That means their earnings are less about their own execution and more about Taiwan Semiconductor Manufacturing Company's (TSMC) ability to produce and package. Specifically, CoWoS advanced packaging. This is the chokepoint. Blackwell B200 uses a dual-die design that requires CoWoS-L packaging. TSMC's CoWoS capacity is the single largest constraint on AI chip supply. The company is expanding monthly capacity from roughly 32,000 wafers to over 60,000, but demand is still outstripping supply. Nvidia holds about half of that capacity. Marvell's custom ASICs for Amazon and Google are equally dependent on the same packaging technology. Based on my experience auditing supply chain risks during the 2020 DeFi yield farming boom, I learned that infrastructure bottlenecks create the most predictable trading signals. The same logic applies here. When a critical input is scarce, the company with the strongest supplier relationship wins. Nvidia is TSMC's most important customer. That gives it pricing power and allocation priority. Marvell is smaller, which means it has less leverage in the queue. Now, the core analysis. The market will focus on Nvidia's revenue guidance. If the company guides above $50 billion for the next quarter, that confirms AI demand is still accelerating. But I'm looking at three specific line items that tell a more nuanced story. First, prepayments. Nvidia has been paying TSMC and SK Hynix large upfront sums to lock in capacity. If those prepayments increase significantly, it signals management's confidence in future demand. It's a direct bet on the AI cycle's longevity. A decrease would suggest they see a plateau coming. Second, gross margin. Nvidia has been running at around 75%. That's extraordinary for hardware. It reflects pure pricing power in a supply-constrained market. If margins hold or expand, the supply-demand imbalance is intact. If they compress, it could mean competition is finally biting or that Nvidia is making concessions to secure more capacity. Third, the CoWoS language. Listen to the earnings call for any mention of packaging capacity. If management says supply is still tight, that's bullish for pricing but bearish for volume. If they say the situation is improving, it means TSMC's expansion is on track, which could lead to higher shipment volumes in the second half of the year. For Marvell, the story is different. The company is the number two player in custom AI ASICs, behind Broadcom. Its growth depends on Amazon's Trainium and Google's TPU programs. The key metric here is AI revenue as a percentage of total revenue. If that number is climbing toward 30% or higher, it confirms that hyperscalers are serious about reducing their dependence on Nvidia. That's a long-term threat to Nvidia's dominance, but it's also a massive opportunity for Marvell. Here's the contrarian angle that most analysts are missing. The market treats Nvidia and Marvell as AI plays. But they're really plays on TSMC's execution. The real risk isn't demand. It's manufacturing. TSMC's CoWoS expansion is the swing factor. If it goes smoothly, both companies beat expectations. If it slips, both face supply constraints regardless of how strong demand is. The chart doesn't lie, but it whispers. The whisper here is that the semiconductor supply chain is more fragile than the stock prices suggest. There's also a geopolitical layer that's underpriced. Export controls on China are already costing Nvidia about 15-20% of its revenue. The company is selling downgraded chips like the H20 to maintain a presence there. But the bigger risk is a full decoupling scenario. If the US tightens restrictions further, Nvidia loses that Chinese revenue entirely. The global demand from the US, Europe, and the Middle East can offset it, but it's a margin hit. The Middle East, particularly Saudi Arabia, is emerging as a new growth area for AI infrastructure. That's a signal worth tracking. Now, let's talk about what this means for crypto. The connection is indirect but real. AI and crypto are competing for the same resources: advanced chips, energy, and data center capacity. When Nvidia reports strong demand, it means more compute is going to AI training and inference. That could squeeze the supply of GPUs available for crypto mining and decentralized AI projects. Conversely, if Nvidia's guidance disappoints, it could free up capacity for other uses. The market hasn't priced this cross-elasticity of demand. Panic sells. Precision buys. The earnings reports this week are not just about two companies. They're about the entire AI infrastructure stack. The signals to watch are prepayments, gross margins, and any mention of CoWoS capacity. Those three data points will tell you more about the AI cycle's health than the headline revenue number. Let me give you a concrete example from my own experience. In 2021, I analyzed the Bored Ape Yacht Club market and identified that the real value was in on-chain provenance and community governance, not the JPEGs themselves. The market eventually caught up to that view. The same principle applies here. The real value in the AI supply chain is not in the GPU design. It's in the packaging technology, the memory supply, and the manufacturing relationships. Those are the moats. For Marvell, the risk is customer concentration. If Amazon or Google decides to bring more ASIC design in-house, Marvell's revenue could drop sharply. The company's top five customers likely account for over 60% of revenue. That's a structural vulnerability. The opportunity is that custom ASICs are becoming more important as AI inference workloads grow. Inference is less demanding than training, which makes custom chips more viable. That's a tailwind for Marvell. For Nvidia, the risk is different. The company's CUDA software ecosystem is a massive moat. Developers are locked in. But the threat from custom ASICs is real. Amazon's Trainium, Google's TPU, and Microsoft's Maia are all designed to handle specific workloads more efficiently than general-purpose GPUs. As these chips mature, they will eat into Nvidia's market share at the margins. The question is how fast that happens. My estimate is that it's a 2026-2027 story, not a 2025 story. The takeaway is straightforward. Watch the supply chain signals, not just the headline numbers. The CoWoS bottleneck is the key variable. If TSMC's expansion is on track, both Nvidia and Marvell will have a strong year. If it slips, expect supply constraints to persist and prices to stay elevated. The chart doesn't lie, but it whispers. Listen for the whispers in the earnings call. One more thing. The regulatory environment is shifting. The SEC's approval of spot Bitcoin ETFs in 2024 opened the door for institutional capital to flow into crypto. That same institutional interest is now driving AI infrastructure investment. The two markets are becoming more correlated. If AI demand stays strong, it's a positive signal for crypto's long-term infrastructure buildout. If AI demand falters, the ripple effects will be felt across both sectors. Signal detected. Action required. The earnings reports this week are the most important data points for the AI trade in 2025. Don't just watch the stock prices. Watch the supply chain. That's where the real signal is.