Nvidia closed FY2025 Q4 with $96.2 billion in revenue. The stock popped at the opening bell. Everyone called it a demand story. They are wrong.
The demand was never the question. The question was whether TSMC could physically build enough silicon to satisfy the order book. The answer, for now, is a qualified yes — but the constraint is not the wafer. It's the package.
The Context: A Fabless Giant With a Fabric Problem
Let me lay out the structure, because the numbers hide the real architecture. Nvidia does not own a single fab. It designs, and TSMC does the heavy lifting. The current Blackwell architecture runs on a 4nm process. Hopper is on 4nm too, but it is the end of its life cycle. The next Rubin architecture will jump to 3nm, scheduled for 2026. Fine. But this is not where the story ends.
Nvidia's supply chain is a concentrated bet on one company. TSMC handles the lithography. It also handles the critical CoWoS packaging. On top of that, HBM memory comes from SK Hynix and Samsung. That is three dependencies, and all three are at capacity. I audited this chain as a trader, not as a fan. And the conclusion is uncomfortable.
CoWoS is the true bottleneck. Nvidia consumes about 60% of TSMC's CoWoS capacity. In a sideways market, this is the kind of structural detail that matters more than price. A factory on the other side of the world with an earthquake or a power issue could trigger a six to twelve month disruption. That is not a headline. That is a supply chain audit. And in my experience, the market rarely prices in such physical fragility.
The Core: Locks, Prepayments, and the Illusion of Control
Let me now get into the actual order flow, because that is where the edge is. Nvidia's book capex to revenue ratio is only 5-8%. That is low. But the real capital commitment is hidden. Nvidia uses prepayments and long-term agreements to lock in TSMC's CoWoS capacity. This is an off-balance-sheet monster. The capacity is not owned, it is rented through a promise. That makes the entire model a function of TSMC's execution.
TSMC's plan is to double CoWoS capacity by 2025. They are targeting a monthly output of 80,000 to 100,000 wafers by the end of next year, up from 40,000 to 50,000 now. This is not idle talk. It is a concrete expansion driven by customer orders. But here is the key signal: the doubling of capacity is a confirmation that AI chip demand is expected to last until at least 2026. The foundry does not expand based on hype. It expands based on booked orders.
Floor sweeps are just data points in motion, and the data here is loud. But there is a secondary effect that most retail traders miss. If Nvidia is buying the supply, the price of the GPU is not set by demand alone. It is set by the cost of the packaging. As N3 production ramps, costs will go up. And Nvidia has the pricing power to pass that through. This is the core of its 70%+ gross margin.
The Contrarian Angle: CUDA Is the Moat, Not the Hardware
Everyone points to the hardware. The chip is big. It is fast. The specs are impressive. But the technical gap between Nvidia and AMD is shrinking. The MI300 and the MI400 are closing in. The real gap is not silicon. The real gap is the CUDA ecosystem. It is a 15-year-old software moat that is built from the ground up.
Smart contracts execute truth, not intent. And CUDA is the same. It is a set of instructions that developers have been writing for over a decade. Migrating away from it is not a simple swap. It is a rewrite of everything. This is why AMD's hardware advantage in some benchmarks is not the relevant metric. The relevant metric is the developer's time to migrate. That is the hidden cost.
But here is the blind spot. The cloud giants are not fighting on the software front. Google has TPUs, Amazon has Trainium, and Microsoft has its own Maia. They are not trying to beat CUDA in the open market. They are building their own islands. They are reducing their dependence on Nvidia, not to become a better GPU vendor, but to reduce a strategic dependency. This is a slow, long-term threat. I see a 50-60% probability that these self-developed chips will take a 10-15% share of the inference market by 2028. That is a significant structural shift.
And there is the China issue. The export controls have driven a "de-China-ization" strategy. Revenue from China has dropped from 25% to about 10-15% in two years. This is a rational decision. It is a way to reduce geopolitical risk. But it is also a way of inviting the enemy to build its own alternative. The Chinese government has a third fund of around $47 billion dedicated to domestic chips. The gap is two to three years. The policy support is real. The question is not whether China will catch up, but when.
Takeaway: The True Metric Is Not the Revenue, It's the Supply Chain
I audited the void and found a backdoor. The void is the AI narrative. The backdoor is the supply chain. The next big signal is not the stock price. It is the CoWoS capacity expansion. If TSMC delivers the 100,000 wafers, the supply constraint will ease and Nvidia will be able to sell everything it can produce. If the capacity falls short, the revenue will be capped.
The market is pricing in a demand scenario. The real variable is the packaging speed. Watch the TSMC monthly revenue reports. Watch the equipment delivery times. That is the true heartbeat of the AI trade. The price will follow the capacity, not the other way around. ** ,