NVIDIA’s fiscal Q2 2025 report, covering the period ending July 28, 2024, was a masterclass in operational dominance. Revenue hit $96.2 billion, up 106% year-over-year, with adjusted gross margins at 74.5%. The market’s reflex is to call this a blowout. But as a macro strategist, I don’t read earnings for the headline. I read them for the stress points—the friction where the model bends. And beneath this perfect surface, there are hairline fractures that tell the real story about the AI infrastructure trade, its liquidity cycle, and the systemic bottlenecks that will define the next two years. This isn't a victory lap. It's a supply chain audit, a look at the financial engineering that keeps the entire machine running, and a forecast of where the stress will appear next.
The context here extends beyond Silicon Valley. NVIDIA is not a chip designer anymore. It is the de facto gatekeeper of the global AI industrialization project. This quarter's data reveals a map of that project: hyperscalers—Microsoft, Google, Amazon, Meta—accounted for roughly 54% of revenue, a $48.7 billion concentration. Their combined AI capital expenditures are projected to exceed $200 billion for 2024. This is the global liquidity map for AI, and NVIDIA is the central bank. The 106% growth isn't just a company number; it is a measure of how fast the real economy is reallocating its capital base towards artificial intelligence, based on my experience stress-testing liquidity pools, this level of concentrated spending is a systematic risk factor, but for now, it's the only engine in town.
Now, let's get to the core of the data—the technical and financial engineering. The margin guidance is the most interesting signal in this report. Management guided Q3 gross margins to 73.5%–74.5%, a dip from Q2's 75%. On the surface, this is a rounding error. But this is a tell. It reveals the ramp-up cost of the new Blackwell architecture. The transition from the Hopper architecture (H100/H200) on TSMC's 4N node to Blackwell (B200) on the more complex 4NP node is the critical path. Based on my audit experience, initial yields for a new chip of this size are always problematic. Industry chatter points to B200 yields sitting in the 60–70% range initially, and the CoWoS-L packaging is far more complex than the previous CoWoS-S. That packaging is the real bottleneck. The B200 uses a larger, two-reticle design with higher interconnect density, which is putting immense strain on TSMC's advanced packaging capacity. The demand for CoWoS capacity is the single most important supply-side variable for the entire AI trade. NVIDIA has reportedly locked in a significant portion of TSMC's capacity, creating a massive barrier to entry. But this strategy has a hidden cost: NVIDIA's prepayments to TSMC and SK Hynix to secure capacity are the reason its free cash flow of $21.34 billion is lower than its net income. They are converting cash into supply chain control.
The contrarian angle here is that NVIDIA is the most predictable asset in the market, and that is precisely what makes it risky. The AI trade is being driven by a first-principles assumption that demand is infinite and that this is a new kind of economic cycle. That is a false premise. This is not a new model; it is an old model with a new vector. The revenue concentration is a vulnerability. When a single customer group controls over half your revenue, the business model is structurally fragile. We saw this in 2021 with NFTs—a digital scarcity illusion. Here, the illusion is that compute demand is solely about utility. It isn't. A huge portion of the spending is speculative inventory building. Every hyperscaler is competing to build out capacity that they hope to monetize, and they are all buying from the same one supplier. If any one of those major players gets hit with a downturn or changes their internal CapEx plan, the entire system’s model breaks. The market is treating a highly leveraged, cyclical capital expenditure cycle as if it were a utility. In my years of stress-testing DeFi liquidity pools, I’ve learned that the only thing that breaks a system is when the margin call comes in the same direction for all actors at the same time.
The takeaway is not that NVIDIA is a bad company. It is a great company with a moat that extends from silicon to software. The CUDA ecosystem is a permanent barrier to entry. But for the macro investor, the question is not about the company; it is about the cycle. We are in the middle of the most intense supply-constrained capital expenditure boom in history. The next leg of the bull market will not be driven by more orders; it will be driven by the execution of those orders. The true test will be whether Blackwell can be produced in volume without further margin dilution. That will signal that the bottleneck is passing. The second signal is when the balance of power shifts from training to inference. Training is a centralized, capital-intensive, concentrated effort. Inference is distributed, lower-cost, and much more diverse. When the inference demand curve overtakes the training curve, this will be the true test of the economic sustainability of the AI trade. Code is law, but man is the loophole. The question is: what happens when the loop is a supply chain?