NVIDIA's Earnings: The Ledger Behind the AI Supercycle

The data shows NVIDIA's H100 GPU commands a gross margin above 70%. The block height does not lie. But the market's lowered expectations for the upcoming earnings report suggest a fracture in the narrative. Over the past 90 days, analysts have trimmed growth forecasts while the company's supply chain remains at full capacity. This is not a story of failure. It is a story of bottlenecks, market concentration, and the quiet shift from hardware dominance to software moats. NVIDIA is the designated first customer for TSMC's most advanced nodes. The B200, built on the 4NP process, is in mass production. The next-generation Rubin platform will move to the N3 node, and by 2027, the N2 process with GAA transistors. This is a one-to-two-year lead over AMD and Intel. But the real constraint is not lithography. It is the packaging. CoWoS, TSMC's 2.5D advanced packaging, is the true bottleneck. NVIDIA consumes over 60% of TSMC's CoWoS capacity, and the utilization rate is close to 100%. The company's shipments are limited by packaging capacity, not by GPU die production. This dependency creates a specific kind of risk. The ledger shows a single source for both logic and packaging. TSMC holds the keys to NVIDIA's revenue. The supply chain is fragile, not because of quality, but because of concentration. A disruption in CoWoS capacity or a shift in TSMC's allocation would directly impact NVIDIA's ability to ship. This is a risk that is not reflected in the gross margin. It is an operational risk, hidden in the capital expenditure plans of a supplier. Institutional compliance alignment dictates that we also examine the financial statements. NVIDIA's gross margin stands at 75%, a figure that rivals software companies. The ROE is over 100%, and the free cash flow conversion rate exceeds 90%. This is the result of a fabless model, a light-asset strategy that avoids the heavy depreciation of a foundry. The capital expenditure is less than 5% of revenue, and all R&D costs are expensed, a conservative policy that underscores the true intensity of the innovation. Stress tests reveal the fractures before the flood. A simulation of a sudden drop in CSP capital expenditure would show revenue growth falling from 100% to 30%, and the valuation would suffer a double blow. The current price-to-earnings ratio of 50-60x is high, but the PEG ratio, at 1.5 to 2.0, is within a reasonable range if the growth persists. The market is underestimating the shift to AI inference. The training phase has been the core driver, but the inference is the next engine. As large language models move from development to deployment, the demand for inference compute will grow exponentially. This is not a prediction. It is an analysis of the installed base and the trend of model deployment. The CSPs have already spent over $200 billion on AI infrastructure, and the plans are extending through 2026. The inference market will be a separate battleground, and the competition will not be AMD or Intel. It will be the custom silicon of the CSPs: Google's TPU, Amazon's Trainium, and Microsoft's Maia. The market is underestimating the threat from these custom chips, not in training, but in inference, where the cost-performance advantage is already visible. The ledger remembers what the market forgets. The data shows that NVIDIA's Chinese revenue has dropped from 20% to below 10% due to export controls. The geopolitical risk is not a binary event. It is a continuous variable. The entity list does not include NVIDIA, but the export restrictions on A100 and H100 have forced the company to design lower-end chips. The long-term effect is not only the loss of sales, but the acceleration of China's domestic AI chip industry. The local chips will not be a threat to NVIDIA's global market in the near term, but they will erode the addressable market in the long run. Formal verification is the only truth in code. The same principle applies to the market. The market's lowered expectations are a form of speculation, but the code of the supply chain, the financials, and the technology roadmap are verifiable. The hidden information is that the moat is shifting. The hardware advantage is narrowing, but the CUDA ecosystem is a durable barrier. The software platform has over 4 million developers, and the switching cost for these developers is extremely high. This is a moat that will take more than five years for a competitor to cross. The system-level solutions, such as DGX and NVLink, are not just hardware. They are an integrated architecture that offers a full-stack value. This is the reason why NVIDIA's valuation, despite being high, has a solid foundation. Chaos is just unverified data. The market's concern is the sustainability of AI demand. The CSPs are investing heavily, but the return on that investment is not yet proven. If the AI return on investment is less than expected, the capital expenditure will be cut. This is a real risk, with a probability of 25-35% over the next 12 to 18 months. However, this risk is not a reason to ignore the structural trend. The AI chip is not a cyclical product. It is a new infrastructure. The demand for computing power is not a bubble; it is a correction of a previous underinvestment. The market expects a decrease, but the technical data shows an increase. This is the premise of a potential expectation gap. Verification precedes value. The coming earnings report is not just a number. It is a signal. The market will be watching the data center revenue growth, the B200 shipment schedule, and the management's guidance on the CoWoS capacity. The block height does not lie, and the earnings will not either. The supply chain is the key, and the CoWoS capacity is the lead indicator. The integration of the data, from the supply chain to the financials, will define the direction. The takeaway is not a forecast. It is a method. The market is a ledger, and we must verify the entries before we value the company. The 2026 outlook is a test of the AI narrative. The numbers will show the truth, and the truth will be the price.