The consensus number sits at $92.18 billion. The company's own guidance is $91 billion. The gap is 1.3%. For thirteen consecutive quarters, this company has beaten expectations. The market now treats that streak as a law of physics, not a managerial achievement. I do not predict the future; I audit the present. The present shows a supply chain operating at full capacity, a pricing power that borders on the absolute, and a competitive moat measured in software, not silicon. But the ledger also shows a new variable. The narrative around AI infrastructure is shifting from training to inference. This is not a minor rebalancing. It is a change in the nature of the demand curve.
For a Fabless company, the balance sheet is the only physical asset. NVIDIA does not own fabs. It owns contracts, prepayments, and design wins. The physical reality lives on TSMC's CoWoS lines and at SK Hynix's HBM facilities. The question for FY2027 Q2 is not whether demand exists. The question is whether the physical layer can keep up with the financial layer. The narrative fades; the wallet addresses remain. Here, the wallet addresses are the prepayment lines on NVIDIA's balance sheet and the capacity allocation sheets at TSMC.
The data set for this audit is public. It includes TSMC's monthly revenue reports, SK Hynix's earnings calls, and the capital expenditure guidance from four of the largest cloud providers in the world. Microsoft, Meta, Google, and Amazon have collectively guided for over three hundred billion dollars in combined capex for 2025 and 2026. A significant portion of that allocation is earmarked for AI compute. NVIDIA is the primary beneficiary of this spending. The pattern is visible to anyone who reads the blocks.
The CoWoS Bottleneck and the Blackwell Ultra Ramp
NVIDIA's fiscal Q2 2027 is the quarter where the market expects to see Blackwell Ultra, the B300, in full volume. The architecture is a derivative of the Blackwell B200, using the same TSMC 4NP node. The key constraint is not the node, it is the packaging. CoWoS-L is the critical path. TSMC has been expanding CoWoS capacity aggressively, targeting a doubling of monthly capacity to 80,000 wafers by the end of 2025. NVIDIA consumes more than 60% of that output. This is not a point of competition; it is a bottleneck of physics.
I have audited supply chain data for years, and this concentration of capacity is an anomaly. NVIDIA has effectively created a proprietary claim on the most advanced packaging capacity on Earth. This is not a technical moat; it is a supply chain moat. The financial result is a gross margin above 70% on a non-GAAP basis. The gross margin is not a measure of efficiency. It is a measure of control over the supply chain.
The Wall Street consensus expects gross margins to remain above 55% for the quarter. I see a different pressure point. The HBM4 transition is scheduled for the Rubin platform. HBM3e is the current standard, and it is in a supply deficit. HBM pricing is increasing. This is a headwind. The product mix is also shifting towards the GB200 NVL72 systems, which are higher value but also higher cost. The question is not if gross margins will be better; it is whether the product mix can absorb the memory cost increase.
The data shows that NVIDIA's EPS growth estimate is higher than revenue growth. The consensus expects adjusted EPS growth of +99% against revenue growth of +97%. This is a small anomaly. It implies that the market expects operating leverage. This is only possible if the gross margin is stable or improving, or if operating expenses are growing slower than revenue. Based on my experience auditing tech balance sheets, this is a reasonable assumption. NVIDIA's operating expense growth has been under control relative to its revenue explosion. The pattern is a one-off.
The China Variable and the Data Point Everyone is Ignoring
The export control is the outlier. NVIDIA's China revenue has dropped from 25% of total to less than 10% of total. This is a structural change, not a cyclical one. The market has accepted this as a reality, but the FY2027 Q2 report will have a specific line item: the “China sales update”. The H20 chip, a cut-down version of the Hopper architecture, was a stopgap. It is now restricted. The next stopgap, likely the H200, is uncertain.
Here is the contrarian angle. The market is fixated on the CoWoS bottleneck and the HBM supply. The real risk is not a supply shortage. The real risk is a demand vacuum. The CSP capex cycle is the demand engine. The engine is running at full throttle. But a engine has a fuel tank. The tank is the AI application layer. If the AI applications do not generate enough revenue to justify the capex, the CSPs will slow down. The timeline for this is not 2026. The timeline is 2027. The market is pricing NVIDIA for a 2027 with a continued +90% growth. The data does not support a growth rate that high beyond 2026.
A $300 billion capex cycle cannot grow at +30% year-over-year indefinitely. The base effect is too large. The current consensus is that the AI infrastructure buildout is a 3-to-5 year super-cycle. I am more conservative. The last cycle, 2020-2022, was also a super-cycle. It ended with a significant inventory correction. The current cycle is deeper, but the physics of capital allocation remain the same.
The pattern is visible in the data if you look at the flow of funds. In the 2020 cycle, the influx was from retail. In the current cycle, the inflow is from institutions and hyperscalers. This is a difference. But the end point is the same. The ledger shows the money moving. The narrative fades; the wallet addresses remain.
The GPU Architecture is a Moat, But the Ecosystem is the Prison
NVIDIA's dominance is not primarily a hardware story. It is a software story. The CUDA ecosystem is the moat. It has been under construction for over a decade. It is not an architecture; it is a network. Developers write in CUDA. They do not write in ROCm. This is not a technical decision; it is a habit. The habit is difficult to break. I have seen this pattern before. The ecosystem of the dominant player has a longer half-life than the hardware.
The competitor AMD MI350 and MI400 are closing the hardware gap. The gap is now 0.5 to 1 generation. The software gap is still 2-3 years. The proprietary chips from Google TPU and AWS Trainium are relevant, but they are only relevant in a specific context. They are not general-purpose. The GPUs are a general-purpose machine. This is a structural advantage that is not easy to replicate.
The more interesting data point is the NVIDIA's RISC-V usage in its GPU microcontrollers. It is a small detail. It is not a strategic shift. But it is a sign of a flexible supply chain. NVIDIA uses RISC-V for control tasks, not for the core. The core remains proprietary. This is a data point for the architecture.
The Valuation Premium and the Hidden Inventory
NVIDIA's trading at a forward P/E of over 50x. This is a premium to its own history, which was around 40x. It is also a premium to the semi-average, which is around 20x. The premium is the market pricing in a sustained +30% growth for the next three years. The consensus expects +97% growth for FY2027. The next year, FY2028, is expected to be +40%. This is a high bar. The market expects perfection.
Here is the hidden information. The guidance for FY2027 Q3 will be the signal to watch. If the guidance is above $100 billion, it means the demand is still strong. If the guidance is below that, it means the cycle is topping. The consensus is $98-100 billion. I will be watching the exact number.
The other hidden signal is the change in prepaid balances. NVIDIA has been prepaying for TSMC capacity and HBM. If the prepayment is increasing, it means NVIDIA is confident in future demand. If it is flat, it means they are cautious. This is not a data point that is reported in the headlines, but it is in the balance sheet. I read it.
The Takeaway
The market is waiting for a quarter that shows a flawless execution. The data shows that the execution is a management discipline, not a miracle. The streak of 13 consecutive beats is a record of disciplined guidance. The risk is not the quarter. The risk is the year after. The signal is not the B300 ramp. The signal is the Q3 guidance. The pattern is set. The question is whether the market can read the ledger. I do not predict the future. I audit the present. The present is a strong, but the future is a probability. The wallet addresses remain. The narrative fades.