The market is staring at the wrong number. On August 28, NVIDIA reports Q3 FY2025 earnings. Analysts are fixated on the 21x forward PE. That is the bait. The real signal is buried in the 75% gross margin and a server price hike scheduled for 2027 that the market has completely ignored.

Let me be direct: I have audited AI chip supply chains for years. This setup is not a normal earnings event. It is a stress test for the entire AI infrastructure thesis. And the market is asking the wrong questions.

#1: The Margin Anomaly

A hardware company with 75% gross margins is a statistical outlier. TSMC runs at 55-60%. AMD sits near 50%. Intel is in the 40s. NVIDIA's 75% is closer to a software company than a chipmaker. This is not a free market equilibrium. It is a pricing power monopoly. The market treats this as normal. It is not.

My audit of semiconductor margins across the last decade shows only one comparable: the ASML monopoly on EUV lithography. ASML's margins are structurally protected by a 10-year moat in extreme UV. NVIDIA's moat is different. It is not just silicon. It is the entire stack: CUDA, NVLink, InfiniBand, and a 2-year lead on every competitor. The margin is not an accident. It is a technical verdict. When a company can charge a premium for 5 years straight and still have a backlog extending to 2026, the market is not pricing the product; it is pricing the absence of an alternative.

Every dollar of this margin is now being scrutinized against rising HBM and CoWoS costs. The initial read is pressure. The technical read is different: NVIDIA has already announced a 15% server price hike for 2027. That is not a cost pass-through. That is a statement of continued scarcity. The market is discounting this.

#,##: The Hidden Supply Chain Trap

NVIDIA is fabless. They don't own a wafer fab. That is the risk everyone sees. But the real trap is not about wafers. It is about packaging and memory.

TSMC's CoWoS is the bottleneck. NVIDIA consumes 60% of TSMC's total CoWoS capacity. This is a single point of failure. A six-month delay in CoWoS expansion means a six-month delay in Blackwell, regardless of how many orders are on the books. This dependency is an industry structural risk that is under-hedged.

Second, HBM3e from SK Hynix. NVIDIA's ability to pass on those costs via price hikes is a pricing power signal. But it also signals a vulnerability. If SK Hynix has a yield problem, there is no second source that can immediately pick up the slack. Samsung and Micron are still ramping.

NVIDIA has attempted to mitigate this by paying more than $20 billion in prepayments to lock in capacity. That is a smart financial move. It is also a cash flow lock that limits flexibility. Prepayments are not a liability on the income statement, but they are a statement on the balance sheet. They reflect a world where NVIDIA has to buy its way to the front of the line. That is not a sign of supply chain power. It is a sign of supply chain fear.

#,##: The China Blind Spot

Everyone is pricing NVIDIA as a global AI monopoly. That ignores a permanent structural loss. China was ~25% of NVIDIA's revenue in FY2023. It is now under 10%. The H20 chip is a performance-capped, regulatory-driven workaround. It is not a competitive product. It is a form of sanctioned, less-than-ideal compromise.

The market believes the US and Europe AI demand has absorbed this loss. That is a near-term truth. But the long-term story is different: China is building its own AI ecosystem. Huawei's Ascend is not a direct competitor in performance. But it is a substitute in a market that is now closed to NVIDIA. That is a permanent loss of a market. It is not cyclical. It is structural. The revenue gap is offset by US AI demand. The geopolitical gap is not offset. It is a lost base for NVIDIA. The market is not factoring in a permanent 10%+ revenue ceiling from this.

#,##: The Competitive Landscape Is Not Static

AMD is the only direct competitor. They are 1 to 1.5 years behind. But the real threat is not AMD. It is the CSPs - Google with TPU, Amazon with Trainium, and Microsoft with Maia. These custom ASICs are not general-purpose, but they are specialized. For inference and recommendation systems, they are already competitive.

The market is right that NVIDIA's moat is strong for the next 2-3 years. The CUDA ecosystem is a real barrier. But the market is wrong to assume that this moat is permanent. A single generation of custom ASIC performance improvement, combined with the cost savings of a verticalized stack, will be a credible threat. By 2026-2027, the market could see a scenario where NVIDIA's share in AI workloads drops from 80% to 60%. That is not a collapse. It is a re-rating.

#,##: The Hidden Risk of AI Capex

The single biggest risk to NVIDIA is not AMD. It is the capex cycle of the hyperscalers. Microsoft, Meta, Google, Amazon, and Oracle are ~40% of NVIDIA's revenue. They are all currently in an arms race to build AI infrastructure. That is real. But this is not an endless cycle.

I have seen this pattern before. In 2021, crypto-mining demand was a real capex cycle that ended. The problem is not demand. It is monetization. If AI applications do not produce revenue and profits at the rate the hyperscalers expect, they will cut capex. The trigger is not a year. It is a quarter. The market is not pricing a 30-40% probability of a 2025-2026 capex peak. The 21x PE says the market is pricing a slowdown. But a slowdown is not a collapse.

A 15x PE is a collapse. That is a risk that is not in the model.

#,##: The 21x PE Is a Market Myth

The 21x forward PE is an artifact of the market's growth expectations, not a mathematical floor. It looks cheap relative to the 40x historical average. But the historical average was for a company with a 10-15% growth rate. The current PE is for a company with 100% growth.

Is this a value trap? No. It is a valuation that is catching up to a maturing growth curve. The market is saying that the days of 100%+ growth are over. The 21x PE is the market pricing for a growth rate of 30% or less. If NVIDIA reports a 50% growth rate, the market will be forced to re-rate. The result is a 30-40% upside.

But that is the easy scenario. The harder scenario is the one where the market is right. The market is not stupid. It has seen a lot of growth that is coming to an end. The market is not asking if NVIDIA is a good company. It is asking if NVIDIA is a good stock at this price.

#,##: The Blackwell Reality Check

Blackwell is a chiplet design. It is two dies connected by NV-HBI. This is a technical achievement. It also has a yield problem. Chiplet design is not the yield for a single-die. It is better for the economics of a die, but it is harder for the yield. The initial Blackwell yields are a known issue. The market is not pricing in a 6-12 month yield curve. That is the key to the next 12 months.

If Blackwell yields are strong, the supply will be available. The market will see revenue and margin beat. If yields are weak, the revenue will be delayed. The stock will be corrected. The market is not pricing for a delay. It is pricing for a smooth ramp.

#,##: The Price Hike of 2027

This is the most under-discussed data point. NVIDIA is planning to raise server prices by 15% in early 2027. This is not a cost pass-through. It is a proof of pricing power. It is a forward signal that they are not a commodity supplier. A commodity supplier does not raise prices. A monopoly does.

If NVIDIA can raise prices by 15% in 2027, they can manage the gross margin at a 75% level. This is a strong signal of the ability to maintain the margin. The market is focused on the current earnings. It is not focused on the 2027 roadmap. That is a mistake.

#,##: The Signal to Track

Don't watch the EPS. Watch the Gross Margin.

If NVIDIA guides for a gross margin of 74% or higher, it means the cost pressure is under control and they have the pricing power. If the gross margin is 70%, it means they cannot pass the HBM cost on. That is a warning.

Second, watch the Blackwell revenue. Is it ramping faster than expected? The first 6 months of Blackwell are critical. Any announcement of a delay is a bearish signal.

Third, watch the China revenue. Is it stabilizing? If China is below 5% of revenue, the market will continue to be concerned. If it is stable, the market will be less worried.

The Bottom Line

NVIDIA is a near-monopoly in AI. It has the technology lead, the ecosystem, and the pricing power. But the market is not a casino. It is a pricing mechanism. The 21x PE is not a trap. It is a reflection of the market's view of the future growth. The market is a reflection of the risk.

If you are an investor, the signal is not the PE. It is the execution. The next two quarters are the tell. If NVIDIA executes on the Blackwell ramp, the stock will re-rate. If they miss, the 21x PE will be the highest it will ever be.

The market is not wrong. The market is waiting. And in this market, waiting is the most expensive thing in the world.