The number sits there, almost offensive to hardware economics: 75% gross margin. For a company that physically ships silicon, that figure belongs in the software realm—think Microsoft, not manufacturing. Yet as NVIDIA prepares to report earnings on August 28, the market's real question isn't whether the company will beat estimates. It's whether this margin is a permanent feature of the AI landscape or a temporary artifact of supply scarcity.
The answer lies in the architecture of the entire AI supply chain. And based on my years mapping systemic dependencies across the crypto and semiconductor stacks, what I see is a company that has transformed itself from a chip vendor into an infrastructure toll booth. The 75% margin isn't just pricing power. It's the financial expression of a structural monopoly that competitors cannot breach within the next 24 months.
The CoWoS Bottleneck Nobody Can Escape
Let's start with the physical layer. Every H100, every H200, every Blackwell B200 passes through TSMC's CoWoS advanced packaging line. This is the 2.5D interposer technology that connects the compute die to HBM memory. It is the single most constrained node in the entire AI supply chain.
NVIDIA consumes approximately 60% of TSMC's total CoWoS capacity. This isn't a market share statistic; it's a preemptive occupation of the battlefield. AMD's MI300 series, which hardware-wise approaches NVIDIA's performance, cannot scale because CoWoS capacity is already allocated. Intel's Gaudi 3 faces the same wall.
The financial mechanism behind this is less understood. NVIDIA has prepaid over $20 billion to lock upstream capacity. On the balance sheet, this appears as prepayments—a liability against future supply. But operationally, it functions as a moat. Any competitor attempting to secure CoWoS capacity today faces a 12-month lead time and a supplier whose production is already committed.
This is the hidden leverage in the 75% margin story. NVIDIA doesn't just design better chips. It has financialized its supply chain into a barrier that capital alone cannot overcome.
The HBM Cost Pass-Through Illusion
The earnings call will inevitably discuss rising HBM memory costs. SK Hynix, the primary HBM3e supplier, has raised prices. Samsung and Micron are ramping but remain behind. The narrative will frame this as margin pressure.
Read the numbers differently. NVIDIA has announced server price increases exceeding 15% by early 2027. The cost increase is being passed through with markup. This is not margin compression; it's margin expansion disguised as cost management.
The pricing power here is extraordinary because the buyer has no alternative. Microsoft, Meta, Amazon, Google, and Oracle—the top five customers representing roughly 40% of revenue—are locked into an AI arms race where compute capacity is the ammunition. They can delay purchases but cannot stop them. Their own capital expenditure plans, exceeding $200 billion combined in 2024, are effectively committed to NVIDIA's roadmap.
This creates what I call a "captive demand" structure. Unlike traditional semiconductor cycles where buyers can substitute or delay, AI infrastructure spending is non-discretionary for hyperscalers. They either buy NVIDIA or fall behind in the model race. That dynamic, not superior chip architecture alone, explains the 75% gross margin.
The China Variable the Market Is Mispricing
The analysis provided by the source material remains conspicuously silent on geopolitics. This is a mistake. NVIDIA's China revenue has collapsed from roughly 25% of total revenue to under 10% following export controls. The H20 "cut-down" chip is a stopgap, not a solution.
The market treats this as a manageable loss, offset by Western AI demand growth. But this ignores a structural transformation. China is not waiting. Huawei's Ascend chips, despite process node disadvantages, are capturing domestic market share. Chinese hyperscalers—Alibaba, Tencent, Baidu—are building their AI stacks on domestic silicon.
This is the early formation of a parallel AI ecosystem. Western AI will run on NVIDIA. Chinese AI will run on Huawei and its domestic supply chain. These two systems will have different software stacks, different developer ecosystems, and different standards. The long-term consequence is not just lost China revenue for NVIDIA but the emergence of a permanent technological competitor insulated from NVIDIA's advances.
The 75% gross margin in the West may be sustainable. But the total addressable market for that margin is shrinking relative to the global AI opportunity.
Valuation: The Market Is Pricing for Decline
Now the uncomfortable part. NVIDIA trades at approximately 21 times forward earnings. Its historical average is 40 times. AMD trades at 40 times. The market has already priced in significant growth deceleration.
This is the classic value trap question: Is the discount justified or is it an opportunity? My assessment leans toward the latter, but with critical caveats.
The market's assumption is that AI capex peaks in 2025-2026. This is a reasonable baseline. But it ignores the second growth curve: inference. Training demand has dominated NVIDIA's revenue, but inference—the deployment of trained models into production—is growing at over 200% annually. By 2025, inference could surpass training as the primary revenue driver.
Inference has a different demand profile. It's more distributed, more persistent, and less cyclical than the build-out phase. Companies don't deploy AI models and then shut them down. They scale them. This creates an annuity-like revenue stream that the current valuation does not fully capture.
The PEG ratio of approximately 0.5 suggests the market expects growth to collapse to single digits. If NVIDIA sustains even 30% growth through 2026, the current valuation is deeply inadequate.
The Real Risk: Not AMD, But The Customer's Own Silicon
The competitive analysis in the source material correctly identifies AMD as the only direct GPU competitor. But the longer-term threat is the hyperscalers' custom silicon. Google's TPU, Amazon's Trainium, Microsoft's Maia—these are not experiments. They are strategic investments to break NVIDIA's grip.
The current generation of custom ASICs remains behind NVIDIA in general-purpose AI performance. But they are competitive in specific workloads: recommendation systems, inference at scale, and vertical applications. The hyperscalers don't need to beat NVIDIA across the board. They need to win in the workloads that constitute their highest-volume, lowest-margin needs.
The 2026-2027 window is when custom silicon reaches credible maturity. If NVIDIA's next-generation Rubin platform fails to deliver a generational leap, or if software ecosystem lock-in weakens, the quasi-monopoly could erode faster than expected.
CUDA remains the defensive wall. After a decade and a half of developer accumulation, the ecosystem is NVIDIA's most underappreciated asset. Switching costs are not just financial; they are cognitive. Millions of developers have trained on CUDA. Rewriting for a new architecture is a multi-year effort that most organizations will not undertake without a compelling reason.
The Bottom Line
NVIDIA is entering this earnings report from a position of technical and financial dominance that has no precedent in semiconductor history. The 75% gross margin is not a bubble metric; it's the output of a well-engineered monopoly across design, packaging, and software.
The risks are real but mispriced. The supply chain concentration—TSMC for manufacturing, SK Hynix for HBM—represents a tail risk that could materialize through geopolitical disruption. The China loss is a permanent structural erosion disguised as a manageable headwind. And the hyperscalers' custom silicon is a slow-burning fuse.
But the market's 21x forward multiple already discounts substantial pessimism. The asymmetry favors the upside if Blackwell ramps as scheduled and inference demand continues its exponential trajectory.
I've seen this pattern before in the 2020 DeFi composability crisis—markets pricing in contagion that never materialized because the underlying demand was structurally stronger than the narrative suggested. NVIDIA's demand is not speculative leverage; it's infrastructure build-out with visible order books extending into 2025.
The earnings report will provide data. But the structural analysis already provides the answer: NVIDIA remains the most important company in AI infrastructure, and the current valuation does not fully reflect its strategic position. Watch the gross margin guidance, watch the China commentary, and watch the Blackwell revenue contribution. These three signals will tell you whether the 75% margin is a peak or a foundation.
