The ledger does not lie, only the interpreters do. Here is the balance sheet for the AI infrastructure era, and it reveals a company whose dominance is predicated on a single point of failure: Taiwan Semiconductor's advanced packaging lines.
Nvidia reported FY2025 Q4 revenue of $96.2 billion, a figure that sent shares rebounding at the start of the earnings call. The market read this as confirmation that AI demand remains insatiable. The data confirms something else entirely: a supply chain concentration so extreme that one earthquake in Hsinchu could erase more value than any competitive threat from AMD or Intel.
The Context: From GPU Vendor to Infrastructure Monolith
The numbers require context. Nvidia's data center segment now accounts for roughly 85-90% of total revenue. This is no longer a chip company. It is an AI infrastructure platform with a GPU attachment mechanism. The gaming division, once the foundation of the business, is now a rounding error at 5-8% of revenue.
Gross margins sit at 70-75%, a figure that approaches software company territory. Microsoft, the world's most valuable software franchise, operates at roughly 68-70% gross margins. Nvidia, a hardware designer, has surpassed that benchmark. The implication is not that Nvidia has become more efficient at manufacturing. It has become more efficient at extracting rent from a captive market.
The product cycle has accelerated to roughly one year. Hopper arrived in 2022. Blackwell followed in 2024. Blackwell Ultra lands in 2025. Rubin arrives in 2026-2027. Each iteration demands more CoWoS packaging capacity, more HBM supply, more TSMC advanced node allocation.
Trust is a bug, not a feature. The market trusts that this cadence can be sustained. The supply chain data suggests otherwise.
The Core Analysis: Deconstructing the Dependency Stack
Let me walk through the structural vulnerabilities that the earnings report obscures. Based on my audit experience examining infrastructure dependencies across crypto and traditional technology, the concentration profile here is extraordinary.
The CoWoS Bottleneck
Nvidia consumes approximately 60% of TSMC's CoWoS advanced packaging capacity. This is not a negotiating position. It is a hostage situation. The packaging technology itself is not the differentiator—the scarcity is.
TSMC plans to double CoWoS capacity through 2025, reaching 80,000-100,000 wafers per month by year-end, up from 40,000-50,000 at the end of 2024. The equipment lead time is 6-12 months. The cleanroom-to-production ramp takes another 6-9 months. This means the capacity doubling announced today was actually decided in early 2024, based on order visibility that extends through 2026.
The hidden information here: Nvidia's "capacity" is not owned. It is leased through prepayments and long-term agreements. The balance sheet shows capital expenditure intensity of only 5-8% of revenue, far below TSMC's 35-45%. But the actual capital commitment, embedded in prepayments and contractual obligations, is far higher than the capex line reveals.
Code is law; intent is irrelevant. The same principle applies to capacity. What matters is not what Nvidia intends to build, but what TSMC physically can deliver.
The HBM Dependency
High Bandwidth Memory supply comes from SK Hynix and Samsung, with SK Hynix holding the dominant position in HBM3e. There is no alternative source at scale. Micron exists but with limited capacity. The concentration here is effectively 100% Korean supply chain.
HBM prices are rising due to supply constraints. Nvidia's pricing power allows it to pass these costs through to customers. But the pass-through mechanism has limits. At some point, the hyperscalers—Microsoft, Google, Amazon, Meta, Oracle—will evaluate whether the total cost of ownership justifies continued premium pricing.
The customer concentration ratio is 50-60% across the top five accounts. Microsoft alone represents 15-20% of revenue. This is not diversified demand. It is a small group of extraordinarily wealthy entities engaged in an AI arms race that has not yet demonstrated sustainable returns on investment.
The Fabrication Monopoly
TSMC's N4 and N3 nodes are the only viable manufacturing options for Nvidia's AI accelerators. There is no second source. Samsung's foundry business has not achieved comparable yields on advanced nodes. Intel Foundry is years behind and has not secured meaningful external customers.
The technical gap between Nvidia and its nearest competitor, AMD, is 1-2 years in hardware. But the actual competitive moat is not the hardware. It is the CUDA software ecosystem, accumulated over 15 years of developer mindshare, libraries, and toolchains. AMD's MI300 series closes the hardware gap. The software gap remains a chasm.
History repeats, but the gas fees change. In 2022, the cryptocurrency collapse created a GPU inventory glut that punished Nvidia's gaming segment. The AI boom absorbed that excess capacity and transformed the company. The question is whether the current AI investment cycle has a similar correction embedded in its trajectory.
The Contrarian Angle: What the Bulls Got Right
I have spent considerable time cataloging risks. Intellectual honesty requires acknowledging the counterarguments.
The AI demand signal is real. Cloud capital expenditure guidance from Microsoft, Google, Amazon, and Meta remains robust through 2025. The inference segment, which will likely represent more than 50% of AI chip demand by 2026, is only beginning to scale. Applications like ChatGPT and Copilot have demonstrated product-market fit. Enterprise adoption is accelerating.
Nvidia's pricing power is extraordinary by any historical standard. Gross margins of 70-75% in hardware are unprecedented at this revenue scale. The company generates approximately $50 billion in operating cash flow with an OCF/net income ratio of 1.2. Free cash flow of $40 billion provides enormous strategic flexibility. ROIC of 60-70% against a WACC of 10-12% represents one of the most efficient capital deployment engines in corporate history.
The product roadmap provides visibility that competitors cannot match. Blackwell Ultra in 2025, Rubin in 2026-2027, and the transition to TSMC's N3 and eventually N2 nodes. Nvidia does not need to be perfect. It needs to maintain a one-generation lead, and the ecosystem lock-in ensures customers wait for Nvidia rather than defect to alternatives.
The "AI bubble" narrative has been wrong so far. Cloud providers are not investing in AI infrastructure for speculative purposes. They are responding to measured demand signals from enterprise customers. The infrastructure build-out is ahead of near-term monetization, but the adoption curve suggests this is a 5-10 year cycle, not a 12-month spike.
The ledger does not lie, only the interpreters do. The current ledger shows a company executing with remarkable precision in a favorable demand environment.
The Takeaway: Accountability in the Age of Concentration
The structural risks are real, but they are not immediate. Supply chain concentration, customer concentration, and geopolitical exposure are liabilities on the balance sheet of the future. They will matter when the demand environment shifts, when cloud providers achieve meaningful cost reduction through custom silicon, or when geopolitical tensions escalate beyond current containment levels.
The signals to track are specific. Nvidia's FY2026 Q1 earnings in May 2025 will reveal whether Blackwell revenue ramp is meeting expectations. TSMC's monthly revenue reports will show CoWoS capacity expansion progress. Cloud provider capital expenditure guidance will indicate whether the demand curve remains steep.
The ledger does not lie, only the interpreters do. The interpretation that matters is whether the AI infrastructure build-out represents a durable economic transformation or a leveraged bet on future monetization that has not yet arrived. The data supports continued growth through 2026. The structural vulnerabilities suggest that the margin for error is thinner than the stock price implies.
The question is not whether Nvidia is a great company. It is whether great companies can sustain premium valuations when their supply chain is a single point of failure. The answer, historically, has been no. The timing of that correction remains uncertain. The direction does not.