Nvidia just dropped a $96.2 billion revenue bomb for FY2025 Q4. The stock bounced at the start of the earnings call. Everyone's celebrating. I'm not.
Let me be clear about what this number actually represents. This isn't a GPU company anymore. This is an AI infrastructure monopoly with a supply chain Achilles heel that could snap at any moment. And the market is pricing it like the growth story will never end.
I've spent the last decade watching chip companies claim dominance. I've audited enough supply chains to know that when a single fabless designer consumes roughly 60% of TSMC's advanced packaging capacity, you're not looking at a moat. You're looking at a single point of failure.
Here's what the earnings report doesn't tell you.
The Blackwell architecture is built on TSMC's 4nm N4P process, and it's already in mass production. The Hopper line—H100 and H200—is winding down on the older N4 node. Rubin, the next-gen architecture, moves to 3nm N3 in 2026. That's the public roadmap. But here's the part that matters: Nvidia doesn't fab anything. They're fabless. Their entire empire rests on TSMC's ability to deliver wafers and, more critically, CoWoS advanced packaging capacity.
Let me break down the packaging situation because this is where the real story lives.
CoWoS is the single biggest bottleneck in AI chip supply right now. TSMC's CoWoS capacity is running at essentially 100% utilization. Nvidia has locked up about 60% of that capacity through prepayments and long-term agreements. The equipment lead time for CoWoS—from ASMPT and K&S—runs 6 to 12 months. TSMC is doubling capacity through 2025, targeting 80,000 to 100,000 wafers per month by year-end, up from roughly 40,000 to 50,000 at the end of 2024.
But here's what I keep coming back to: Nvidia's capex-to-revenue ratio sits at just 5-8%. TSMC's is 35-45%. Nvidia is effectively renting TSMC's balance sheet to secure capacity. That's smart. It's also fragile.
The supply chain concentration is a rational choice, not a management oversight. I've seen this pattern before. When you're the dominant player with 80-90% market share in AI training chips, you don't diversify. You double down on the best supplier and lock in capacity. TSMC's process leadership and CoWoS scale are simply unmatched. Samsung and Intel can't replicate it in the near term. So Nvidia's strategy is: concentrate, lock, and pay whatever it takes.
The risk profile is what keeps me up at night. If TSMC's fabs go down—earthquake, geopolitical conflict, whatever—Nvidia faces 6 to 12 months of supply disruption. That's tens of billions in lost revenue. And there's no Plan B. No alternative foundry. No backup packaging partner at scale.
Now let's talk about the demand side, because the numbers are staggering.
Data center revenue now accounts for roughly 85-90% of Nvidia's total revenue. Gaming is down to 5-8%. Professional visualization is 2-3%. Automotive is 1-2%. This is not a diversified semiconductor company. This is an AI compute infrastructure company wearing a GPU company's skin.
The AI training demand is still exploding. OpenAI, Anthropic, Google—they're all buying H100, H200, and GB200 systems as fast as Nvidia can ship them. But here's the shift I'm watching: AI inference demand is starting to accelerate. As applications like ChatGPT and Copilot move into production, inference workloads are growing faster than training. I expect inference to represent over 50% of AI chip demand by 2025-2026.
That's a problem for Nvidia's margins. Inference chips like the L4 and L40S carry lower gross margins than training chips like the H100 or GB200. The current 70-75% gross margin—which is closer to a software company than a hardware company—will likely drift down to 65-70% as the product mix shifts. That's still excellent. But it's a trend worth watching.
The competitive landscape is where the narrative gets uncomfortable. Nvidia holds roughly 80-90% of the AI training chip market. AMD is second at about 10%. Intel is a rounding error. In inference, Nvidia's share drops to 60-70%, with AMD at 15% and Google's TPU at 10%.
Here's the thing that most analysts miss: the hardware gap between Nvidia and AMD is narrowing. AMD's MI300 and upcoming MI400 are competitive on raw specs. The real moat is CUDA. Fifteen years of developer ecosystem, libraries, and toolchains. That's not something AMD can replicate in a year or two. But it's also not something Nvidia can take for granted forever.
The cloud giants are the long-term threat. Google's TPU, Amazon's Trainium, Microsoft's Maia—they're all getting better. In specific inference workloads, they're already competitive. I estimate they could capture 10-15% of the market by 2027-2028. That's not existential. But it's a slow bleed.
The geopolitical layer adds another dimension of risk. Nvidia's China revenue has dropped from about 25% of total revenue in 2022 to roughly 10-15% now. The export controls on AI chips with compute power exceeding 4,800 TOPS have forced Nvidia to create downgraded products like the H800 and H20 for the Chinese market. The company has effectively executed a "de-China" strategy, reducing exposure to manage geopolitical risk.
But here's the hidden angle: China's domestic AI chip push is a long-term threat that the market is underpricing. Huawei's Ascend and Cambricon are 2-3 years behind Nvidia technically. But with the Big Fund III injecting roughly $47 billion into the sector, that gap will narrow. I give it 3-5 years before Chinese AI chips become a meaningful competitive force.
The supply chain for HBM memory is another concentration risk. SK Hynix and Samsung are essentially the only suppliers, and HBM is critical for AI chip performance. If either has a factory incident—like the SK Hynix fire in 2013—the impact on Nvidia's ability to ship would be immediate and severe.
Let me talk about the financials, because the valuation story is more nuanced than the headlines suggest.
Nvidia's gross margin is 70-75%, with operating cash flow around $50 billion and free cash flow near $40 billion. The OCF-to-net-income ratio is about 1.2, which is healthy. Return on equity is 80-90%. Return on invested capital is 60-70%. The WACC is 10-12%. This is the most value-creative semiconductor company in history.
The valuation is where I have to be careful. The trailing PE is 30-35x. That's actually below the historical average of 40-50x. The PEG ratio is 1.5-2.0, which is reasonable given the 50%+ earnings growth. But here's the catch: the market is pricing in 30%+ annual profit growth for the next three years. If AI demand slows—if we hit an AI bubble moment like the 2000 internet crash—the PE could compress to 20x. That's a 30-40% downside from current levels.
I've seen this movie before. In 2022, when crypto crashed, GPU inventory ballooned. Nvidia's gaming revenue collapsed. The stock dropped 60% from peak to trough. The AI demand eventually absorbed the excess inventory, but the lesson is clear: when the narrative shifts, the multiple compresses fast.
The product cycle is accelerating, and that's both a strength and a risk. Hopper launched in 2022. Blackwell in 2024. Blackwell Ultra in 2025. Rubin in 2026-2027. The cadence is now roughly one year per major architecture. That puts enormous pressure on competitors who can't match the pace. But it also means Nvidia's customers are constantly upgrading, which creates a treadmill effect. The installed base doesn't have time to fully depreciate before the next generation arrives.
Here's my contrarian take: the market is focused on the wrong risk. Everyone's worried about AMD, about cloud giants building their own chips, about export controls. The real risk is the CoWoS bottleneck. If TSMC can't expand capacity fast enough—if the equipment delivery slips, if yields don't improve—Nvidia's growth hits a wall regardless of demand. The company has locked in capacity with prepayments, but that's not the same as owning the fabs.
I've audited enough supply chains to know that prepayments don't guarantee delivery. They guarantee priority. And priority only matters if the capacity actually exists.
The inference shift is the opportunity the market is underpricing. As AI applications move from training to inference, the demand profile changes. Inference is more distributed, more diverse, more price-sensitive. Nvidia's L4 and L40S chips are positioned for this market, but the margins are lower. The company's software stack—NVIDIA AI Enterprise, CUDA-X—is designed to maintain stickiness even as hardware margins compress.
The automotive and robotics angle is a longer-term play. Nvidia's Orin and Thor chips are in autonomous driving and humanoid robots. This is a 2026-2030 story, not a 2025 story. But the total addressable market is enormous.
Let me give you the signals I'm tracking.
Short-term (1-3 months): Nvidia's FY2026 Q1 earnings in May. I'm watching revenue growth, gross margin trajectory, and Blackwell shipment volumes. TSMC's monthly revenue reports will show CoWoS capacity expansion progress. Cloud capex guidance from Microsoft, Google, Amazon, and Meta will signal whether the AI spending spree continues.
Medium-term (3-12 months): Blackwell Ultra shipments in the second half of 2025. AMD's MI400 launch. Cloud giant custom chip progress—Google's TPU v6, Amazon's Trainium 3.
Long-term (12+ months): Rubin architecture launch in 2026. AI inference demand inflection. China's domestic AI chip progress.
The bottom line is this: Nvidia is the most dominant company in the most important technology sector of our time. The moat is real. The financials are exceptional. But the concentration risks—supply chain, customer base, geopolitical exposure—are real too.
The question isn't whether Nvidia is a great company. It is. The question is whether the current valuation already prices in perfection. At 30-35x trailing earnings with 50% growth, the market is paying for flawless execution. Any stumble—a CoWoS delay, an inference margin miss, a cloud capex cut—could trigger a repricing.
I've been through enough cycles to know that the most dangerous moment is when everyone agrees. When the stock bounces at the start of the earnings call and the analysts all nod in unison, that's when I start looking for the cracks.
The cracks are there. They're just not where most people are looking.
Watch the CoWoS capacity numbers. Watch the inference margin trajectory. Watch what the cloud giants do with their custom silicon. And remember: in this industry, the biggest risk is always the one you're not thinking about.