There is a moment in every technology cycle when the numbers stop being about products and start being about power. Nvidia's FY2025 Q4 earnings, with revenue landing at $96.2 billion, was such a moment. The stock rebounded at the opening bell, and the analysts nodded approvingly, but I found myself thinking less about the revenue figure and more about the invisible scaffolding beneath it. From code audits to community heartbeats, I have spent nearly three decades watching how infrastructure shapes human behavior. What Nvidia has built is not merely a chip company. It is the closest thing our industry has to a public utility for intelligence itself.
Let me be clear about what this quarter actually reveals. The headline number is staggering, but the deeper story is in the concentration. Data center revenue now accounts for roughly 85 to 90 percent of Nvidia's total. The gaming business, which once defined the company, is now a rounding error at five to eight percent. Professional visualization and automotive together barely register. This is not a diversified semiconductor portfolio. This is a single, massive bet on AI compute as the new global infrastructure.
As a cryptographer who has spent years auditing consensus mechanisms and incentive structures, I recognize this pattern. When a system reaches this level of concentration, the risks are no longer technical. They are systemic. The question is not whether Nvidia can keep shipping chips. The question is whether the entire global economy can absorb what these chips represent, and at what cost to the communities that rely on them.
The Blackwell Architecture and the Rhythm of Obsolescence
The technical details of this quarter matter less than the cadence they reveal. Blackwell, built on TSMC's 4nm N4P process, is in full production. Hopper, the previous generation on N4, is winding down. And Rubin, the next architecture on 3nm, is slated for 2026. What this means is a product cycle of roughly one year, compressing what used to be a three-year rhythm into something far more aggressive.
I have seen this before, in the early days of GPU computing when the annual cadence was a survival tactic. But there is a difference between iterating for competitive advantage and iterating because your customers demand ever-increasing compute to justify their own capital expenditures. Nvidia is now locked in a dance with its largest customers. Microsoft, Meta, Amazon, Google, and Oracle together represent 50 to 60 percent of revenue. These companies are not just buying chips. They are buying the narrative that AI requires exponentially more compute every year. And Nvidia is happy to sell them that story, because the story is true, at least for now.
The architecture itself is formidable. The Blackwell B200 uses a dual-die design integrated through TSMC's CoWoS packaging, with HBM3e memory tightly coupled to the compute dies. This is not just a chip. It is a system designed to solve the memory bandwidth bottleneck that has plagued AI training since the beginning. The die size is enormous, roughly 800 square millimeters, which means yield rates matter enormously. TSMC's N4 process is mature, above 90 percent, but the sheer size of these dies makes cost control a constant pressure. Blackwell Ultra, expected in the second half of 2025, will push this further with CoWoS-L packaging, allowing even larger chiplet configurations.
Here is the insight most analysts miss. Nvidia does not actually manufacture anything. It is a fabless designer, which means TSMC bears the yield risk and the manufacturing complexity. But Nvidia's dependence on TSMC is absolute. The 4nm and 3nm processes are exclusively available through TSMC. The CoWoS advanced packaging, which is the true bottleneck in AI chip supply, is also almost exclusively TSMC. Nvidia consumes roughly 60 percent of TSMC's CoWoS capacity. This is not a supply chain. It is a stranglehold, and both companies know it.
The CoWoS Bottleneck and the Illusion of Capacity
Let me speak plainly about the packaging bottleneck because it is the single most underappreciated constraint in the entire AI supply chain. CoWoS, TSMC's 2.5D advanced packaging technology, is the glue that holds AI chips together. Without it, you cannot integrate HBM memory with compute dies. And right now, CoWoS capacity is running at nearly 100 percent utilization. TSMC is doubling its CoWoS capacity in 2025, with monthly output expected to reach 80,000 to 100,000 wafers by the end of the year, up from roughly 40,000 to 50,000 at the end of 2024.
This expansion is not happening in a vacuum. TSMC's capital expenditure for 2025 is estimated at $40 to $50 billion, and a significant portion of that is dedicated to advanced packaging. But here is the hidden dynamic. Nvidia's book capital expenditure is only five to eight percent of revenue, which looks remarkably efficient for a company growing at this pace. The reality is that Nvidia is using prepayments and long-term agreements to lock in TSMC capacity. The true capital commitment is hidden off the balance sheet, embedded in prepaid assets and contractual obligations.
This is a rational strategy, but it is not without risk. If AI demand were to slow, Nvidia would be left holding commitments for capacity it no longer needs. The equipment lead times for CoWoS are six to twelve months, and the production ramp takes another six to nine months. This is not a flexible system. It is a system built for a world where demand is certain. And certainty, in my experience, is the rarest commodity in any technology cycle.
The CUDA Moat and the Real Competitive Landscape
The competitive analysis of Nvidia's position is often reduced to hardware specifications, but that misses the point entirely. Nvidia's true moat is CUDA, the software ecosystem that has been accumulating for over 15 years. This is not a programming language or a library. It is a complete stack, from low-level kernels to high-level application frameworks, that has become the default language of AI development.
When I audit a protocol, I look at the network effects. CUDA has network effects that are almost impossible to replicate. Every researcher who learns CUDA, every library that is optimized for Nvidia hardware, every paper that assumes Nvidia as the baseline, adds to the moat. AMD's MI300 and MI400 series are competitive on raw specs, and Intel's Gaudi line is improving, but they are not competing on the same playing field. They are trying to build a new field while the existing one is already fully developed.
The market share numbers tell this story. Nvidia holds 80 to 90 percent of the AI training chip market and 60 to 70 percent of the inference market. The second-place competitor, AMD, holds roughly 10 percent in training and 15 percent in inference. This is not a close race. It is a procession. But the long-term threat is not AMD. It is the cloud providers themselves. Google's TPU, Amazon's Trainium, and Microsoft's Maia are all designed to reduce dependence on Nvidia. In inference workloads, these custom chips are already competitive on a cost-per-inference basis.
I have seen this dynamic before. In the early days of networking, Cisco dominated with proprietary protocols. The market eventually shifted to open standards, and Cisco's dominance eroded. The question is whether CUDA becomes the TCP/IP of AI or the Cisco IOS of AI. If CUDA becomes a true standard, Nvidia benefits from the network effect. If the market fragments, Nvidia loses its lock-in.
The Margins, The Valuation, and The Uncomfortable Questions
Nvidia's gross margins are extraordinary, running between 70 and 75 percent. This is not a hardware company margin. This is a software company margin. It reflects the scarcity of AI compute and the pricing power that comes from being the only game in town. The gross margin has climbed steadily from roughly 60 percent in FY2023 to 65 percent in FY2024 to the current level. This is a trajectory that would make any CFO weep with joy.
But here is the contrarian angle. The margin structure is not sustainable. As inference workloads grow as a percentage of AI compute, and as custom silicon from cloud providers takes share in inference, Nvidia's product mix will shift toward lower-margin parts. My estimate is that gross margins will gradually decline to 65 to 70 percent over the next two to three years. That is still an excellent margin, but it represents a compression that the market has not fully priced in.
The valuation, at 30 to 35 times trailing earnings, looks reasonable on a PEG basis of 1.5 to 2.0, given the growth rate. But this valuation embeds an assumption of sustained 30 percent plus earnings growth for the next three years. If AI investment slows, if the bubble narrative proves correct, the multiple will compress. I have lived through the dot-com crash and the crypto winter. I know what happens when the market collectively realizes that growth rates cannot persist forever.
There is also the question of return on invested capital, which is the metric I care most about. Nvidia's ROIC is 60 to 70 percent, against a WACC of 10 to 12 percent. This is exceptional value creation. The company is not just growing. It is growing profitably, with operating cash flow of roughly $50 billion and free cash flow of $40 billion. The balance sheet is pristine. This is not a company in financial distress. It is a company in financial dominance.
The Geopolitical Tightrope and the China Question
The geopolitical dimension of Nvidia's position is often misunderstood. The export controls imposed by the US government have reduced China's share of Nvidia's revenue from roughly 25 percent in 2022 to about 10 to 15 percent today. This is a deliberate strategy of de-China-ification, reducing dependence on a market that is increasingly closed to advanced AI chips. The H800 and H20 variants, designed to comply with export restrictions, are a stopgap. They do not represent a long-term solution.
China's response has been predictable. The Big Fund III, with roughly $47 billion in capital, is funding domestic AI chip development. Huawei's Ascend and Cambricon are the most advanced domestic alternatives, and while they lag Nvidia by two to three years in performance, the gap is narrowing. The long-term threat is real, not because Chinese chips will be better, but because the Chinese market will increasingly be closed to Nvidia, and the scale of Chinese AI development will eventually produce viable alternatives.
TSMC's geographic diversification, with fabs in Arizona, Dresden, and Kumamoto, provides some supply chain resilience. But the advanced packaging capacity, the CoWoS capacity that is the true bottleneck, remains concentrated in Taiwan. A disruption in Taiwan, whether from natural disaster or geopolitical conflict, would halt Nvidia's supply for six to twelve months. This is a tail risk that is difficult to hedge and impossible to ignore.
The Structural Shift and What It Means for Web3
Now let me bring this back to the world I inhabit, the world of Web3 and decentralized systems. There is a deep irony in Nvidia's dominance. The AI revolution, which promises to democratize intelligence, is being built on infrastructure that is more centralized than anything that came before it. A single company, fabless but dominant, controls the compute that powers the most important technology of our era. This is not decentralization. This is the opposite.
Building bridges where DeFi once built walls, I have always believed that technology should serve communities, not the other way around. The AI infrastructure stack, as it currently exists, serves the interests of a few hyperscalers and their shareholders. The communities that provide the data, the labor, and the context for AI systems are largely excluded from the value they create. This is a failure of design, not a failure of technology.
The Web3 response to this centralization has been to build decentralized compute networks, marketplaces for GPU resources, and protocols for verifiable inference. These efforts are laudable, but they face an uphill battle. The CUDA moat is not just a technical advantage. It is a social advantage, a network of developers, researchers, and enterprises that have built their careers on Nvidia's stack. Breaking that network effect is not a technical problem. It is a sociological problem.
Trust is not a protocol, it is a practice. And the practice of building decentralized AI alternatives requires more than technical competence. It requires a deep understanding of the communities that AI serves and the willingness to build systems that serve those communities first. This is the work I have dedicated my career to, and it is the work that will define the next decade of the industry.
The Signals to Watch and the Questions That Remain
The immediate signals to watch are clear. Nvidia's FY2026 Q1 earnings, expected in May 2025, will show whether the growth rate can be sustained. TSMC's monthly revenue reports will reveal the pace of CoWoS expansion. And the capital expenditure guidance from the hyperscalers will indicate whether the AI buildout continues at its current pace. These are the short-term signals, and they will be noisy.
The medium-term signals are more important. Blackwell Ultra shipments in the second half of 2025 will test whether the product cycle can sustain its accelerated pace. AMD's MI400 launch will show whether the competitive gap is narrowing. And the progress of cloud provider custom chips, Google's TPU v6 and Amazon's Trainium 3, will reveal whether the threat to Nvidia's inference dominance is real.
The long-term signals are the ones that matter most. The Rubin architecture, expected in 2026, will define Nvidia's trajectory for the second half of the decade. The growth of AI inference demand will determine whether the market is as large as the optimists believe. And the progress of Chinese AI chip development will shape the geopolitical landscape of the industry.
These are the questions that keep me up at night. Not whether Nvidia will beat earnings estimates, but whether the infrastructure we are building serves the people it claims to serve. Whether the intelligence we are creating is distributed or concentrated. Whether the trust we are building is a protocol or a practice.
The numbers are impressive. The revenue, the margins, the market share, they are all extraordinary. But the real question is not what Nvidia has achieved. It is what we, as an industry, will do with the power that this infrastructure represents. Liquidity flows, but culture remains. And the culture we build around AI will determine whether this technology liberates or entrenches.
I have spent my career auditing code, but the audit that matters most is the one we conduct on ourselves. The audit of intent, not just invoice. The audit of community, not just consensus. The audit of the soul behind the smart contract. That is the work that remains. And it is the work that will define whether the AI era is one of abundance or extraction.