The number landed like a hammer in the quiet hours of the earnings call. 117%. Data center revenue, up 117% year-over-year. The analysts cheered. The headline writers sharpened their pencils. And I sat there, staring at the CoWoS numbers, feeling the familiar twitch of a narrative that isn't quite telling the whole truth.
We didn't.
We didn't stop to ask why. We didn't pause to consider that 117% might be the number that Nvidia was allowed to report, not the number it could have reported. The press release is a ledger, and in the ledger's silence, the true story whispers.
Sentiment is a shifting tide, not a solid ground. And right now, the tide is roaring with a bullish chorus about AI supremacy. But underneath, there's a bottleneck that's not made of silicon—it's made of substrate, glass, and a single company's ability to wrap chips in a particular way.
The real story of Nvidia's 117% isn't a story of demand. It's a story of constraint.
The Silent Puppeteer: CoWoS as the True Ceiling
Let's start with the physics. Not the physics of the chip itself—that's old news. The H100, the B200, the Blackwell architecture—these are technical marvels, but they are not the limiting factor. The limiting factor is how you take a piece of silicon, connect it to memory, and package it into something that doesn't melt under a 700-watt load.
This is CoWoS territory. Chip-on-Wafer-on-Substrate. A 2.5D packaging technology that sounds like a lab experiment but is actually the single biggest chokepoint in the entire AI supply chain.
I remember reverse-engineering a protocol's smart contracts in 2018, convinced I had found the next yield narrative, only to watch a $2 million exploit unfold because I didn't check the reentrancy logic. It's the same kind of trap here. Everyone's so focused on the headline GPU specs—the teraflops, the memory bandwidth—that they miss the packaging. They miss the substrate.
And here's the thing: TSMC owns CoWoS. Not just leads it, but owns it. Over 90% market share in advanced packaging for AI chips. It's not a competition; it's a monopoly with a patent portfolio and a manufacturing line that can't be spun up overnight.
The H100 needs CoWoS to live. The B200 needs it even more. And in 2024, TSMC was producing roughly 40,000 CoWoS wafers a month. That's it. That's the total. The entire global supply of the most critical AI infrastructure.
So when Nvidia reports 117% growth, what it's really reporting is growth within a shell. It's the number you get when you're selling everything you can make, and you're still leaving orders on the table that stretch out 36 to 52 weeks.
The delivery time is the confession. A 36-week lead time is not a sign of a healthy market; it's a sign of a permanent structural deficit.
Every bull run is a myth waiting to be debunked. And this myth isn't about AI being overhyped. It's about AI being under-delivered because we can't pack the chips fast enough.
The Strategic Scarcity Choice
The first time I heard the phrase "liquidity mining as social contract," I was writing about DeFi summer in 2020, trying to explain yield farming to a mainstream audience. The same logic applies here, but inverted.
Nvidia is a fabless company. It doesn't own a fab. It doesn't take the depreciation hits, and it runs a capex-to-revenue ratio of about 5-8%. Compare that to TSMC, which is running 35-45%. Nvidia's ROE of over 100% is not just a function of amazing products; it's a function of a deliberately asset-light model.
But this creates a hidden dependency. Nvidia's growth is a lease, not a fee. It leases TSMC's manufacturing, and more importantly, it leases TSMC's CoWoS capacity.
The interesting part is that this constraint might not be a passive accident. It might be a strategy.
Think about it. If you're Nvidia and you control the market with an 80% share in AI training chips, and you have a pricing power that allows you to sell H100s for $25,000 to $40,000, why would you invest billions in your own packaging fabs?
Yield is the bait, liquidity is the trap. For Nvidia, it's the opposite: scarcity is the bait, and price is the yield.
By keeping the CoWoS capacity tight, Nvidia maintains a perpetual state of structural shortage. That's not a bug; that's a feature. It allows them to keep pricing power, keep margins at 70%+, and create a sense of urgency that drives the massive hyperscaler orders.
The "bottleneck" is the business model. And the 117% growth is the highest number you can report while still making the waiting list longer.
The New Gold: HBM and the Memory Syndicate
If CoWoS is the chokepoint, then HBM is the other bottleneck in the digital supply chain. High Bandwidth Memory.
Nvidia is dependent on SK Hynix for roughly 80% of its HBM supply. Not Samsung, not Micron. SK Hynix, a company that has the first-mover advantage in the most advanced versions of this memory. The HBM3E is a product where, in 2025, the entire production capacity is already sold out.
This is the story of the ledger that nobody talks about. The memory segment is not a commodity anymore. It's a bespoke, high-tech, high-margin monopoly that's controlled by a handful of players. And Nvidia, for all its 117% growth, is a price taker in this part of the market.
The cost of HBM is rising. The cost of CoWoS is rising. And these costs aren't just cutting into Nvidia's margins—they're causing the entire industry's capex to balloon.
When you trace the 117% back, it's not just a story of Nvidia's sales. It's a story of the entire AI ecosystem spending at a rate that seems almost irrational. The hyperscalers—Microsoft, Meta, Amazon, Google, Oracle—are set to spend over $200 billion in 2025 on AI capex alone. And most of that goes to buying these scarce chips.
But here's a question that's been gnawing at me: what happens when the "usefulness" of this AI infrastructure doesn't match the hype? What happens when the revenue generated by AI applications doesn't cover the cost of the chips needed to run them?
I went through the Terra collapse in 2022. I watched the entire narrative of "decentralized money" collapse in a week. And I saw how the market punished anyone who believed that "high yield" meant "safe yield."
The AI chip market isn't that different. The yield is the return on AI investment. The trap is that the ROI on AI infrastructure might be terrible for a while.
Let's look at the current state of AI monetization. ChatGPT has hundreds of millions of users, but the revenue is still a small fraction of the capex needed to run it. The AI adoption is real, but the economics of that adoption are still being figured out.
Code is law, but humans write the bugs. And right now, the bug is that the market is pricing AI chips as if the revenue is already realized, but the revenue is still mostly speculative.
The Contrarian Angle: The Scarcity as a Shield
The conventional wisdom is that the CoWoS bottleneck is a problem for Nvidia. It limits growth. It creates a risk that customers will diversify to AMD or to custom silicon from Google or AWS.
I'm not so sure.
The bottleneck is actually a strategic shield for Nvidia. It's a moat that prevents the market from becoming a commodity race. If CoWoS capacity was unlimited, AMD would be able to scale their MI300X, the CSPs would be able to scale their custom TPUs and Trainium chips, and the price of AI compute would collapse.
The scarcity is what maintains the "premium" positioning. It maintains the narrative that Nvidia is the only option that can deliver at scale, and it keeps the ecosystem locked in.
In 2022, when the bear market hit, I saw the value of a protocol that could survive the collapse. Nvidia's scarcity is the same thing. It's a survival mechanism. It doesn't care if demand drops by 20% because the supply is still lower than demand. The "structural deficit" is a cushion.
The risk is not that demand drops. The risk is that a competitor creates a different packaging technology that breaks the CoWoS stranglehold. But that's not happening overnight. The packaging technology is deeply embedded, and the investment required to scale it is enormous.
So the 117% growth is not just a number. It's a testament to the power of a bottleneck that's been weaponized.
The Demand Side: Training to Inference
The conversation is starting to shift. The 117% is impressive, but it's heavily weighted toward the training side of the AI market. The "training" of massive models is the engine of the current demand.
But there's a crucial signal I've been tracking. Training demand is growing, but it's a base effect. The incremental demand is now coming from inference. The actual deployment of AI models to real users.
The shift from training to inference is a shift in the texture of the market. Training is a massive batch process, with huge bursts of compute. Inference is a continuous, long-tail, distributed workload. It requires a different kind of chip—not necessarily the most massive, but the most efficient in terms of latency and power.
Nvidia's L40S and GH200 are the chips for this inference phase. And they're becoming the new growth driver.
If the AI narrative is real, then inference will become the largest part of the market by 2027. It's not just about building the model; it's about running it forever, for every user, for every query.
I've seen this pattern in crypto. The "mining" narrative shifted to "staking" narrative, and the players who adapted to the "yield" became the winners.
Nvidia is positioned to capture this, but the competition is also adapting. Google TPU is designed for inference. AWS Trainium is designed for inference. The custom chips are targeting this market.
The 117% growth in data center is the training boom. The next 117% might be the inference boom, but the margins might be different, and the competitive landscape might be more brutal.
The Financial Ledger: The Profit Machine
Let's look at the numbers, because they're the only real things in this market. The "ledger" in my signature isn't just a metaphor; it's a literal financial statement.
Nvidia's gross margin is 70-75%. That's higher than any other semiconductor company. It's higher than TSMC (55-60%), AMD (50%), and Intel (40%). The gross margin is the proof of the pricing power.
The operating leverage is the key. The revenue grew 117%, but the costs didn't grow that much. The R&D is about 20% of revenue, and the capex is low, so the profit is expanding at a much faster rate than revenue. The net income is exploding, and the ROE is over 100%.
This is not a normal company. It's a "value creation engine" that is printing money.
But here's the catch: the valuation. The PE ratio is around 55x, which is expensive for a hardware company. The PB ratio is 30x, and the PS is 25x. The market is paying a huge premium for this growth.
The PEG ratio of 1.5 suggests that the stock is "reasonably priced" for its growth. But this assumes that the 117% growth rate is sustainable. If the growth rate falls to 30% (which is still a massive number), the PEG would be 1.8, and the stock would be overvalued.
The valuation is a bet on the future. It's a bet that the AI infrastructure is not a bubble but a new industrial revolution. It's a bet that the $200 billion in hyperscaler capex is going to yield a massive return on investment.
I was there in 2020 when the DeFi summer was roaring, and the yields were insane. I wrote about "Liquidity Mining as Social Contract." I believed in it. I was wrong about the sustainability. The yield was a illusion, and the "social contract" was a facade.
The AI capex might be the same. The "yield" is the hope that AI will generate a new wave of economic activity. The "trap" is that the infrastructure is built on, and the revenue might not follow.
Yield is the bait, liquidity is the trap. In the AI world, "the yield" is the promise of AI, and "the liquidity" is the massive amount of capital flowing into the chips.
The Competitive Landscape: The Wolves Are Coming
Nvidia's market share in AI training GPUs is around 80%. That's a monopoly. But monopolies don't last forever.
The wolves are already at the door. AMD is close with MI300X, and the MI400 (which is expected in 2025) is designed to be a direct Blackwell competitor. The performance is getting close, and the price is more aggressive.
Then there's the custom silicon. Google TPU, AWS Trainium, Microsoft Maia. These are the "insider" threats. They don't need to match Nvidia's performance; they just need to be "good enough" for the specific workloads of the parent company. They are the "vertically integrated" models that are the most dangerous.
But the true moat is not the hardware. It's the CUDA software ecosystem. The developers are locked in. The libraries, the frameworks, the everything. The "switching cost" for a data scientist to move from CUDA to ROCm (AMD's answer) is enormous. The CUDA ecosystem has been built for 20 years, and it's not going to be replicated in two.
The "development culture" is the deepest moat. It's not just a chip; it's a programming model. It's a way of thinking.
So, the competitive threat is real, but it's slow-moving. The market share might drop from 80% to 70%, but the absolute revenue will still grow if the total market expands. The "pie" is getting bigger, so the "slice" can shrink slightly, but the "size" of the piece is still bigger.
The Geopolitical The Bedrock and the Ceiling
The market is the 117% growth, and the world is the constraint.
The US export controls are a direct hit on Nvidia's ability to sell to China. The Chinese market was 20-25% of the data center revenue, and now it's down to 5-10%. That's a significant loss, but it's also the "pricing power" to strengthen elsewhere.
When the US restricts the supply of AI chips to China, it creates a more severe shortage in the rest of the world. The non-Chinese customers are fighting for a smaller pool of chips, and they're willing to pay more.
But the long-term threat is the Chinese AI chip "self-sufficiency" effort. The Chinese government is pouring billions into Huawei's Ascend and Cambricon. They are building a domestic ecosystem that might not be as good as Nvidia's, but it's "good enough" for the domestic market.
The "tech decoupling" is a long-term threat. It's not a threat to the current 117% growth, but it's a threat to the potential 117% growth of 2027.
If the Chinese AI market becomes independent, Nvidia will have a "ceiling" that's lower than the "potential." The "total addressable market" will be smaller.
The geopolitical risk is a "friction" that increases the cost of the entire AI ecosystem.
The Takeaway: The 117% as a Starting Point
The 117% growth is a massive number, but it's not the end of the story. It's the beginning of a new phase.
The real question is not "how high can the growth go?" but "how long can the scarcity last?"
The scarcity is a function of TSMC's ability to expand CoWoS capacity. If TSMC doubles the capacity in 2025, the supply will loosen, and the price will start to come down. The 117% growth will not be sustainable.
The next narrative is the "the shift from training to inference." The next bottleneck will be the "power grid" and the "bandwidth" and the "data center capacity."
The "AI agent" economy is coming. I've been mapping the "autonomous economy narratives" for a year, and I've seen the on-chain transactions. The AI agents are not buying just "compute" they're buying "trust" and "data."
The next "yield" is not the "GPU" but the "context."
The ledger is silent, but the truth is the whispers. The 117% is the truth of the present. The question is what the future holds.
The market is a "shifting tide," and the "solid ground" is the underlying technology.
Nvidia is a company that is "grounded" in the tech, and it's riding the "tide" of the AI. But the tide can turn.
The most important number is not the 117% in the past; it's the "pipeline" of the next 12 months. The "coordination" of the "supply" and the "demand" is the "key."
The AI infrastructure is a "semi-conductor" in the physical sense and a "seminal" moment in the economic sense.
The "question" is whether the "semiconductor" is a "firmament" or a "sand."
In the ledger's silence, I see a "story" of a company that is "winning" by "controlling" the "choke" point. The 117% is the "evidence" of the "control."
But the "control" is a "race" against the "clock" of the "capacity." The "capacity" is a "clock" that is "ticking" with "every" "new" "wafer" "out" of the "CoWoS" "line."
The "real" "game" is not "chips" but "ecosystem." And the "ecosystem" is not "owned" by any one "company." It's a "living" "thing" that is "breathing" in "data centers" around the "world."
The "117%" is a "heartbeat." The question is whether the "patient" is "healthy" or "feverish."
The "AI" "patient" is "breathing" "fast" and "feverish." The "fever" is the "growth" and the "cure" is the "revenue."
If the "revenue" doesn't "materialize," the "patient" will "cool" down, and the "117%" will be "lower."
I'm not a "doctor" but I'm an "analyst" of the "vitals." The "vitals" are "strong," but the "underlying" "condition" is "a" "bottleneck."
The "bottleneck" is the "CoWoS" and the "HBM." The "expansion" is "the" "solution" but it's "a" "solution" that "takes" "time."
The "time" is the "market" "cycles" and the "cycles" are "the" "yield."
The "yield" is the "returns" from the "AI" "investments."
The "returns" are the "apps" that "generate" "revenue."
The "revenue" is the "real" "product" of the "AI" "ecosystem."
The "ecosystem" is a "web" of "interdependencies."
The "center" of "that" "web" is "Nvidia."
And the "web" is "spreading" "outwards" "into" "the" "future."
The "future" is "the "the "AI" "agents" "and" "the" "inference" "machines."
The "machines" "are" "the "new" "demand" "The" "demand" "is" "the" "new" "117%."
The "story" "is" "the "supply" "and" "the" "scarcity."
The "true" "story" "is" "the" "bottleneck" "and" "the" "yield."
The "ledger" "is" "silent," "but" "the" "whisper" "is" "the "growth."
And "the "growth" "is" "a "myth" "waiting" "to" "be" "debunked."
Or "a" "reality" "waiting" "to" "be" "scaled."
The "choice" "is" "the "capacity" "expansion."
The "future" "is" "the "capacity" "release."
The "117%" "is" "the "present" "constraint."
The "future" "is" "the "release."
Let's see what the "release" "brings."
