Here's the data point nobody wants to discuss: Nvidia's H100s were 'sold out' for eighteen months straight. Revenue tripled. Guidance crushed estimates by $4 billion. The stock went nowhere fast. That's not a demand story. That's a capacity ceiling wearing a bull mask.
I've spent the last decade trading around semiconductor supply chains, and I've learned one thing: when a fabless chip designer says 'sold out,' they're telling you their upstream partner's CoWoS line is maxed. Not that the market can't absorb more. Nvidia's Q2 beat, the $108 billion forward guidance, the 80-90% AI training share — all of it is downstream of a single bottleneck: TSMC's advanced packaging capacity.
Let's be clear about what's actually happening here.
The Context: A Monopoly Wrapped in a Bottleneck
Nvidia designs the best AI silicon on the planet. That's settled science. The Hopper architecture, the Blackwell transition, the CUDA moat — none of that is in dispute. But Nvidia doesn't fab a single wafer. TSMC does, on 4nm and 3nm nodes. And critically, TSMC's CoWoS 2.5D packaging line is the single most constrained resource in the entire AI supply chain.
I audited this exact dependency in early 2023 when I was evaluating restaking protocols that depended on AI compute demand. What I found then still holds: CoWoS capacity is running above 100% utilization. TSMC is doubling it, but that expansion doesn't hit full output until 2026. Meanwhile, HBM supply — dominated by SK Hynix — is its own separate choke point.
Here's what that means in practical terms. Nvidia's revenue growth over the next 12 months is not a function of demand. It's a function of how many advanced packages TSMC can push out the door. The 'sold out' status isn't a signal that the AI trade is healthy. It's a signal that the physical layer of the AI trade is structurally constrained.
The Core: Three Choke Points, One 'Sold Out' Status
Let's break down the supply chain math, because this is where the real analysis lives.
Choke Point 1: TSMC Advanced Node Capacity
TSMC's 4nm is mature — yield above 90%. 3nm is still ramping, somewhere in the 80-85% yield band. But the issue isn't yield. It's allocation. TSMC's advanced node capacity is effectively sold out across all customers. Nvidia gets priority, but priority isn't infinite supply. When your foundry is at 95% utilization and your packaging line is above 100%, your growth ceiling is someone else's CapEx schedule.
Choke Point 2: CoWoS Packaging
This is the real story. CoWoS is 2.5D advanced packaging that stacks HBM memory alongside the GPU die. It's what makes an H100 an H100. And TSMC basically owns this technology. Samsung's I-Cube and Intel's EMIB are years behind in yield and capacity. The result: Nvidia's ability to ship AI accelerators is directly proportional to TSMC's CoWoS output.
Here's a number that should concern you: TSMC's CoWoS capacity is set to roughly double by 2025. But AI compute demand is doubling every 3-4 months. The math doesn't close. The bottleneck persists through 2026.
Choke Point 3: HBM Supply
SK Hynix dominates HBM3e. Samsung and Micron are scrambling to catch up. HBM prices are rising. Supply is allocated years in advance. Nvidia has locked in supply, but that lock-in comes at a price premium that gets passed down the chain.
Now here's the part most analysts miss. Nvidia's gross margin sits at 60-65%. That's the highest in the semiconductor industry. But that margin is partially a function of scarcity. When capacity normalizes — when TSMC's Arizona fab comes online, when CoWoS doubles, when Samsung's HBM yield improves — the pricing power softens. The question isn't whether Nvidia's margin compresses. It's whether the compression happens before or after the valuation repricing.
The Contrarian Angle: 'Sold Out' Is a Double-Edged Sword
Here's the counterintuitive take that I haven't seen in the mainstream coverage.
The 'sold out' status is actually a strategic choice disguised as a supply constraint. Nvidia controls allocation. By keeping supply tight, they maintain pricing power, lock in customers through the CUDA ecosystem, and create a narrative of scarcity that supports a 60x P/E ratio. It's a beautiful flywheel — until it isn't.
But there's a darker side to this dynamic. Every customer who can't get Nvidia chips is incentivized to find alternatives. AMD's MI300 is close to H100 performance. Google's TPU v6 is shipping. Amazon's Trainium is improving. And OpenAI, Microsoft, and Anthropic are all exploring custom silicon. The 'sold out' status is simultaneously Nvidia's greatest pricing weapon and their greatest competitive vulnerability. Every quarter of constrained supply is a quarter of competitor validation.
The second blind spot is geopolitical. US export controls have actually helped Nvidia in a perverse way. By cutting off the Chinese market, Nvidia has redirected supply to US and allied customers, intensifying the scarcity in those markets. But the long-term cost is real. China is the world's largest semiconductor market. Huawei's Ascend chips are improving. The national AI chip push is funded to the tune of $50 billion through the Big Fund Phase III. Nvidia's exit from China isn't a withdrawal — it's a forced retreat that creates a vacuum competitors are already filling.
And here's the part that keeps me up at night. The current 'sold out' status masks the true demand picture. We don't know how much of the AI chip demand is real and how much is speculative hoarding. CSPs are stockpiling. Microsoft, Meta, Amazon, Google — they're all buying ahead of need. If even 20% of that hoarding normalizes into actual deployment timelines, we could see an inventory correction by 2026-2027. And semiconductor corrections, historically, are 30-50% drawdowns.
The Takeaway: Track the Bottleneck, Not the Narrative
Here's my position. Nvidia is the best semiconductor company of this cycle. The CUDA moat is real. The technology lead is 1-2 years ahead of AMD and 2-3 years ahead of custom ASICs. The financials are pristine — 80% ROE, 70% ROIC, $28 billion in operating cash flow, all R&D expensed conservatively. This is a world-class operation.
But the stock is priced for perfection, and the 'sold out' narrative is the linchpin of that pricing. If you're going to hold Nvidia through this cycle, you need to stop watching Nvidia's earnings calls and start tracking three things: TSMC's monthly revenue reports, CoWoS capacity announcements, and CSP capital expenditure guidance. The bottleneck is the trade. When the bottleneck breaks, the narrative changes.
The question isn't whether Nvidia can grow. It's whether the market is paying for the growth that's physically possible, or the growth that's narratively promised. Right now, those two numbers are very different. And when they converge — whether through capacity release or demand destruction — the repricing will be violent in both directions.
I know which side of that trade I'm positioning for.
