The Chokepoint Trade: Why ARK's NVIDIA and TSMC Buys Are Not a Bet on AI Hype
Credtoshi
ARK's daily trade disclosures just hit the wire again. NVIDIA added. TSMC added. Same session. Same fund family. The immediate reading was obvious: Cathie Wood is doubling down on AI after Meta's earnings miss made the whole narrative wobble. That reading is lazy. This is not a portfolio manager expressing gentle confidence in the AI application layer. This is a supply-chain operator moving capital into the two chokepoints where demand meets physical capacity. Meta missed. The market flinched. ARK bought the shovels. I have been parsing on-chain and semiconductor flows long enough to know that the real story is not "AI is back." The real story is "AI is physically constrained." And whoever owns the constraint owns the margin.
To understand why this matters, you have to understand what ARK usually buys and what TSMC actually is. ARK's core mandate is disruptive innovation: genomics, fintech, robotics, blockchain. The typical ARK holding is a high-growth, capital-light software company with a network effect and a story that sounds good at a dinner party. TSMC is none of those things. It is a kilometer-deep, capital-heavy, physically grounded manufacturing enterprise that has spent decades building a moat in lithography, process control, and advanced packaging. The fact that ARK is now treating TSMC like a growth stock is not an anomaly. It is a thesis.
The thesis is simple: Moore's Law stopped being a free lunch. Transistor scaling still happens, but the cost per node has exploded, and the value in the AI supply chain has shifted from raw transistor count to the ability to deliver a complete, packaged, high-bandwidth compute system. NVIDIA designs the architecture. TSMC manufactures it. TSMC also packages it with CoWoS, stuffs it with chiplets, and coordinates with HBM suppliers. When you buy both NVIDIA and TSMC, you are not getting two versions of the same AI bet. You are locking in both sides of the only contract that matters: design that cannot ship without manufacturing, and manufacturing that cannot command premium pricing without a killer design. That is a vertical arbitrage. ARK has quietly become a participant in it.
Let me be precise about the technical stack. NVIDIA's Hopper generation, the H100, sits on TSMC's 4N process, a customized 5nm-class node. The Blackwell generation, the B200, moves to 4NP, an enhanced version of the same 5nm-class family, and uses two chiplet dies in a single package. This is not a minor packaging detail. Blackwell's double-die design means it consumes twice the advanced packaging capacity and roughly twice the advanced process capacity per unit compared to a monolithic chip. TSMC's 3nm family, N3, went into production at the end of 2022. N3E and N3P are already in the field. TSMC's 2nm node, N2, is scheduled for second-half 2025 volume production and marks the first full transition to Gate-All-Around nanosheet transistors. NVIDIA's next platforms, Blackwell Ultra and Rubin, will ride TSMC's N3 and N2 derivative processes. The two companies are not separate entities in any operational sense. They are one vertically integrated compute machine split by a corporate firewall.
Now look at the yield picture. TSMC does not publish node-specific yield numbers. Publicly traded foundries almost never do. But the industry consensus is that TSMC's 5nm and 3nm yields have ramped faster and stabilized higher than Samsung's equivalent nodes. Samsung reached 3nm GAA before TSMC, but yield and performance lagged enough that most high-end AI customers stayed on TSMC. This yield gap is not a footnoted technical story. It is the foundation of ARK's TSMC position. A foundry with weak yields cannot deliver enough Blackwell-class packages to meet hyperscaler demand. TSMC can. That delivery reliability is what justifies the premium pricing TSMC has been able to push through in 2024 and 2025.
Here is the number that matters more than any AI revenue forecast: advanced packaging capacity. The real bottleneck in the AI supply chain is not photolithography. It is CoWoS. CoWoS is TSMC's 2.5D and 3D packaging technology that places multiple chiplets and HBM stacks on a silicon interposer. Every Blackwell B200 needs CoWoS. Every H100 needs CoWoS. AMD's MI300 series needs CoWoS. Broadcom's custom accelerators need CoWoS. TSMC was estimated to have around 40,000 CoWoS wafer starts per month in 2024 and was targeting roughly double that in 2025. Even with that expansion, orders are pre-committed by NVIDIA, AMD, Broadcom, and a handful of other design houses. There is no spare capacity. There is no second source. Samsung has an equivalent technology, but it has not achieved the same yield and design-win traction at the high end. Intel's Foveros is interesting but late in the HPC race. TSMC's CoWoS is the single most concentrated chokepoint in the entire AI hardware stack.
That concentration is why ARK's move feels different from the standard "buy the leader" trade. ARK is not betting that the AI hype cycle continues. ARK is betting that even if AI revenue forecasts get cut in half, the physical scarcity of advanced capacity will still command a premium. This is the same logic that made me start treating protocol audits as liquidity signals in 2017. When I flagged the Parity multisig vulnerability, I was not making a statement about Ethereum's long-term value. I was saying: there is a fragile point in the chain, and the people who understand it will price it correctly. The same applies to TSMC's production floor. The fragility is not in the demand curve. It is in the supply chain.
Now let's walk through the profit pool distribution. NVIDIA's gross margin is above 70%. TSMC's gross margin is in the mid-50s to low-60s, depending on mix and utilization. Both are at the top of their respective segments, but the gap between them tells you exactly where pricing power lives today. NVIDIA can charge $30,000 plus for a Blackwell GPU because there is no competitive alternative with its software ecosystem and performance. TSMC can charge rising wafer prices because there is no competitive alternative at the leading edge. The 17-point gross margin gap reveals the true cost of trust. The market trusts NVIDIA to keep delivering the best architecture. The market trusts TSMC to keep delivering the only manufacturing process that can realize that architecture. Without the second trust, the first one is worthless.
On the upstream side, NVIDIA has weak bargaining power against TSMC, HBM suppliers, and CoWoS capacity. During HBM shortage periods, NVIDIA effectively queues behind SK Hynix and Samsung allocation. Against its hyper scaler customers, though, NVIDIA holds almost absolute pricing power. TSMC faces a different structure. It depends on ASML for EUV equipment, which is a single-source monopoly, and on Japanese material suppliers for photoresist and silicon wafers. But because TSMC controls the leading-edge manufacturing and packaging capacity that everyone needs, it can negotiate favorable terms even with ASML. The equipment delivery timeline for EUV is already 12 to 18 months. TSMC gets priority on high-NA EUV tools because it is the customer that can actually use them. This is not a symmetrical market. It is a hierarchy of dependency.
Let me put the capex number in perspective. TSMC's 2025 capital expenditure was projected in the $38 billion to $42 billion range. That is roughly 35% to 45% of its revenue, a massive reinvestment level for a company its size. The expansion includes the Arizona complex, approximately $65 billion of total investment across three fabs; the Kumamoto facility in Japan, around $8.6 billion for the first phase, with the second phase already in motion; and continued CoWoS capacity expansion in Taiwan. Arizona Fab 1 is slated for production in 2025, but yield ramping at a new fab with American workers takes four to six quarters. Japan is further along, with first-phase production already running. In Taiwan, advanced process and CoWoS expansion will lag at a faster pace, roughly three to four quarters for a full ramp. The depreciation impact is real. New fabs carry heavy depreciation schedules, traditionally straight-line over five years. That will drag TSMC's overall gross margin by two to four percentage points during the ramp. But the AI premium process mix, and the fact that CoWoS pricing is effectively set by the highest bidder, partially offsets that drag.
ARK has never been a fan of heavy-asset balance sheets. It has spent a decade telling investors that software eats the world and that fixed assets are old-world thinking. So when ARK buys TSMC, the firm is making an implicit admission: in AI, the scarce asset is not code. It is the physical ability to turn code into shipped silicon. The "shovel seller" analogy gets thrown around too casually, but here it is exact. TSMC sells the tools and manufacturing capacity that every AI designer needs. NVIDIA also sells a necessary capability, but it is still one design among many. TSMC's manufacturing capacity is a rental market with no vacancy. When you can raise prices every quarter without losing customers, that is not a cyclical stock anymore. That is a royalty stream.
The demand side of this trade is more robust than the market fears. Look at the split: cloud data center AI training accounts for over 60% of relevant AI compute demand. AI inference is over 20% and growing at a faster rate. Smartphones and PCs represent around 10%, with AI phones and AI PCs pulling demand for TSMC's N3-class process. Automotive and industrial add the rest. NVIDIA's Hopper and Blackwell order books were visible through 2025. Inference demand is expanding even faster than training because every deployed model needs continuous computation. TSMC's AI-related revenue was roughly 10% to 15% of total revenue in 2024 and was on track to exceed 20% by 2025. That is not a niche. That is a structural rotation.
Inventory cycles matter here. The AI semiconductor segment is in a persistent under-supply state. GPU channel inventory is low because demand outruns supply. Traditional semiconductor segments, especially consumer PC and mobile, are closer to balanced inventory. The historical comparison is the 2017-2018 cloud capex expansion, but the AI cycle has higher barriers to entry and a deeper demand pool. The risk is not 2025 destruction. The risk is 2026, when hyperscalers might pause to absorb all the compute they have ordered. That risk exists. But ARK's recent buying suggests the firm believes the absorption phase will be shorter than the market fears. Why? Because the capex race is not a discretionary spending decision. It is an arms race. If Meta decides to cut AI capital spending while Google and Microsoft keep building, Meta will lose the next model-ability cycle. Nobody wants to be the first hyper scaler to blink. That is the hidden bid under the entire AI trade.
Geopolitics is the part that most retail commentary misses because it is slow and bureaucratic. But it changes the risk profile of every long-term position. U.S. export controls have blocked NVIDIA from selling its highest-end GPUs like A100, H100, and B200 to Chinese customers. NVIDIA's China data center revenue dropped from about 20% of the total in 2022 to single digits by 2024 and 2025. The company now sells reduced-capability chips like the H20 to that market, and even those are under pressure. TSMC, by contrast, is not subject to export controls in the same way because it manufactures outside China and has no China-facing high-end process sales. But TSMC is concentrated in Taiwan, and that is the single largest tail risk in the entire global semiconductor supply chain. If cross-strait tensions escalate, the world's AI buildout stops. Not slows. Stops. The value of TSMC's Arizona and Japan fabs is not just financial. It is insurance. But even a fully built Arizona fab cannot replace Taiwan's capacity overnight. ARK's position is a bet that this tail risk does not materialize, but also a bet that the risk is underpriced in TSMC's stock.
China's countermeasures, including export controls on gallium and germanium, matter for niche semiconductor materials but have limited impact on high-volume silicon logic. The bigger effect is acceleration of dual-supply dynamics. Both the U.S. CHIPS Act, with roughly $52.7 billion, and the European Chips Act, with around $43 billion, are trying to create regional fabs. Japan is subsidizing TSMC's Kumamoto facility as part of its semiconductor renaissance plan. China is pouring an estimated 344 billion yuan into its third Big Fund to push domestic equipment and materials. None of this replaces TSMC in the next two to three years. EUV lithography is a hard monopoly controlled by ASML under government oversight, and China has no access to leading-edge EUV. The result is a bifurcating world: one AI ecosystem built on TSMC, NVIDIA, and American cloud giants, and another, slower, more isolated ecosystem inside China. For now, the former gets superior efficiency. The latter gets strategic autonomy. Neither cancels out ARK's thesis.
Here is the contrarian angle that most AI commentary will not give you. ARK's move is not primarily a bet on NVIDIA. It is a bet on TSMC's capacity to tax the AI industry. NVIDIA is the more famous name, but it is also the one with more competitive exposure. AMD is close in hardware. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft is designing custom accelerators. The ecosystem is trying to disintermediate NVIDIA at the silicon level. There is no comparable attempt to disintermediate TSMC at the leading edge. Samsung and Intel are trying, but neither has the yield, packaging, and customer trust combination that TSMC has accumulated. That is the true moat. Yield farming taught me this in 2020: the highest-yielding strategy is usually the one that owns the underlying primitive rather than the one that leverages it. Yield farming is not the only place where APY lies; read the long-term capex line instead. TSMC's capex line is telling you that the company is spending billions to make itself more indispensable. NVIDIA's capex line is tiny because it outsources all the pain. ARK buying both is a way of saying: I want the margin from the design, and I want the margin from the moat.
The second contrarian point is about Meta. Meta's earnings miss spooked people because it called into question the return on AI investment. That is a reasonable short-term concern for Meta's stock. It is a terrible reason to sell TSMC or NVIDIA. The application layer can disappoint while the infrastructure layer compounds. Look at history. The dot-com bust destroyed Pets.com but Cisco and Corning, the shovel sellers, kept building networks. The overbuilding eventually hurt them, but for a multi-year window, the physical infrastructure names outperformed the retail applications. AI is doing the same thing. Meta might not know how to monetize its AI capex. Google and Microsoft are still writing checks. NVIDIA's order book is not tied to one customer's monetization timeline. It is tied to the collective fear of being left behind. That fear is still rising.
There is a darker version of the trade. The BAYC crash wasn't a floor price failure; it was a liquidity illusion that took 48 hours to reconcile on-chain. I have never forgotten that lesson. Liquidity perception changes faster than physical reality, and in 2022 it took exactly 48 hours for the crypto market to realize that Terra's collateral was fake. When you buy TSMC, you are buying real physical liquidity: wafers, machines, power, cleanrooms, and contracts. When you buy NVIDIA, you are buying a thinner layer of real liquidity: tape-outs, software, and brand. If AI funding suddenly pauses, NVIDIA's stock will drop faster than TSMC's because the market will reassess design optionality before it reassesses manufacturing irreplaceability. That asymmetry is the trade. ARK is loading up on the side of the trade that has the least downside in a funding winter.
Now let me walk through the technical roadmap and what I am watching. TSMC's N2 process with GAA nanosheet transistors is scheduled for second-half 2025 production. The industry is watching two things. First, yield learning speed. Every new node transition has a dangerous early period where yield is low and costs are high. TSMC has historically managed this well, but N2 is a structural change from FinFET to GAA, not just another shrink. Second, high-NA EUV adoption. ASML's high-NA tools are expected to arrive in production fabs after 2026. TSMC will get the first units because it is the only customer whose process technology can absorb the capital cost. NVIDIA's Rubin platform, expected in 2026, will ride these advanced nodes. If N2 ramps on time, TSMC extends its lead. If N2 slips, every AI roadmap slips with it. The market is not pricing that risk because it is a binary technical event, and binary events are hard to sell as narratives.
What about NVIDIA's own technology risk? Blackwell's dual-die architecture is a stress test for packaging. By splitting a GPU into two chiplets, NVIDIA reduces the risk of a single large die defect ruining a full wafer, but it increases the need for high-bandwidth interconnects and CoWoS interposer area. The technical challenge is thermal density. Two chiplets plus HBM stacks on one interposer generate enormous heat. Thermal management is not a marketing problem. It is a packaging and materials problem. TSMC's advanced packaging team is the hidden hero here. If TSMC cannot ship enough CoWoS or handle the thermal load, NVIDIA cannot meet its guidance. ARK buying both names is a hedge against this dependency. It is like owning both the airline and the airport. If the airport has too few runways, it does not matter how good the airline's planes are.
The data I want to see next is not NVIDIA's headline revenue. It is TSMC's monthly revenue breakdown and CoWoS pricing changes. TSMC revenue disclosures are the most underrated leading indicator in the AI trade. When TSMC announces monthly revenue, it gives a real-time picture of leading-edge utilization and packaging demand. If TSMC's monthly revenue keeps beating even when hyperscaler earnings wobble, you know the chokepoint is holding. If it starts missing, the physical constraint is loosening and the entire AI pricing structure comes under pressure. Speed without precision is just noise; the difference is execution. ARK has executed this trade with precision, and the market will not understand it until the next software earnings miss forces people to ask, again, why these two names are still rising.
The institutional angle also cannot be ignored. In 2025, I built an arbitrage framework around spot Bitcoin ETFs, mapping the latency difference between TradFi custody settlement and decentralized liquidity pools. The framework taught me a simple lesson: when institutions buy physical infrastructure, they are not participating in the story. They are participating in the settlement. ARK's buying of TSMC is not about believing the AI story. It is about believing that AI compute will settle into a physical grid, and TSMC owns the grid. NVIDIA owns the most valuable tenant on that grid, but TSMC owns the land, the power lines, and the parking lot. In a bull market, the tenant gets the headlines. In a correction, the landlord gets the cash flow.
The final piece is the regulatory shift. Compliance and export control frameworks are no longer a side issue in this trade. They are the price of admission. NVIDIA is learning to operate inside a rulebook that changes every quarter. TSMC is being forced into multi-geography manufacturing to satisfy U.S. government demands. This raises costs for everyone. But it also raises barriers to entry. A startup cannot suddenly build a fab in Arizona. A Chinese vendor cannot suddenly build an EUV-based AI chip. Regulation is a moat-building machine, and the incumbents are on the right side of the machine. Trust no one. Audit everything. Repeat. That is how I think about this market now. The balance sheet is the smart contract, and the export license is the oracle. ARK is reading the contract carefully.
Let me be clear about what this article is not. It is not a prediction that NVIDIA or TSMC will always go up. NVIDIA's valuation already reflects years of perfect execution. If AI model training demand saturates earlier than expected, NVIDIA could suffer a painful de-rating. TSMC's exposure to multiple customers provides some cushion, but even TSMC is not immune to a synchronized global AI capex slowdown. The point is that ARK's simultaneous buying is a structural signal, not a cyclical one. It says that the management team behind one of the most famous innovation funds believes the AI era will be governed by physical scarcity, not by software dreams. The investment implications are straightforward: watch TSMC's capacity announcements more carefully than NVIDIA's developer conferences; watch CoWoS pricing more carefully than GPU benchmark scores; watch export control policy more carefully than AI chatbot app rankings.
A final thought on the often forgotten side of this trade. The people who will actually be hurt by the AI chokepoint are not the hyperscalers and not the chip designers. They are the companies two layers down, the ones buying AI inference from NVIDIA at premium prices and reselling it to small businesses. Those companies are leverage on someone else's bottleneck. They have no pricing power, no manufacturing exclusivity, and no ability to control their own unit economics. ARK is not buying those companies. ARK is buying the two entities that collect the toll every time a model is trained or served. That is the cleanest, most brutal expression of the chokepoint trade I have seen in this cycle.
So the next time a headline says ARK is buying AI, read it more carefully. ARK is buying the machine, not the dream. The dream is volatile. The machine is slow, expensive, and increasingly scarce. In 2017, I learned that trust is the most expensive item on a blockchain. In 2020, I learned that yield curves hide real operational leverage. In 2025, I learned that institutional flows into physical infrastructure tell you more than a hundred tweets from AI influencers. The question now is not whether ARK is right about AI. The question is whether the rest of the market will recognize, before the next earnings season, that the only thing real in this entire cycle is the silicon that actually ships.