Citi raised its price target on TSMC to NT$4,000. The market read this as a signal. It was a number, not a machine.
A number can be revised in an afternoon. A fab takes eighteen months to arrive. That distinction matters more than usual right now, because the same Citi note that lifted Taiwan's foundry giant also lifted a far noisier complex: the crypto tokens selling "decentralized AI compute" as a commodity. Render, Akash, io.net, and a dozen thinner imitators repriced on a shared narrative — that AI demand is effectively infinite, and that the supply to meet it is elastic.
The supply is not elastic. It is gated by a single chokepoint: TSMC's CoWoS advanced packaging lines, which sit downstream of a near-monopoly on the most advanced logic nodes on earth. The code spoke, but the logic was a lie. You cannot tokenize your way around a lithography machine.
What follows is a teardown of what the Citi note actually claims, and where the claim stops being a forecast and starts being a faith.
TSMC manufactures roughly 90% of the world's most advanced semiconductors. That is not a marketing figure. It is a structural fact. Apple, Nvidia, AMD, Broadcom, Qualcomm — the top five customers account for 60 to 70% of revenue, and each depends on leading-edge nodes for which there is no credible alternative. Samsung's SF3 yields have crawled. Intel's 18A is still proving itself. The gap is not closing. In the most advanced nodes it runs roughly six to twelve months against Samsung and twelve to eighteen against Intel.
The Citi thesis, as transcribed by Jinshi, is straightforward. Maintain Buy. Raise the target from NT$3,800 to NT$4,000. Forecast revenue growth above 40% through 2027. Forecast capital expenditure of $81 billion in 2027 and $90 billion in 2028.
Read the capex numbers again. In 2024, TSMC spent roughly $29.8 billion. The 2025 guidance is $38 to $42 billion. Citi asks you to believe that within two years the company will spend more than double its 2024 figure — exceeding Samsung's peak annual capex of roughly $40 billion and Intel's peak of roughly $25 billion. It would be the largest single-company annual capital expenditure in semiconductor history.
It is worth stating plainly that the Citi note is a single sell-side source, transcribed second-hand, and not cross-verified against TSMC's own filings. Sell-side targets reflect a base case. Tail risks are excluded by construction. That is not a criticism of Citi. It is a description of the genre. The same genre has been absorbed by crypto's AI complex without friction. The tokens trade on the headline. The headline is a projection.
Start with first principles. Capital expenditure leads demand by two to three years. A fab ordered today produces wafers in 2028 or 2029. So when Citi forecasts $81 billion for 2027, it is not forecasting 2027 demand. It is underwriting 2029 and 2030 demand. The capex number is a promissory note on a decade that has not been written.
That matters because depreciation arrives on schedule regardless of whether the demand does. TSMC depreciates on a five-to-seven-year straight line. An $81 to $90 billion annual capex program translates into roughly $12 to $18 billion of incremental annual depreciation. To absorb that without compressing gross margin, the company must hold utilization above 85% and defend its advanced-node pricing premium. If AI demand disappoints in 2027, the same depreciation that reads as a growth signal today becomes a margin anchor. TSMC has run this play before. Never at this scale.
Now the physical bottleneck. CoWoS is the hardest constraint in the AI supply chain, and it is not a chip — it is a packaging process. An Nvidia Blackwell GPU or a Broadcom ASIC is not a single die. It is a set of dies bonded onto silicon interposers through CoWoS. You can print all the logic dies you want; without CoWoS capacity, they never become accelerators. TSMC has been doubling CoWoS capacity and it is still short. That shortage is why AI GPU lead times stretched through 2024 and 2025.
This is where the crypto compute narrative fractures. Networks like Render and Akash aggregate GPU supply. But the GPUs they aggregate are downstream of the same CoWoS chokepoint. A decentralized compute market is a marketplace, not a fab. It can redistribute scarcity. It cannot manufacture scarcity away. When a token's pitch is "we make AI compute abundant," and the abundance is gated by a packaging line in Taiwan, the pitch is not a supply solution. It is a demand aggregator with a token attached.
The customer mix tells the same story in a different register. HPC and AI are roughly 53% of revenue and growing fast. Smartphones are roughly 33% and growing slowly. IoT and automotive are smaller still. Advanced nodes run full; mature nodes face a wave of Chinese capacity from SMIC and Hua Hong that pressures pricing. The result is a K-shaped foundry market: the leading edge reprices upward while the trailing edge deflates. TSMC is levered to the top arm of the K. It is also, by extension, exposed to the assumption that the top arm keeps climbing.
There is a second-order story the note gestures at without developing: CPO, co-packaged optics. Citi lists it as one of three growth drivers alongside AI compute and AI entity growth. The logic chain is worth spelling out. As AI clusters scale, compute stops being the sole bottleneck. Interconnect bandwidth and power become the constraint. Copper hits a wall around 800G and 1.6T. The answer is optical — TSMC's COUPE platform, which co-packages a photonic engine with the switch ASIC. This shifts value from the logic die to the packaging layer. Advanced packaging stops being a support function and becomes a revenue engine in its own right.
I have audited infrastructure that makes this kind of claim. In 2025, I spent roughly 150 hours simulating attack vectors against an AI-agent wallet protocol and found its oracle validation carried no cryptographic signatures — a feed any sufficiently clever agent could bend. The lesson generalizes. When a system's value migrates into a layer nobody is auditing, the vulnerability migrates with it. CPO is not a vulnerability. But it is a value migration into a layer — optical-electrical co-design — that most analysts, and virtually every token whitepaper, do not model.
Now the growth math. Consider the arithmetic. TSMC's 2024 revenue was roughly $90 billion. A 40% CAGR to 2027 implies a top line near $247 billion within three years. The entire global semiconductor market is projected to grow at 10 to 12% CAGR under optimistic assumptions. TSMC would be capturing a wildly disproportionate share of that expansion — possible only if it takes share from Samsung and Intel while simultaneously raising prices. Both can happen. Assuming both happen at once, for three consecutive years, is a modeling choice, not a law.
A 40%+ revenue CAGR on a company of this size is, historically, almost unheard of. TSMC's best years ran 25 to 30%. To reach 40% sustained, you need volume growth, share gains, and price increases stacked simultaneously. Volume alone cannot get there. The note implicitly requires AI revenue to nearly double year over year, CoWoS and CPO to add high-value content per wafer, and pricing to hold. That is not a base case. It is the optimistic tail wearing a base case's clothing.
In 2024, I spent roughly 200 hours comparing the custody architectures in the spot Bitcoin ETF filings against Ethereum's decentralized node infrastructure. The finding was unremarkable and, for that reason, instructive: control concentrated in a handful of custodians. Institutional adoption does not distribute trust; it relocates it. The same dynamic governs the AI trade. The ETF wrapper made Bitcoin legible to Wall Street by making it centralized. The AI compute trade makes TSMC legible to the same capital by concentrating an entire technological epoch into one company on one island.
Here is what the bulls get right, and it is not nothing.
The demand is real. I have spent enough time in protocol code to distrust narrative, but AI training and inference are not narrative. They are compute consumption with a measurable footprint. Inference demand is now outpacing training demand, and inference is the more durable of the two — it scales with usage, not with a race to the frontier. If Citi's growth forecast is wrong, it is more likely wrong on timing than on direction. The compute is coming. The question is whether it arrives in 2027 or 2029.
The bulls are also right that TSMC is the purest way to own the theme. It is the shovel seller. It does not carry the business-model risk of the companies buying its wafers. If AI monetization disappoints at the application layer, TSMC still gets paid at the fab gate. That asymmetry is genuine.
And the crypto bulls are right about one thing the traditional note ignores entirely: demand for compute is fragmented, and fragmentation creates market structure. A hyperscaler will not rent from a competitor. A small research lab cannot get an allocation. That gap is where decentralized compute has a real, if narrow, claim. It is not a substitute for TSMC. It is a secondary market for what TSMC's customers cannot absorb.
So the fault line is not TSMC's capex forecast. It is the assumption that the physical and the tokenized are the same trade. They are not. One is a bet on lithography, packaging, and depreciation schedules. The other is a bet on a narrative that borrows the first one's credibility without inheriting its constraints. Trust is a variable you cannot hardcode. Data does not lie, but it does not care. The note will be revised. The fab will not.