The story crossed my desk from a crypto outlet, which was the first anomaly. A project called Terafab is reportedly recruiting Korean suppliers to build a self-sufficient semiconductor supply chain from scratch. That is the entire claim. No process node. No capital figure. No capacity target. No timeline. No customer. No technology roadmap. No financials.
I have spent enough years reading announcements to know that the shape of what is missing usually matters more than the shape of what is claimed. In 2017, as a nineteen-year-old cybersecurity undergraduate in Tallinn, I spent eight weeks cross-referencing Ethereum transaction hashes from the Parity wallet breach against ICO whitepapers. Four thousand transactions. Three distinct layers of funneling, investor capital diverted to private wallets instead of project treasuries. The whitepapers described utility. The ledger described something else. Following the money, always — and when there is no money to follow, follow the silence.
Here is what can actually be verified: a report says Terafab is courting Korean suppliers. That is it. Public semiconductor registries contain no known, scaled wafer fabrication entity operating under that name. Three candidate explanations exist, none confirmed. First, a systems or terminal vendor's mega-fab — the industry has occasionally used the shorthand tera-fab for trillion-scale capital expenditure. Second, an emerging foundry startup or specialty process project, though nothing supports the capacity, node, or funding such a project would require. Third, a compute or crypto-infrastructure capital extension, listed only because the source was Crypto Briefing — and the body text contains no Web3 content at all.
Because the entity cannot be identified, the only honest method is scenario-based analysis: high confidence on industry regularities, low confidence on anything specific to Terafab. That is not analytical caution. It is a description of the information base.
The regularities are well established. An advanced-logic giga-fab at 3nm and 100,000 wafers per month costs $20–30 billion and takes four to six years from decision to volume production. A mature-node plant at 28nm and 40,000–50,000 wafers per month costs $3–6 billion and takes three to four years. Advanced packaging capacity in the CoWoS class runs $1–3 billion over two to three years. Specialty processes — SiC, GaN, power, analog — run $500 million to $1.5 billion over two to three years. Equipment depreciates on a five-to-seven-year straight line; buildings over twenty. Those numbers are the skeleton every real fab announcement eventually has to show.
Terafab shows none of them.
Start with the only technically legible clue: the word Korean.
Korea's semiconductor ecosystem is formidable, but its comparative advantage is specific. Memory — Samsung and SK Hynix. Packaging substrates, FC-BGA in particular — Samsung Electro-Mechanics, Daeduck. Back-end packaging equipment, thermocompression bonders — Hanmi Semiconductor. CMP and deposition tooling. Electronic materials — Dongjin Semichem, Soulbrain. Glass substrates for advanced packaging — SKC and Absolics. What Korea does not offer at scale is a leading-edge logic front-end supply chain; that domain belongs almost entirely to Samsung, whose own supply chain leans heavily on American, Japanese, and Dutch vendors.
If Terafab genuinely intended advanced logic, the correct courtship would run through Applied Materials, Lam, and KLA in the United States; Tokyo Electron, SCREEN, Shin-Etsu, and JSR in Japan; ASML in the Netherlands. Recruiting Korean suppliers points somewhere else: memory, advanced packaging, back-end, materials. On-chain evidence > Hype applies to silicon as much as to wallets — the supplier list is the evidence, and it says this is probably not an advanced logic fab.
The second tell is the phrase self-sufficient supply chain. In semiconductors, that is a fiction. No company and no country has achieved it. Samsung operates an IDM, a foundry, a memory business, an equipment subsidiary in SEMES, and a materials footprint — and still depends on ASML, Applied Materials, and Shin-Etsu for things it cannot make itself. When an announcement promises self-sufficiency, the accurate translation is usually partial vertical integration plus supplier diversification. The phrase is financing language or policy language. It is not an engineering target.
The third constraint is physical. Any node below roughly 7nm requires EUV. ASML ships on the order of 50–60 EUV systems a year, and that output is locked by TSMC, Samsung, Intel, SK Hynix, and Micron well into 2027 and beyond. Export requires Dutch government licensing. A new entrant has essentially no path to allocation before 2027, and if it sits outside the allied club, no path at all. Advanced immersion DUV — NXT:2000i and above — is licensed too. This is not a scheduling inconvenience. It is a locked door.
The fourth constraint is arithmetic. Take a $10 billion fab. Annual depreciation lands around $1.5–2 billion. If that plant generates $3 billion of revenue at ramp, depreciation consumes 50–65% of it. Gross margin sits at negative to single digits until utilization passes 80%. EBITDA breakeven typically arrives in year three or four; net income turns positive after year five. That is the observed curve at SMIC and at GlobalFoundries. A newcomer with no yield history, no pricing power, and no scale should expect worse.
The fifth constraint is procurement. A greenfield plant's dependence on lithography, etch, deposition, and metrology tools, plus photoresists, large silicon wafers, specialty gases, and CMP slurries, is close to total. A new buyer has no volume, no long-term agreement, and no priority allocation — bargaining power near zero, sometimes negative, since jumping the queue costs a premium. The design side is no softer: three EDA vendors hold more than 80% of the toolchain. Upstream, the newcomer is a price taker in every direction.
The sixth constraint is the market itself. Between 2024 and 2026, mature-node capacity — 28nm and above — is being expanded aggressively by SMIC, HuaHong, and Nexchip. The result is structural oversupply and a price war. Advanced nodes are moving the other way; TSMC raised prices again in 2025. A greenfield fab entering mature nodes opens in a segment that is already oversupplied and dominated by low-cost incumbents. Building a giga-fab to chase commodity work is the semiconductor version of hauling cargo in a Rolls-Royce: it insults the machine and it does not carry much.
The seventh constraint is the one that would kill the project outright in a diligence review. There is no customer. Fabs are heavy-asset B2B businesses; without a take-or-pay commitment, financing does not close. Build first, find customers later has never worked — Wuhan Hongxin is the cautionary case, and GlobalFoundries' early Abu Dhabi expansion is another. The paths that do work share three ingredients: an anchor customer, government backing, and a technology licensor. Rapidus has IBM's 2nm process, Toyota and Sony as customers, and trillions of yen in Japanese subsidy. Terafab, as described, has none of the three. The absence of any mentioned subsidy is itself a negative signal, not a neutral omission.
The eighth constraint is geopolitical, and it sits inside the project's own core narrative. Korean suppliers have compliance departments. Their first questions will be about end user, end use, and whether controlled technology is involved. If the entity behind Terafab carries Chinese capital, Korean suppliers will very likely self-exclude under American pressure, and the recruit Korean suppliers story collapses on its own logic. The Foreign Direct Product Rule reaches widely; Japan has controlled 23 categories of semiconductor equipment since July 2023. China's countermeasures — gallium and germanium in 2023, antimony and graphite in 2024, rare-earth items in 2025 — raise costs and uncertainty but do not constitute a supply cutoff. Even China's Big Fund III, at RMB 344 billion, is a reminder of how slow catching up actually is: an economy with a whole-of-nation approach, trillion-level funding, and more than a decade of effort still imports most of what it needs.
Finally, the timeline. A greenfield fab takes three to five years even when everything goes right: eighteen to twenty-four months of construction, twelve to twenty-four months from tool move-in to volume production, another twelve to twenty-four months of yield ramp. A project starting today benchmarks against the 2028–2030 frontier, when TSMC is expected to be at A14 or A10. The gap does not narrow. It widens. And yield is not an equipment problem — it is process integration and defect-density engineering, built on thousands of process engineers and a deep pipeline of wafer experiment data. TSMC's N3 family needed twelve to eighteen months to reach acceptable yield. A newcomer has no historical data pipeline at all.
Now the part I have to police in myself: correlation is not causation, and the venue of a story is not the story.
Because the report surfaced on a crypto outlet, the reflex is to read Terafab as a crypto-compute play, a sovereign mining or AI-infrastructure bet dressed in fab language. The body text does not support it. Over-attribution is the same error I made during DeFi Summer in 2020, when I scripted impermanent loss across 150 Uniswap V2 positions for six months and found that 68% of retail LPs finished negative despite advertised APYs. The yield number was real. My inference from it was wrong. A headline number and a conclusion are different objects. I made a version of the same mistake in 2025, mapping 50,000 wallet interactions behind BlackRock's ETF flows into Ethereum Layer 2s and finding 40% routed through privacy-preserving mixers for compliance reasons — the clean narrative of transparent institutional adoption was simply not what the chain showed.
The more interesting reading is about the venue, not the subject. Crypto media now covers semiconductor supply chains because compute has become the reserve asset of this cycle. The fact that a fab rumor clears a crypto newsroom is itself a datapoint about where capital attention sits in a bear market.
And the deeper pattern is a narrative template I have seen before. Self-sufficiency functions exactly the way RWA tokenization has functioned for three years: a promise of vertical integration that omits the counterparty. Institutions never needed a public chain; they needed settlement finality and compliance rails. A fab does not need self-sufficiency; it needs an anchor customer and a yield curve. I have watched rollup teams budget blob space as though it were infinite — the accounting holds right up until it doesn't, and then fees reprice. Capacity planning without demand data is not a strategy. It is a mood.
The next datapoint worth watching is not another memorandum of understanding. Supplier intent produces no wafers. Watch for the first purchase order, the first capex figure with a node attached to it, the first take-or-pay contract, the first named customer with a qualification timeline. Any one of those would change the confidence level of this entire analysis.
Until then, Terafab belongs in the concept-stage drawer, next to every other announcement that led with ambition and omitted arithmetic. The ledger remembers everything — including what an announcement leaves out. Silence is suspicious.