The announcement contained no yield data. No packaging roadmap. No EUV delivery schedule. No transistor architecture disclosure. No technology roadmap beyond a single node. What it contained was one number: "over 1 trillion watts" of annual AI compute capacity.
I parsed that number. It is not an engineering metric. If interpreted as power draw, one terawatt equals the sustained output of ten large nuclear reactors โ physically impossible to house in one building. If interpreted as compute, it implies roughly 500,000 NVIDIA B200-class accelerators drawing 20 kilowatts apiece. Either reading renders the figure rhetorical, not operational. In my audits โ ICO token distribution algorithms in 2017, DeFi yield aggregators in 2020, NFT royalty enforcement schemes in 2021 โ the first red flag is always the same. A large round number where a technical specification belongs.
The second red flag is the timeline. Terafab targets 2nm chips, with first shipments "after 2028." TSMC N2 enters mass production in 2025. Samsung SF2 enters mass production in 2025. Intel 18A โ the equivalent node โ enters mass production in 2025. When you announce a node three years after your competitors have shipped it, you are not entering a race. You are announcing an ambition and calling it a schedule.
The Context: A Fabless Pioneer Declares Itself an IDM
Based on the reporting, the Terafab project is a joint Tesla-SpaceX venture in Texas, spanning nine million square meters of planned floor space. Phase 1 carries a $16.8 billion price tag. The full build-out vision reaches $119 billion. Musk's stated justification: existing global chip manufacturing capacity covers only about 2 percent of his ecosystem's future needs.
This is the moment a fabless system integrator announces it will become an IDM โ integrated device manufacturer. The transition is not incremental. It is a structural jump from designing chips to manufacturing them at the most advanced logic node in commercial existence. The industry has no precedent. No company has independently erected a 2nm-capable foundry from a zero installed base.
The timing deserves attention. We are in a bull market for AI compute. The demand narrative is loud, capital is cheap, and every major player is announcing capacity. That is precisely when technical due diligence goes quiet. It happened with ICOs in 2017. It happened with liquidity mining in 2020. It is happening now with fabs.
The reporting's own confidence scores reflect evidence quality. Technical process analysis: 3/10. Industry chain: 5/10. Capex: 5/10. High confidence appears only where the data is negative. The "1 trillion watts" metric is flagged as rhetorical at 9/10 confidence. The "2nm" label as target rather than capability sits at 7/10. The baseline is established: the vision is loud. The technical infrastructure is silent.
What follows is a systematic teardown. I am not evaluating whether Musk believes the vision. I am evaluating whether the plan survives contact with semiconductor economics.
The Node Gap Is a Manufacturing Experience Gap
The headline gap is three years and roughly one node. Using the upper-bound estimate โ 2028 first shipments โ Terafab trails TSMC and Samsung's 2nm entry by three years. Slip to 2030 and the gap stretches to five years. The temporal gap is not the binding constraint. The experiential gap is.
2nm nodes require gate-all-around (GAA) transistor architecture. TSMC deploys GAA Nanosheet. Samsung deploys GAA MBCFET. Both spent the better part of a decade converting architecture into manufacturable process recipes. Tesla and SpaceX have no public GAA manufacturing experience. No disclosed process integration team. No yield engineering pipeline. No cleanroom construction history. The reporting's 3/10 confidence on the GAA assumption is an inference from industry convention, not from a Terafab technical disclosure.
The missing roadmap is the sharper signal. Every credible foundry publishes a two-generation forward plan before its current node reaches volume. TSMC has N2 and beyond. Samsung has SF2 and SF1.4. Intel has 18A and 14A. Terafab's reporting reveals no 1.4nm or 1nm path. A line without a next-generation plan is not a line. It is a demonstration project.
The most probable reading โ 7/10 confidence โ is that "2nm" is a strategic objective, not a process capability. Its function is to attract policy support, capital, and engineering talent. It is a label with a purpose: maximized signaling value at the announcement stage. In crypto terms, it is a token with no mainnet.
Yield Is a Learning Curve, Not a Construction Budget
No yield data was disclosed. The industry benchmark: TSMC's N3 ramp consumed six to nine months to reach profitable yield โ above 80 percent. New greenfield fabs generally require two to four years. A new entrant with zero installed fab base faces two to three years to reach even 70 percent yield on a transferred process, assuming complete technical transfer and unconstrained equipment access. That is assumption-heavy arithmetic.
I have watched this dynamic in adjacent territory. Cryptocurrency mining ASIC manufacturers announce hashrate targets; the delivered hashrate is determined by manufacturing yield, not by the press release. The announcement is not the product. The cumulative learning curve is the product.
Terafab's realistic trajectory from groundbreaking to economically viable production runs five to seven years. The "2028 first shipments" framing is consistent with a pilot wafer, not a scaled line. Nothing in the reporting contradicts this. Nothing in the reporting addresses it.
EUV Allocation Is the True Gate
Consider the lithography math. A 50,000-wafer-per-month 2nm facility requires 15 to 25 EUV scanners. Standard EUV โ ASML NXE:3800E โ lists near $180 million per unit. High-NA EUV โ EXE:5000 โ runs $350 million to $400 million. Lithography tooling alone for a meaningful 2nm line costs $5 billion to $10 billion. That is the capital barrier. It is not the binding barrier.
ASML builds roughly 60 to 70 EUV machines per year. TSMC takes 20 to 25. Samsung takes 15 to 20. Intel takes 10 to 15. That leaves fewer than ten machines per year for every other entity on the planet that wants one. Terafab, a new foundry customer with no allocation history, enters the queue behind incumbents who hold multi-year purchase agreements. Standard EUV delivery runs 12 to 18 months; high-NA runs 18 to 24 months. If orders have not been placed, equipment arrival lands in 2027 at the earliest, more plausibly 2028.
There is no alternative supplier. Nikon and Canon cannot produce EUV. The technology originated in the United States through the EUV LLC consortium, but commercial manufacturing is a Dutch monopoly.
Material dependencies compound the equipment problem. EUV photoresist: 100 percent Japanese. Twelve-inch silicon wafers: 90 percent-plus Japanese โ Shin-Etsu and SUMCO. Specialty gases: split between American and German suppliers. The EDA layer: fully controlled by Synopsys, Cadence, and Siemens, all accessible to a US entity. The allied supply chain is functional. The point is not that Terafab will be cut off. The point is that Terafab joins a queue for every critical input โ equipment, materials, EDA seats โ and the queue is already long.
The CHIPS Act adds a compounding constraint. Its subsidy provisions attach equipment-arrival deadlines. Delayed tooling can trigger clawback of disbursed funds. A project simultaneously at the back of ASML's allocation queue and subject to federal delivery deadlines sits between two contradictory obligations. The reporting flags this as a real risk. It is not a tail risk. It is the base case.
The Omitted Paragraph: Advanced Packaging
The reporting does not mention advanced packaging. That omission is the functional tell. For AI and HPC silicon, packaging is co-equal with the process node. NVIDIA, AMD, and Intel all route through TSMC's CoWoS; monthly capacity exceeded 60,000 wafers in 2025 and remains insufficient. Tesla's Dojo program uses TSMC's InFO_SoW wafer-level integration โ a related but distinct path. Silicon interposers, through-silicon vias, micro-bump technology: these are separate industrial capabilities with separate learning curves and separate supply chains.
A shippable AI chip requires both a manufactured die and an advanced package. If Terafab produces 2nm dies but cannot package them, the output is unsellable silicon. The reporting's silence implies continued dependence on incumbent packaging providers โ including TSMC โ for the back end. That directly contradicts the stated rationale for the project: reducing dependency on external foundry capacity. The project, as disclosed, does not solve the problem it claims to solve.
The Depreciation Trap
The capex structure cannot sustain the narrative. The full-vision figure of $119 billion equals 123 percent of Tesla's 2024 annual revenue โ $96.9 billion. Tesla's 2024 net income was roughly $7.1 billion. SpaceX carries a $350 billion to $400 billion valuation but publishes no transparent financials. TSMC's annual capex is approximately $30 billion, funded by more than $40 billion in free cash flow. Terafab's full vision represents three to four years of TSMC's total annual capital spending, executed by an entity without TSMC's cash generation.
The depreciation arithmetic is unforgiving. Semiconductor equipment depreciates over five to seven years under US GAAP. On a $119 billion asset base, annual depreciation runs $17 billion to $24 billion. Assume the completed facility generates $30 billion to $50 billion in annual revenue โ the plausible envelope for captive AI silicon consumption. Depreciation alone consumes 34 to 80 percent of revenue. Gross margin in the depreciation window trends toward zero or negative. TSMC operates at 25 to 30 percent depreciation-to-revenue and still needs scale and pricing power to make the equation work. Terafab would run the same equation without the scale, without the customer base, without the cost base.
The reporting's own numbers expose the narrative gap. A theoretical run of 100,000 AI chips per year at an average $25,000 to $30,000 per unit produces $250 billion to $300 billion in top-line revenue. But that requires 70 percent-plus yield, 20,000-plus wafers per month, a resolved packaging pipeline, and zero equipment slippage. The distance between the "1 trillion watt" headline and the $30 billion to $50 billion realistic revenue envelope is not a rounding error. It is the difference between a strategic vision and a financially operable facility.
The 7x Signal
Phase 1 allocates $16.8 billion. The full vision is $119 billion. That is a 7x delta between announcement and completed build. Berlin Gigafactory required about three years from announcement to production and absorbed roughly $7 billion. Terafab's initial capital plausibly covers one building and a fragment of one production line. More than 85 percent of the stated capital is contingent on future funding events. The most probable bridge: a SpaceX public listing that converts a $350 billion to $400 billion private valuation into liquid capital.
This is the SpaceX playbook. Sell the long-dated vision. Raise incrementally. Deliver the early milestones. Refinance on the back of the narrative. It worked for launch vehicles. The open question is whether the model transplants into semiconductor manufacturing โ a yield-driven, capital-saturated, commoditized industry with no tolerance for schedule narratives. Ledger balances do not lie; they only wait. The ledger for a $119 billion fab is not waiting. It is accumulating interest.
Market Demand: Real Appetite, Wrong Pricing Power
The demand side deserves scrutiny. The reporting estimates the application split: HPC/AI training at roughly 60 percent of Terafab's projected output, AI inference at 20 percent, automotive FSD at 15 percent, aerospace at 5 percent. These are analyst constructs, not company disclosures. But they point to a structural mismatch.
The market for AI training silicon is dominated by NVIDIA at 80 percent-plus share. Dojo was designed to break that dependence. It has not. The ecosystem is skeletal; the software stack does not compete with CUDA. The 2nm process advantage โ roughly 30 to 40 percent power reduction or performance gain versus 5nm โ matters only if the silicon can actually ship into workloads. Tesla's FSD real-time inference demand is real but price-sensitive. Optimus edge AI is real but early. Starlink satellite chips are niche. The internal order book is real; the sustainable pricing power is not.
Consider the cycle. The semiconductor industry runs a three-to-four-year inventory cycle. The reporting places the current position in late restocking or early destocking territory. AI chips remain supply-constrained through 2026-2027, but that constraint is driven by CoWoS packaging and HBM capacity, not wafer starts. Terafab will reach volume in a different cycle phase. A facility that comes online in 2030-2032 faces a mature AI chip market, established competitors with amortized cost bases, and pricing pressure from the capacity overhang that new entrants typically misunderstand. The global AI chip market is projected to grow from about $60 billion in 2024 to $300-400 billion by 2030 โ a 35 percent CAGR. That growth is real. It is also why incumbents are already expanding. Terafab's differentiation is not technology; it is timing. And the timing is late.
The Hidden Variables
Three variables could change the teardown's conclusion. First, acquisition. If Musk acquires a functioning semiconductor manufacturing team โ poaching core process engineers from Samsung, Intel, or GlobalFoundries, or acquiring a second-tier fab outright โ the technology base becomes real. The reporting scores this possibility at 6/10 confidence but discloses no acquisition plans.
Second, partnership. Samsung Semiconductor or Intel Foundry would be natural collaborators. Samsung already supplies Tesla's storage silicon. Intel has US manufacturing experience and federal goodwill. A technical partnership would materially raise the project's credibility. Neither is disclosed.
Third, supply chain geopolitics. As a US entity, Terafab faces no export-control prohibitions on EUV procurement. China's gallium and germanium export controls have limited direct impact on 2nm silicon logic. Yet any escalation in the broader technology war lengthens every lead time in the chain, and China's aggressive equipment bidding tightens global supply further.
The counterfactuals matter because the announcement is silent on all of them. The reporting identifies the internal-demand thesis as the project's structural advantage โ 100 percent captive demand from Tesla, SpaceX, and xAI eliminates customer acquisition risk. What the announcement does not address is whether captive demand justifies captive supply when the depreciation schedule alone would strain a publicly traded balance sheet.
The Contrarian Angle: What the Bulls Get Right
Now the counter-intuitive side. The bulls are not entirely wrong. The internal demand thesis is structurally sound. Tesla's fleet โ roughly seven million vehicles by 2025 โ needs FSD silicon. Optimus, at scale, requires edge AI inference. Starlink needs satellite chips. xAI's Colossus cluster demonstrates an insatiable appetite for training compute. Captive demand means zero customer acquisition risk. That advantage is real.
Tesla has also shipped working silicon. FSD SoCs and the Dojo D1 โ a 7nm part โ prove the design-side competence exists. The reporting places design capability one to two generations behind NVIDIA's Blackwell line, but that is a solvable gap. Design v1.0 is the hard part; v2.0 is incremental refinement.
The macro argument has merit. TSMC's Arizona complex โ Fab 21 at 5/4nm, Fabs 52/53 at 2nm/3nm โ reaches roughly 100,000 wafers per month by 2028. That is insufficient for the stated AI demand curve. The industry genuinely faces a capacity wall. If AI compute demand doubles every two years, the incumbents' expansion plans do not close the gap. Musk's "2 percent" claim is unverifiable, but the direction of the underlying scarcity is real.
Musk's execution record also deserves weight. SpaceX landed boosters when the industry said it was impossible. The pattern โ ignored expert consensus, relentless vertical integration, story-driven financing โ has produced real infrastructure before. Terafab is the same pattern applied to semiconductors.
The project's success, then, hinges not on whether the vision is plausible โ it is โ but on whether the execution timeline can survive contact with ASML's allocation queue and the depreciation ledger. Trust the delivery schedule, not the headline. Hype evaporates; receipts remain. There are no receipts yet.
Takeaway: Price It as a 2032 Facility
Terafab, as disclosed, is not a fab-on-schedule plan. It is a story-driven financing vehicle with a technical appendix. The 2028 first-shipment target will slip. The yield curve will consume years. The packaging question is unanswered. None of this means the project fails โ acquisition, partnership, or SpaceX proceeds could alter the trajectory. It means the project should be priced as a 2030-2032 possibility, not a 2028 commitment.
The variable to watch is ASML's allocation queue. Equipment deliveries are the undeniable proof point. Follow the wafer starts, not the press release. Verify against the ledger. Volatility is not risk; opacity is. And the opacity around Terafab is considerable.
