The arithmetic fails on first pass. Terafab — the SpaceX, Tesla, and Intel joint venture in Grimes County, Texas — announces a 100-million-square-foot facility, a $16.8 billion initial tranche, and a theoretical $119 billion total envelope. Coverage translates one terawatt of AI compute per year into 100 billion to 200 billion custom chips. That figure demands roughly 100,000 12-inch wafer starts per month. No single fabrication plant on Earth runs at that volume. The number is not a forecast. It is a structural artifact of marketing. In my practice, a headline metric that fails first-order arithmetic is a liability line until proven otherwise. This project warrants the same scrutiny.
The entity launched under a JETI agreement in August 2026, with $10 million in non-refundable funds already exchanged. Its mandate is vertically integrated semiconductor production: logic, memory, packaging, and test. Intel's role as process partner points to Intel 18A — a 1.8nm-class GAA node built on RibbonFET. Texas added a $30 million Enterprise Fund grant. Gibbons Creek Reservoir supplies the water.
This is not a merchant foundry. It is a captive supplier for three consumers: Tesla's Optimus and Cybercab fleets, and SpaceX's Starmind orbital network. The distinction determines the evaluation method. A merchant foundry survives on external pricing power. A captive fab survives on utilization and depreciation outcomes. Ledger integrity precedes market sentiment — and the ledger begins with physical capacity, not conference-room narrative.
The Node Gap Does Not Close on Paper
Intel 18A carries nominal parity with TSMC's 2nm-class production. Nominal is not operational. TSMC's N3 held mature yields of 80–90% by 2025. Intel 18A, in early ramp, is estimated at 50–70% depending on the wafer stack. A new facility on a node in ramp needs two to three years from equipment move-in to yield stability. Assume 2027 move-in, 2028 trial, 2029–2030 volume. Terafab sits one to two years behind the credible frontier while wearing the same generational label.
Yield is not aesthetic. Sub-70% yield collapses the wafer-per-month math and inflates unit cost. Every failed wafer carries the same depreciation weight as a passing wafer. The captive customer absorbs it, but it does not exit the system.
The scope also includes memory. If that means HBM-class stacks, the distance to the front is larger than logic: SK hynix, Samsung, and Micron hold a decade of stacked-memory integration experience. If it means commodity DRAM, Terafab competes on margins it cannot defend.
Depreciation Never Defaults
The initial $16.8 billion, depreciated straight-line over five years, generates $3.4 billion in annual depreciation before the first wafer moves. The theoretical $119 billion envelope implies annual depreciation above $20 billion. TSMC runs 15–25% depreciation-to-revenue. Terafab, at plausible early revenue, runs 40–60%. Internal transfer pricing hides the damage from consolidated statements. Stability is a calculated illusion. Cash leaves the system regardless.
The Packaging-First Read
The official scope includes packaging and test. I treat that as the priority deliverable. Tesla and SpaceX do not face an EUV shortage. They face a CoWoS allocation problem. TSMC's advanced packaging capacity is the binding constraint on every AI accelerator in volume. A packaging line serving captive chips is the fastest route out of a queue Terafab does not control. The strategy is rational — and it is not the story the announcement told.
Advanced packaging from zero is brutal. Process qualification, thermal validation, wafer-level integration discipline. If Terafab licenses Intel's Foveros or EMIB stack, the packaging is Intel's packaging. The vertical-integration narrative softens with every signed license.
Demand Is the Real Liability
Tesla's AI5 is already scheduled for Samsung's Taylor fab in the second half of 2027. Migration from a mature Samsung or TSMC process design to Intel 18A is not a recompile. It is six to twelve months of IP-level re-optimization per chip, and that is where designs go to die. Terafab must win that allocation away from a qualified supplier — and then survive the transplant.
Optimus volume is a forecast, not a contract. Starmind is strategically important and operationally small. Customer concentration is 100% captive, with Tesla dominating. If Optimus slips six quarters, the fab does not pause. Hype evaporates; solvency remains.
The implied consumption is another impossibility test. The annual target of 100 to 200 billion units requires a captive market that does not exist. Global automotive-grade SoC shipments total in the hundreds of millions per year. Filling the fab means Optimus alone must ship at a scale comparable to the entire smartphone industry. No humanoid-robot production forecast supports that.
Supply Chain Geometry
US soil neutralizes direct export controls. It does not solve allocation. ASML's EUV queue runs 12–18 months, with leading customers holding priority. Terafab enters behind existing Intel, TSMC, and Samsung commitments. Japan's dominance in EUV photoresist and high-purity silicon is structural. China's gallium and germanium controls are a medium-low risk — which only means the threat sits elsewhere in the chain.
The water calculus is the most under-appreciated constraint. Gibbons Creek Reservoir solves the initial phase. At full buildout, a complex at this scale consumes an estimated 20 to 40 million gallons of ultrapure water per day. Texas does not carry a water surplus. Expansion phases tied to new water rights will move slower than any technology roadmap.
What $119 Billion Actually Is
The full envelope equals three to five years of combined annual capex for the industry's largest incumbents. It is more than double the entire CHIPS Act incentive pool. When a number exceeds a national industrial policy budget, it stops being a capital budget and becomes a sovereignty statement — possibly with defense procurement undertones given SpaceX's posture. The only binding capital on the table is the $10 million alongside the initial tranche.
The envelope also depletes the parents. A decade of full drawdown approximates the combined projected free cash flow of SpaceX and Tesla. Staged capital releases are the only governance mechanism protecting the core businesses — and staged releases are also the mechanism by which the project can be quietly abandoned.
One more scale check. One hundred million square feet is roughly 2,300 acres of roof. A semiconductor fab phases cleanroom in increments of 50,000 to 100,000 square feet. The disclosed dimension belongs to a master plan, not a factory. Either Terafab is a multi-decade landbank play, or the figure is a reporting artifact.
I have seen this pattern before. My Curve 3Pool audit found an elegant invariant whose fee parameter leaked risk under volatility. My Bored Ape floor work showed twelve percent of a published price was wash-traded scaffolding. My early Geth work found an optimistic propagation path understating state divergence risk. The failure mode is universal — from L2 rollups announcing throughput they cannot settle, to physical fabs announcing wafers they cannot run. When a protocol announces capacity it cannot physically schedule, that announcement is a liability item. Audits reveal what code conceals — and press releases conceal more.
The Bullish Counterweight
The bulls are not wrong. Vertical integration for AI inference is the deepest moat available to a firm that controls its end market. If Optimus and Cybercab mature on schedule, fixed-cost fabrication converts a chronic external expense into a depreciating internal asset.
The packaging-first read flatters the project further. Targeting the CoWoS bottleneck is precise strategic positioning. And Intel's participation is real: this is not greenfield semiconductor construction. It is a joint-operating overlay on an existing process, an existing supplier network, and qualified vendors. That compresses the timeline.
The geopolitical tailwind is genuine. In a sustained decoupling regime, a US-flagged fab serving a defense contractor receives capital prioritization, regulatory deference, and subsidy protection a foreign-owned Arizona plant cannot fully access. Terafab's actual field of competition is not TSMC's process technology. It is TSMC's allocation queue. On that field, the project is not obviously wrong.
The legal structure deserves grudging respect. By framing the commitment as a framework agreement rather than a binding capital obligation, the partners retain an exit without a liquidation event. A venture that can be wound down in phases is a real option, not a suicide pact. That discipline is the most prudent decision in the entire announcement.
The Only Metric That Matters
The calendar is marked. Equipment move-in: 2027. Then track quarterly 18A yields at Terafab scale. Below 70% by 2029 — a fortress of negative carry. Above that, the captive model works precisely because it never needs an outside customer. Precision is the only risk mitigation. Follow the wafers per month, not the $119 billion headline.