Musk's $25 Billion Chip Factory Is a Bet Against Physics

The most audacious pitch deck ever delivered from a decommissioned power plant

Elon Musk has never been accused of thinking small. But on Saturday evening, standing inside the husk of the defunct Seaholm Power Plant in downtown Austin, he unveiled a project so staggeringly ambitious that it made the original Gigafactory look like a lemonade stand. Terafab — a joint $25 billion semiconductor fabrication venture between Tesla, SpaceX, and xAI — aims to produce one terawatt of AI computing power annually from a single campus on the North Campus of Giga Texas. Musk called it “the most epic chip building exercise in history by far,” and for once the hyperbole might be directionally accurate, even if the execution probability is another matter entirely.

The scope of the announcement is breathtaking. Terafab is designed to consolidate every stage of semiconductor production under one roof: chip design, lithography, fabrication, memory production, advanced packaging, and testing. Tesla is targeting 2-nanometer process technology — the most advanced node currently entering commercial production at TSMC’s facilities in Taiwan — with an initial output of 100,000 wafer starts per month and ambitions to scale to one million wafer starts per month at full capacity. That final target would represent roughly 70 percent of TSMC’s entire current global output. The facility will produce two categories of chips: inference processors for Tesla vehicles and Optimus robots, and a novel D3 chip custom-designed for orbital AI satellites that SpaceX intends to deploy in the coming years. Musk stated that 80 percent of Terafab’s compute output would be directed toward space-based orbital AI infrastructure, with only 20 percent serving ground-based applications.

The timing and structure of this announcement reveal as much as the technical ambitions. Terafab is a joint venture between Tesla, SpaceX, and xAI — the artificial intelligence company that SpaceX recently acquired in an all-stock deal. The consolidation of Musk’s three most compute-hungry entities under a single semiconductor roof is a strategic masterstroke if it works and a catastrophic capital misallocation if it doesn’t. Musk acknowledged his dependence on existing suppliers during the presentation: “We’re very grateful to our existing supply chain, to Samsung, TSMC, Micron and others,” he said, before delivering the knife twist, “but there’s a maximum rate at which they’re comfortable expanding. That rate is much less than we would like.” The subtext is unmistakable. Musk believes the global semiconductor supply chain is structurally insufficient for his ambitions, and rather than wait in the queue behind Apple, Nvidia, and every other customer clamoring for cutting-edge silicon, he intends to build the queue itself.

The competitive implications ripple through the entire semiconductor supply chain. If Terafab were to achieve even a fraction of its stated capacity, it would fundamentally alter the bargaining dynamics between hyperscaler customers and foundry suppliers. Today, Apple, Nvidia, Qualcomm, and AMD compete for limited allocation at TSMC’s leading-edge nodes, and that scarcity gives TSMC extraordinary pricing power. A vertically integrated Musk empire that fabricates its own silicon would not only escape that dependency but could theoretically sell excess capacity to other customers, creating a new entrant in the foundry market backed by the world’s most valuable company. The Korean semiconductor press, covering the announcement for Seoul Economic Daily, noted that both Samsung and TSMC executives are treating the announcement with cautious skepticism — acknowledging the ambition while questioning whether financial resources alone can substitute for generational manufacturing expertise.

The presentation also unveiled speculative renderings of SpaceX’s planned orbital data center satellites, compact modules capable of 100 kilowatts of power each. In January, SpaceX requested a license from the Federal Communications Commission to launch one million data center satellites into Earth orbit. The D3 chip at the heart of these satellites must operate at higher temperatures than conventional terrestrial silicon, adding yet another layer of process engineering complexity to a project already stacked with unknowns. When you combine a greenfield semiconductor fab, a novel chip architecture optimized for space, and a constellation of a million orbital compute nodes, you get either the most transformative infrastructure project of the century or the most expensive science fiction ever committed to a corporate earnings call.

Follow the money, then check the receipts

The financial architecture of Terafab deserves forensic scrutiny because the numbers don’t just stretch credulity — they shatter it. Industry estimates from TrendForce’s analysis indicate that a single 2nm fab capable of 50,000 wafer starts per month costs roughly $28 billion in the current capital expenditure environment. Tesla’s announced $25 billion budget for a facility targeting double that initial capacity and eventually twenty times that scale is, by any reasonable semiconductor accounting, wildly insufficient. The math simply does not close without massive external capital injections, government subsidies, or a fundamental misunderstanding of what advanced node fabrication actually costs.

For context, consider the competitive landscape. TSMC spent decades and tens of billions of dollars accumulating the institutional knowledge, process engineering expertise, and supplier relationships required to achieve viable yields at leading-edge nodes. Intel, once the undisputed king of semiconductor manufacturing, has invested over $100 billion in recent years attempting to regain its fabrication edge and still trails TSMC in yield rates at advanced geometries. Samsung’s foundry business, despite pouring comparable capital into its process technology, continues to struggle with yield consistency at 3nm and below. These are organizations with thousands of specialized process engineers versed in lithography, etching, chemical-mechanical planarization, yield management, and EUV equipment operation — disciplines that Tesla has never employed at any scale.

The technological challenge is equally daunting. The transition to 2nm requires a fundamental architectural shift from FinFET to Gate-All-Around (GAA) transistor structures, demanding broad upgrades in materials, equipment, and process modules across the entire fabrication pipeline. Even minor deviations at any stage can cause catastrophic yield drops. According to TrendForce’s semiconductor analysis, it takes approximately 38 months just to physically construct a fab of this caliber in the United States, and that timeline assumes you already possess the engineering workforce to operate it. Tesla currently possesses zero semiconductor manufacturing experience. The gap between announcing a fab and producing a functional chip at a leading-edge node is measured not in ambition but in the accumulated knowledge of millions of engineering hours.

Tesla’s stock slipped overnight following the announcement, and analysts were notably cautious. Multiple semiconductor industry observers noted that only three firms on Earth — TSMC, Samsung, and Intel — currently possess the capability to manufacture chips at this scale, and all three required decades to build that capability. The idea that Tesla can leapfrog this institutional knowledge through sheer capital deployment is either visionary or delusional. There is no middle ground. The market, at least initially, is leaning toward the latter interpretation. But Musk has defied similar skepticism before with SpaceX’s reusable rockets and Tesla’s gigafactory ramp, which lends the project just enough historical precedent to prevent outright dismissal.

The ASML bottleneck deserves its own paragraph. Every leading-edge fab on Earth depends on extreme ultraviolet lithography machines manufactured exclusively by ASML in the Netherlands. Each machine costs approximately $380 million, weighs over 150 tons, and requires teams of specialized engineers to install and calibrate. ASML’s production capacity is roughly 60 to 70 EUV systems per year, and its order book is committed multiple years in advance to TSMC, Samsung, Intel, and a handful of other qualified customers. Terafab’s one-million-wafer-start ambition would require dozens of these machines. Whether ASML will even accept Tesla as a customer — given the company’s complete absence from the semiconductor manufacturing ecosystem — is an open question that no amount of capital can automatically resolve.

One quantified insight that emerges from stitching together the announced production targets, the known cost structures, and the competitive benchmarks: if Terafab achieves even 25 percent of its stated one-million-wafer-starts-per-month capacity at 2nm within five years, it would represent a capital efficiency roughly four times worse than TSMC’s historical spending per wafer start. That inverse efficiency ratio is the price of starting from zero in a field where institutional knowledge compounds exponentially.

Battery Day all over again, or something genuinely new

The most potent criticism of Terafab draws a direct line to Tesla’s Battery Day presentation in September 2020. On that stage, Musk promised a revolutionary 4680 battery cell, a proprietary dry electrode manufacturing process, and a pathway to three terawatt-hours of annual production capacity. Five and a half years later, as industry analysts have documented extensively, the 4680 program has been a bitter disappointment. Tesla’s own top battery supplier publicly stated that Musk “doesn’t know how to make battery cells.” The dry electrode process required six or seven fundamental revisions, took years longer than promised, and the 3 TWh target remains a distant fantasy. When Musk describes Terafab as a revolution in semiconductor manufacturing, every informed observer mentally overlays the Battery Day timeline and adjusts their confidence interval accordingly.

The parallels are uncomfortably precise. Both projects involve Tesla entering a mature, capital-intensive manufacturing domain with zero prior expertise. Both feature Musk’s characteristic first-principles reasoning — the assumption that physics is the only constraint and that institutional inertia is merely a speed bump. Both were announced with theatrical presentations, aggressive timelines, and budgets that industry veterans immediately flagged as insufficient. And both require mastering process engineering at the atomic level, where the difference between a viable product and an expensive paperweight is measured in nanometers. Critics argue that semiconductor fabrication is, if anything, harder than battery manufacturing because the tolerances are tighter, the supply chains are more fragile, and the required expertise is more specialized. If Tesla couldn’t master wet chemistry for battery cells after five years of trying, the argument goes, what possible basis exists for confidence in mastering extreme ultraviolet lithography?

But the counterargument has genuine force. Unlike Battery Day, Terafab arrives in a fundamentally different geopolitical and industrial context. The global chip shortage of 2021–2023 exposed structural vulnerabilities in the semiconductor supply chain that governments and corporations have been scrambling to address ever since. The CHIPS Act has allocated tens of billions in federal subsidies specifically designed to incentivize domestic semiconductor manufacturing. Musk’s political proximity to the current administration — and his companies’ enormous leverage as employers of American workers building American technology — positions Terafab to capture significant public funding that was unavailable during Battery Day. Furthermore, Tesla’s AI chip design team, which produced the D1 chip for the Dojo supercomputer and the AI4 inference processor for Full Self-Driving, is legitimately world-class at chip design even if the company has never operated a fab. The question is whether design excellence can compensate for manufacturing inexperience when the fab floor is the battlefield.

There is also a strategic dimension that Battery Day lacked. Tesla’s chip demand is not hypothetical. The company needs inference processors for millions of vehicles, training chips for its internal AI workloads, and — through xAI — frontier-model-scale compute for its ambitions in artificial general intelligence. SpaceX needs radiation-hardened, high-temperature chips for its satellite constellation. The combined internal demand from these three entities could theoretically justify a captive fab even at yields that would be commercially unviable on the open market. TSMC’s model depends on selling capacity to dozens of customers; Terafab only needs to serve three. That captive-demand advantage meaningfully changes the break-even calculus, even if it doesn’t eliminate the technical risk.

The orbital gambit and the operator’s playbook

The most underexamined element of the Terafab announcement is the one that could prove most consequential: the orbital AI satellite constellation. Musk’s claim that 80 percent of Terafab’s output would serve space-based infrastructure initially reads as absurd. Why would you build the world’s most advanced terrestrial chip factory primarily to serve satellites? The answer lies in the convergence of three trends that most analysts are evaluating in isolation but that Musk is attempting to synthesize into a single strategic vector.

First, the economics of terrestrial data centers are approaching physical limits. Power consumption, cooling requirements, water usage, and real estate costs are all escalating faster than the efficiency gains from new chip architectures can offset. Google, Microsoft, and Amazon are collectively spending hundreds of billions on data center infrastructure and facing genuine constraints on where they can build and how much power they can procure. Space, by contrast, offers unlimited solar energy, natural cooling via radiative heat dissipation, and zero land-use conflicts. The engineering challenges are severe, but the economic incentives are becoming increasingly compelling as terrestrial constraints tighten.

Second, SpaceX’s Starship program has fundamentally altered the cost curve for orbital payload delivery. If Starship achieves its projected cost targets, launching a hundred-kilowatt AI compute satellite becomes economically feasible at a scale that would have been science fiction five years ago. The combination of cheap launch capacity and custom silicon optimized for the space environment creates a new category of compute infrastructure that no competitor is currently positioned to replicate. Amazon’s Project Kuiper is focused on broadband, not compute. Google and Microsoft have no launch capability. Only SpaceX possesses the full vertical stack — launch vehicles, satellite manufacturing, ground stations, and now, potentially, custom silicon — required to build an orbital AI cloud.

Third, the regulatory and geopolitical landscape strongly favors space-based compute for certain applications. Data sovereignty laws, export controls on advanced AI chips like those at the center of the Supermicro smuggling scandal, and the strategic vulnerability of terrestrial data centers to physical attack or natural disaster all create demand for compute infrastructure that exists outside traditional jurisdictional boundaries. A constellation of AI satellites operating in low Earth orbit would be simultaneously accessible from every country on Earth and subject to none of their territorial restrictions — a legal and strategic gray zone that Musk is uniquely positioned to exploit. The White House’s recent national AI legislative framework explicitly prioritizes domestic AI infrastructure resilience, and orbital compute could become a compelling answer to the federal government’s growing anxiety about concentrated data center targets.

The operator checklist for anyone watching Terafab unfold over the next three to five years is stark:

• Track the hiring. If Tesla begins aggressively recruiting process engineers from TSMC, Samsung, and Intel — particularly specialists in EUV lithography, GAA transistor fabrication, and advanced packaging — the project has real momentum. If the job postings remain vague and generalist, the project is still in the PowerPoint phase.
• Watch the CHIPS Act applications. Terafab’s financial viability almost certainly depends on securing billions in federal subsidies. The timing, size, and conditions of any CHIPS Act award will reveal whether the government views this project as credible infrastructure or political theater.
• Monitor the SpaceX FCC filings. The one-million-satellite license application is the canary in the orbital coal mine. If SpaceX begins deploying prototype compute satellites within 18 months, the D3 chip program is real. If the filings stall, the orbital vision is vaporware.
• Benchmark against TSMC Arizona. TSMC’s own Arizona fab has experienced significant delays and cost overruns despite possessing the world’s best semiconductor manufacturing expertise. If TSMC struggles to build a fab in Arizona, Tesla’s chances of doing so from a standing start are proportionally lower. The delta between TSMC’s Arizona timeline and Terafab’s timeline is the most honest measure of Musk’s credibility gap.
• Follow the supplier contracts. Terafab requires EUV lithography machines from ASML, the sole global supplier. ASML’s order book is committed years in advance. Whether Terafab secures delivery slots — and when — will determine whether the project can meet any reasonable timeline.

The semiconductor industry operates on geological time scales of institutional learning and has an unforgiving physics department that respects no billionaire’s timeline. Musk is betting that his first-principles approach, his captive demand across three companies, and his political leverage can compress decades of institutional learning into a handful of years. History suggests he is wrong. His personal track record suggests he might not be. The only certainty is that the $25 billion price tag is merely a down payment on an invoice that nobody has yet calculated in full, and the global semiconductor market will spend the next five years deciding whether Terafab is the Wright Brothers’ workshop or the most expensive science project since the Superconducting Super Collider.

In other news

OpenClaw creator joins OpenAI as agentic platform goes viral — OpenClaw, the open-source autonomous AI agent platform built by Austrian developer Peter Steinberger, hit 318,000 GitHub stars in just 60 days, prompting Nvidia CEO Jensen Huang to call it “definitely the next ChatGPT.” OpenAI CEO Sam Altman announced that Steinberger was joining the company, with OpenClaw continuing as an open-source project under OpenAI’s stewardship.

Britannica and Merriam-Webster sue OpenAI over RAG copyright infringement — Encyclopedia Britannica and Merriam-Webster filed a federal lawsuit alleging OpenAI scraped nearly 100,000 copyrighted articles to train ChatGPT without permission. The case is notable for being one of the first major lawsuits to specifically target Retrieval-Augmented Generation as a separate vector of copyright infringement beyond training data.

Tencent pledges to double AI investment to $5.2 billion in 2026 — China’s largest tech company announced it will at least double its AI product development spending from 18 billion yuan in 2025, despite U.S. chip export restrictions constraining its ability to acquire advanced AI processors. Total capex reached 79 billion yuan in 2025, below internal targets due to chip supply constraints.

GPT-5.4 mini brings advanced reasoning to free ChatGPT users — OpenAI rolled out GPT-5.4 mini and nano variants to free-tier ChatGPT accounts, delivering improved coding, reasoning, and multimodal capabilities at more than double the speed of GPT-5 mini. The full GPT-5.4 Thinking model remains gated behind Pro, Business, and Enterprise subscriptions.

Shopify bets the platform on agentic commerce — Shopify President Harley Finkelstein declared that AI shopping agents will act as commission-free personal shoppers, surfacing products across platforms like ChatGPT and Perplexity. CEO Tobi Lütke announced that every Shopify store would be “agent-ready by default,” targeting the 82 percent of U.S. retail purchases that still happen offline.