The Secret Ring Inside Tesla's $16.8 Billion Mega Factory

The first-phase Terafab plant is planned for Grimes County, Texas. On August 6, 2026, Gov. Greg Abbott and Tesla/SpaceX announced a first-phase capital investment of more than $16.8 billion and about 3,000 jobs. That is an announced commitment, not cash already spent. Texas JETI filings describe a multi-application, multi-phase project. One Anderson-Shiro recommendation lists about $10.3 billion for that district's first-phase slice. Later phases have been described in public in the $55 to $119 billion range.

SpaceX and Tesla describe a planned campus of more than 100 million square feet. That building does not exist. 100 million square feet is about 15 times the Pentagon's official 6.64 million square feet of gross floor area. Elon posted that it would be "50 times the size of the Pentagon when complete." That comparison does not match Pentagon floor-area figures.

After a circular feature showed up in the Terafab render, Elon replied "FEL FTW" to a post inferring a particle accelerator. That is an endorsement of free-electron-laser tech, not a detailed confirmation that the ring is an operating FEL EUV source. Tesla and SpaceX have not published a technical disclosure of a FEL ring. The hint is still the part almost nobody priced.

The light bulb you cannot buy

Every advanced chip starts as a silicon wafer. The factory deposits thin films, prints a pattern, etches, implants, polishes, inspects, and does it again hundreds of times until you have billions of switches. The print step is lithography. Once the features get small enough, ordinary light is too blunt.

Production EUV uses 13.5 nanometer light. You cannot buy a 13.5-nanometer bulb. ASML remains the only company selling production EUV scanners. Industry reporting in 2026 puts Low-NA tools around $200 million and High-NA around $350 to $400 million each.

Inside one of those machines, a generator fires molten tin droplets through a vacuum. ASML's production EUV source fires about 27-micron droplets at about 50,000 times per second. A first laser pulse flattens the drop. A second, much hotter carbon-dioxide laser turns it into plasma. That plasma throws off a burst of EUV. Mirrors, not glass lenses, collect a sliver of useful 13.5-nanometer light and throw a pattern onto the wafer. EUV is absorbed by almost everything, so the whole path lives in vacuum.

Usable source power is the throttle. A 250-watt source historically enabled about 125 wafers per hour. Later 0.33-NA tools moved to about 500 watts and 220 wafers per hour. ASML has shown a lab path to about 1,000 watts, publicly reaffirmed in February 2026. The company has said a commercial 1,000-watt source still takes years. Do not date that 1,000-watt demo to April 2025. The tin-plasma source is still a debris and droplet-timing problem at industrial scale.

One laser, many scanners

An FEL makes the same 13.5-nanometer light a different way. An electron gun. Accelerators that push the beam near the speed of light. A row of alternating magnets, undulators, that force the electrons to wiggle and emit light. The light and the beam lock together into a tight, intense wavelength. Some designs loop the spent electrons through an energy-recovery linac so the next bunch gets most of that energy back. On a site plan, that path looks like a racetrack. That is why people read a ring into the render.

The industrial dream is one FEL next to, or under, the fab, feeding EUV to several lithography scanners the way a plant feeds electricity or process gas. Every scanner still needs ASML's mirrors, wafer stage, sensors, and controls. The FEL does not replace the machine. It replaces the part that makes the light. If one source can light several tools, you get more wafers from the same scanner fleet, and you throw out a lot of the tin-droplet mess.

xLight, a Palo Alto company working on FEL sources, claims its system can deliver about four times today's EUV power, cut the EUV-cost slice of wafer cost by about half, raise productivity about 50% in existing fabs, and up to 2x at a greenfield site. Those are vendor claims, not measured Terafab results. Treat them as a direction of travel.

Cheaper intelligence does not mean fewer chips

In July 2026, Moonshot AI released Kimi K3 (API on July 16, open weights on July 27), a 2.8-trillion-parameter open-weight model. Independent coding arenas put it at or near the top on frontend and code tasks. Overall leaderboards still had it behind the best closed U.S. models. Stanford's 2026 AI Index, using data through March 2026, says the U.S.-China model performance gap has effectively closed. As of March 2026 the top U.S. model led by 2.7%, a single-digit gap that moved around over the prior year. That is a benchmark gap, not a compute or chip-supply gap. Smaller specialized models keep beating giant general models at one job.

The easy conclusion is that we need fewer chips. The opposite is what happens when a useful unit gets cheaper. You do not pocket the savings. You buy more of the unit. A company that would not pay $10 per agent task will run a thousand of them at 10 cents. A robot that needed a remote supercomputer for every hard move can keep a smaller model on board all day. That is Jevons: cheaper use, more total consumption, if the thing is actually useful. Intelligence is useful in software and in the physical world.

Tesla already has two products that eat inference silicon. Tesla's Q2 2026 shareholder update says Cybercab production began at Gigafactory Texas, with more than 125,000 units a year of installed capacity listed there. That is installed capacity, not current run-rate. The first unit was shown in Q1. Tesla's 2024 IR aim was at least 2 million Cybercabs a year across more than one factory at full design capacity. That is a company target, not measured Texas output.

Optimus is the human-shaped robot, because factories, doors, stairs, and tools were built for people. Tesla and Elon have described Optimus scale as eventually millions, then tens of millions, and in long-horizon talk 100 million or more units per year. That is ambition language, not a dated production schedule. Near-term IR figures are a 1 million-a-year Fremont line and a 10 million-a-year Texas line still being built. If either product works at scale, chip demand leaves the data-center chart. Competitors will copy the form factor. Then you have cars, humanoids, drones, boats, construction gear, and a pile of machines nobody has named yet, each with a computer that has to decide in the vehicle, not in Memphis.

Lithography is not the only wall

An AI processor still needs high-bandwidth memory stacked close to the compute die. Capacity is the size of the workbench. Bandwidth is how fast parts move on and off it. A huge bench is slow if the parts arrive late. Then advanced packaging: compute dies, memory stacks, and smaller chips in one tight package, moving huge data without cooking the part.

TSMC has said advanced packaging stays tight. Micron said this year that server demand was constrained by supply, and that demand for some data-center storage products exceeded what it could provide for the foreseeable future. A fab is a stack of factories: lithography, memory, packaging, power gear, cleanrooms, materials, and process knowledge. One better light source does not dissolve that stack.

Tesla and SpaceX say Terafab is meant to put logic, memory, and advanced packaging (and, in Texas filings, mask and design) in one loop. SpaceX's S-1 describes two families: edge and inference chips for vehicles and Optimus, and space-optimized chips for orbital compute. Radiation, thermal-swing, and solar-to-compute tradeoffs beyond that sentence are commentary, not a published spec. The point is a chip designed around the machine that will run it, then run through the fab again when the model changes.

Apple already designs silicon around its software and devices. Tesla and SpaceX want more of the manufacturing loop in-house instead of waiting in someone else's queue.

Power, politics, then orbit

Those chips also run in data centers. SpaceX bought xAI in February 2026. Grok and Colossus now sit inside SpaceX, not Tesla. Greater-Memphis Colossus and Colossus 2 is the main cluster. SpaceX reported about 1.0 gigawatt nameplate in Q1 2026 and 1.4 gigawatts in Q2, aiming for more than 2 gigawatts by year-end 2026 and closer to 10 gigawatts than 5 gigawatts by the end of 2027. That is a target, not a measured 10x of one campus.

On the August 4, 2026 SpaceX earnings call, Elon called data centers a trivial problem versus daily rocket engineering. He said they are taking a small amount of rocket and satellite expertise and applying it to terrestrial data centers. That is his boast, not an engineering result.

They are still getting harder. IEA's April 2026 Key Questions on Energy and AI projects global data-center electricity roughly doubling from 485 terawatt-hours in 2025 to 950 in 2030, with AI-focused centers' use tripling. LBNL's 2025 Update, published June 2026, has a reference case of U.S. data centers at 11.8% of U.S. electricity by 2030, with a scenario range of 9.5% to 15.3%. Both are projections. Even if you could print every chip you wanted, you still need megawatts, substations, and neighbors who will let you build. Data-center siting is already a political fight in a lot of the U.S.

That is the logic for compute in space. Starmind is SpaceX's name for the planned orbital AI constellation: satellites built around processors, wrapped in solar arrays and radiators, linked by lasers. Elon confirmed the name in June 2026. FCC filings discuss up to a million orbital data-center sats. None of that is on orbit yet. Think of it as racks that never need a county permit.

Starlink is already the network. By mid-March 2026, trackers counted about 10,000 Starlink satellites on orbit. Starlink's 2025 Progress Report said the mesh had more than 24,000 lasers. Elon on March 15, 2026 said more than 30,000 lasers in space. Lasers are not the same as laser links. A compute sat could run a job, hand the result through that mesh, and drop it to a car or a robot through Starlink. The real prize is solar without waiting on a grid interconnect, and a way around the political wall on the ground.

Starship is what makes the mass affordable. Elon's 2026 target is more than 10,000 Starship flights a year, more than once an hour. He has also said about 30 a day by 2030. That is not current operations. On August 4, 2026, after Flight 13 data, he said he would call the heat-shield problem solved at this point, and that he did not want to jinx it. Flight 13's ship was inspected after an ocean splashdown, not caught. Rapid reuse is not demonstrated. If the cost of a kilogram in orbit falls toward the cost of fuel, orbital compute stops being a stunt.

The loop, and why the Pentagon should care

Terafab makes space-optimized chips. Earth data centers prove the architecture and train the models. Starship launches the sats. Starlink ties them to each other and to users. More sat demand raises launch rate. Higher rate spreads fixed costs. Lower launch cost makes more orbital compute worth flying, which pulls more chips. Tesla sits on the other side with a pile of machines that need inference in the physical world.

Semiconductors sit inside almost every modern military and critical-infrastructure system. In 2022 the Pentagon said virtually every Defense Department system and critical-infrastructure component depends on microelectronics. If the United States can design a powerful AI chip and cannot manufacture and package it at volume, the design is a slide deck. The COVID chip shortage shut vehicle lines and delayed other electronics. Put that same fragility next to a Taiwan shock, a blockade, or a bad earthquake under TSMC's main base, and the problem is national rather than commercial.

TSMC Arizona has been in high-volume N4 production since late 2024. N3 and N2 U.S. volume is still ahead. Intel is running 18A in Arizona and Oregon. The Ohio megafab is not in production. Terafab is another onshore attempt, with the extra bet that an FEL can loosen the EUV-source bottleneck without waiting for ASML to print a light source for every scanner.

ASML still makes the scanner. An FEL does not magic that monopoly away tomorrow. Over time, owning the light source inside the United States is the piece Tesla and SpaceX are actually trying to grab.

That is also why the merger chatter around SpaceX and Tesla will not die. It is not an announced deal. On Tesla's Q2 2026 call Elon said there is more and more overlap around Terafab, then that they cannot talk about combining companies and that there is an appropriate process. Gwynne Shotwell said combining might make Elon's life a little easier. Treat the merger talk as a consequence of one intelligence supply chain (chips, launch, connectivity, cars, robots, orbital compute), not a press release.

The ring on the site plan looks like architecture. It is an attempt to turn the scarcest light in computing into something a factory can pipe. If it works, Terafab is not one product. It is the onshore print shop for the rest of the stack.

Check the video here.