Terafab is a semiconductor factory that Tesla and SpaceX confirmed on August 6, 2026 for Grimes County, Texas. The first phase is a 16.8 billion dollar investment, with a potential build-out toward 119 billion dollars across four phases, and it is pitched as the largest building on Earth at over 100 million square feet. It will make inference chips for Tesla's Optimus robots and Cybercabs and high-power chips for SpaceX's planned space-based data centers.

What is the Musk playbook for owning bottlenecks?

It is the pattern of building the thing that makes a scarce input rather than buying it: launch through SpaceX, connectivity through Starlink, power storage through Tesla Megapack, compute through xAI's Colossus supercomputer, and now chips through Terafab. Each company owns a bottleneck and their outputs feed each other - power runs the compute, chips feed the compute, and cheap launch can carry it to orbit.

The Lens · How the Builders Think

The Musk Playbook: Own the Bottleneck, Don't Rent It

Q: How should an ordinary investor interpret Musk's bottleneck strategy?

As a map rather than a stock tip. The purest bottleneck companies, SpaceX and xAI, are private, and Tesla is a blended car, energy, and robotics bet. The transferable lesson is where the AI buildout's value pools - power, chips, memory, and launch - and that the suppliers who own those scarce layers tend to get paid regardless of which brand wins on top.

Ignore the personality for a moment and Elon Musk's companies start to rhyme. Each one grabs a different bottleneck in the AI and energy buildout — and they feed each other. Tesla makes the power, xAI makes the intelligence, the new $16.8B Terafab will make the chips, and SpaceX makes the launch and the connectivity. This isn't a “buy Tesla” piece. It's about the pattern — and what it tells the rest of us about where value pools. Hover or tap any underlined term.

Dragonfly Lens · August 10, 2026 · First in a “How the Builders Think” series — analysis, not admiration.

Updates to this piece

2026-08-13: Colossus turbine status updated. Since publication, SpaceX has agreed (in an order with Mississippi regulators) to remove all 69 temporary gas turbines by July 2027, replacing them with a permitted 1.2 GW plant — though the fight continues (NAACP/environmental-law-center suit; DOJ intervention). The caveat below reflects this. Source: Gizmodo / TechCrunch.

The short version

The pattern: whenever Musk needs something scarce, he doesn't rent it — he builds the thing that makes it. Rockets, batteries, chips, power, connectivity. He buys the bottleneck.
The newest proof (Aug 6, 2026): Tesla and SpaceX are building Terafab, a chip factory in Grimes County, Texas — a $16.8B first phase (up to ~$119B across four), pitched as the largest building on Earth (100M+ sq ft), to make inference chips for Optimus robots, Cybercabs, and space-based data centers.
It's a system, not a scatter: Tesla's batteries + solar power xAI's Colossus supercomputer (555,000 GPUs, 2 gigawatts); Colossus trains the AI; Terafab will feed all of it cheaper chips; SpaceX launches the satellites (and, eventually, data centers in orbit). Each company hands the next one its missing organ.
The Lens takeaway: you can't buy most of this directly (SpaceX and xAI are private) — but the map is the lesson. The AI buildout's value pools at the bottlenecks: power, chips, memory, launch. Musk is just the loudest example of a rule that names the winners. Bet the bottleneck, not the horse.
The honest caveat: vertical integration this wide is brutally capital-hungry and fragile; several claims are self-promotional (verify before you believe); and Colossus was partly powered by unpermitted gas turbines — real regulatory and social risk, not a footnote.

What actually happened

On August 6, 2026, Tesla and SpaceX jointly confirmed Terafab — a semiconductor plant in Grimes County, Texas, with a $16.8 billion first phase and a potential build-out to roughly $119 billion across four phases. Musk calls it the largest building on Earth: more than 100 million square feet. Its job is to stamp out inference chips for Tesla's Optimus robots and Cybercabs, plus high-power chips for SpaceX's planned space-based data centers.

Read alone, it's just another giant factory headline. Read alongside everything else Musk is building, it's the missing piece of a machine that has been taking shape for years.

Believe the direction, not the superlatives. “Largest building on Earth,” the four-phase $119B figure, and the timelines are plans and projections, announced by the companies themselves. Phase one is real money and real intent; the rest is a roadmap. And the “50× the Pentagon” line making the rounds overstates it — 100M sq ft is closer to ~15× the Pentagon's floor area. Anchor on the $16.8B first phase; treat the rest as ambition.

The plain-English version: he builds the bottleneck

Most companies rent what's scarce. Need chips? Buy them from Nvidia. Need power? Buy it from the grid. Need launch? Book a rocket. Musk's habit is the opposite: when a scarce thing sits between him and what he wants, he builds the factory that makes the scarce thing.

Don't queue for the well — own the well. In a drought, everyone lines up at the one well and pays whatever the owner asks. Musk's move is to go dig his own well — and then sell water to the people still standing in line. Rockets were the well for space; Starlink was the well for connectivity; batteries were the well for storing power. Terafab is him digging the well for chips.

Why does that matter far beyond Musk? Because it's the same lesson the Lens keeps landing on from the other side: in a gold rush, the durable money is in the picks and shovels — the scarce inputs everyone needs. Musk is that lesson lived at full scale: he doesn't pan for gold, he becomes the shovel store.

The machine: how the companies feed each other

The empire looks scattered — cars, rockets, robots, a chatbot — until you line the companies up by which bottleneck each one owns. Then they click together:

Power — Tesla. Batteries (Megapack) and solar. Electricity is the gating bottleneck of the AI era — you can't run chips you can't power. Tesla makes the power gear that feeds the compute.

Compute — xAI. The Colossus supercomputer: ~555,000 GPUs, 2 gigawatts, ~$18B — and partly powered by on-site gas turbines and Tesla Megapacks. One Musk company literally keeps another one's AI running.

Chips — Terafab. The newest link: instead of buying every AI chip from others, make the inference silicon in-house — the same “escape the chip bottleneck” move the whole industry is racing on (see the memory wall).

Launch & connectivity — SpaceX. Cheap launch made Starlink possible; Starlink funds more launch. And the next idea — data centers in orbit, running on unlimited solar and cooled by space — would fuse all of it: SpaceX lofts the compute, Terafab supplies its chips, and the bottleneck moves off-planet.

The insight: each company is a bottleneck, and each one's output is the next one's scarce input. Power feeds compute; compute needs chips; chips and compute ride to orbit on cheap launch. It's not five bets — it's one bet, five times, on owning the scarce middle of the whole stack.

The opportunity: what this teaches the rest of us

Here's the trap: the obvious reaction is “so buy Tesla.” But the two most bottleneck-pure pieces — SpaceX and xAI — are private, and Tesla the stock is a car company, an energy company, and a robotics bet all mixed together. The real value isn't a ticker. It's the map.

The bottleneck	Musk's version	The general lesson (who gets paid)
Power	Tesla Megapack + solar	Electricity gates the whole AI buildout. The people who make power and grid gear get paid whoever wins the model war. (See the power bottleneck.)
Chips & memory	Terafab	Everyone's escaping the chip bottleneck at once. The designers, the memory makers, and the one factory underneath get paid. (See the memory wall.)
Launch & orbit	SpaceX / Starlink	Cheap launch is its own bottleneck; whoever owns it controls space-based connectivity — and maybe space-based compute next. (See the off-earth economy.)
The moonshot 🌙	Data centers in orbit	Clearly speculative. If compute really moves to space for free cooling and solar, it rewrites the map again. Early, unproven — we'd only act on data.

The through-line: Musk isn't magic — he's a very loud demonstration of one rule. In every scarce layer, the value flows to whoever owns the scarce layer. You don't have to buy his companies to use the map: it points at power, chips, memory, and launch as the places the buildout's money pools — the same own-the-bottleneck idea we keep coming back to.

The risks — named, not buried

Vertical integration is a high-wire act. Owning every layer means paying for every layer, up front, at once. It's staggeringly capital-hungry, and if one link (say, Terafab's build-out) slips, it strains the others. Wide is powerful and fragile.
The claims run ahead of the concrete. “Largest building on Earth,” $119B, aggressive timelines, “heat-shield problem solved” — these are announcements, not audited results. Musk's direction is often right and his dates often wrong. Lead with what's built.
Real-world friction is real. Colossus ran on unpermitted gas turbines near Memphis, drawing regulatory and community pushback. In mid-2026 SpaceX agreed (in an order with Mississippi regulators) to remove all 69 temporary turbines by July 2027, replacing them with a permitted 1.2 GW gas plant — but the fight isn't over (the NAACP and an environmental law center are suing; the Justice Department moved to intervene on national-security grounds). Building the bottleneck yourself means owning its externalities — pollution, permits, politics.
You mostly can't buy the purest pieces. SpaceX and xAI are private; Tesla is a blended bet. “Own the bottleneck” as an investing idea usually points at the suppliers underneath, not at Musk's cap table.
Not advice. This is how we read the pattern, not a recommendation. Company names illustrate a structure, not buy calls — do your own work and size to your own risk.

The viral take and the true take are rarely the same trade

Don't watch the man. Watch the bottlenecks he keeps buying.

Dragonfly Lens maps the AI buildout as one connected chain — and finds the layer where the value actually pools. Plain English, every claim sourced and flagged. When we're wrong, we say so.

Join the Lens →

Not ready to subscribe? Get the free weekly →

More: The memory wall · The power bottleneck · All explainers

Quick answers

What is Terafab? A semiconductor factory that Tesla and SpaceX confirmed on August 6, 2026, for Grimes County, Texas. The first phase is a $16.8 billion investment, with a potential build-out toward $119 billion across four phases, and it's pitched as the largest building on Earth at over 100 million square feet. It will make inference chips for Tesla's Optimus robots and Cybercabs and high-power chips for SpaceX's planned space-based data centers.

What is the “Musk playbook”? The pattern of building the thing that makes a scarce input rather than buying it: rockets and launch (SpaceX), connectivity (Starlink), power storage (Tesla Megapack), compute (xAI's Colossus), and now chips (Terafab). Each company owns a bottleneck, and their outputs feed each other — power runs the compute, chips feed the compute, cheap launch carries it to orbit.

How should an ordinary investor read this? As a map, not a stock tip. The purest bottleneck companies (SpaceX, xAI) are private, and Tesla is a blended bet. The transferable lesson is where the AI buildout's value pools — power, chips, memory, and launch — and that the suppliers who own those scarce layers tend to get paid regardless of which brand wins on top.

Sources: Terafab: Tesla/SpaceX $16.8B first phase (up to ~$119B / 4 phases), 100M+ sq ft, Grimes County TX, inference chips for Optimus/Cybercab + space data centers — TechCrunch, Electrek, Fox Business. xAI Colossus: ~555,000 GPUs, 2 GW, ~$18B; gas turbines + Tesla Megapacks; unpermitted-turbine controversy; SpaceX agreed (2026) to remove all 69 temporary turbines by July 2027, replacing them with a permitted 1.2 GW plant, amid NAACP/SELC litigation + DOJ intervention — Introl, Gizmodo, TechCrunch. Industry-wide move to custom/inference silicon — see our memory-wall explainer and its sources.

Educational research, not personalized investment advice. Dragonfly Lens is not a registered investment advisor. Figures are as reported by the sources above and were accurate at publication; forward figures (phase totals, timelines, “largest building”) are company projections, not results. Company names illustrate a structural pattern in the AI and energy buildout, not buy recommendations — verify against primary filings before acting. Past performance does not guarantee future results.