99.9% Of People Have No Idea What's Coming: Why Machine Hours Rewrite Labor

Five ideas about the Musk stack keep getting read wrong. People stare at the flashy product. The money and the strategy sit one layer down. ## RoboTaxi is the side show. FSD cash is the main event

Everyone watches the dedicated RoboTaxi network and the Cybercab with no wheel and no pedals. Dig into running that network and RoboTaxi looks like a red herring.

A real RoboTaxi fleet needs charging depots, cleaning depots, overnight parking, insurance, and someone who owns the car when a passenger spills a burrito in the back seat at 2 a.m. On top of that you need regulation that barely exists city by city. That is why unsupervised FSD in RoboTaxi vehicles, as of July 2026, still sits in a small fleet across Austin, Dallas, Houston, and Miami. Scaling that grind takes years at this pace.

Now compare that to a normal Tesla with a steering wheel and pedals. What does that car need? Almost nothing new. It is already legal in most of the world. The charging depot is the garage. The cleaning depot is the owner. Regulators get the thing they like most: a licensed human still sitting in the seat. That is why the Netherlands cleared FSD in April 2026 as the first country in Europe, with wider EU approval lined up behind it. Scale is software on hardware people already bought, not a multi-year depot war.

Tesla already tipped its hand on the business model. On November 14, 2026, the one-time FSD purchase died. You cannot buy it outright anymore. The only path is $99 a month for the life of the car, for supervised self-driving where you still have to pay attention. Most people treat $99 as the forever price. It is the floor.

Right now you are paying for a very good assistant you still babysit. The day the car drives and you are only in that seat for regulatory reasons — reading, answering email, taking a nap — the product changes. You are buying time back. An hour a day. Maybe two. People already pay crazy money to get their time back. A full-time driver runs thousands a month. Unsupervised self-driving at a few hundred dollars a month, even $500, starts to look cheap against that, not expensive against today's $99.

The margin story is worse for skeptics and better for Tesla. What does it cost to deliver that subscription to a car already on the road? Almost nothing. The computer is in the car. The customer paid for it at purchase. The cameras, the training stack, the monitoring systems — Tesla built all of that for work it was already doing. A new subscriber is close to pure margin. One over-the-air update and millions of cars that already exist start wiring software cash every month. Run that forward: millions of cars a year, $100-plus a month, then $200, $300, $500 near pure margin as the fleet grows. That recurring machine can outgrow the car business it sits on.

The hardest part of Optimus is already on the road

On July 1, 2026, Elon posted a photo walking the Optimus line at Fremont. That line is going in where Model S and Model X used to be built, with Gen 3 production starting sometime this year. His own line on the ramp: slow at first, because everything is new. They show up in businesses first, then homes — faster than most timelines allow.

People get the timing wrong because they watch the wrong part of the robot. Walking and balance look hard. Hands look hard. The load-bearing problem is the brain. The machine has to see a messy physical world in any condition, understand it, and decide what to do next in real time. One company has been training that exact loop for years across millions of camera-heavy Teslas. That stack moves straight into Optimus: same vision approach, same networks, same training computers. The hardest piece of the robot is the piece Tesla already has.

That leaves the hand. Gen 3 has 22 degrees of freedom, up from 11 on Gen 2. A humanoid that can walk but cannot manipulate is a statue with good balance. The hand is what turns it into a worker. Actuators are hard: tiny motors, gearing, sensors packed into finger space. Better actuators come from manufacturing reps — build a thousand, see what breaks, build the next thousand better and cheaper. Verticalizing parts that do not exist as catalog components is muscle Tesla and SpaceX have used for years.

Who gets these robots first? Businesses. Safety decides the order. Thousands of units under internal test inside Tesla factories come before stairs, dogs, toddlers, and clutter. A full-size robot falling near a kid is an event Tesla cannot afford even once. Factory first, then warehouse, then — way later — the kitchen.

Economics pulls the technology into the market hard. Estimates floating around in 2026 put fully ramped humanoid labor around $3 to $5 an hour all-in against roughly $22 to $28 an hour for U.S. warehouse labor, trending toward $2 to $3 as production scales. If you run a plant, the choice is blunt: about $25 an hour or about $3 an hour for the repetitive, boring, dangerous work. Undercut human labor from day one and businesses buy hand over fist.

The Musk drone is the obvious product almost nobody prices

A drone needs three things: propulsion, battery, and a brain. SpaceX has lived in propulsion and aero since 2002 — orbital-class rockets landing on boats, catch towers with giant arms. Going from rocket engines to electric fans is a step down in difficulty, not up. Tesla builds packs at a cost curve almost nobody can touch. The brain is the same FSD stack. A drone at 400 feet does not deal with pedestrians, cyclists, or four-way stops. If that brain can drive a car through downtown Austin, flying a cargo drone is the easier problem.

First use case writes itself: point-to-point hauling. The global drone market is about $69 billion in 2026, and America barely builds the hardware. Domestic drone manufacturing is hollowed out. A drone whose propulsion, batteries, and autonomy are all built in the United States is a massive unlock — and Pentagon contests for drone autonomy already read like demand knocking. The AI brain that runs cars, humanoids, or drones does not care what the body looks like. Count outside suppliers for batteries, chips, motors, and autonomy software in this stack and the number trends toward zero.

One machine: make intelligence, then run it

Why bring the important tech in house? Because this group of companies is building one machine with two jobs: produce intelligence, then execute it in the real world.

Production starts with power — Tesla solar and Megapacks feeding grids and data centers. On top of that sits compute. Colossus in Memphis was running somewhere around 550,000 GPUs as of early 2026, first phase around 230,000 on the way to a million, pulling about 2 gigawatts when the expansion finishes. Tesla runs Cortex at Giga Texas in parallel. Silicon goes vertical too: AI5 at about 150 to 250 watts, samples late 2026, volume in 2027. In April 2026, Intel joined the Terra joint venture aimed at terawatt-scale compute hardware per year.

Then the stack leaves the planet. AI1 orbital compute satellites average about 120 kW (peak near 150) on roughly 70-meter wingspans — wider than a 747. First launches are slated for late 2027. Filings describe up to a million of them. A million at 120 kW average is 120 gigawatts overhead — on the order of a quarter of average U.S. electrical load floating in sunlight. No grid connection. No water cooling. No county permit fight. Data moves on Starlink laser links already active today. Starship is the freighter.

Digital execution is Grok and agents — knowledge work without a body. Physical execution is FSD cars, Optimus off the Fremont line, and the drone path above. Same brain class, different bodies. The ecosystem stops looking like a car company and starts looking like intelligence infrastructure — something you live on without thinking about it, the way last century lived on roads and power lines.

Earth is Mars beta

Why design infrastructure that works where nothing exists yet? Because the destination has nothing for you when you land. Everything this stack builds on Earth is also for the Moon and Mars.

The FSD brain is autonomy for anything that moves. Optimus lands ahead of humans to start building. Starlink becomes Moonlink or Marslink. Gwynne Shotwell has publicly endorsed Boring Company tunnels for Mars habitats. Starship hauls mass. Tesla solar and batteries power the base. Park AI1 compute shells around the Moon or Mars and robots stop waiting five to twenty minutes one way for Earth to answer. Anything working down there has to think where it is.

Nobody will pay Earth wages plus life support plus a ride home to pour foundations on Mars. Robot-priced labor is the only spreadsheet that closes. So every piece ships and earns on Earth first, gets better under real cash pressure, then transfers. Earth is Mars beta. Earth is Moon beta. Next time this stack launches something that makes zero sense for a car company or a rocket company alone, ask what it does on the Moon and on Mars. The picture usually snaps into focus. Machine hours — not human hours — rewrite labor on this planet and the next. Most people are still staring at the wrong product.

Check the video here.