Five Misread Bets Across Musk’s Empire Point to Software Margins, Humanoids, and Multiplanetary Infrastructure

From $99-a-month FSD on cars people already own to Gen 3 Optimus on the old Model S/X floor, the stack that matters is recurring intelligence - not the flashiest hardware headline.

The product people argue about is almost never the one that compounds. Across Tesla, SpaceX, and the wider Musk stack, the durable bets sit under the noise: software margin on existing cars, a shared physical-world brain for robots and eventually drones, and an end-to-end intelligence machine that only makes full sense if Earth is the testbed for the Moon and Mars.

Key Takeaways

• RoboTaxi still needs depots, cleaning, insurance, overnight parking, and city-by-city rules - unsupervised fleets as of July 2026 remain limited to Austin, Dallas, Houston, and Miami.
• Supervised FSD on ordinary Teslas already fits the legal model regulators prefer: a licensed human in the seat, garage as charger, owner as cleaner - the Netherlands greenlit it in April 2026, first in Europe.
• On February 14, 2026 Tesla killed one-time FSD purchase; the path is $99 a month for life of the car, with unsupervised capability looking like a higher floor, not a permanent price.
• Delivering that sub costs Tesla almost nothing once the car is sold - computer, cameras, and training stack already exist - so each new subscriber is close to pure margin after one over-the-air flip.
• Gen 3 Optimus production is lined up at Fremont where Model S and Model X once ran; the hand jumps to 22 degrees of freedom from 11 on Gen 2, which is what turns a walker into a worker.
• Fully ramped humanoid labor estimates land around $3 to $5 an hour all-in versus roughly $22 to $28 for U.S. warehouse labor, with a path toward $2 to $3 as volume climbs.
• A drone only needs propulsion, battery, and a brain - SpaceX aerodynamics, Tesla batteries at scale, and the same vision autonomy stack that already drives cars through dense cities.
• Colossus in Memphis sits near 550,000 GPUs in early 2026 (first phase about 230,000), aiming at a million GPUs and roughly 2 GW; Tesla’s AI5 samples late 2026 with volume in 2027 at about 150 to 250 watts per chip.
• AI1 orbital compute filings sketch satellites averaging 120 kW (peak near 150) on ~70-meter wingspans, first launches late 2027, up to a million units - on the order of 120 GW of spaceborne compute - linked by Starlink lasers and lofted by Starship.

---

RoboTaxi Grabs the Headlines. FSD Subscription Grabs the Cash

Dedicated robotaxi fleets look like the destiny product because they photograph clean: no wheel, no pedals, a pure network. The economics fight that story. A real network still needs charging and cleaning depots, someone to own and insure the cars, somewhere to park them overnight, and a regulatory map that barely exists outside a handful of U.S. cities. That is why unsupervised RoboTaxi operation remains a small-footprint grind as of mid-2026.

A regular Tesla with controls - or even a Cybercab-style vehicle still tested with a wheel - needs almost none of that infrastructure. The car is already legal in most of the world. The “depot” is a garage. The cleaner is the owner. Regulators get the thing they like most: a licensed driver still in the seat. That is the same logic that opened the Netherlands to FSD in April 2026 and put EU-wide approval on a short fuse. Scale here is software attach on hardware already sold, not a multi-year depot war.

$99 a Month Is the Floor, Not the Forever Price

Killing the one-time FSD buy in February 2026 forced the model into pure subscription. Right now the product is still supervised: a very good assistant you babysit. Once attention stops being required - when the seat is mostly regulatory theater and the driver can work, rest, or ignore the road - the product is time back, not lane-keeping polish.

People already pay huge money to buy hours. A full-time driver runs thousands a month. An unsupervised package at a few hundred dollars a month starts to look cheap against that, not expensive against today’s $99. I know how aggressive that sounds. But the cost side is what makes the thesis hard to dismiss: Tesla already paid for the hardware and the brain for other reasons. A new sub is mostly margin. Millions of cars a year times $100-plus a month, then $200, $300, maybe $500 when supervision drops, compounds into a cash machine that can dwarf the car business it rides on. Cars become robots on wheels with a monthly wire.

The Hardest Part of Optimus Is Already Running on the Road

Humanoids arrive faster than most timelines allow if you stop staring at legs and start staring at perception. Walking and balance are hard. Hands are hard. The piece that still breaks general robots is the brain: see a messy world in any condition, understand it, decide in real time. Tesla has been training that exact loop for years across a fleet of camera-heavy cars - kids and balls in the street, weird object motion, surprise obstacles - and that stack is moving straight into Optimus.

The hand is the remaining hardware bottleneck. Gen 3’s 22 degrees of freedom (double Gen 2’s 11) is the difference between a balancing statue and something that can manipulate tools and parts. Actuators that fit in a finger do not show up as catalog parts; they get better through manufacturing reps - build a thousand, break them, build the next thousand cheaper. Verticalizing components that do not exist yet is muscle Tesla and SpaceX already have.

Factories First, Kitchens Later - and the Wage Gap Does the Rest

Safety sets the order. Thousands of units under internal test on controlled factory floors come long before stairs, dogs, toddlers, and clutter. A full-size fall near a child is a risk class Tesla cannot absorb early. So the path is factory, then warehouse, then home years later.

Economics still pulls the technology into the market hard. Humanoid all-in labor in the $3 to $5 range against mid-twenties human warehouse wages - trending toward $2 to $3 - turns the plant manager’s choice into arithmetic. Amazon already automates with expensive robots; a humanoid that undercuts human labor from day one does not need a long payback story. Businesses buy that hand over fist. Homes wait for reliability and liability to catch up.

The “Obvious” Product Almost Nobody Prices: the Musk Drone

Propulsion, battery, autonomy. SpaceX has lived in hard aerodynamics and recovery since 2002; electric fans are a step down from orbital-class engines, not a step up. Tesla builds batteries at a cost curve few can match. The brain is the same FSD-class stack: a drone at a few hundred feet does not face pedestrians, cyclists, or four-way stops the way a car in downtown Austin does. If the car problem is tractable, the air cargo problem is the easier one.

Domestic U.S. drone hardware has been hollowed out while the global market sits around $69 billion in 2026. A platform whose fans, packs, and autonomy are all built in-country is both a commercial and a defense unlock - the Pentagon’s autonomy contests already read like a demand signal. Count the products on one foundation - cars, humanoids, drones - and count outside suppliers for batteries, motors, chips, and autonomy software. For this stack the answer trends toward zero. The body changes. The intelligence does not.

One Machine With Two Jobs: Make Intelligence, Then Run It

Strip the branding and the empire is a production line for intelligence plus actuators for the physical world. Power starts in-house: Tesla solar and Megapacks feeding grids and data centers. Compute sits on that power - Colossus in Memphis already near half a million GPUs and climbing toward a million and multi-gigawatt draw, plus Cortex at Giga Texas. Silicon goes vertical with AI5 at a fraction of big-GPU wattage, samples late 2026, volume 2027, and TerraFab with Intel (joined April 2026) aimed at terawatt-scale compute hardware per year.

Then the stack leaves the ground. AI1 orbital compute - ~120 kW average per satellite, 70-meter class wingspans, launches targeted late 2027, filings that scale toward a million birds - skips grid interconnects, water cooling, and county permit fights. Starlink’s laser mesh already interconnects the sky. Starship is the freighter. Digital execution is Grok and agents: knowledge work without a body. Physical execution is FSD cars, Optimus on the Fremont line, and the drone path above. Same intelligence class, different limbs. Tesla stops looking like a car company and starts looking like infrastructure for a civilization that runs on produced intelligence the way the last century ran on roads and power lines.

Earth Is the Beta for the Moon and Mars

Design that works where nothing exists yet only makes sense if the destination is empty. Autonomy for anything that moves, Optimus ahead of crews to pour foundations, Starlink as Moonlink or Marslink, Boring-style tunnels for radiation-hard habitats, Starship for mass transport, Tesla solar and batteries for base power, orbital compute so robots do not wait 5 to 20 minutes one-way for Earth to answer - that is a single multiplanetary kit list, not five unrelated product lines.

Nobody staffs Mars with Earth-wage crews plus life support plus a return ticket to pour concrete. Robot-priced labor is the only spreadsheet that closes. Earth pays for the R&D in cash-generating products first - FSD sub revenue, factory robots, power and compute sold into markets that already exist - so the same hardware arrives on another world already battle-tested. When a launch or product looks absurd for a carmaker or a rocket company alone, the useful question is simple: what does it do on the Moon, and what does it do on Mars?