Robotics as a Service · Part 2
Optimus
The model employee
Investments · September 2026 · 10 min read
In the second quarter of 2026 Tesla shut down the lines in Fremont, California that had built the Model S since 2012 and the Model X since 2015, and began installing lines for a humanoid robot. The first ones off it, Tesla says, will not be sold. They go to what the company calls the Optimus Academy, where they are used to collect the data that teaches the next ones. The product is a worker. Its first job is learning to be one.
The Number That Sells It
The Bureau of Labor Statistics counts what an employer pays for an hour of a manufacturing worker's time, wages and benefits together, and in June 2025 the figure was $46.30. That number is the reason every humanoid robot company exists. JPMorgan has estimated that a humanoid could run for $10 to $12 an hour against about $30 for the factory labor it would replace, with the caveat that today's robots are far less productive than the people. Musk has said Optimus could be worth more than the rest of Tesla combined. Hold the claim. The arithmetic underneath it is short enough to do here.
What the Robot Costs
Musk's figure is $20,000 to $25,000 per unit once production reaches a million a year, and more recently "below $20,000." Outside estimates of what one costs to build today run $50,000 to $100,000, with the first commercial units expected at $100,000 to $150,000. No one outside Tesla has a bill of materials, and Tesla has not sold one outside the company, so every number in this paragraph is either a target or a guess. This piece uses two: $25,000, Tesla's target, and $150,000, the outside estimate for an early unit.
The design has moved fast enough that the price target keeps company with a new body. Version 3, the robot the Fremont lines are being built for, is the one Musk said in early 2026 he was finalizing; its reveal was promised for the first quarter and slipped to the summer. Version 4 will be built in Austin. The line for the first is not finished and the second already has a factory assigned.
The Hourly Arithmetic
A robot does not draw a wage. Its cost per hour is its purchase price spread across the hours it works before it wears out, plus what it costs to run. Spreading a capital cost across a useful life is depreciation. The simplest kind, straight-line, takes the price, divides by the years, and charges that much every year.
Assume five years of life and two shifts a day, 300 days a year: 4,800 hours a year, 24,000 hours in all.
| Line | Arithmetic | Per hour |
|---|---|---|
| Capital cost, at Tesla's target | $25,000 over 24,000 hours | ≈ $1.04 |
| Capital cost, at an early unit's price | $150,000 over 24,000 hours | ≈ $6.25 |
| Power, maintenance, software | Placeholder. Tesla publishes nothing | $1 to $2 |
| All in, at target price | ≈ $2 to $3 | |
| All in, at early price | ≈ $7 to $8 | |
| The person it stands beside | Manufacturing, total compensation, June 2025 | $46.30 |
At either price the robot is cheaper than the person by an order of magnitude, and the payback is quick. A $150,000 robot that replaced one worker on two shifts would displace about $220,000 a year of compensation and pay for itself in eight months. A $25,000 one, in six weeks.
Which is where the arithmetic gets too easy, and the article has to slow down.
The Swing Variable
The table assumes one robot does one worker's job. Nobody has shown that. Tesla had about 1,000 Optimus units in its own factories as of early 2026, and Musk told analysts in January that they were there "primarily for learning and data collection rather than performing productive tasks," and were "not in usage in our factories in a material way." The units that exist are students. Their output per hour is unpublished.
So put output in the table. If a robot delivers 30% of a worker's output in an hour, then the $3-an-hour robot costs $10 for an hour of human-equivalent work. At 10%, it costs $30, and the early $150,000 unit at 10% costs $75 an hour of work, more than the person. The line where the model stops making sense is a productivity ratio rather than a price, and it is the one number in this article that nobody outside Tesla can estimate.
The record on timing belongs here too, dated. In 2022 Musk said Optimus could be in production the following year. In 2024 he said several thousand would be working in Tesla's factories in 2025; the count reached roughly 1,000 in January 2026, and they were learning rather than working. The Version 3 reveal planned for the first quarter of 2026 arrived in the summer. Production in Fremont, promised for 2026, began in the third quarter of 2026. The dates slip and the product arrives. Both halves of that sentence have held for a decade, on cars, and the arithmetic in this piece does not depend on which half you weigh more.
Sell It or Rent It
This is the RaaS decision, and Optimus faces it more starkly than the Cybercab, because a Cybercab in Tesla's fleet earns fares from strangers and a humanoid in a customer's warehouse earns whatever the contract says.
Sold, at $25,000: revenue once. At Tesla's current gross margin that is about $4,200 of gross profit per robot, and then it belongs to the customer, whose problem its maintenance and obsolescence become.
Rented at $10 an hour, a rate this piece assumes rather than one anyone has published, for 4,800 hours: $48,000 a year, every year, from the same robot, with its cost on Tesla's books and its upkeep Tesla's job. Over a five-year life that is $240,000 of revenue from a machine that cost $25,000, against $25,000 from the sale. The rental earns nearly ten times the sale and requires Tesla to carry the capital, the maintenance, and the risk that the robot stops working before the five years are up. Xerox made the same choice in 1959 and Rolls-Royce in 1962.
The rest of the field has already chosen. 1X offers both prices on its NEO home robot: $20,000 to buy, or $499 a month on a six-month minimum, which recovers the sticker in about 40 months of subscription. Agility's Digit works shifts in GXO's warehouses under a multi-year contract priced by the hour, at a rate neither company discloses. Unitree sells its G1 outright for $13,500, which is the price a rented Optimus has to beat per hour, since a customer can always buy the cheaper robot instead of renting the dearer one.
Tesla has not said which seat it will take. The 2025 pay package counts Optimus "delivered," which reads like a sale, and Musk has talked about selling the robot to households and about businesses running fleets of them. Both are possible. The seat decides whether Optimus revenue arrives once or every year.
Scaled to the Milestone
One million Optimus delivered is one of the operational milestones in the 2025 award. Sold at $25,000, a million robots is $25 billion of revenue, once. Rented at $10 an hour, a million robots on two shifts is about $48 billion a year. The award also pays on $400 billion of adjusted EBITDA over four consecutive quarters. EBITDA is earnings before interest, taxes, depreciation, and amortization; "adjusted" means with whatever else the company chooses to add back. In plain terms, operating profit before the cost of the machines is counted. A million rented robots at $10 an hour, costing $3 to run, would contribute about $34 billion toward that number. The Cybercab piece put the fares from a million robotaxis at $70 billion. The rest has to come from somewhere else, or the machines have to number more than a million, or earn more than $10 an hour. Which of those Tesla is planning on, the milestones do not say.
The cost curve is the one thing moving in the model's favor. Goldman Sachs found that the manufacturing cost of humanoid robots fell about 40% during 2025, double what analysts had forecast, and Unitree's price cuts say the same thing from the other side. A robot that costs $150,000 today and falls 40% a year costs about $25,000 in three and a half years. That is the trajectory Musk's price target assumes. It has held for one year.
What Has to Be True
- The unit cost reaches $25,000, which needs volume, and volume needs buyers, and there are none yet outside Tesla.
- A robot does a measurable fraction of a human's work per hour, and somebody publishes the fraction.
- It lasts five years, or the hourly cost above is wrong by however much it falls short.
- Someone outside Tesla pays for it, by the unit or by the hour, at a price that survives a $13,500 competitor.
- The rental model, if Tesla chooses it, survives carrying a million machines on its own balance sheet, which is $25 billion of capital before the first hour is billed.
The Cybercab is a car that earns. Optimus is a worker that costs. The arithmetic of the second is easier to write and harder to believe, because the only number in it that has been measured is the wage of the person it is meant to stand beside.
Sources
- Tesla, Q2 2026 Update, July 2026: Fremont conversion, Optimus Academy, production outlook. Call remarks via Not a Tesla App.
- Bureau of Labor Statistics, Employer Costs for Employee Compensation: manufacturing total compensation of $46.30 per hour worked, June 2025.
- Musk's price targets and factory deployment remarks, Q4 2025 earnings call, January 2026, as reported by Optimusk and 36kr.
- Outside cost estimates and JPMorgan and Goldman Sachs figures, as reported by RoboZaps and Revolution in AI, August 2026.
- The Robot Report on 1X NEO pricing; GXO on the Agility agreement; Unitree store pricing.
- 2025 CEO Performance Award milestones, SEC filing.