Completion
14%
range 10–18%
est. from public artifacts
Since last quarter +2 pts · last official Dec 31 2027
Labor trench · merge relevance 92
Optimus — the robot that is not a labor product
Gen 3 design frozen. Fremont S/X converted to Academy. Useful factory labor is still ~0. One million bots is a pay-package number.
Completion
14%
10–18% of a general-purpose labor product. A line is not a labor product.
Merge relevance
92
Mars/Moon labor is the sharpest Tesla–SpaceX shared story.
Pay tranche
1M bots
Elon’s number. Not a shipment forecast. Not Tesla revenue.
Daily desk · live + forensic
Today’s update
Sep 15 2026
Completion tape
- Q1'24 4%
- Q3'24 6%
- Q1'25 8%
- Q3'25 10%
- Q1'26 12%
- Q3'26 14%
IN TRENCHES
T10
Optimus (general-purpose labor product)
Merge relevance 92
Child trenches · % complete
Today’s still
Today’s video
Stills
Reconstruction — idle bot on a car line. Academy units are not useful factory labor. 
Optimus Gen 2 still. Design-frozen Gen 3 has not had a public reveal. 
We, Robot (Oct 2024). Stage demos, including teleoperation. Not unsupervised labor. 
Popcorn, not a stamping press. Honest picture of public “useful work.” 
Hands are the trench. Gen 3 doubles DoF (11 → 22). A demo catch is not a factory shift. 
Fremont floor. S/X capacity mined for Optimus Academy — not customer sales.
Videos
Reconstruction · Idle line recon — Reconstruction. Standing still on a line is not a shipment and not useful work. Official · Optimus navigating around — Official Tesla. Navigation demo. Musk: not in usage in factories in a material way. Official · Optimus Gen 2 — Official Tesla. Gen 2 is the public face. Gen 3 is the mass-production intent, unshown. Official · Optimus catching a ball — Official Tesla. New hand, closed-loop catch. A trick is not a labor product. Official · Tesla Bot update — Official Tesla. Actuators and gait. Trackers counting Academy units as “delivered” would be the tell.
Promised
A general-purpose humanoid that does the work humans do not want to do. A labor product, not a demo.
Shipped
Gen-3 design frozen. Internal 'Optimus Academy' units from a converted Fremont S/X line. That is a factory doing R&D at slightly larger batch size. Fremont design capacity ~1M/yr and Texas talk of ~10M/yr are aspiration, not output.
Missing
No customer sales. No humanoid supply chain (actuators, sensors, hands). Musk has said the early S-curve is 'flat and long.' 1 million robots is a pay-package number, not a build plan.
IN TRENCHES
Not a labor product
Stage demos, popcorn, teleop fights. Musk said the bot is not in usage in factories in a material way. A line is not a shipment. A shipment is not useful work.
Optimus Academy
Internal units for imitation learning. All 2026 Fremont output is described as self-use. External sales talk earliest 2H 2027. Academy units counted as “delivered” would be the pay-package tell.
Generations
We, Robot · Oct 2024
Gen 2
11 DoF hands. Stage demos, including teleoperation fights. Walks, waves, serves popcorn. This is the public body. It is not a labor product.
Design frozen · 2026
Gen 3
First design Tesla describes as meant for mass production. 22 DoF hands, actuators in the forearm, ~22% lighter. Public reveal repeatedly slipped. Fremont Academy units 2H 2026 — internal, not customer sales. External B2B talk earliest 2H 2027.
Texas slide · not this year
Gen 4
The 10M/yr Gigafactory Texas story. Hardware base for a future home SKU. A slide. Do not score it as 2026 progress.
Design
- Generation
- Gen 3 design frozen
- First design Tesla describes as meant for mass production. Public reveal repeatedly slipped.
- Height / mass
- ~1.73 m · ~57 kg
- Gen 3 cited ~22% lighter than prior. Payload cited ~20 kg.
- Hands
- 22 DoF / hand (Gen 3)
- Gen 2 was 11 DoF. Actuators in the forearm; tendons to the fingers. Hands are the trench.
- Actuators
- ~25 per forearm · 50-class next hand
- No existing humanoid supply chain. Musk: it moves as fast as the least lucky part in ~10,000 unique.
- Battery
- ~2.3 kWh class
- Cited 8 h light-duty / 4–6 h intensive. A pack, not a Megapack. Same cell family as the car.
- Compute
- AI4 now · AI5 roadmap
- Same inference family as the car. Cortex trains. End-side is not independence from Nvidia.
- Parts
- ~10,000 unique
- A car company knows parts. A humanoid supply chain does not yet exist at auto scale.
- Plant
- Fremont S/X → Academy
- Line converted ~late Aug 2026. Design capacity 1M/yr Fremont, 10M Texas — aspiration, not output.
Protections / dual-use
Workplace safety / OSHA + ISO
A 57 kg biped with a 20 kg payload in a stamping shop is a robot next to humans, not a screensaver. ISO 10218 (industrial robots) and ISO 13482 (personal care robots) are the actual standards; ANSI/RIA R15.06 is the US collar. Academy units that “learn” in Fremont are a safety file before they are a labor file. Speed, separation, and a watchable e-stop — not a keynote.
Product liability
No consumer SKU. The current collar is internal. The day a bot ships to a customer, Tesla inherits a new product-liability class — not Autopilot, not a car. Hands that can hold an egg can also crush one. Insurance and residual risk sit on Tesla, not on a supplier Tesla does not have.
Export on actuators / AI
Fine manipulation + onboard inference is dual-use whether the slide says “factory helper” or not. EAR already watches AI chips. Actuator packs at this density will get a jurisdiction look if they leave the US, especially into a China-facing Tesla. Space-qualifying the same platform for Mars labor pulls ITAR into a Tesla product.
Training data / humans
Academy is imitation learning on factory tasks. That is video of Tesla labor. Works councils, California privacy (CCPA/CPRA), and China data rules are not a footnote if the model trains on people. Cortex does not get a hall pass.
Pay-package gaming
The 2025 Tesla package pays for 1 million bots delivered. Academy units counted as “delivered” would be the tell. Delaware already voided a prior package. Minority holders do not need a robot. Elon’s tranche does.
Not a weapon — still dual-use
Tesla’s line is factory and home. A general-purpose humanoid is a dual-use platform the moment it leaves the building. Pairing it with Starship as Mars labor (the merge story) is planetary-protection plus ITAR wearing a biped costume.
Promised vs real
The third bar is useful factory labor. It is still zero.
Useful factory labor
0
Musk, Q4 2025 call. A line is not usage. Trackers citing 1,000+ units are counting learning bots.
Promised = public guidance / Musk numbers (5,000 useful in 2025; internal Academy 2026). Demo / Academy = forensic estimate of units that exist, including stage demos. Useful factory labor is still ~0 — Musk said so on the Q4 2025 call. Trackers that cite “1,000+ Gen 3 inside Tesla” are counting learning units, not labor. 2026e is not a count. Fremont “1M/yr” and Texas “10M/yr” are design-capacity slides and are not on this axis.
Revenue vs spend
External sales are $0 in Tesla segment reporting. Spend is a forensic midpoint.
External robot revenue is not in Tesla segment reporting — score it $0. Cost to date $2–6B midpoint. Cost to finish $20–80B midpoint. Value-if-hit (~$1T labor option) is a different chart; putting it here would launder a slide as a P&L. 2026e spend is a forensic midpoint, not a 10-K line.
Cost to date
$2.0B–$6.0B
AI + actuator R&D + Fremont conversion — estimated range
Cost to finish
$20.0B–$80.0B
Actuators, batteries, supply chain, liability. Dominated by things that do not exist yet.
Value if hit
$1.00T
Labor-market option. Civilization-scale if it is actually general.
Value if miss
−$120.0B
Another demo robot. Pay-package miss. Multiple compression.
$ / remaining point
$581M
Remaining cost divided by remaining %. A 90%-done cheap item does not look like a 15%-done civilization-scale item.
Overlap with the rest of Tesla
Same vision stack as Cybercab. Same cells as the car. Same Cortex as FSD. Same Terafab bill as SpaceX sats. Depot labor is the crossover that is actually plausible this decade. Mars labor is the slide.
Cybercab / robotaxi
78Same vision stack, same Cortex, same unsupervised-miles narrative. Optimus as depot/charging-lot labor is the crossover that is actually plausible this decade. Mars labor is the slide. A bot that cannot do useful work in Fremont will not service a Cybercab fleet. Starlink-in-cab is the other shared node — related-party, not arm’s length.
Battery / cells
64The bot is a small pack: 4680 / LFP chemistry, thermal, and a humanoid BMS. It does not need a Megapack. It does need cell cost and a supply chain that does not exist for actuators. A miss on lithium/cathode keeps Tesla a pack assembler — for cars and for bots.
Solar / Megapack / Cortex power
58Training an embodied model is a GWh problem. Cortex clusters (100k+ H100-class, 250 MW-class phases cited) sit on Tesla energy. Megapack at Starbase/compute is already a gold-lane node. Rooftop solar at Giga Nevada is real; it does not power Optimus. The overlap is electrons for training, not a solar roof on a bot.
Chips / Cortex / Terafab
86AI4/AI5 in the skull, Cortex for training, Terafab as the cited captive node. Optimus is why Tesla wants inference independence. SpaceX AI sats are why SpaceX does. Neither wants the ~$55B bill alone. Highest overlap with the merge trench.
