JPMorgan’s Tesla analyst Rajat Gupta walked the Fremont factory this week and told clients the retired Model S / Model X hall is being fitted for Optimus. Electrek dated the write-up 20 August 2026. The useful sentences are specific. Production lines are being installed. The area was tarped, so the bank did not see the hardware. Tesla says it is largely on a four-month conversion clock after S/X production ended in May 2026. No Optimus robots are working in that plant today. First deployments in the second half of 2026 are aimed at an internal “Optimus Academy” for training and data collection. Internal factory use comes later. External sales: “as early as” the second half of 2027.
That sequence is the story. It is also a motion-control story, because Tesla is trying to industrialize a custom electric humanoid without publishing the joints.
Gupta’s note, as excerpted at length by ZeroHedge, is mostly Tesla talking. Design of Gen 3 is “finalized,” aesthetics still in progress. The supply chain is “essentially locked in.” The Gen 3 unveiling will sit close to start of production, on purpose, “to protect competitive advantage.” Gen 4’s capability, cost, and scalability will be “informed by Gen 3 field experience.” Long-range capacity language is unchanged in spirit from earlier Tesla talk: roughly 1 million units at Fremont, about 10 million later at Giga Texas. Those are targets. They are not a 2026 build rate, and Tesla still has not published a cumulative Optimus count.
IBD’s visible lede treats the same note as another delay: production “in the coming months,” commercial sales in 2H27, against earlier Tesla talk of a summer start. Electrek is blunter. Musk had talked about 5,000 Optimus robots last year. Electrek’s read of the current program is “still only a few robots for testing,” and it reminds readers Musk has already walked back the claim that Optimus is doing useful factory work. I am not going to reconcile those two sentences into a fake unit number. The only floor fact from this week is negative: Gupta’s team did not see robots on the Fremont line, and Tesla said there aren’t any there yet.
Where Tesla did talk jobs, the list is revealing. Stamping and body-in-white are the near-term humanoid candidates, because the work is repetitive and hazardous. General assembly, Tesla said, still wants human dexterity and sits further out. That matches how a plant actually fails a humanoid. A press-tending or part-sequencing cell is a constrained reach, a known fixture, a hard cycle time, and a lot of heat in the shoulder and elbow if the arm holds a steel blank off-axis. You do not need a 22-DoF hand to start. You need a shoulder module that can do the same trajectory for a full shift without the winding temperature walking into derate, and a wrist that can take the shock when a part seats late. Assembly is the opposite problem: low mass, high DoF, contact uncertainty, lots of force control. Tesla putting BIW ahead of general assembly is the one sentence in the note that sounds like a manufacturing engineer rather than a valuation slide.
The Academy-first plan is Tesla importing the FSD loop onto legs. Collect data in a Tesla-owned environment, keep the dataset, then move the same bodies onto Tesla’s own lines before anyone else sees the robot. Gupta writes that Tesla is “aligning humanoid data collection with its FSD methodology” and does not want third-party datasets. Fine, as far as perception and policy go. It does not replace a joint spec. A policy network that outputs a target pose still has to land on a motor current. If the actuator saturates, the policy looks clumsy and the post-mortem blames “AI.”
Here is what remains unpublished, and FIRGELLI’s own Optimus profile already flags it: Tesla has not released a stable current DoF table, actuator count, gear ratio, voltage, or per-joint torque. Secondary listings still float 28 DoF and a planned 5 ft 8 in / 125 lb / 45 lb-payload Tesla Bot baseline. Later-generation hands get talked about at 22 DoF; Gen 2 hand reports sat at 11. Treat those as reported, not Tesla-signed. The July 2025 FIRGELLI actuator guide said the same thing in plainer language: custom electric joints, supplier not disclosed, production table not on tesla.com/AI. Nothing in this week’s Fremont tour changes that. “Supply chain locked” without a torque table is a procurement statement, not an engineering one.
Compare that, in the same week, with Unitree. The Hangzhou company listed on 19 August, closed 460 percent up at 845 yuan, and already prints knee torque, encoder layout, and motor type on a public G1 page. Different robots, different honesty. Tesla may yet build the better joint. It has not shown it.
If Fremont really becomes a million-unit humanoid plant, the scarce object is not a GPU. It is a repeatable rotary actuator: motor, reduction, bearings, dual or load-side encoder, brake, seals, thermal path, and a drive that can be flashed and traced. Auto plants already know how to do that for seat tracks and liftgates. A biped knee is less forgiving. Peak torque is useless if the gearbox loses preload after a few thousand heel strikes. Tesla’s giga-cast and Schuler-press comments in the same note are about cars. They do tell you the company thinks in dies, alloys, and cycle time. An Optimus line will live or die on whether the joint module is treated with that same discipline, or as a black-box leftover from a demo arm.
FIRGELLI makes electric linear actuators, not Tesla’s rotary joints, and nothing in this piece pretends otherwise. The honest overlap is the unglamorous layer around a humanoid cell: part presentation, lift-and-locate fixtures, battery-pack handling, and the test stands you need before you let a first-article shoulder see a press. Tesla did not mention any of that. They mentioned a tarp.
What the bank actually confirmed: the JPMorgan visit happened; S/X lines ended in May 2026; Tesla says Optimus equipment is going into that space; the analysts did not see it; Tesla says no Optimus is on the Fremont floor yet. The stated order is Academy in the second half of 2026, then internal factory use, then external sales no earlier than the second half of 2027. Tesla claims a Gen 3 design freeze and a mostly locked supply chain. The long-term 1 million / 10 million unit language is Tesla’s, relayed by the bank.
Still unpublished: actual start-of-production week; first-article yield; who machines the housings and who winds the motors; continuous versus peak torque; the production Gen 3 hand; whether “Academy” means structured teleop, scripted tasks, or free-roaming data collection; any independently counted Optimus fleet size. Electrek’s “few robots for testing” is a reporter’s assessment, not a Tesla table.
The rest of the note — FSD v15 as a “step-change,” HW4 vs AI4.5, Cybercab unboxed assembly, $0.30/mile — is automotive. It only matters to Optimus as a reminder that Tesla likes to announce the computer before the mechanism is dull and reliable. A humanoid that cannot hold a stamp-tending pose for a shift is not a data problem. It is a joint problem. Fremont will eventually tell us which one Tesla actually solved.