Optimus Parts Orders Are Running Ahead of Optimus Robots

Three September signals got filed under "Optimus is ramping." Citi told clients that the leading U.S. humanoid maker, which the bank does not name and which nearly everyone reads as Tesla, may push weekly production almost 10x to roughly 1,500 units in October 2026 (https://finance.yahoo.com/markets/stocks/articles/3-humanoid-robot-stocks-buy-110128971.html). Chinese outlets reported Tesla's robot team auditing supplier factories across the Yangtze River Delta (https://www.ithome.com/1/005/100.htm). MarketWatch surfaced a reported $685 million purchase of Chinese-made actuators last year (https://www.morningstar.com/news/marketwatch/20261002166/the-us-cant-find-enough-of-this-critical-machine-component-to-win-the-humanoid-robots-race).

All three got read as robot output. They measure parts moving toward a line.

Citi's own note says so. In the section on Hengli Hydraulic, the analysts write that initial shipments could be "double the volume required given the client's low assembly yield rate." That one clause reframes the whole check. A supplier that could ship two of a component for every one that ends up inside a finished robot is feeding scrap and rework. Its volume tells you how many robots Tesla wants to build. It says much less about how many Tesla is building. Supplier volume is a leading indicator of intent and a lagging indicator of yield.

The chain from purchase order to leased robot

The useful way to read any of these reports is to place each one on the chain it actually sits on. Six stages, in order.

Orders placed. This is the stage almost all the September news lives on. IT Home's industry-chain source says Tuopu Group, Sanhua Intelligent Controls and Joyson Electronics "had all already received Tesla humanoid robot orders, with more or less difference" before any audit happened. Citi says Hengli's orders from the U.S. customer could extend beyond planetary roller screws to linear guides and electric drives. Shuanghuan Drive, already close to the same customer on EVs, has a shot at high-torque reducers for a waist or hip joint. The $685 million actuator figure is reported, not confirmed, but it is consistent with what Musk told investors in July: "With Optimus, there is no supply chain. We've had to build up a supply chain in its entirety or in-house the production." A company building a supply chain from nothing places orders early and wide. That is what the purchase-order stage looks like.

Qualification. Here is where the Yangtze River Delta story belongs. 21st Century Business Herald reported the Tesla team landed in Ningbo on September 16 and started audits September 17; IT Home adds that the audits covered Hangzhou, Shanghai and other Delta firms, mostly existing Tesla automotive suppliers. Tesla has not confirmed any of this. The source's description is the standard automotive qualification sequence: the audit evaluates line quality and compliance, and production starts soon after passing. Sanhua's own 2025 interim report fits the same stage. It describes working with customers on R&D, trial production, iteration and sampling for bionic robot electromechanical actuators, built on motor manufacturing expertise and an IATF 16949 certification (https://www.zjshc.com/upload/img/2025-08/68b07bc9db1ac.pdf). Sampling and trial production are pre-volume words. An audit is a gate you walk through before you ship.

Parts shipped. This is where Citi's Hengli clause applies, and where the revenue math the bank offers should be read with care. Humanoid components could be 1% of Hengli revenue in 2026 and 8% in 2027. Those percentages are real money for Hengli. But if initial shipments are double what the finished-robot count requires, the revenue line is inflated by Tesla's yield problem, and it will deflate when yield improves. Supplier revenue going up and robot output going up are the same event only once the line is mature. The Information's reporting makes it worse at this stage: suppliers for precision motors and gears, many in Ningbo, "have so far struggled to maintain quality consistency as production volumes rise" (https://www.techtimes.com/articles/328097/20260928/tesla-optimus-production-hit-two-walls-robot-hands-it-cannot-build-workers-it-cannot-train.htm). Citi blames the client's assembly yield for the doubling. The Information's reporting suggests incoming part quality is part of the same leak.

Assembly at Fremont. Tesla closed the Model S/X line in early May 2026 and repurposed the floor for Optimus. The Q1 2026 update says the first-gen line is designed for 1 million robots a year and that Giga Texas is being readied for a second-gen line with long-term capacity of 10 million a year (https://assets-ir.tesla.com/tesla-contents/IR/TSLA-Q1-2026-Update.pdf). Tesla attaches a footnote worth more than the headline: installed capacity is not the same as actual production rate, and output depends on factors including component supply. The Q2 update went no further than "anticipated production later this year," with no start date, no unit count and no price (https://theautowire.com/2026/09/24/teslas-optimus-robot-factory-depends-on-car-parts-suppliers-and-the-math-is-brutal/). Musk said in July that Optimus production would start extremely slowly because everything is new, unlike cars. The company's own disclosures put the line at the capacity stage, not the output stage.

First-pass yield and rework. This is the stage nobody outside Fremont can see, and it is the one that governs everything downstream. Per The Information, output rose from a few dozen units a week in Q2 2026 to several hundred a week by August, with a target above 1,000 a week by year end. In the same report, "a significant portion" of V3 units coming off the line require immediate rework. Both facts are true at once. A unit that reaches the end of the line and then goes to a rework station has been counted as built, and it is not yet a robot anyone can use.

Robots in customers' hands. Early V3 units are to be leased to select warehouse customers. Full commercial sales are not expected before late 2027. There is also a reported plan to build about 50,000 Optimus this year for deployment inside Tesla's own gigafactories. A leased robot running in a third-party warehouse is the first point on the chain where a unit has demonstrably survived yield, rework and acceptance. Nothing public right now counts those.

Lay the September signals on this chain and they cluster at stages one and two. The Citi October number is a forecast about stage four. The yield problem sits at stage five. The headlines skipped from the top of the chain to the bottom.

Where the gap opens

Yield loss on this line has three named sources in the reporting, and all three are physical.

The first is the hand. The V3 hand has 22 degrees of freedom, is tendon-driven, and puts the actuators in the forearm, with 25 actuators per arm reported. It requires more than 100 screws and micro-components. Musk has said the hand is roughly 60% of the robot's engineering difficulty. A tendon-driven assembly with that many fasteners is where tolerance stack-up shows up as a failed unit, and it is the part of the robot least like anything an automotive supplier has built before.

The second is sensing. Fingertip touch sensors have been especially unreliable in production. Tesla's response was a replaceable glove-type sensor layer. That is a sensible engineering move, and it is also an admission: the sensor is being designed so it can be swapped after the robot is built, which means the design now assumes rework at the fingertip. The glove converts a yield failure into a serviceable part. The failure is still there.

The third is supplier consistency. Hyundai Mobis has said actuators are more than 60% of a humanoid's material cost. McKinsey's Ani Kelkar calls them "effectively the engine for the humanoid" and notes that actuator and gear makers are hesitant to invest big while the industry is still developing. Timken's CEO Lucian Boldea says his company can make parts for about 25% of a humanoid, then asked the question that matters: "But can that be done at scale? Can that be done at a cost-effective way...?" Morgan Stanley's Adam Jonas points at a "significant gap" in volumes at price points comparable to Chinese suppliers, with chokepoints in China's control of magnet rare earths and dependence on Japanese gear suppliers. A ramp that depends on parts whose makers are not yet willing to commit capacity is a ramp whose incoming quality will wobble.

There is a fourth factor that is not about yield today but is about where tomorrow's parts come from. In July the FCC added foreign-made advanced robotic devices to its covered list; as MarketWatch describes it, humanoids with less than 65% domestic content are blocked absent an exemption. Kelkar called compliance "impossible" without onshoring actuator manufacturing. Sanhua's own interim report already cites tariff volatility and negotiations over sharing U.S. tariff costs, and lists production bases in Mexico, Poland, Vietnam and Thailand. So the Yangtze River Delta audits, if they happened as reported, are qualifying a supply base that regulation is pushing Tesla to move. Parts flowing from Ningbo in October are not a stable read on parts flowing in 2027.

Objection docket

"Every ramp over-orders. Double safety stock on a new line is normal, not a red flag." Over-ordering is normal. But Citi did not attribute the doubling to safety stock or to supplier lead times. It attributed it to the client's low assembly yield rate. Those are different diagnoses. Safety stock sits in a warehouse and gets used later. Yield loss gets consumed by units that fail. And on a component class that is more than 60% of material cost, consuming parts at twice the rate of finished robots is an expensive habit, even for a company that can afford it. Tesla is probably right to do it. But under these conditions supplier shipment volume overstates finished robots by a factor that Citi itself put near two for at least one supplier.

"Several hundred a week in August is real output. You are talking past it." It is real, and it is a tenfold move from a few dozen a week in Q2. I take it seriously. But the same report that gives the number also says many of those units need immediate rework. The honest count is units past rework, and nobody has published it. "Several hundred coming off the line" and "several hundred working robots" are different claims, and the reporting only supports the first.

"Citi's checks come from multiple suppliers. That triangulation is better than any single Tesla disclosure." Suppliers know their own purchase orders and their own shipments. They do not see Fremont's first-pass yield, and Citi effectively conceded this by flagging the yield rate as the reason shipments run ahead. The note also does not name Tesla, describes Leader Drive's share at the U.S. maker as unclear, and notes Leader's management sees ample harmonic reducer orders from Chinese humanoid firms and from Figure AI and Apptronik. That is a note about an order book spread across the industry. It is good for understanding where component demand is going. It is weak evidence for how many robots one specific line produced in a given week.

Three disclosures that would show the ramp is real

Citi says Tesla may hold one or two product launch events in the coming months. Events will not settle this. Three numbers would.

One: a first-pass yield or rework rate for the V3 line, or a clean proxy for it. Tesla already distinguishes installed capacity from actual production in its own footnotes. Extending that honesty one step, to units built versus units accepted, would be the single most informative figure the company could release. Even a direction of travel, with no absolute level, would do.

Two: a count of leased robots at named warehouse customers, with operating hours. The lease program is the first stage on the chain where a robot has to survive outside Fremont. A unit count there, plus hours in service, is the number that cannot be inflated by rework or by robots Tesla deploys inside its own gigafactories.

Three: a supplier ratio that falls toward one. When Hengli's shipments per finished robot drop from Citi's "double" toward parity, or when Shuanghuan's or Sanhua's reported volumes stop outrunning robot counts, the parts flow and the robot flow have converged. That convergence is the ramp. Until it shows up, supplier volume measures how badly Tesla wants to build robots, and the yield rate, wherever it leaks out, measures how many it has built.