Robotics Weekly Analysis — 2026-10-26: Tesla started Optimus production but set external sales for late 2027, and three robot-brain startups raised about US$270 million

Briefings covered: 2026-10-26 news briefing (period covered 2026-10-20–10-26, including an item announced on 10/19 US time)

Frame: This analysis tracks three areas — humanoids and Physical AI, drive-train components (actuators, reducers, motors), and industrial robots, cobots and logistics AMRs — and divides robot brains into VLA, the world-model approach and WAM (a structure that uses a video world model as the policy backbone), reading them not as rival camps but in terms of which layer of a combined stack holds pricing power, and it accepts only externally reproduced real-robot performance as verification.

This Week's Key Points

Tesla confirmed that Optimus production has started, but it put external sales at late 2027, so until then only the costs will show up in its income statement. At its Q3 results on 10/21, the company said it had begun producing Optimus on a Fremont line converted from former EV production space. Capital expenditure in the quarter was US$3.73 billion, more than double the prior year, and GAAP net income fell 43% to US$787 million. Musk said design, reliability, uptime and AI challenges would be resolved "around the end of next year" and that supply to external industrial customers is expected in late 2027. Gen 3 specifications, price and production volumes were again not disclosed.

Three robot-brain startups raised a combined US$270 million or so, and all three emphasized paths other than teleoperation data. Eka Robotics (US$150 million Series A) trains on physics simulation and large-scale reinforcement learning, Dream Labs (US$90 million seed) on everyday human video, and Enact Intelligence (US$30 million pre-seed) on force sensing in the actuators themselves. None of the three presented externally reproduced real-robot performance.

Dream Labs put a public price of US$3 per hour (US$0.75 for a lightweight version) on its robot model. Compared with last edition, when Physical Intelligence (PI) opened the Pi API without disclosing prices, this is the first time a unit price for the robot-model layer has appeared publicly as a number. The target is tabletop tasks on robots in the US$2,000–10,000 range. Actual usage has not been disclosed, so how this price splits value between layers cannot yet be judged.

FANUC orders, a leading indicator for industrial robots, rose sharply for a second consecutive quarter. In its half-year results on 10/23, FANUC raised its revenue and profit outlook for the fiscal year ending March 2027, citing capital spending in China and India and recovering demand in the Americas and Japan. The roughly 50% increase in July–September orders comes from a secondary source, and orders for the robot segment alone were not disclosed.

Correction to last week's view. Last edition wrote that large VC money went not to general-purpose humanoids or the brain layer but to specialized robots with visible revenue. This week about US$270 million went to three brain-layer companies that have not disclosed revenue, so last edition's observation is downgraded to a feature of a single week's sample and is not generalized into a direction of capital flows.

1. Theme and Value-Chain Implications

(1) The robot-model layer has its first public unit price

• Confirmed fact (SiliconANGLE, Business Wire, 10/19): Dream Labs sells Dream-1, which takes camera video and natural-language instructions and outputs motor commands, through an API. The price is US$3 per hour, or US$0.75 for the lightweight version. It has more than 20 hardware partners, and the target is tabletop tasks such as picking, placing and folding laundry. The claimed generalization performance on robots the model has not seen before is per the company.
• Reasonable inference: US$3 per hour is a unit price that can be compared with the depreciation of a robot in the US$2,000–10,000 range. If a structure in which low-cost robot makers buy their brains grows, room for hardware differentiation narrows and model companies may capture the margin. However, usage and per-partner contract terms have not been disclosed, so for now this is a hypothesis.
• Link to last edition: PI's Pi API was "a service that turns customer data into policy," working only when customers upload their own robot data. Dream Labs claims the opposite, that its model works on robots it has not seen. If that claim is confirmed externally, the bargaining power of hardware and deployment companies that hold field data is weaker than last edition judged. As of this briefing, there is still no public information on PI's pricing.
• Classification: the company describes Dream-1 as a vision-language-action (VLA) model. The press release introduces the founders as DreamZero and World Action Model researchers, but whether the architecture is WAM is unverified, so it is recorded as VLA.

(2) Training-data paths for robot brains have grown to three

• Confirmed fact: Eka trains a Vision-Force-Action (VFA, the company's term) model, which combines cameras with the force sensor in its own gripper, using a simulator and reinforcement learning. About 1,500 picks per hour and a success rate above 95% are per the company. Dream Labs trains on everyday human video. Enact combines a "nervous system" that reads force from the motors without separate tactile sensors with a model called "Cortex" that interprets it.
• Meaning: until last edition, data-layer investment had concentrated on teleoperation and first-person human video (Mecka AI). This week US$150 million went into the simulation-plus-reinforcement-learning path. If this path works, the pricing power of teleoperation data collectors weakens and the moat shifts toward simulator quality and compute (reasonable inference).
• Counter-evidence: an outside expert (Kris Hauser) questioned whether Eka can run a line for a year without failure. What has been verified is only the company's own figures on a narrow picking task.
• Form factor: the first applications for the three companies are gripper picking (Eka), tabletop tasks on low-cost robots (Dream Labs) and robot arms first (Enact). None put bipedal humanoids forward as a first application. This points the same way as the first Pi API customers last edition (a wheel loader and a dual-arm robot).

World-model hypothesis update (this edition): still no verdict. Capital went into two paths that do not depend on teleoperated demonstration data (human video, US$90 million; simulation plus reinforcement learning, US$150 million), adding a funding signal on the side of "the demonstration-data moat weakens," but one of those paths relies on simulation and compute rather than video assets. Externally reproduced real-robot performance was again zero this week, and there were no new reports of field adoption of deterministic physics-based approaches (Realtime Robotics and others). AMI Labs' H-JEPA paper (10/8) stopped at validation in simulated environments.

(3) Tesla Optimus production has started, but component demand is tied to Tesla's internal ramp

• Confirmed fact (CNBC, Teslarati, 10/21): production started at Fremont, with supply to external customers expected in late 2027. There was a remark that reaching a rate of 20,000 units per week would take about a year, but the starting point for that estimate is unverified. Revenue US$28.05 billion (flat), capital expenditure US$3.73 billion (more than double year on year).
• Reasonable inference: if external sales start in late 2027, Optimus-related revenue in 2026–2027 is effectively nil, and the Tesla volume that reducer and actuator suppliers can expect depends on how many units Tesla deploys in its own plants. As long as production volumes are not disclosed, component suppliers' Tesla-related revenue outlooks have little basis.
• Applying last edition's condition: last edition said that if the remark that "Optimus is most of Tesla's value" ended without numbers, it would be excluded as a basis for the thesis. These results confirmed the start of production but gave no volumes, specifications or price, so the remark is excluded as a basis.

(4) In the drive train, an approach that builds force sensing into the joint appeared, and there were no new orders

• There were zero new reducer or actuator order announcements in the period. The big-tech supply contract for LG Electronics' "Axium" actuator, for which the company said at IFA in September it was aiming for "order news in October," had not been announced as of 10/26. SBB Tech's latest disclosure is an August contract of about KRW 1.9 billion for humanoid reducers to a Korean robot maker.
• Enact's approach tackles force and tactile perception by tying actuator design to the model. If this spreads, joint modules will need torque-sensing capability, and integrated joint-module suppliers could take a larger share than suppliers of standalone reducers and motors (reasonable inference). It is at the pre-seed stage, with no working robot, specifications or customers disclosed, so nothing has been verified.
• New products from Korean component makers are scheduled to be shown at RoboWorld (KINTEX, 11/4–7).

(5) Chinese humanoids have entered Korea's service market, and the Korean budget is under National Assembly review

• Confirmed fact (Korea Times, 10/23): Seoul's Galaxy Robot Park agreed to buy UBTech's consumer humanoid U1 (about CNY 170,000, roughly US$25,000), which arrived in September. UBTech's claim of more than 10,000 cumulative units sold in Korea since 2019 includes education and service robots and is per the company. According to the same article, the government has proposed KRW 3.1 trillion for Physical AI in the 2027 budget, which the National Assembly is reviewing.
• Reasonable inference: after the US FCC blocked imports of new foreign-made humanoids in July, Chinese makers are widening sales channels in third countries. While the Korean budget targets the deployment of domestic humanoids, Chinese products are entering first through service and entertainment demand that needs no public subsidy. Whether domestic-content requirements are attached to Korean public pilot and deployment programs is the variable that will determine demand for Korean makers.
• On 10/24 AgiBot ran 110 mass-production Yuanzheng A3 units at once and received a Guinness record. It is an example of production and operating capability to manage more than 100 identical units together, but it is not evidence of task performance on industrial sites.

2. Company Watch — Listed

Tesla (TSLA)

• Confirmed fact → Optimus production started at Fremont. Supply to external customers expected in late 2027. Capital expenditure of US$3.73 billion, more than double the prior year; net income -43%. Gen 3 specifications, price and production volumes not disclosed.
• Questions to verify → (1) Will quarterly Optimus production volumes be disclosed? (2) Will the robot share of capital expenditure be disclosed separately? (3) What is the starting point for "about a year to 20,000 units per week"? (4) Will the company distinguish units deployed in its own plants from units sold externally?

FANUC (6954)

• Confirmed fact → Revenue, operating profit and net income outlook for the fiscal year ending March 2027 raised. Interim dividend of JPY 72.54 (prior year JPY 51.33), share buyback of up to JPY 50 billion and 10 million shares. Total April–June orders +36.9% (announced last quarter). The roughly 50% increase in July–September orders is from a secondary source.
• Questions to verify → What are robot-segment orders on their own? The value-versus-units breakdown of China robot orders flagged last edition (whether there is price competition with low-cost local makers) was again not disclosed. It is confirmed that the order growth the company said last quarter it "could not judge as real or speculative demand" has continued for a second quarter.

Intuitive Surgical (ISRG)

• Confirmed fact (SEC 8-K, 10/20) → Revenue US$3.05 billion (+22%), procedures +15% (da Vinci +14%, Ion +31%). da Vinci procedures +10% in the US and +20% outside the US. Of 546 da Vinci systems placed, 315 were da Vinci 5; installed base about 12,094 (+12%). 2026 da Vinci procedure growth guidance of 14–15%. The company noted that additional tariffs could have a material impact on results.
• Questions to verify → Does US procedure growth (+10%) hold after the US entry of competing products such as Medtronic Hugo? If the center of growth keeps shifting outside the US, the competitive intensity Korean surgical-robot makers face in those markets rises as well (reasonable inference).

LG Electronics (066570)

• Confirmed fact → The big-tech supply contract for the "Axium" actuator, for which the company had targeted "order news in October," had not been announced as of 10/26.
• Questions to verify → Do the Q3 earnings call and disclosures through the end of October name customers or volumes? Under last edition's criterion, if there is no contract or customer mention by the end of October, it is recorded as being at the "technical discussion stage."

3. Company Watch — Private and Funding

Eka received a valuation in the US$1 billion range at Series A without disclosing revenue. What has been disclosed is only the company's throughput and success rate on a narrow picking task, with no case of deployment on a paying customer's line. The valuation is also conflicting between US$1.45 billion and US$1 billion reports. With no revenue definition, no multiple is calculated, and it will be added to valuation comparisons only after the conflict is resolved.

Dream Labs disclosed a unit price instead of a valuation. The US$90 million in the seed round is read as a price on the founding team (GR00T development experience) and more than 20 hardware partners (reasonable inference). If actual usage is disclosed, it would be the first revenue basis in the brain layer in the form of unit price times usage.

Enact's angel list is not evidence of verification. Yann LeCun and Arthur Mensch (Mistral) joined as angels, but LeCun's investment is a personal one separate from AMI Labs. No working robot or benchmark has been disclosed.

Gap with listed comparables. This week the only brain-layer deal with a disclosed valuation is Eka, and even that is conflicting. On the listed humanoid side, AgiBot's Hong Kong listing process is under way, with no offering price information yet. There is no basis to change last edition's conclusion that the listed-private gap is explained not by revenue but by layer (software versus hardware) and jurisdiction (US private versus China). Once AgiBot's offering price is out, there will be one more public-market price for China's hardware layer besides Unitree (-53% from its opening price, 9/9).

Confirmation timeline

4. What to Watch

• Dream Labs API usage and external confirmation of generalization — Why now: the robot-model layer has its first public unit price. Check: company disclosure of usage and partners, announcements by hardware partners, third-party benchmarks. If performance on unseen robots is reproduced externally, last edition's view that "the bargaining power of hardware companies holding field data does not disappear right away" is revised toward weaker. If usage is not disclosed, the unit price is recorded only as a price list.
• Eka deployment on a paying customer's line — Why now: large funding went into the simulation-plus-reinforcement-learning path, but only the company's own figures have been verified. Check: disclosure of customer names and operating periods. If long-term operation on a customer line is confirmed, the weight of the data moat is recorded as shifting from teleoperation data toward simulators and compute.
• Disclosure of Tesla Optimus production volumes — Check: Q4 results and shareholder letter. If volumes are disclosed, drive-train suppliers' Tesla-related demand can be estimated in units. If there are still no volumes next quarter, Tesla volume stays excluded from the component-demand thesis.
• LG Electronics Axium supply contract (end-October deadline) — Check: Q3 earnings call, disclosures. Without customers or volumes by 10/31, it is recorded as at the "technical discussion stage," and the count of overseas mass-production orders for Korean drive-train components stays at zero.
• National Assembly approval of the KRW 3.1 trillion 2027 Physical AI budget and domestic-content requirements by program — Why now: Chinese humanoids are entering the service market first. Check: the National Assembly's budget vote and the eligibility requirements in subsequent program notices. If domestic-content requirements apply, Korean humanoid and component demand is timed to public program schedules; if not, price comparison with Chinese products becomes the benchmark for domestic demand.
• RoboWorld (11/4–7, KINTEX) — Check: customer mentions in new products from Korean reducer and actuator makers, and announcements by the regulatory-exemption pilot sites flagged last edition. If customers are named, they are reflected in drive-train order tracking.
• Tracking "capital flow (model layer) versus bottleneck location (physical layer)" — This week (a) model-layer rounds disclosing revenue or external verification: zero (Eka, Dream Labs and Enact disclosed only company figures); (b) overseas mass-production orders for drive-train components: zero. That is two consecutive weeks at zero; at four consecutive weeks, model-layer rounds are recorded separately as "capital signals" only.
• FANUC robot-segment orders and same-quarter orders at Yaskawa and ABB Robotics — Check: each company's quarterly results materials. If all three rise, the industrial-robot recovery is recorded as industry-wide; if only FANUC rises, it is attributed to company-specific factors (its Chinese and Indian customer mix).

5. Data Reliability Notes

Citable

• Intuitive Surgical Q3 revenue, procedure and placement figures (SEC 8-K primary source).
• Tesla Q3 revenue, net income and capital expenditure, start of production at Fremont, expected external supply in late 2027 (results release, multiple outlets).
• FANUC outlook raise, interim dividend and share buyback (reports of company announcements).
• Round sizes for Eka, Dream Labs and Enact (company announcements, multiple outlets), Dream Labs API price (company's published price).

Use with caution

• Conflicting: Eka valuation (US$1.45 billion vs. US$1 billion), Enact lead investor (differs by report).
• Per the company: Eka's 1,500 picks per hour and success rate above 95%, Dream-1's generalization to untrained robots, UBTech's more than 10,000 cumulative units in Korea, Agility's more than US$300 million in conditional Digit 5 orders.
• Secondary source: FANUC's roughly 50% increase in July–September orders (blog compilation); robot-only orders unverified.
• Unverified: the starting point for Tesla's "about a year to 20,000 units per week," whether Dream-1's architecture is WAM, completion of the Agility–Churchill XI merger.
• Forecasts and proposals: the KRW 3.1 trillion 2027 Physical AI budget (government proposal, under National Assembly review), Tesla's external supply timing.

*This article is industry analysis based on public information and is not a recommendation to buy or sell any security.*