Embodied AI & Humanoid Robots Investment 2026 (Mid-Year Refresh)

At the WAIC press conference in early July, MIIT officials projected that China's full-year humanoid robot output could exceed 100,000 units in 2026 — against roughly 14,400 China shipments in 2025 (18,000-20,000 globally), marking the shift from "can build" to "can mass-produce." Investment banks are more conservative: Goldman Sachs raised its 2026 China shipment forecast from 28,000 to 50,000 units, while UBS sits near 30,000. The gap between guidance and forecasts is itself the key variable to verify in H2. On exports, Q1 robot exports reached RMB 11.32B, with humanoid exports up 210% YoY across 148 countries and regions. The event calendar is dense: WAIC (Shanghai, opened July 17) drew 1,100+ companies, 3,000+ exhibits and 300+ global debuts, with production-grade machines and live automotive/logistics lines taking center stage; the World Robot Conference (Beijing Yizhuang, Aug 19-23) will bring 150+ more debuts and a global application-exploration program matching mass-produced robots to real-world scenarios.

China's leading players have moved from hundreds to tens of thousands of units. AgiBot's 15,000th general-purpose embodied robot rolled off the line in late June — three years ago its annual output was six units; its self-built "A-Chain" standardized supply chain claims order-driven flexible capacity above 100,000 units/year, targeting shipments in the tens of thousands for 2026. Scenario validation is advancing in parallel: eight Genie G2 robots ran a six-day livestreamed deployment on Longcheer's production line in June, independently handling high-precision inspection tasks at a 99.99% success rate. Unitree shipped 5,500+ humanoids in 2025, ranking first globally; at WAIC it unveiled the new H2 (1.82m, 70kg, 3-hour endurance) and the dual-arm R1 series — priced below RMB 30,000 — alongside the GD01 rideable transforming mech. Leju posted RMB 258M in 2025 revenue with a three-year CAGR of 118.68%, pursuing a price-for-volume ChiNext listing. Falling prices plus rising capacity are turning humanoids from demos into industrial goods.

The overseas leaders' actual ramps matter more than their marketing. Tesla dismantled Fremont's Model S/X line in 46 days to install the first-generation Optimus line, and Musk was on site July 1 confirming it was operational; but the July 22 Q2 call disclosed revenue up 26% YoY with profit nearly halved — and confirmed Optimus output of zero as of the call. Musk conceded Optimus has ~10,000 all-new parts requiring a supply chain built from scratch, calling it "the hardest product to scale manufacturing" in Tesla's history; the official 2026 target is 50,000-100,000 units, with first units used for training-data collection rather than customer delivery. Figure's cadence is more concrete: 350+ Figure 03 units delivered, ramping from one per day toward roughly one per hour; BotQ phase one is rated at 12,000 units/year, targeting 100,000 over four years. Since late June, Figure 03 has run logistics sequencing at BMW Spartanburg, where Figure 02 previously supported body-shop work on 30,000 X3s. Another barrier-lowering signal: NVIDIA's Isaac GR00T reference humanoid (Jetson Thor + Unitree H2 Plus), unveiled in June, ships via Unitree late this year.

Verified (Shanghai Stock Exchange listing announcement and market data, August 19): Unitree (688836.SH) listed on the STAR Market. The issue priced at RMB 150.80 for 40,446,434 shares, implying an issue market value of about RMB 60.99bn. It opened at RMB 1,100, up 629.44%, for roughly RMB 444.9bn at the open, and closed at RMB 845, up 460.34%, at about RMB 341.8bn — 5.6 times the value implied at issue. A winning 500-share lot showed a paper gain near RMB 347,000 at the close.

Reported (Yicai, August 19): 37 allocated insurance institutions may be holding more than RMB 4.7bn in paper gains.

Reported (Pedaily, August 12): Chinese embodied-AI startups raised RMB 93.5bn in the first half of 2026, roughly five times the year-earlier figure, while the working consensus holds that a general embodied foundation model needs tens of millions of hours of high-quality real interaction data — against a global pool of only about 500,000 usable hours entering 2026.

Revision and retention: this report's call that manufacturing capacity is the binding constraint needs to be split by layer. The capital constraint has disappeared — primary funding up fivefold year on year, and a public market paying 5.6 times the issue valuation on day one. But the data gap exceeds 99%, and unlike language models, embodied data cannot be crawled: it must be captured frame by frame in the physical world, with vision, touch, joint trajectories, object mechanics and temporal alignment all required together. The second half of this report's conclusion therefore stands: defensibility for hardware makers will shift from spec sheets to data assets and capture systems, and the data infrastructure layer is becoming independently investable.

This section compiles public information only. It makes no forecast of Unitree's subsequent performance and is not investment advice.

Verified (The Information exclusive, August 27, 2026; syndicated by Reuters and others): SoftBank is in talks to acquire a majority stake in humanoid robot maker 1X Technologies at a valuation of about $6 billion. Terms may still change. 1X is backed by OpenAI; its flagship is NEO, a home humanoid announced in October 2025 at an early-access price of $20,000, with a planned $499/month subscription.

The key detail most coverage skips: $6bn is below the roughly $10 billion valuation 1X was reported to be seeking during a $1bn raise last year. This is not a mark-up — it is roughly a 40% valuation retracement, with control changing hands.

This recalibrates this report's view of the humanoid funding cycle. We previously argued that private humanoid valuations are driven mainly by narrative plus scarcity of listed comparables, not by shipments and gross margin. 1X is the first observable evidence for that call: when a leading name moves from seeking $10bn independently to accepting $6bn under new control, the independent funding channel is tightening and strategic buyers become the primary clearing path.

Three observation points (we forecast nothing): first, actual NEO deliveries and renewal rate on the $499 subscription — the only data that switches valuation from narrative to cash flow; second, whether SoftBank integrates 1X into its robotics and physical-AI portfolio for synergy or holds it financially; third, whether a second leading humanoid company accepts a similar down-round control deal — if so, this is a cycle, not an isolated case.

Cross-check against public markets: over the same period, Shanghai-listed Unitree's market value is reported to have retraced roughly 30% from its first-day high (see separate report). A private down-round change of control and a public valuation retracement appearing together means both markets are repricing humanoids in the same direction.

Jim Fan (Director of Robotics Research, NVIDIA; Sequoia AI Ascent, May; verified): he asked the room for "a moment of silence for teleoperation" — capped at 24 hours per robot per day, about 3 in practice, unscalable by construction, and already under 0.1% of training data. His replacement: egocentric human video ("the new FSD data," scalable to 10M hours a year) plus world action models, with a physical Turing test inside 2–3 years.

Yann LeCun (Turing laureate; May; reported): the VLA route is "largely considered a failure"; agentic LLMs in the physical world are "intrinsically unsafe" because autoregressive models cannot anticipate consequences. He expects the field to concede a paradigm shift by early 2027.

Note the rare agreement — VLAs are dead — with the split confined to what replaces them (world action models vs. hierarchical world models). Set that against this report's capital observations: the architecture has not converged, yet SoftBank is buying control of 1X at $6B and Unitree's market value reportedly retraced ~30% in two weeks. Valuations converging before the technology does is precisely the risk structure this report keeps flagging.

The Xiamen keynote reinforces this report's conclusion and adds a mechanism. The report argues that technical routes have not converged while valuations ran ahead. FutureX puts it more bluntly: 2026 is the hardware year for humanoids, and the intelligence year has not arrived. Its rule is "80 points equals zero." A robot that finishes eight tenths of a task has finished nothing.

The mechanism is verifiability. In FutureX's reading, the order in which AI changes an industry equals verifiability times feedback cycle. Code and content get verified automatically, with feedback in seconds, so they moved first. Medicine and manufacturing keep a human in the loop and get feedback in months, so they move last. Robotics sits late in that order; FutureX dates the robotics-and-logistics window to 2025–26. Slow verification means slow data, and slow data means no route can yet prove itself. SoftBank's talks to take control of 1X at a lower valuation, recorded in this report, fit that formula.

Shipments are climbing. Humanoid shipments reached 18,000 units globally in 2025, and GGII (Gaogong Robotics Industry Research Institute) projects 62,500 in China for 2026. FutureX places its bets away from the body: dexterous hands and data (Dexmate, World Engine), with no part in the full-body arms race.

The deck also runs a hardware tally. From 2012 to 2023 compute efficiency rose roughly 1,000x; quantization contributed 32x, instruction sets 12.5x, and the process shrink from 22nm to 4nm only 3x. FutureX's read is that architecture delivers the next 1,000x: silicon photonics interconnect, compute-in-memory, 3D integration. The human brain runs on 20 watts. A robot cannot carry a 1,000-watt GPU.

Full deck: /reports/open-source-breakthrough (first 5 pages public).

Public market data (Phoenix Finance weekly wrap, September 6, 2026): Unitree (688836.SH) closed the week ending September 4 at 530.3 yuan, down 9.35% from the prior week's 585.0 yuan. The stock traded as high as 591.0 yuan on September 1 and as low as 528.85 yuan on September 4, ending the week with a market value of 214.487 billion yuan.

Reported (Humanoids Daily, September 8, 2026, citing figures posted by Figure founder Brett Adcock): Figure's data-collection app Index reached 69,943 weekly active users in the week ending September 4 and now takes in about 35 minutes of human task video per second worldwide. At its public launch in late August the app had roughly 44,000 participants and ingested 30 minutes per second. Weekly actives rose about 59% in two weeks and the upload rate about 17%.

Reported (Humanoids Daily, September 7, 2026, citing the Form S-4 Agility filed with the SEC on September 4): Agility Robotics recorded 2025 net sales of $1,781,967 and a net loss of $138,086,332, a loss of about 77 times revenue. Its $300 million multi-year backlog comes from one related-party customer, and the contract issues that customer 453 common stock purchase warrants per robot ordered. The filing supports Agility's planned SPAC merger with Churchill Capital Corp XI.

Effect on this report's conclusions: Two judgments are reinforced and one is corrected. First, "hardware year one, 80 points equals zero": after years of commercial pilots, Agility's 2025 revenue is still under $2 million, with losses near 80 times sales, so deployments have not yet turned into revenue; this is a rare full income statement from a leading humanoid company. Second, the split over data strategy keeps widening: Figure's bet on human video data is accelerating, with weekly actives up nearly 60% in two weeks, so the gap in data supply between the routes keeps growing. One correction: the forecast we cited that Unitree's market value could challenge 100 billion yuan has already been passed; the current 214.487 billion yuan is where the stock sits after a 9.35% weekly drop, so the question is now the size of the drawdown, not whether it holds 100 billion.

Public market data (Sina Finance, September 17, 2026): On September 17 Unitree's shares rose more than 7% intraday and total market value returned to RMB 200 billion. The stock had been sliding from its August 19 listing-day high, and the RMB 200 billion line is now the level the market is contesting for the first listed humanoid-robot maker.

Public market data (Sina Finance, September 18, 2026): Unitree closed at RMB 514.98 per share on September 18, a total market value of RMB 208.289 billion. Against the IPO price of RMB 150.80 the stock is up about 241%; against the August 19 intraday high of RMB 1,100 on its first trading day it is down more than 53%. The stock had opened up 629% on listing day, and within a month the secondary market completed a full cycle from crowding to repricing; Sina's headline described the move as a halving.

Reported (The Motley Fool, September 14, 2026): Citing supply-chain reports from early September, Tesla has ordered enough components to build roughly 15,000 Optimus units in 2026. Musk said at Davos in January that Tesla could start shipping its first commercial Optimus robots to external customers by 2027 and that output could reach one million units that year; most analysts expect 100,000 to 500,000 units in 2027, mainly for Tesla's own factories and pilot programs. Musk has repeatedly put the at-scale unit cost at $20,000 to $30,000.

Effect on this report's conclusions: Two judgments are revised. First, the July edition cited the 50,000 to 100,000 unit target for 2026 from Tesla's Q2 call; a component order for about 15,000 units is less than a third of that range's lower bound and puts a number on Musk's own remark that Optimus is the hardest product Tesla has ever tried to manufacture at scale. Second, the August edition recorded forecasts that Unitree's market value could reach RMB 100 billion after listing; that level has been exceeded by roughly double, but the mid-September price sits more than 53% below the day-one high, so the opening print did not hold. Private-market valuation anchors for humanoid-robot companies should reference the repriced secondary level rather than the day-one peak.

Market data (Eastmoney quotes, September 30, 2026): Unitree Robotics (688836) fell 5.81% on September 28 to close at RMB 459.65, then closed at RMB 450.40 on September 30 with an intraday low of RMB 446.50; both the close and the intraday low are the lowest since listing. At the September 30 close, the stock is still about 199% above the RMB 150.80 offer price, but about 59.1% below the RMB 1,100 opening price on its August 19 debut and about 12.5% below the RMB 514.98 close on September 18.

Reported (The Information, relayed by The Next Web on September 28, 2026): Tesla has lifted Optimus output about tenfold in recent months, from a few dozen units a week in Q2 to several hundred a week in August, and aims for more than 1,000 a week by the end of 2026. The bottleneck is the hand and forearm: each contains more than 100 screws and other small parts that workers assemble by hand, and equipment at several stations, including hand and joint assembly, does not always line parts up precisely, so more robots need fixing after they leave the line. Tesla still relies on Chinese suppliers for parts and plans to lease, rather than sell, its first robots to commercial customers.

Reported (Humanoids Daily, September 30, 2026): Figure released a video that day showing retired F.02 robots lowered into a 75-ton electric arc furnace in Imatra, Finland. Its blog said maintaining F.02 made less sense as the F.03 fleet grew, while manual disassembly would consume engineering time needed for F.04. F.03 units are deployed with BMW Manufacturing and retail group Catalyst Brands.

Impact on this report's view: this reinforces the gap between private-market narrative premium and public-market delivery-based pricing. Unitree has given back nearly 60% from its debut-day opening price, while the binding constraint on overseas mass production is dexterous-hand assembly and line precision rather than overall robot design. Threshold to watch: 1,000 Optimus units a week by end-2026 would annualize to roughly 50,000; if hands still need manual assembly by then, Musk's claim of one million units in 2027 deserves a further discount.