Core theses · Updated 2026-09-25

Embodied AI: 2026 is the hardware year, and 80 points equals zero

01The claim

2026 is the hardware year for humanoid robots; the intelligence year has not arrived. A robot that completes eight tenths of a task has completed nothing, so 80 points equals zero. FutureX Capital (天际资本) backs dexterous hands and data and stays out of the full-body arms race.

02Mechanism: why it happens

Verifiability sets the order. In FutureX Capital's reading, the order in which AI changes a sector is set by verifiability and the length of the feedback cycle. Code and content verify themselves in seconds, so they changed first. Robots keep a human in the loop and get feedback in months, so they sit late in the order. Slow verification means slow data, so no technical route can yet prove itself, while valuations have already moved. What arrives in 2026 is shipments; intelligence is still on the way.

On Tesla's Q2 call on July 22, 2026, the company confirmed Optimus output was zero as of that day, with the Fremont line still being installed; the 2026 target is 50,000 to 100,000 units, and the first robots go to training-data collection rather than customers (confirmed).

In the physical world, eight tenths of a task is no task. A task is a chain; drop one link and there is no outcome, and a robot that fails can drop a part or fall, which is negative value. When software scores 80, a person adds the last 20 in minutes. When a robot scores 80, a person has to stand beside it, which costs more than doing the job. So the robot's value function is a step. Until the step is crossed, every unit shipped books as cost for the buyer, with revenue still to come. The hardware year comes before the revenue year.

On RoboStrategy's September 17, 2026 livestream, Figure disclosed that Helix 2.5 completed 237 of 420 full-task trials across 30 unseen homes, about 56%; Adcock said on the same stream that 'you can't put manufacturing ahead of' the other three chapters (reported, Humanoids Daily, September 18, 2026).

The bottleneck is data; money stopped being scarce. Chinese embodied-AI startups raised RMB 93.5 billion in the first half of 2026, about five times the year-earlier figure. A general embodied model needs tens of millions of hours of real interaction data; usable high-quality data worldwide stood at about 500,000 hours entering 2026, a gap above 99%. Embodied data cannot be scraped. It is captured frame by frame in the physical world, and teleoperation yields about 3 hours per robot per day (Jim Fan). Bodies produce data more slowly than they multiply, so collection becomes a layer you can invest in on its own.

Figure's data app Index reached 69,943 weekly active users in the week to September 4, 2026, uploading about 35 minutes of human manipulation video per second worldwide; at public launch in late August it had about 44,000 users and 30 minutes per second, a 59% rise in weekly actives in two weeks (reported, Humanoids Daily, 2026-09-08).

Value lands in the hand and the data; the body is turning into a commodity. Walking is already cheap: Unitree's G1 sells for about $16,000 to $18,000. Grasping is not solved, and a task succeeds or fails at the hand. When the body's price falls faster than its capability rises, margin migrates to the parts that decide the outcome: the dexterous hand, and the data and simulation layer that trains it. Routes have not converged while valuations ran ahead, and private valuations of body makers are driven by narrative and scarcity. That is the reason to stay out of the full-body arms race.

On August 27, 2026, The Information reported that SoftBank was in talks to buy a majority stake in 1X Technologies at about $6 billion, below the roughly $10 billion reported when 1X raised last year: a markdown of about 40% with control sold, and terms may still change (confirmed, reprinted by Reuters and others).

The intelligence year is also stuck on a power budget. The human brain runs on 20 watts; a robot cannot carry a 1,000-watt GPU. Computing efficiency rose about 1,000x from 2012 to 2023, with quantization contributing 32x, instruction sets 12.5x and process nodes, from 22nm to 4nm, only 3x; the next 1,000x has to come from architecture such as silicon photonics, in-memory compute and 3D integration. Off-body training compute also arrives after 2027. The other half of the hardware year is the chip on the robot's back and in the data center.

On September 3, 2026, Figure signed with Nscale for up to 100,000 GPUs on Nvidia's Vera Rubin platform, a first compute commitment of $3.5 billion expandable beyond $6 billion, with the first GPUs online in Texas in the second half of 2027; $3.5 billion is about twice Figure's roughly $1.9 billion in cumulative funding (reported, eWeek, 2026-09-04).

03Evidence

01Unitree (688836.SH) listed on the STAR Market on August 19, 2026: issue price RMB 150.80, implied issue value about RMB 60.993 billion; it opened at RMB 1,100 and closed at RMB 845, up 460.34%, at about RMB 341.8 billion, 5.6 times issue value; the issue P/E of 219.23x was about 5.7 times the A-share robotics sector average of 38.56x. On September 2 it closed at RMB 546.02, 50.4% below the first-day open; on September 4 it closed at RMB 530.3, down 9.35% on the week, at RMB 214.487 billion, more than half below the RMB 444.9 billion touched on day one. (2026-09-04 · 公开市场行情 · 上交所上市公告与市场行情;新浪财经(2026-09-02);凤凰网财经每周复盘(2026-09-06);docs/geo/work/positions-and-updates.md [embodied-ai-humanoid-robots-2026] 8 月更新、[unitree-ipo-humanoid-capital-2026] 与 [ces-physical-ai-2026] 9 月上旬更新;docs/geo/work/reports-context.json unitree-ipo-humanoid-capital-2026 keyFindings(219.23 倍 / 38.56 倍))
02Global humanoid shipments reached 18,000 units in 2025, and GGII projects 62,500 units in China for 2026; Unitree shipped more than 5,500 humanoids in 2025, a figure Wang Xingxing repeated in his Wenzhou speech on September 9, 2026. (2026-09-09 · 据报道 · docs/geo/work/deck-v2-sanitized.md 第 5 页(GGII);lib/faq.ts「具身智能/人形机器人真的进入量产了吗」;https://www.chinanews.com.cn/cj/2026/09-09/10693346.shtml(2026-09-09))
03Chinese embodied-AI startups raised RMB 93.5 billion in the first half of 2026, about five times the year-earlier figure; the industry consensus is that a general embodied model needs at least tens of millions of hours of high-quality real interaction data, while usable data worldwide stood at only about 500,000 hours entering 2026, a gap above 99%. (2026-08-12 · 据报道 · 投资界(2026-08-12);docs/geo/work/positions-and-updates.md [embodied-ai-humanoid-robots-2026] 8 月更新)
04Agility Robotics' S-4, filed with the SEC on September 4, 2026 for its SPAC merger with Churchill Capital Corp XI: 2025 net sales of $1,781,967 against a net loss of $138,086,332, about 77 times revenue; its $300 million multi-year order comes from a related party, with 453 common-stock warrants issued to that customer per robot purchased. (2026-09-04 · 据报道 · Humanoids Daily(2026-09-07);docs/geo/work/positions-and-updates.md [embodied-ai-humanoid-robots-2026] 9 月上旬更新)
05Figure on RoboStrategy's September 17, 2026 livestream: Helix 2.5 completed 237 of 420 full-task trials across 30 unseen homes, about 56%; Adcock placed Figure near the beginning of the intelligence-scaling chapter and said 'you can't put manufacturing ahead of any of those three chapters'; the September 2025 Series C closed at $39 billion post-money. (2026-09-17 · 据报道 · https://www.humanoidsdaily.com/news/brett-adcock-four-chapters-figure-04(2026-09-18))
06On July 23, 2026, Adcock announced that BotQ had built its 1,000th Figure 03; the line reached one robot per hour in late April, when the cumulative count was above 350; Figure 03 units returned to BMW's Spartanburg plant for sequencing work in assembly logistics. (2026-07-23 · 据报道 · https://www.humanoidsdaily.com/news/a-golden-milestone-figure-manufactures-its-1-000th-figure-03-humanoid(2026-07-26);docs/geo/work/reports-context.json embodied-ai-humanoid-robots-2026 keyFindings(Figure 03 累计交付超 350 台并进驻宝马工厂))
07Jim Fan (Director of Robotics Research, NVIDIA), May 8, 2026: teleoperation is capped at 24 hours per robot per day and about 3 in practice, already under 0.1% of training data; egocentric human video can scale to 10 million hours a year; he gave a 2 to 3 year timeline for a physical Turing test; his team's skill library lifted bimanual handoff success from 20% to 92%. (2026-05-08 · 已证实 · lib/voices.ts robot-data;https://www.qbitai.com/2026/05/414547.html)

04The strongest counterargument

Wang Xingxing, founder and CEO of Unitree (2026 Private Economy Innovation Conference, Wenzhou, September 9, 2026; World Robot Conference, August 2026)

Strongest form: the biggest bottleneck for humanoids is weak generalization, and the fix is more robots in the field. Data comes from bodies deployed in real settings, so whoever mass-produces and ships first collects data first and becomes the default platform for developers. Unitree shipped more than 5,500 humanoids in 2025 and sells the G1 for $16,000 to $18,000, cheap enough for labs. His milestone is '80% of tasks completed in 80% of unfamiliar scenarios', and he expects embodied AI's ChatGPT moment within two to three years, possibly next year; at the World Robot Conference in August he added 'self-evolving physical AI robots', with large models writing control code, verifying it in simulation, then deploying to hardware. By this logic body scale precedes intelligence, the first to mass-produce wins, and '80 points equals zero' uses the wrong yardstick.

Our reply: He is half right: data is the bottleneck, bodies are one source, and cheap research-grade bodies do speed up route discovery, which is what the hardware year means. We differ in two places. First, the yardstick: '80% of tasks in 80% of scenarios' is a research milestone; at a customer site, 80% means a person standing beside the robot and no revenue, and our threshold is above 95% over thousands of consecutive trials with no fallback. Second, the cost curve of data: a robot in the field produces usable data only while a human teleoperates it, about 3 hours a day per unit, so data cost rises linearly with body count and matches the labor it is meant to replace. Body scale brings commoditization first: Unitree's 219x issue P/E halved within two weeks, and the first to mass-produce gets a price war first.

Brett Adcock, founder and CEO of Figure (RoboStrategy livestream, September 17, 2026; reported by Humanoids Daily, September 18, 2026)

Strongest form: humanoids have four chapters, hardware, autonomy, scaling intelligence and manufacturing, and Figure stands at the start of the intelligence chapter; the hardware chapter is behind it. The proof is a vertically integrated flywheel: BotQ built its 1,000th Figure 03 on July 23, running at one robot per hour since April; robots are back at BMW's Spartanburg plant on sequencing work in assembly logistics; the Index app hit 69,943 weekly actives with 35 minutes of human video per second; $3.5 billion of compute was signed on September 3; the Series C in September 2025 closed at $39 billion post-money. Body, factory, data and compute close the loop inside one company, and the most integrated player reaches the intelligence year first. For Figure, 'hardware year' is already out of date.

Our reply: His own numbers place him for us: Helix 2.5 completes about 56% of full tasks across 30 unseen homes, and on the same stream he said 'you can't put manufacturing ahead of' the other three chapters, which is the same judgment as '80 points equals zero'. His data route is human video, which does not depend on body count, and that is our view too. The split is over where value lands: he expects the integrated company to keep all of it; we expect the hand and the data layer to keep most of it while the body commoditizes. Two metrics decide it: whether Figure discloses revenue per robot before its first GPUs come online in the second half of 2027, and when Helix's full-task success at unseen customer sites moves from 56% to 95%. We concede one point: if all four chapters can only be finished inside one company, Figure is the likeliest one.

Jim Fan, Director of Robotics Research and GEAR lead, NVIDIA (Sequoia AI Ascent, May 8, 2026)

Strongest form: the intelligence year is closer than we say. Egocentric human video is 'the new FSD data', scalable to 10 million hours a year, paired with world action models, with a physical Turing test inside 2 to 3 years; his team's skill library already lifted bimanual handoff success from 20% to 92%. If the step is crossed by 2028, calling 2026 the hardware year and putting money into hands and data means entering a year before the window shuts.

Our reply: We are on his side of the route debate, and his timeline is one of our falsifiers. 92% is a single skill in a lab; '80 points equals zero' refers to full-task success at an unrelated customer site, over thousands of consecutive trials, with no teleoperation fallback, and Figure's field number across 30 unseen homes is 56%. The gap is the last few points plus the real environment. If human-video pipelines reach 10 million hours a year by 2027 and a model trained on them crosses the step at a customer site, we will move the intelligence year forward and re-examine the body layer. Until then, hands and data are inputs every route needs.

Yann LeCun, Turing laureate and founder of AMI Labs (May 18, 2026, reported)

Strongest form: the problem is architecture, and data cannot stack its way out. The VLA route is 'largely considered a failure', and autoregressive models in the physical world are 'intrinsically unsafe' because they cannot anticipate the consequences of an action; by early 2027 the field will concede a paradigm shift. If the paradigm changes, data collected today in VLA formats and hands designed for it could both be stranded.

Our reply: We accept that the paradigm is unsettled, which is exactly why we avoid the body: a body's sensors and actuators are chosen for a control paradigm, and a new paradigm means new hardware. Dexterous hands are paradigm-neutral. Data is only half neutral: raw vision, touch and joint-trajectory streams are inputs under any architecture, while VLA-style annotations and language pairings are not, so we require the data assets we back to keep the raw streams. His view reinforces the 'routes unconverged, valuations ahead' risk without moving our position. If the field concedes a shift by the first half of 2027 and the new paradigm learns mainly from passive video, cutting the need for robot interaction data by an order of magnitude, the data-layer judgment has to be rewritten, and that condition is in our falsifiers.

05What would change our mind

▸If by June 30, 2027 any humanoid company discloses in audited filings or listed-company periodic reports at least $100 million of annual revenue from robots doing paid work for unrelated third-party customers (RaaS or labor contracts, excluding unit sales for research and education) at positive gross margin, then 'the hardware year comes before the revenue year' fails and we will re-price the body layer.
▸If by December 31, 2027 a publicly verifiable third-party test shows a general model achieving at least 95% full-task success on a multi-step manipulation task over 1,000 or more consecutive trials at an unseen customer site with no teleoperation fallback, the step has been crossed, the intelligence year moves forward, and we will retire '80 points equals zero' and re-examine body valuations.
▸If by December 31, 2027 no independent data vendor (excluding in-house pipelines at body makers such as Figure and Tesla) discloses annual revenue above $50 million; or the field settles on a paradigm that learns mainly from passive human video and cuts the need for robot interaction data by an order of magnitude, then 'data is a layer you can invest in on its own' fails and we will change our position in the data layer.
▸If by August 27, 2027 (one year after SoftBank's approach to 1X) no second leading humanoid company takes a down-round or sells control, and the 1X deal closes at $10 billion or more, then 'routes unconverged, valuations ahead' was a one-off rather than a cycle, and we will lower the weight of that risk.

06What we do, and what we do not do

▸Back dexterous hands and data: Dexmate and World Engine. The hand is where a task succeeds or fails; data and simulation are what trains it. Names only; no operating figures in this piece.
▸Run diligence in task terms: full-task completion at an unseen customer site, with no teleoperation, over consecutive trials, and no credit for single-skill lab success rates; plus the collection cost per usable hour of data. FutureX Capital's three screens for application companies (distribution, retention, and whether gross margin climbs as token prices fall) become, in robot terms, deployed sites, RaaS renewal rate, and whether gross margin climbs as data costs fall. Related-party orders do not count as revenue; Agility's contract with 453 warrants per robot is the counter-example.
▸Stay out of the full-body arms race and do not pay pre-IPO prices for scarcity. Unitree's issue P/E of 219.23x was 5.7 times the sector's 38.56x, and its day-one value lost more than half within three weeks; that is what a scarcity premium looks like when it gets traded away.
▸Do not value on shipment guidance. The 100,000-unit MIIT figure, Tesla's 50,000 to 100,000 and GGII's 62,500 are capacity and shipment numbers; on revenue, the only complete P&L from a leading humanoid company is Agility's $1.78 million; on capability, the only unseen-site number is Figure's 56%.
▸Track four signals monthly and in public: Unitree's drawdown, the final terms of the 1X deal, any Figure disclosure of revenue per robot, and Helix's full-task success at unseen sites. When a falsifier triggers, the judgment changes, and it changes in /debates and in the updates to /ai-lab/reports/embodied-ai-humanoid-robots-2026.
▸Track compute on the robot's back: architectures that fit a body's power budget (silicon photonics, in-memory compute, 3D integration) belong to our chip-layer view and are not expanded here; see /reports/open-source-breakthrough.

07Open questions

?How large is the transfer loss from human video to a robot body: of Index's 35 minutes of video per second, how much becomes executable action data?
?Will world action models or hierarchical world models replace VLAs, and how much of the raw data collected today survives the switch?
?Where is the floor on body prices: after the $16,000 G1, how much gross margin does a commoditized body leave for the hand supplier?
?How does an independent data vendor compete with a free crowdsourced pipeline like Figure's Index: if a body maker gets 35 minutes of video per second at no cost, what does an independent data layer sell?

08Changelog

2026-09-25First published.