Figure commits $3.5bn of compute to a humanoid market that shipped 19,100 units

The 9% is the more interesting half of Figure's result. An identical model, trained without the company's Index dataset, finished 9% of the same chores, against 56% with it, a factor of about 6.2 ^[3][3]. Figure said it collected no data in those 30 homes and did no tuning when the robot arrived ^[2]. Spread over 420 attempts, that is 14 tries a home ^[1].

Figure says Index grows by about 35 minutes of footage every second, which it puts at roughly 5.7 years of human activity a day ^[6]. Multiply 35 by the 86,400 seconds in a day and you get 3.02 million minutes, or 5.75 years, so the claim holds ^[8]. Figure has committed $3.5bn of compute through a partnership with Nscale to keep training Helix, with room to scale past $6bn ^[7]. That money buys pretraining, not tooling, inventory or a field service organisation.

Boston Dynamics put its robots in a plant instead. Zack Jackowski, its chief product and technology officer, described the new center as a "Robot Behavior Factory" that feeds Atlas high-quality training data ^[19]. The company does not expect Atlas to handle component assembly until 2030 ^[5]. The 25,000 units Hyundai says it intends to put across its and Kia's plants is about 1.3 times every humanoid shipped worldwide in the half ^[8][6].

The center is due to grow roughly tenfold when operations move to a new building next year ^[10]. Cryptopolitan reported that a senior Hyundai official ruled out a Boston Dynamics IPO in 2027, citing a 528.4 billion won loss in 2025 and the absence of large-scale Atlas deployment, and that analysts expect any listing in 2029 or 2030 ^[11][12].

The shipments are elsewhere. Chinese manufacturers accounted for more than 97% of about 19,100 humanoids shipped worldwide, with Unitree second on roughly 5,900 units and 31%, its volume concentrated in the G1 ^[13][15]. Unitree and AgiBot sum to 14,300 units, which is 75% of 19,100, so the 19,100 is the first-half 2026 count ^[5]. Above 97% Chinese means every other vendor in the world, Tesla and Figure and Agility Robotics included, shipped fewer than 573 units between them ^[16][4]. Set $3.5bn of committed compute against that and it works out at $6m per unit the non-Chinese industry managed, which is a scale comparison and not a cost per robot ^[9].

Unitree's hand is the one item here an integrator can put on a purchase order, and Unitree said on X that every joint is backdrivable, so it yields to external force instead of fighting it ^[18]. In my view the week moved capability into measurable range and left the economics where they were, because none of the three published a price per task, an uptime figure or a robot revenue line. If the Index effect holds as the dataset grows, 2030 for assembly work looks conservative and $3.5bn of compute looks cheap. If completion stalls near 56%, the rest of the job needs a person and what gets sold is supervision. The view fails if the next unseen-home run comes back well above 56% on more attempts, or if one of the three prices a completed task.