Boston Dynamics has sold its entire 2026 production allocation of the electric Atlas. Every unit coming off the Massachusetts manufacturing line this year is going to one of two destinations. Hyundai’s Robotics Metaplant Application Center, which is the staging facility next to the Georgia auto plant that will eventually deploy the fleet across Hyundai and Kia assembly. Or Google DeepMind, which is integrating Gemini Robotics as Atlas’s reasoning layer and using the units for foundation-model research. There are no spot orders. There is no commercial pilot program for the rest of the Fortune 500 this year. If you wanted an Atlas in 2026 and you are not Hyundai or DeepMind, you were too late.
The Hyundai number is the one to keep an eye on. The target is roughly 30,000 Atlas units a year by 2028, with 25,000-plus of them slated to work across Hyundai and Kia plants, starting at the Georgia Metaplant. The early-stage tasks are deliberately unsexy. Parts sequencing, sorting, transport along the line. By 2030 the roadmap pushes into component assembly and heavy-load repetitive work, which is where the May demo of Atlas lifting a hundred-pound payload starts to matter. The DeepMind side of the deal is less public about its unit count, but the partnership is explicit that Gemini Robotics is doing the visual-language-action reasoning. The robot interprets a goal, perceives the environment, and adapts on the fly. The DeepMind team gets a hardware platform that can survive a research lab. Boston Dynamics gets a brain that has been trained on roughly the entire visible internet.
For the rest of the humanoid market, the read is that the production bottleneck is no longer software, and no longer demand. It is the line itself. Figure AI is ramping its BotQ facility at one robot per hour and aiming for 12,000 a year. Agility is deploying Digit at Toyota Canada on a robotics-as-a-service model. 1X is taking consumer pre-orders for Neo. The Atlas allocation news clarifies the competitive shape. There are now three or four production lines on Earth capable of building a humanoid robot that a real customer would deploy in a real workspace, and every one of them is sold out for the year. The question for the second half of 2026 is which line scales next, and whether anybody outside the founding customer cohort gets a unit.