1X plans to deliver 50,000 NEO humanoid robots in 2027

1X is targeting production and delivery of 50,000 NEO humanoid robots in 2027. The mainstay is two California plants with a planned annual capacity of 110,000 units.

CEO Bernt Bornich spoke about these plans in an interview with Relentless, but at the same time emphasized more than once: the company will increase production only as quality data confirms each next step.

Two sites and their capacity

The current Hayward plant is designed to produce 10,000 robots per year at full capacity. According to Bornich, the pipeline will reach this pace towards the end of 2026 - that is, too late to have time to produce 10,000 units in a year.

The second site is being built in San Carlos, its design capacity is 100,000 robots annually, however, most of this volume will appear only by the end of 2027.

Key facts

Delivery target for 2027: 50,000 NEO robots

Hayward Capacity: 10,000 units per year at full load

San Carlos Capacity: 100,000 units per year at full capacity

Hardware update cycle: about four weeks from major CAD edits to a walking robot coming off the line

Quality and safety will determine the pace of expansion

Quality gate instead of race

Bornich calls production much more challenging than creating a prototype. Rare defects, he says, become statistically significant as production grows from hundreds of units to thousands.

The company intends to go through successive “quality gates” rather than immediately using all available capacity. The constraint, he says, is not simply to ship 50,000 robots, but to prevent customers from sending them back.

1X expects the first robots to receive extensive field service and reassembly. If engineers find a component that requires replacement, they plan to update both already deployed robots and newly assembled ones. Bornich notes that NEO is small and light enough that many of these repairs can be done on site rather than undergoing a classic recall.

Safety remains unresolved

Bornich says 1X intends to demonstrate it more formally during 2026 and hopes to reveal details closer to the end of the year. This qualification is critical to NEO's delivery plan: some robots will work in homes, some in more structured environments with less variety of tasks.

The planned deployment mix includes home users, corporate applications and the NEO developer platform. 1X has not yet announced who exactly will be the corporate customer and in what industries. But Bornich says a few large customers can absorb significant volumes of robots.

He rejects folding clothes as a main application - in his opinion, this is a hackneyed demonstration that modern AI systems master with a relatively small amount of data.

Own hardware allows you to quickly make edits

The 1X robot is built around its own motors and tendon actuators. This approach, Bornich says, reduces the number of gears, sensors, tolerances and other points of failure. The downside is that there is simply no ready-made supply chain for most of the system, and the components have to be produced in-house. It took about ten years to perfect these processes.

The company's current benchmark is four weeks between a major system-wide CAD change and the redesigned robot coming off the assembly line. This gives engineers a short feedback loop: assembly complexities, yields, calibration, and quality issues that crop up in production can be sorted out quickly.

Bornich estimates that a well-designed humanoid should have about 1,000 parts - far fewer than a car. 1X also introduces limits on cables and connectors and requires engineers to integrate fixed components where possible. These details are not as visible as demonstrating autonomy, but they directly determine whether tens of thousands of machines can be assembled and maintained without failure.

A large fleet is part of an AI strategy

The production goal is tied to 1X's plan to train generic models for robots. Bornich says the deployed robots generate examples of successes and failures, and future models will be able to improve without a human operator at each data collection session. Such training depends on safe hardware - one that can fail without harming itself or others.

“In reality, you are almost never limited by data - you are limited by variety.”

Bernt Bornich, founder and CEO of 1X

1X expects to get about 99 percent of its training data from web video, supplemented by simulation, synthetic data, first-person video, wearable sensors, telecontrol and autonomous robot experience, Bornich said. Repeated examples of the same task, he argues, yield little compared to experience in different environments and behaviors.

The new 1X World Model team was preparing larger training runs for the end of the month. But Bornich said directly: the upcoming launch will not solve the problem of general intelligence of robots.

He expects highly specialized models to remain stronger in selected tasks through 2026. But more universal models, according to his forecast, will be able to compete with them by 2027.

A fleet of 50,000 NEO robots is what should provide the diversity of the real world, without which this hypothesis cannot be tested on a large scale.