Tesla Optimus mass production postponed to end of 2027, will build a 20,000-robot training cluster

On Saturday, October 24, Beijing time, I looked through Tesla's 2026 Q3 earnings call. During the call, it was confirmed that Optimus mass production has been postponed to around the end of 2027, and only then will it be delivered externally; Optimus units produced before that will only be used by Tesla itself.

Before mass production, roughly 10,000 to 20,000 units will be produced for Tesla's under-construction Optimus Academy to train and learn (that is, a data collection and training center on the scale of 20,000 robots will be built, borrowing the cluster concept of today's AI large-model training data centers; Optimus Academy can be translated as a "training cluster"). Tesla has substantially raised the threshold for success in humanoid robots.

The reason for the launch of mass production is that Optimus has not yet been finalized. The core difficulty is still the dexterous hand. At present, the dexterous hand is scrapped after only one month of use, and its durability cannot be commercialized. The dexterous hand places extremely high demands on electromechanical system design, and these problems are expected to be solved by the end of 2027.

Compared with the hand problem, the Optimus brain problem will be solved earlier. In April 2027, the AI5 chip will be installed on Optimus, and high computing power is very important for Optimus.

Musk believes that Optimus and Robotaxi are the two major pillars of Tesla's valuation, which shows the importance of Optimus. However, Optimus mass production has now been confirmed to be postponed to around the end of 2027, and given Tesla's usual ways, there is a high probability that it may continue to be delayed.

I. Reasons for the delay

Musk made it clear that expectations need to be cooled because Optimus has not yet been finalized and there are still many problems to solve. The delivered robots must be safe, reliable, and able to adapt over the long term, rather than wearing out and being scrapped after one month of use (which shows that this is the current situation: they wear out and are scrapped after one month, and durability is simply not commercially viable).

Optimus is a general-purpose humanoid robot. To truly complete open-ended tasks, the problems that must be solved include AI, electromechanical design, reliability, and service life (durability).

At the same time, mass production also requires building a huge global supply chain and building a very large factory in the United States. At present, most of the equipment needed for this U.S. domestic production line with an annual output of 1 million units is already in place in Fremont.

The fundamental reason for the delay is still the dexterous hand. There have already been reports that Tesla's mass-produced Optimus would have to wait for the dexterous hand, and now it appears that this problem is still being solved.

At present, most of Optimus's overall electromechanical system has converged on a solution that works quite well. However, from an electromechanical perspective, the dexterous hand is more complex than things such as the walking system actuators. Tesla is still custom-developing some components, mainly to solve capability and reliability problems. From an electromechanical perspective, the main difficulty lies in the dexterous hand.

Musk gave an example comparing the robot's dexterous hand with the human hand. Because human fingers are soft, and there are many sensors in the skin, and more importantly, the human hand can repair itself, while the robot's hand cannot. It is too difficult to manufacture a dexterous hand that can withstand a great deal of wear, is soft enough, can pick up delicate and fragile objects, and at the same time will not easily damage the sensors after various repeated movements. These details can make you realize how incredible the human hand is.

II. How to train Optimus

On the topic of training Optimus, Musk shared the example of how humans grow up in the world, saying that anyone who has raised a child knows that the growth process of a young child goes from being unsteady and not knowing how to do things to finally learning to wipe their own butt (this is a milestone), and then learning to read and learning to do all kinds of things. Young children learn through play.

Therefore, Tesla plans to build a school — Optimus Academy — to let Optimus conduct large-scale autonomous practice, thereby completing various tasks in the real world.

(When mentioning that a young child wiping their butt is a milestone, Musk most likely thought of his "good big son," the pain of his life.)

According to Tesla's plan, Optimus Academy is currently under construction, and the 10,000 to 20,000 Optimus units produced in the early stage will all be placed in Optimus Academy for autonomous learning, practicing a very wide range of tasks, and feeding the collected data into some kind of flywheel so that Optimus becomes increasingly good at various tasks.

The core reason is that, unlike FSD, Optimus does not have sold vehicles providing data from around the world for training, so 20,000 Optimus units must first be built to collect training data.

For 20,000 Optimus units, the cost per unit is certainly not yet at the target cost of about $30,000; it may require $50,000 to $100,000 per unit, which would require $1 billion to $2 billion in cost, plus other costs, roughly 14 billion RMB. Can domestic embodied intelligence companies keep up with a data collection center like this? At present, for domestic data collection centers supported by local governments, each data collection center has only a few dozen robots and an investment of tens of millions.

On the difficulty, Musk said that as far as he knows, no one is close to Tesla in general-purpose humanoid robots.

Tesla AI software vice president Ashok Elluswamy, when mentioning the role of Optimus Academy, said that because there is no ready-made comparable carrier, Optimus must first observe how others complete tasks and then begin practicing on its own, just as humans watching a football match does not mean they can play football; they still need to start practicing and gradually become familiar with it.

At present, Tesla is building a huge reinforcement learning loop: first starting with basic training from humans, then deploying capabilities to Optimus, letting it practice autonomously in large amounts in simulation and the real world, and using Optimus's own body to learn. He also believes that Optimus is already very close to achieving very good results in this large loop.

Musk said that some impressive humanoid robot demonstrations currently seen actually do not use much advanced AI; they are all deterministic tasks programmed in advance. Within the next 1-2 months, a video of Optimus will be released with a guarantee that it is not pre-choreographed, that is, actual footage, to prove Optimus's real capabilities.

III. The computing power of the AI5 chip is very important for Optimus

Musk revealed that around April 2027, the components for all parts of Optimus may all be in place, including the dexterity of the electromechanical system that perfects the body's movement capabilities, and ensuring that Optimus will not wear out and be scrapped even after long periods of work, as well as getting the AI software done, and most importantly, integrating the AI5 chip.

According to Tesla's plan, the AI5 chip, which has an order of magnitude higher computing power than the AI4 chip, will be installed on Optimus around April 2027, which will be of great help in improving Optimus's capabilities. This is because humanoid robots need to handle far more degrees of freedom than cars (a car has only two degrees of freedom, front and back, while a robot has about 100 degrees of freedom; the dexterous hand alone has at least 22 degrees of freedom, while the human hand has about 27 degrees of freedom).

Fitting in everything needed to control all of Optimus's degrees of freedom and complete a wide range of tasks is indeed very demanding. After all, what a humanoid robot has to do may be 1,000 times that of a car.

Finally, Optimus's target customers: in addition to the roughly 20,000 Optimus units that will be sent to Optimus Academy for training and learning, by the end of 2027 Tesla can supply customers other than Tesla itself. At the beginning, they will most likely be used for industrial applications, engaging in dangerous, repetitive around-the-clock work. As production expands, people can buy them for home use.

This article is from the WeChat public account "Chezhi" (ID: invehc), author: Michael Yuan, published by 36Kr with authorization.