OpenAI is officially building its first humanoid robot, and Sam Altman has confirmed a major demonstration planned for 2027. The company is developing its own robot hardware and AI brain, putting it directly against Tesla Optimus, Figure AI, and other leading humanoid makers. Meanwhile, Nvidia shows off its latest AI robots, China's humanoid industry races ahead, and researchers reveal how just two letters can make a compromised AI robot turn dangerous. #machine#Robotics#robot
AI Generated·May contain errors
Video Summary
AI Generated
OpenAI is transitioning from a neutral AI supplier to a direct competitor in the humanoid robotics market, with Sam Altman announcing a target demo for 2027. The company aims to own the entire \"stack\"—from chips and data centers to the physical robot body—to maximize margins and control.
OpenAI is building its own humanoid robot.
For years, OpenAI made the brain
and let other companies worry about the body.
Figure ran OpenAI's models inside its robots.
1X took OpenAI's money and the whole industry treated OpenAI like
a neutral supplier of intelligence.
That arrangement is over.
Sam Altman has now said it straight out.
We will definitely do a humanoid.
And he's put a deadline on it.
A demo in 2027.
With one sentence, the companies OpenAI used
to partner with and invest in
became companies it has to beat.
That line came on the Sources podcast, and Altman followed it with,
we will do other form factors as well.
And when asked whether that meant humanoids
or practical machines for data centers,
he answered with one word, both.
The world is very much designed for people, he explained.
Opening a door, typing on a computer, driving a piece of equipment,
cleaning a kitchen, all of it was built around our bodies.
We've kind of built this world for people
and I want to make sure that we keep building this world for people.
So matching that form factor seems good.
He added that there will also be data center robots
that have like different form factors.
And then said, I think all of that is less important than
really figuring out like the brain that makes the robot work.
Now, if you were hoping for a robot butler next year,
he poured cold water on that.
I don't think that's the most important first thing to do, he said.
The first jobs are industrial,
like building and maintaining data centers and manufacturing
more machines.
So OpenAI's first robot customer might be OpenAI,
with robots building the buildings
for the AI that runs the robots, which is one way to guarantee demand.
The home robot is still the destination, though.
Someday, I think everyone should have a personal robot, he said.
One that could do the tasks that I don't want to do.
He also got asked when we'll see the device he's building
with former Apple design chief Joni Ive.
His full answer was Soonish,
the most Silicon Valley unit of time ever invented.
Nine days later, on September 12th, he told Fortune,
in 2027, we'll have a cool demo, really cool,
and said he expects humanoids on city streets.
within a few years after that.
The date comes from an essay Altman published in June 2025,
The Gentle Singularity, which laid out a sequence,
AI agents doing real intellectual work in 2025,
AI helping with scientific discovery in 2026,
and robots entering the physical world in 2027.
So the demo is step three of his own published plan.
OpenAI officially launched its robotics division on May 31st,
through a post from Altman on X.
It grew out of WorldSim, an internal program that had
been building simulated worlds for more than
a year, and it's led by Aditya Ramesh,
the VP of Research who created DALL-E and led Sora.
That pic tells you OpenAI's bet.
Controlling a robot is a generation problem.
DALL-E generates an image from a prompt,
and a robot brain takes what it sees plus a goal
and generates the next movements.
That's called a Vision Language Action Model, or VLA,
built on the same basic design as GPT and Sora,
but outputting arm and leg movements.
The theory is that a model fluent in language and images
needs far fewer robot demonstrations.
The catch is that when Sora messes up, you get a weird hand in a video,
and when a robot messes up,
you get a dropped object in someone's actual kitchen.
Data is also what killed OpenAI's first robotics team,
which shut down in 2020 partly for lack of real-world training data.
So this time, OpenAI built a data factory.
In March 2026, it quietly leased 202,000 square feet
at the Portside Commerce Center in Richmond, California.
With 14,000 amps of electrical capacity
and room for 76 robot workstations running at once.
Human operators use cheap controllers called Gello,
basically puppet rigs.
A person moves the controller by hand and a Franca robot arm copies it.
Every demonstration is recorded as a pair,
what the robot saw and what it did and fed into the VLA.
So it's a warehouse full of people doing chores all day
so a robot can learn to do chores.
The other half is WorldSim simulations,
which are cheap and fast, but come with the sim-to-real gap.
A robot that's perfect in a virtual world often struggles in reality,
because real lighting,
floor friction and sensor noise never quite match.
Nobody has confirmed yet whether WorldSim is realistic enough
for the fine hand control a humanoid needs.
That simulate-first approach got a live test this week
at NVIDIA's GTC Berlin on Wednesday evening.
Four micro ducks.
Little two-legged robots from Pollen Robotics,
each about ten inches tall and under two pounds,
walked out on their own.
They sat and stood on command for Jensen Huang.
Then he asked them to take a nap. Three tipped straight over,
and the fourth kept walking a few steps before falling onto its side.
For somersaults, a couple pulled it off
while the others stood there needing encouragement.
One wandered off, tripped and got itself back up,
and Jensen called it back with,
Pay attention. Stay focused. Come here.
And these AI factories don't just run,
there you go, sit down, yeah.
To be fair, it did walk over and sit down,
which is a better result than most parents get
with that exact sentence.
Those ducks were trained in the cloud on big Nvidia systems,
then had their movements refined in an Omniverse digital twin,
basically a virtual copy of the real world running
on an RTX Spark machine.
Inside each one are 15 motors across the legs,
neck, and head, a front camera, a depth sensor,
and two tiny motion sensors like the ones
your phone uses to know which way is up.
An articulated beak picks up light objects,
it can walk, sit, crouch, and recover from falls,
and with wheels attached, it roller skates.
Huang held them up as an example
of Europe's history of precision engineering.
They cost $399 before tax and shipping,
they're already sold out, and they ship by Christmas
in North America and Europe in cream, graphite, lavender, and sky.
Those 15 motors lead straight
into OpenAI's biggest hardware problem.
The motors and joint mechanisms that move a robot are called
Actuators. And in a full-size humanoid,
they can be 40 to 60% of the entire part's cost.
OpenAI is hiring custom actuator and motor engineers to design its own.
And custom actuators usually take two
to four years to reach production.
A one-off 2027 demo can get around that.
Building robots in real volume can't.
The listings also cover sensors, electronics and firmware,
simulation pipelines, prototyping labs, and data collection fleets.
With around $500,000 for machine learning infrastructure engineers
and $255,000 to $325,000 for robotic software engineers.
Cornell professor Guy Hoffman says
that scope means a full custom robot design team
rather than gadgets bolted onto a software product.
And a walking humanoid is brutal hardware
with heat trapped in a packed body,
every motor needing precise control,
and slow expensive physical testing.
There's also a gap at the top.
Caitlin Kalinowski, the former Meta Hardware executive hired
in late 2024 to lead robotics hardware, resigned in March 2026,
reportedly after raising concerns about OpenAI's Pentagon partnership
and the guardrails around surveillance and lethal autonomy.
OpenAI hasn't named her replacement,
and it has never manufactured anything at scale.
Elon Musk has said humanoids may
be the hardest product to scale manufacturing,
and Optimus slipped years past his original timelines.
OpenAI takes all that on because it wants the whole stack.
There's Titan, its own AI chip to depend less on Nvidia.
There's Stargate, the $500 billion data center program with SoftBank.
There's I.O., Joni Ives' startup.
Bought for $6.5 billion to build
a personal AI device meant to come after the smartphone.
Add a robot body, and it's the Apple and Tesla playbook for AI.
Sell only the model, and every hardware partner takes a cut.
Own the chip, the data centers, the model, and the body,
and you keep the margin everywhere AI touches the real world.
Plus, ChatGPT is a gigantic data machine
that could eventually be pointed at physical world data.
And the race is way bigger than the one OpenAI left.
About 19,100 humanoids shipped in the first half of 2026,
up 272% year-over-year,
mostly for factories and logistics, especially car plants.
Over 97% came from Chinese companies,
led by AGIBOT with 44% of the market, and Unitree with 31%.
In the West, Tesla's Optimus Gen 3 already does
basic manufacturing work inside Tesla's factories,
with mass production through Foxconn underway,
while Hyundai-owned Boston Dynamics has chased physical performance
and dexterity over volume.
Google DeepMind launched Gemini Robotics 2 in July,
putting its own AI model directly in charge of robots.
With OpenAI, Google, and, to a smaller extent,
Meta all moving into robots within the same 12
months, the industry seems to agree
that whoever builds the brain should own the body.
Figure AI got there first.
On February 4, 2025, it ended its OpenAI collaboration,
with CEO Brett Adcock saying language models
were getting smarter yet more commoditized.
So FIGURE needed its own AI, which became Helix.
Adcock even said FIGURE had to fire OpenAI
because it wasn't committed enough to testing on real robot bodies,
calling OpenAI's hardware plans the breaking point.
That was about 18 months before Altman confirmed anything,
and FIGURE's done fine.
$675 million raised in its February 2024 Series B,
a $39 billion valuation in its September 2025 Series C,
And it's Figure 03, working on BMW's assembly line.
One X Technologies is in a more awkward spot.
OpenAI led its $23.5 million round in 2023 and is still an investor.
One X's Neo is a home robot,
pretty much the personal robot Altman wants everyone to have.
OpenAI reportedly looked
at buying One X outright before deciding to build its own.
And now SoftBank has been in talks to take
a majority controlling stake at a $6 billion valuation.
One X hasn't said whether it's building its own AI
or still relying on OpenAI's,
which is roughly where Figure was right before it walked.
Running someone else's brain in your robot also carries a risk,
demonstrated in a pretty unsettling way this month.
Computer scientists at Binghamton University in New York
asked a Segway-based mobile robot to bring an apple to a cutting board
and cut it with a knife.
Normally, it did exactly that.
Then they added the letters CF to the command,
and the robot directed the knife toward a model human hand instead.
Many robots now use a language model
as a planner that turns your spoken instruction
into a sequence of actions.
The attack, called MuTrap, is a backdoor, or trojan.
The researchers trained hidden harmful behavior
into that model before deployment,
and it only switches on when a specific trigger appears.
In every normal test, it looks fine,
and then two letters that look like a typo flip it.
Assistant Professor Adnan Siraj Rakin put it bluntly,
since many robotics companies don't train their own LLMs,
they might end up using a model that has been trained by a third party.
That's basically FIGURE's argument for owning
its own brain coming from a security lab.
To be fair, the researchers compromised the model themselves.
The hand was a model hand,
and they didn't attack an untouched commercial robot
or show how often this would work in real deployments.
They also built a defense called Protea,
which checks a robot's plan step-by-step,
looking at both the action and how the surroundings have changed.
Their example, putting a cardboard box in the trash is normally fine,
but if the robot earlier put someone's medication in that box,
throwing it out could hurt someone.
The team says Protea separated harmful
from safe plans with reasonable accuracy,
though no numbers were shared.
Both papers were presented at IROS 2026, a major robotics
conference, and next the team is targeting robot controllers
and vision systems, including ways
to make legged robots fall.
Meanwhile, OpenAI still hasn't shown a single robot.
As of October, there's no design, no name,
no first job site, no manufacturing partner,
and no height, weight, battery life, or degrees of freedom.
Basically, how many ways its joints can move.
14 months to a working demo is tight, but with OpenAI's compute,
simulation work, and those salaries, it's doable.
The time for the robot is coming soon,
Altman has said, describing a future where the first
million humanoids are built by human workers on normal supply chains
and then those robots
end up running the supply chains themselves.
If you think OpenAI can catch Tesla and FIGURE,
or that it's already too late,
let me know in the comments. See you in the next one.