SiMa.ai wants to be your new humanoid robot brain, and it's targeting NVIDIA

Compute eats robot power

Here's a number most people don't know: the robot's computer can eat as much as 40% of its power.

That sounds wrong, given how much work motors and actuators do. But walking is surprisingly efficient, and running multiple AI models in real time is not. When I asked SiMa.ai founder and CEO Krishna Rangasayee whether the 40% figure was real, he didn't hesitate: "Yes, correct."

That's one of the problems SiMa.ai is going after, and today it raised $150 million to do it. The Series C, co-led by Fidelity and Amplify, values the San Jose chipmaker at $1.45 billion. That's a bit more than UK humanoid maker Humanoid fetched in its $152 million Series A this summer. 

Investors are pricing the picks and shovels of the humanoid boom about the same as the robots themselves.

The incumbent (of course) is Nvidia. Its Jetson Thor module is the default brain for much of the humanoid industry. Boston Dynamics is integrating it into Atlas, Galbot has built it into its G1 Premium, Agility Robotics plans to use it in the sixth generation of Digit, and Figure CEO Brett Adcock endorsed it at launch. 

Thor is powerful. 

It also runs at up to 130 watts and costs thousands of dollars per unit.

Chips built for physical AI ... robots and humanoids

Rangasayee's argument is that Nvidia's robotics chips are built from data center architectures. SiMa.ai's aren't.

"We don't want to down bin a GPU and say, hey, this is the best you got," he told me.

(Down-binning is the chip industry's term for selling lower-spec versions of a flagship part. It's efficient for the chipmaker. It's not necessarily optimal for a robot running on a battery.)

SiMa.ai's current-generation chip, Modalix, is in production on TSMC's 6nm process. The next generation, funded by this round, targets twice the performance of Thor at 60 watts, with volume hardware slated for the first half of 2028. That timing isn't an accident.

"We believe that by 2028 you'll start seeing meaningfully good volume in humanoid," Rangasayee says. "And so we want to build a platform for scale at the right cost, at the right power efficiency and with enough headroom and performance."

Cost matters, but so does software

Cost matters as much as watts. Actuators already make up 40% to 60% of a humanoid's bill of materials, and a good pair of dexterous hands can run $13,000. SiMa.ai won't share pricing, but says it's pricing to be competitive with Nvidia. My sense in my conversation with Rangasayee is that Sima.ai's chips will be significantly cheaper, but I'm only guessing here.

The more interesting angle might be the software.

Most humanoids run a stack of intelligence: low-level motor control, perception, task planning, often a language model for talking to people, and fleet management on top. Humanoid AI's KinetIQ platform, for example, has four layers. Running all of that on one chip is hard. SiMa.ai says it can run multiple vision-language models concurrently, which matters for safety when a robot needs independent opinions from different sensors.

"Here's a radar based model, here's a camera based model, and I'm going to make a decision," Rangasayee says.

SiMa.ai supports ROS 2 natively and is building reference libraries on top: "Here's a perception stack, here's a SLAM stack, here's a fleet management stack." But it says it isn't trying to build the robot brain itself.

"We do not want to be in the business of competing with our customers, but we want to enable them to go to production quicker," he says.

To make switching easy, SiMa.ai offers a pin-compatible replacement for Nvidia's system-on-module, plus agentic tools that port CUDA-based code. In theory, a humanoid maker could pull the Nvidia board, slide in SiMa.ai's and keep the rest of the robot unchanged.

In theory, of course.

Partners and customers

SiMa.ai's named customers and partners (Bosch, Emerson, Micron, TRUMPF, Intrinsic and others) include no humanoid maker as of yet. And, naturally, Nvidia isn't standing still: its lower-end Jetson T4000 module already runs at 40 to 70 watts.

Rangasayee sees the humanoid industry at the stage the data center was a decade ago, obsessed with training. He's betting on what comes after.

"What I think the world misses today is really two things: data for training, big problem," he says. "Second, the right inference vehicle to deploy, at the right cost, right power, right performance capability. And I would say we are far better at the second problem."

One of his customers told him that "robotics is run at the pace of China speed."

Rangasayee's view is that Western humanoid makers can't win on cost for years, so they have to win on iteration: "being able to iterate 20 different designs within a month. That's going to eventually decide who's going to win."

Right now, humanoid makers mostly have one serious option for onboard AI compute. SiMa.ai just raised $150 million to become the second. 

Whether they can win against the $5 trillion behemoth in the space is another question.