Nvidia’s new T3000 matches flagship inference at half the power

Signals Inbox·July 15, 2026·AI Chips

Nvidia says its new Jetson T3000 can run the same kinds of advanced robot models as its flagship T5000 while using roughly half the space and power. The headline is not simply a smaller chip. It is Nvidia trying to turn robot-grade AI compute from a bulky premium component into something manufacturers can realistically fit inside large fleets of humanoids and autonomous machines.

Q1What did Nvidia actually announce?

According to Nvidia’s official announcement, it introduced two Blackwell-powered Thor modules called the Jetson T3000 and T2000. The T3000 has 865 FP4 teraflops, 32GB of memory and 273 GB/s of memory bandwidth. The smaller T2000 offers 400 FP4 teraflops and 16GB for visual AI, mobile robots and industrial machines.

Q2Does the T3000 really match Nvidia’s flagship?

Not in raw compute. The flagship T5000 reaches 2,070 FP4 teraflops, more than twice the T3000’s 865. Nvidia’s narrower claim is that the T3000 delivers similar inference performance for large language models, vision models, robot action models and world models. It also keeps the same 273 GB/s memory bandwidth, which helps feed those models quickly.

Q3Where does the half-power claim come from?

The current T5000 can run at roughly 140W, while the T3000 is listed around 70W. That matters inside a robot because every extra watt creates heat and drains the battery. Cutting the computer’s power can reduce cooling, free space for a larger battery and let the robot work longer without making its body heavier.

Q4Why launch a weaker chip when the T5000 already exists?

Because most production robots do not need 128GB of memory and maximum peak compute. They need enough local intelligence at a size, power and cost that can fit into thousands of machines. The T3000 drops memory to 32GB while preserving the flagship’s bandwidth. The T2000 goes down to 16GB. Nvidia is basically breaking Thor into practical tiers instead of forcing every robot maker to buy the biggest possible brain.

Q5Is this really about humanoid robots?

Humanoids are the loudest use case, but the market is wider. Nvidia lists autonomous mobile robots, industrial arms, visual AI agents and machines working near humans. It also announced an IGX T3000 version with functional safety features. That is important because factory robots need more than impressive demos. They need predictable behavior and safety systems that companies can certify.

Q6So are mass-market robots arriving now?

Not yet. The T3000 and T2000 are expected to become available in the first quarter of 2027, so this is still a roadmap signal. But it attacks a real deployment problem. Nvidia’s first Thor module already delivered up to 7.5 times more AI compute and 3.5 times better energy efficiency than Jetson Orin. Now Nvidia is shrinking that newer architecture into modules designed for volume. The next proof will be actual robot shipments, prices and independent performance tests.