NVIDIA Jetson T3000 and T2000: Thor Edge AI Modules Arrive in 2027

NVIDIA has introduced two new edge AI compute modules, the Jetson T3000 and Jetson T2000, built on its Thor architecture. The T3000 delivers 865 FP4 teraflops of AI compute in roughly half the size and power footprint of NVIDIA's existing T5000, while the T2000 offers 400 FP4 teraflops as a lower-cost entry point. Both modules are scheduled for general availability in Q1 2027. Partners including Boston Dynamics, Amazon Robotics, and FANUC are already building on the Jetson Thor platform.

What happened

Detail	Fact
T3000 AI compute	865 FP4 teraflops
T3000 memory	32GB LPDDR5X, 273GB/s bandwidth
T3000 connectivity	25 GbE
T2000 AI compute	400 FP4 teraflops
T2000 memory	16GB
Cosmos 3 Edge parameters	4 billion
General availability	Q1 2027
T3000 emulation mode	Later this month, via JetPack 7.2.1

NVIDIA announced the Jetson T3000 and T2000 at the same time as an expansion to its Cosmos 3 model family. The T3000 pairs an NVIDIA Blackwell GPU with an eight-core Neoverse Arm CPU. Despite being roughly half the footprint and power draw of the T5000, NVIDIA claims it achieves comparable inference performance on multimodal workloads: large language models, vision language models, vision language action models, and world foundation models.

The T2000 is positioned as the accessible entry point. At 400 FP4 teraflops and 16GB of memory, it targets visual AI agents, autonomous mobile robots, and industrial manipulators. Together, the two new modules extend NVIDIA’s Jetson platform to cover a range from 70 TOPS to 2,000 teraflops across the full lineup.

Confirmed platform partners include 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot. Hardware ecosystem partners such as ADLINK, Advantech, AAEON, Connect Tech, and Seeed Studio are already building Thor-based solutions.

What are the new Jetson agent skills?

Alongside the hardware, NVIDIA released Jetson agent skills: software tools that automate memory optimization and system configuration across the entire Jetson portfolio, including Jetson Thor and Jetson Orin. According to NVIDIA, developers can achieve meaningful memory savings in days rather than weeks.

Real-world results cited by NVIDIA include:

• UBTech and Agile Robots reduced memory usage by up to 15GB, moving from the Jetson AGX Orin 64GB to the 32GB module.
• SandStar (smart retail) cut memory usage by up to 4GB, enabling deployment on the Jetson Orin NX 8GB instead of the 16GB version.
• NoTraffic (intelligent transportation) reduced memory usage by 30% on Jetson TX2 NX, freeing headroom for additional AI capabilities.
• GROOVE X, maker of the LOVOT companion robot, uses Jetson’s heterogeneous AI accelerators to cut memory usage without swapping hardware.

The practical effect is that businesses running Jetson hardware may be able to move down one memory SKU within the same product tier, reducing per-unit hardware cost without a performance penalty.

Cosmos 3 Edge: on-device world modelling

NVIDIA also released Cosmos 3 Edge, a 4-billion-parameter lightweight version of its Cosmos 3 frontier world foundation model, tuned to run on Thor platforms. The model is designed to let an embodied system (a robot or autonomous machine) perceive its environment, reason over it in real time, and generate actions on-device rather than relying on cloud inference.

Developers can fine-tune Cosmos 3 Edge for specific hardware setups and sensors in approximately one day using the open Cosmos framework. NVIDIA positions this fine-tuning step as the key to closing the simulation-to-real-world gap before deployment.

Why it matters

Edge AI compute has historically required either expensive, power-hungry modules or a sacrifice in model capability. The T3000 attempts to close that gap: Blackwell-class GPU performance in a half-sized, half-power envelope means more capable AI can fit into constrained form factors like robot arms or retail cameras.

The agent skills story is arguably more immediately useful for businesses already running Jetson hardware. Cutting memory requirements by 15GB without rewriting core software is a direct cost saving, especially in high-volume deployments where per-unit hardware bills add up quickly. This kind of software-driven hardware optimization is the kind of AI integration that pays for itself fast.

For developers building physical AI products, the Q1 2027 timeline gives reasonable runway to test using emulation mode now, which NVIDIA says shares the same chip architecture and software stack as production modules.

Our take

NVIDIA is doing something smart here: instead of just selling more powerful chips, it is selling tools that let customers spend less on chips. The Jetson agent skills approach (automate memory optimization, drop a memory SKU, save money) is a strong commercial argument that works regardless of whether a customer is excited about foundation models or just wants to ship a product.

The Cosmos 3 Edge model is worth watching closely. A 4-billion-parameter model that can be fine-tuned for a specific robot in roughly a day is a real change in the economics of custom robotics. A day of compute instead of months of data engineering is a meaningful shift, though the “about a day” claim will need verification against real-world sensor configurations beyond NVIDIA’s own benchmarks.

For businesses exploring AI integration in physical products or operational environments, the T2000 at 400 teraflops is the one to watch. It is the price-accessible module, and if the agent skills deliver on memory savings, the total cost of a Jetson-based deployment could come down noticeably before the hardware even ships. For broader context on where edge AI sits in the current landscape, see our AI news coverage.

What to do about it

1. If you run existing Jetson Orin hardware, test the new agent skills now. The memory savings are available today across the current portfolio, not just on future Thor modules.
2. Download JetPack 7.2.1 (releasing later this month) to start emulating T3000 performance on the existing Jetson AGX Thor developer kit.
3. Evaluate whether Cosmos 3 Edge fits your use case. If you need on-device vision-and-action inference without cloud latency, the 4-billion-parameter model is worth a proof-of-concept against your sensor setup.
4. Plan hardware procurement for Q1 2027 if your product roadmap requires T3000 or T2000 production modules. Partner availability through ADLINK, Connect Tech, and others may vary.

The clearest near-term action: run the agent skills against your current Jetson deployment and measure actual memory reduction before committing to a hardware upgrade cycle.

Source: NVIDIA Blog