NVIDIA Jetson: The Decade-Old Hardware Lineage Behind Physical AI

01What Jetson is

Compute that lives inside the machine

Jetson is NVIDIA's family of small, power-efficient computer modules built to run AI models on-device — inside a robot, a drone, a security camera, a piece of factory equipment — rather than in a data center. Where a data-center GPU processes AI workloads for thousands of users at once over a network, a Jetson module runs entirely inside the machine it's embedded in, under a power budget measured in watts rather than kilowatts.

The lineage under GR00T and DRIVE Thor

Jetson is the hardware lineage both Isaac GR00T and DRIVE Thor build on.

02Where Jetson actually shows up

"Robot" and "physical AI" get used loosely in coverage of this space. What's actually shipping today, running on Jetson hardware in commercial volume rather than as a demo, is narrower and in some ways more mundane than humanoids — the clearest examples are in logistics and delivery, where running AI on-device instead of over a network connection has an obvious, easy-to-explain payoff: a delivery robot or drone can't wait on a cellular round-trip to decide whether to brake for a pedestrian.

Jetson also underpins the humanoid and quadruped robots covered elsewhere on this site — Isaac GR00T's Unitree reference platform, Boston Dynamics' Atlas, and others — but those remain closer to pilot and research deployments than the delivery and inspection examples above, which are running at commercial volume today.

03Hardware history (2014–2025)

Jetson's product history is unusually long-running for a category that only became a mainstream investment theme in the last few years — NVIDIA had a decade of iteration and an installed base of developers before "humanoid robot" and "physical AI" became pitch-deck language.

04The current generation: Jetson AGX Thor

Thor's specs, at a glance

Jetson AGX Thor, NVIDIA's current flagship module, reached general availability in August 2025. Built on NVIDIA's Blackwell GPU architecture, it delivers up to 2,070 FP4 teraflops of AI compute in a 130-watt power envelope — NVIDIA states roughly 7.5x the AI compute and 3.5x the energy efficiency of the prior Orin generation. The developer kit sells for $3,499; production modules start lower.

2,070

FP4 teraflops (peak)

130W

Power envelope

$3,499

Developer kit price

Aug 2025

General availability

Who signed on at launch

Named early-access and evaluation partners at launch spanned a wide slice of the robotics industry, shown below.

05Market position and competition

NVIDIA's estimated share

NVIDIA's Jetson platform is estimated to hold roughly 39% of edge AI computing revenue, by one market-research estimate worth treating as directional rather than precise — no research firm with a fully public, checkable methodology was found publishing a comparable figure, though it's broadly consistent with Jetson's widely reported position as the default choice for high-end robotics compute.

The organized challengers

The clearest organized challenge comes from Qualcomm, whose Dragonwing IQ10 series (introduced at CES 2026) is explicitly positioned as a lower-power alternative for robots that don't need Thor's full compute ceiling; Intel has taken a different approach with a dedicated Robotics business unit built around more than 130 commercial design partnerships and an open-platform pitch that avoids competing head-on with NVIDIA's brand; and NXP is building out edge-AI capability through its Kinara acquisition and a new "Neural Axis" chip architecture.

06Where Jetson fits in the business

Jetson revenue isn't broken out on its own — NVIDIA reports it inside a combined "Automotive and Robotics" line that also includes the much larger DRIVE automotive business, and that whole combined line was worth about 1% of NVIDIA's total revenue in fiscal year 2026. NVIDIA's Isaac GR00T and DRIVE articles on this site cover that financial picture in full; the short version is that Jetson, like the rest of NVIDIA's physical-AI push, is a long-run platform bet layered on top of a data-center business that dwarfs it, not something NVIDIA needs to monetize today.

07Challenges ahead

Generational fragmentation

Generational fragmentation is a real, if unglamorous, cost of an eleven-year hardware lineage. Each Jetson generation (TK1, TX1/TX2, Xavier, Orin, Thor) uses a different JetPack software version with its own support window, meaning developers who build on an older module eventually face a forced migration to newer hardware to keep receiving updates — a maintenance burden that a genuinely new entrant, without a decade of legacy modules to support, doesn't carry.

Independent benchmarks are still scarce

Independent, apples-to-apples benchmarks against Qualcomm's and Intel's newer offerings are scarce. Most of the performance and market-share figures in this article — NVIDIA's own compute numbers, the roughly 39% share estimate, the developer-adoption claims — trace back to NVIDIA's own releases or to research firms without a fully public methodology, not to a neutral third-party lab running the same workload on all three platforms.

Is Thor's tradeoff right for every robot?

Whether Thor's price and power draw are the right tradeoff for every use case is genuinely contested. Qualcomm's explicit pitch for Dragonwing is that many robots don't need Thor's full compute ceiling and would rather have a cheaper, lower-power chip — a bet on the market fragmenting by workload rather than everyone converging on NVIDIA's top-end module.

08FAQ

09Sources