Field survey · June 2026
Embedded AI
hardware, 2026
Twelve platforms that can be a robot's onboard brain right now, scored on the four axes that matter, and sortable by any of them.
The thesis
Compute, memory, power, price. Onboard intelligence is a four-way trade. No platform wins all four.
A field survey of the current embedded AI hardware menu, from flagship Jetson modules to USB accelerators. Star ratings are normalized within this set: five stars is best in class on that axis: fastest, biggest, most frugal, cheapest. Filter by family, search, or sort any column to see the trade-offs move.
The hardware menu
Twelve ways to give a robot a brain.
Click a column header to sort. The first click ranks best first. Stars are relative to this set, not absolute.
| Jetson AGX Thor | 2025 | ★★★★★2070 FP4 TFLOPS | ★★★★★128 GB | ★★★★★40 to 130 W | ★★★★★$3,499, ≈€4,190 EU |
|---|---|---|---|---|---|
| Jetson AGX Orin | 2022 | ★★★★★275 TOPS | ★★★★★64 GB dev kit | ★★★★★15 to 60 W | ★★★★★$1,999 launch price |
| Jetson Orin NX 16 GB Super | 2023, Super 2025 | ★★★★★157 TOPS | ★★★★★16 GB | ★★★★★10 to 40 W | ★★★★★≈€900 to €1,050 kit or bundle |
| Jetson Orin Nano Super | 2024 refresh, 2023 HW | ★★★★★67 TOPS | ★★★★★8 GB | ★★★★★7 to 25 W | ★★★★★$249 official |
| Raspberry Pi 5 barebone 16 GB | 2023, 16 GB in 2025 | ★★★★★No AI accelerator | ★★★★★16 GB | ★★★★★2 to 10 W board, 27 W PSU budget | ★★★★★$305 current official 16 GB price |
| Raspberry Pi 5 plus AI HAT+ 2 | 2026 | ★★★★★40 INT4 TOPS | ★★★★★16 GB Pi plus 8 GB AI HAT | ★★★★★≈5 to 15 W system class | ★★★★★$130 HAT, Pi extra |
| Google Coral USB Edge TPU | 2019 | ★★★★★4 INT8 TOPS | ★★★★★No usable model RAM class | ★★★★★≈2 W | ★★★★★≈€90 to €130 |
| Hailo 8 M.2 | 2019 | ★★★★★26 TOPS | ★★★★★On-chip memory, no external DRAM | ★★★★★2.5 W typical | ★★★★★≈€165 to €200 |
| Hailo 10H / AI HAT+ 2 class | 2024 chip, 2026 Pi product | ★★★★★40 TOPS | ★★★★★8 GB on Pi HAT+ 2 | ★★★★★Under 5 W chip class | ★★★★★$130 via Pi AI HAT+ 2 |
| Qualcomm RB3 Gen 2 Core Kit | 2024 | ★★★★★12 TOPS | ★★★★★6 GB kit class | ★★★★★≈7 to 12 W class | ★★★★★$399 core kit |
| Arduino VENTUNO Q | 2026 | ★★★★★40 TOPS | ★★★★★16 GB | ★★★★★Unknown, assume 10 to 25 W class | ★★★★★Just under $300 expected |
| RK3588 SBC class | 2022 to 2025, board dependent | ★★★★★6 TOPS | ★★★★★Up to 32 GB | ★★★★★≈5 to 15 W board class | ★★★★★≈€220 to €500 |
My synthesis
Memory is the new TOPS.
Sort by any column and four tiers fall out. The deeper pattern: compute is no longer the scarce resource. Memory gates what class of model can live onboard at all (Thor's 128 GB exists precisely for VLA-class models), and watts decide whether the platform can ride on a battery.
-
01
Flagship JetsonVLA-class models onboard, paid for in watts and euros
-
02
Mid JetsonThe System 1 sweet spot: perception + compact policies
-
03
Accelerator + hostFixed-function perception at single-digit watts
-
04
Kits & SBCsPrototype cheap; rarely the final reflexes
Strategic implication
Silicon follows the System 1 / System 2 split.
Note 002's dual-system framing decides the shopping list: System 1 must run onboard in real time, so the robot's silicon is sized for the reflexes: perception and control at a power budget the chassis can carry. System 2 can burst to off-board or cloud compute when deliberation is needed.
Fleet economics push the same way: onboard compute multiplies with every unit shipped, while shared System 2 amortizes across the fleet. The cheapest TOPS are the ones you don't put in the robot.
TOPS are cheap. Memory and watts decide what runs onboard.