Edge AI for drones, delivery & HAPS
Frontier AI inside the drone power budget.
Onboard vision-language reasoning, real-time perception, and comms-denied autonomy — running on the platform within a sub-10W envelope, so the model survives the same flight as the airframe. No tethered GPU. No cloud round-trip.
Use cases
Where on-platform AI earns its keep
GPS-degraded & BVLOS autonomy
Visual-inertial navigation, terrain-relative localization, and route planning running on-platform with deterministic latency — beyond the link, beyond GNSS.
Onboard vision-language perception
Run a VLM on the drone. Identify targets, classify scenes, and ground free-form instructions in what the camera sees — without offloading frames.
On-platform model inference
LLM and vision-language models inside the AeroScale V1 envelope. The model lives where the camera lives — no chase vehicle, no GPU rack on the ground.
Comms-denied operation
Decisions stay onboard when the link drops. No degraded behaviour when the uplink does.
Battery-budget AI
A SWaP-tuned compute envelope: 4.5 W typical, 24 g module — designed to disappear into a payload bay or strap to an airframe.
Aerial-grade qualification
MIL-STD-810H vibration, 30,000 ft altitude qualified, −40 to +85 °C operating. Secure boot and per-module attestation.
The matched module
AeroScale V1
FPGA-based · aerial-grade
AeroScale V1 puts large-language and vision-language models into autonomous drones, delivery platforms, and HAPS systems — within a sub-10W envelope so the model survives the battery, not the other way around. Built on the Invotet Unified Engine running in an FPGA fabric you can buy today — not a chip-down ASIC waiting on tape-out.
- Throughput
- 38 GOPS
- Power envelope
- 4.5 W
- Operating range
- −40 to +85 °C
- Weight
- 24 g
Why this module, this vertical
The properties that map to your platform.
Up to 20× efficiency
A unified compute engine — systolic and vector processing in one — purpose-built for transformer workloads. Smallest logic footprint, highest utilization, up to 20× more efficient than NVIDIA Jetson.
Sustainable autonomy
Frontier-class models inside a sub-15W envelope. AI fits inside the battery or solar budget — Size, Weight, and Power optimized for every module.
Transformer-grade fidelity
BF16-native execution preserves training-equivalent accuracy at 95% sustained utilization, with native flash attention and hardware tensor-parallel sync. A cycle-accurate hardware trace buffer and compile-graph-to-hardware specialization make every inference verifiable and tuned to the workload.
GPT-native logic
Matrix multiplication, softmax, element-wise operations, and the rest of the transformer operator set run natively in purpose-built logic — no general-purpose emulation tax.
Invotet SDK
Compile once. Deploy to every Invotet module.
A unified Python SDK that ingests PyTorch, ONNX, and HuggingFace checkpoints, quantizes for Invotet modules, and ships a deterministic runtime to the device. No CUDA in the loop.
Framework
PyTorch
Trace or torch.export checkpoints compile directly with no rewrite.
Framework
ONNX
Standards-based interchange — compile any ONNX-exported model.
Framework
HuggingFace
transformers checkpoints land on Invotet through a one-line loader.
Talk to the team that ships the modules.
Most aerial conversations start with a sample unit and a flight envelope. Tell us the airframe, the payload budget, and the workload — we will line up an AeroScale V1 eval kit and route the right datasheet the same day.