DEEPX DX-M1 NPU Explained: What This Edge Chip Changes

The DEEPX DX-M1 NPU is a low-power neural processing unit — a dedicated AI inference chip — that runs computer-vision and detection models directly on a device at roughly the power of a light bulb, no data-center GPU required. On June 2, 2026 at Computex Taipei, that chip stopped being a demo and became a supply chain: Korean AI-silicon maker DEEPX and industrial hardware vendor AAEON signed a three-year mass-production agreement to put DX-series NPUs into standard, off-the-shelf edge hardware.

This page is the plain-English reference for what the DEEPX DX-M1 NPU is, what actually changed at Computex, why it matters now, and where it honestly falls short. No equations, no hype — just what shipped and what it means for small and mid-size operators, as of June 2026.

TL;DR

• According to DEEPX, the DX-M1 delivers 25 TOPS of INT8 compute while drawing just 1 to 5 watts.

• According to PR Newswire, DEEPX and AAEON signed a 3-year mass-production MOU on June 2, 2026 at Computex Taipei.

• The chip ships in M.2, mPCIe, PCIe-card, and COM Express form factors — the same slots integrators already use.

• It targets machine vision, defect inspection, object detection, robotics, and automation at the edge, without a cloud round-trip.

• The honest catch: it runs vision and detection models well; it is not a server for large language models.

What the DEEPX DX-M1 NPU actually is

A neural processing unit is a chip built for one job: running a trained AI model's math (inference) fast and efficiently. The DX-M1 is DEEPX's flagship edge NPU. Rather than ship a giant accelerator that lives in a rack, DEEPX packages the silicon onto a module that drops into common slots — an M.2 stick the size of a stick of gum, or a small PCIe card.

The point is on-device inference. Instead of a camera streaming video to a cloud GPU and waiting for an answer, the model runs locally on the DX-M1, and the answer comes back in milliseconds. That removes three recurring costs at once: bandwidth, cloud-inference fees, and latency. It also keeps the image data on the premises, which matters for plants and warehouses with privacy or connectivity constraints.

According to DEEPX, the DX-M1 delivers 25 TOPS of INT8 compute inside a 1-to-5-watt envelope. TOPS — trillions of operations per second — is the rough yardstick for how much AI work a chip can do; INT8 is the compact numeric format most vision models run in. The headline isn't the raw TOPS number, which is modest next to a server GPU. It's the ratio: enough vision throughput to watch a production line, at a power budget low enough to run fanless inside a sealed industrial box.

The specs that matter

Sources: DEEPX.

The DX-M1 supports models from TensorFlow, PyTorch, ONNX, Keras, and Ultralytics, according to DEEPX, which lists 5 framework families. A team that already trained a defect-detection model in PyTorch or exported it to ONNX can target the DX-M1 without rewriting it — the same way teams already routing documents through US Tech Automations workflows can swap an inference target without rebuilding the pipeline around it.

The memory figure defines the chip's lane. According to DEEPX, the M.2 module carries 4GB of LPDDR5 (5600 MT/s) and the chip supports up to 8GB — enough to hold compact vision and detection models for inspection and counting, but not the tens of gigabytes a large language model needs. That's a feature, not a flaw: by specializing in perception, the chip hits power and form-factor targets a general-purpose accelerator never could. The right mental model is a dedicated set of eyes for a machine, not a brain that does everything.

What happened at Computex 2026

The technology existed before June. What changed is the path to buying it at volume. According to PR Newswire, DEEPX and AAEON signed a 3-year Mass Production Cooperation MOU at Computex Taipei on June 2, 2026. AAEON is part of the ASUS Group and a long-standing maker of industrial PCs, single-board computers, and edge gateways — the boxes that already sit on factory floors and inside delivery vehicles.

AAEON's role is to embed DEEPX NPUs across its product lines in the four standard form factors, so an integrator can order a vision-capable edge box the way they order any other AAEON unit. According to PR Newswire, DEEPX's partner ecosystem grew from roughly 15 companies to more than 30 globally over the past year. That doubling is the real signal: edge NPUs are moving from one-off design wins to a stocked, mass-produced supply line.

Sources: PR Newswire.

DEEPX CEO Lokwon Kim framed AAEON as the route to market, calling it "the ideal partner, combining a globally proven hardware platform with the ASUS Group's extensive distribution network," in remarks reported by PR Newswire.

Why now — the constraint that broke

Edge AI has been "almost ready" for years. Two constraints kept it stuck: power and packaging. A GPU good enough to run vision models in real time drew too much power and heat for a fanless box on a dusty factory floor, and the hardware came as bespoke design projects rather than catalog parts. The DX-M1's 1-to-5-watt envelope solves the first; the AAEON mass-production deal in standard form factors solves the second.

The market backdrop explains the timing. According to Global Market Insights, the edge AI software market was $3.7 billion in 2025 and rises to $4.5 billion in 2026, on a 28.3% CAGR through 2035. Computer vision is the workhorse of that market: according to Global Market Insights, it held a 37% share of the edge AI software market in 2025 — the single largest segment.

The Computex stage itself underlines the shift. According to Intel, the company used its June 2, 2026 keynote to frame inference moving to the edge and the rack as the defining theme of the year. The DX-M1 is the small-end bookend of that same story: inference moving out of the training-era cloud and onto the device.

What it changes for operators

For a small or mid-size operator, the DX-M1 changes the unit economics of "putting a smart camera on it." Today, a vision-inspection pilot often means a workstation with a discrete GPU, a cloud-inference subscription, or both. A DX-M1 module folds that into the edge box that's already there. Below, the DX-M1 power figure is the sourced spec; the other cells are directional categories, not vendor pricing.

Sources: power envelope per DEEPX; approach categories per PR Newswire.

The practical change is that vision inference becomes a line item on the bill of materials instead of a project. A team can specify an AAEON box with a DX-M1, deploy a defect-detection model, and run it offline. For organizations that already orchestrate downstream steps — routing a flagged defect into a work order, a hold, or an alert — through US Tech Automations, the NPU becomes the sensor and the automation layer becomes the response, with no cloud GPU in the loop.

Signal vs Speculation

Signal (sourced fact, as of June 2026): According to DEEPX, the DX-M1 ships with 25 TOPS INT8, a 1–5W envelope, and 4GB LPDDR5 (5600 MT/s) on the M.2 module, with the chip supporting up to 8GB. According to PR Newswire, a 3-year mass-production MOU with AAEON was signed June 2, 2026 and the partner ecosystem passed 30 companies.

Our read (forecast, not fact): An MOU is intent, not a shipped volume commitment, and the press materials disclose no unit pricing or per-SKU ship dates. If the AAEON form-factor strategy holds, the most likely 12-to-36-month outcome is that vision inference becomes a checkbox option on industrial edge hardware rather than a custom build — pushing real-time defect detection and object counting into the reach of operators who could never justify a GPU workstation. The risk: edge NPUs only matter if the models stay small. The DX-M1 is a vision and detection engine, not an LLM server, so the cloud doesn't disappear — it specializes. We'd expect a split where on-device NPUs handle perception and large models stay central. Our read is that the operators who win are the ones who treat the NPU as a new input into existing automation, not as a reason to rebuild.

For the role-by-role version of this forecast, see the implications for manufacturers, for logistics operators, and for small businesses.

The honest limits

The DX-M1 is purpose-built and proud of it — which means it is not general-purpose. It excels at INT8 vision and detection models. It is not the chip for running a 70-billion-parameter chatbot locally. According to DEEPX, the M.2 module's memory is 4GB LPDDR5 (the chip supports up to 8GB), which bounds model size. And an MOU, however large, is a plan: production cadence, pricing, and availability per region will determine how fast this reaches a given buyer. None of those numbers were published with the announcement, so anyone claiming a price or a ship date is guessing. We won't.

Key Takeaways

• According to DEEPX, the DX-M1 is a 25-TOPS INT8 edge AI chip running at 1–5 watts.

• According to PR Newswire, a 3-year mass-production MOU with AAEON, signed June 2, 2026, turns it from a part into a supply line.

• Standard M.2, mPCIe, PCIe, and COM Express form factors mean integrators buy it like any other module.

• According to Global Market Insights, edge AI software is a $4.5 billion market in 2026 with computer vision the largest slice.

• It is a vision and detection engine, not an LLM server — plan it as a sensor that feeds your automation, not a replacement for the cloud.

Frequently Asked Questions

What is the DEEPX DX-M1 NPU in one sentence?

It is a low-power edge AI accelerator that runs vision and detection models on-device. According to DEEPX, it does so at 25 TOPS and 1–5 watts, so cameras and machines decide locally without sending data to the cloud.

What did DEEPX and AAEON actually announce?

A three-year mass-production cooperation MOU. According to PR Newswire, it was signed June 2, 2026 to embed DEEPX NPUs across AAEON's industrial PCs, single-board computers, and edge gateways.

How much power does the DX-M1 use?

Between 1 and 5 watts depending on load. According to DEEPX, that envelope is low enough to run fanless inside a sealed industrial enclosure.

Can the DX-M1 run large language models?

No. According to DEEPX, the M.2 module carries 4GB of LPDDR5 (the chip supports up to 8GB) and targets INT8 vision and detection — not enough to serve large language models. It is a perception engine, not an LLM server.

Why does this matter for small and mid-size businesses?

Because vision inference becomes a catalog part instead of a custom GPU project. According to Global Market Insights, the edge AI software market reached $4.5 billion in 2026, and mass production puts that capability inside boxes operators already buy.

When can I buy a DX-M1-based device?

According to PR Newswire, the mass-production MOU was signed June 2, 2026, but no per-SKU pricing or regional ship dates were published, so availability depends on AAEON's rollout.

Put it to work

The chip is the easy part. The value shows up when a detection result triggers the next action — a hold, a re-route, a ticket. That's the layer to design now, before the hardware lands. See how to wire on-device inference into a response path with agentic workflows from US Tech Automations, and read the role-specific playbooks for manufacturers and logistics operators.