DeepX books $13M of NPU orders in year one, at roughly $169,000 per order

DeepX said on August 14 that its DX-M1 on-device AI chip has collected commercial purchase orders worth more than $13 million over the past year across more than 10 countries and regions ^[1]. That is the first full year of results from a low-power NPU built on a leading-edge process, and the shape of the order book matters more than the headline number ^[1][8].

Start with the arithmetic. Cumulative POs reached 77 as of July, and 48 of those came in the four months from April ^[2]. That implies 29 orders in the preceding eight months, or about 3.6 per month, against 12 per month in the April-to-July stretch ^[1][2]. It also implies an average order size of roughly $169,000 ^[3]. These are not volume design wins. They are the size of pilot builds, first production lots and module-level integrations, which is what you would expect from a company whose customers span robotics, smart factories, smart mobility and intelligent transport, mission-critical and aerospace work, AIoT, healthcare, edge computing and video surveillance ^[3].

The geography is the more durable signal. Orders have been booked in South Korea, the United States, China, Europe, Taiwan, Hong Kong, Japan, Singapore and India ^[4], and DeepX says more than half of cumulative revenue through the end of July came from overseas ^[5]. CEO Kim Nok-won said the most meaningful development of the first year was that orders and repeat orders are occurring across multiple markets rather than being confined to one country or industry ^[6]. Repeat orders from small first orders is the correct pattern for an edge silicon business; the source does not disclose revenue in absolute terms, so the overseas share is a mix statistic rather than a scale one ^[5].

The distribution build-out explains part of the April inflection. DeepX spent two to three years running sample-based evaluations with more than 400 companies before mass production ^[7], then signed 27 semiconductor distribution partners and worked with more than 50 hardware and module companies after production started ^[9]. It is also linked to Raspberry Pi, Baidu PaddlePaddle and Ultralytics ecosystems ^[10]. Against that base, 77 POs is a low conversion count, though the source does not say how many distinct customers those orders represent ^[2][7].

The capital question is the next chip. The DX-M1 is mass-produced on Samsung's 5nm process ^[8]; the DX-M2 targets Samsung's 2nm process, with prototype wafers planned for March next year, bring-up in April and May, and early customer supply preparation from June ^[11]. The stated goal is generative AI at as little as 5 watts at chip level, aimed at power-constrained physical AI such as robots, drones, industrial equipment and battery-powered devices ^[12]. Moving from 5nm to 2nm is a mask-set and yield commitment that a $13 million annual order book does not fund by itself ^[1][11].

What to watch: whether the 12-orders-per-month pace holds past July or reveals itself as a pull-forward ^[2]; whether any single PO grows into a disclosed multi-million-dollar volume ramp rather than another six-figure lot ^[3]; and whether the March prototype wafer date slips ^[11]. A second year of $169,000 orders would say the evaluations are not converting into production programs ^[3].