DeepX DX-M2: $100 Million-Plus on Samsung SF2P 2nm AI Chip
DeepX will spend more than $100 million on the DX-M2, an AI chip designed for about 5 W for robots, on Samsung's second-generation 2nm SF2P node. First silicon is planned for March 2027.

DeepX's own product-page picture of the DX-M2, a chip that has taped out but is not yet fabricated; this is a company illustration, not a production sample.
Image: DeepX
Source · Press image from the DeepX website, used for news reporting.
South Korean chip startup DeepX said it will spend more than $100 million on the DeepX DX-M2, an AI chip for robots and industrial machines to be made on Samsung Foundry's second-generation 2 nm process, which TheElec calls SF2P, with first silicon planned for March 2027 (the chip has taped out but is not yet fabricated).
DeepX said on Nov. 10 that the DX-M2 design is complete, according to TheElec, and the mask tape-out (the hand-over of final design data to the foundry) was done in early November. Photomask production comes next. The $100 million-plus budget, which also covers performance verification and production engineering, was reported by Bloomberg on Nov. 13 from a DeepX statement, as carried by The Korea Herald.
What a 5 W On-Device AI Chip Means for Data Centers
Samsung describes its 2 nm order wins as centred on high-performance-computing customers, including unnamed major hyperscalers. The DX-M2 goes the other way: DeepX says it is designed to run large language models and vision-language models at about 5 W, aimed at robots, industrial equipment and AI-enabled PCs. That is a bet that some inference moves out of the data center, which would reduce the power and cooling load there if it happens. Bloomberg framed the project as a large wager on on-device AI, and noted that running models on devices can reduce the energy use, cost and delay of sending data to the cloud and back. DeepX names Samsung's 2 nm process for the DX-M2; Samsung had not commented in the reports we read. Samsung's own third-quarter 2026 results release spoke of 2 nm order wins centred on HPC customers, including unnamed major hyperscalers, while the 2 nm capacity expansion at Giheung, as Voltgrid reported, is still unconfirmed by the company.
Why Tape-Out Is Not Yet a Chip
Tape-out is the point where a finished design goes to the foundry as data, which turns it into photomasks, the stencils that print each layer onto a silicon wafer. DeepX says the DX-M2 reached that point in early November, so no working chip exists yet. First silicon, the first wafers back from the fab, is planned for March 2027, with evaluation samples in the second quarter and initial shipments in the fourth quarter of 2027, TheElec reports.
Until that silicon is tested, the about 5 W figure is a design target, not a measurement. It matters for power because a part that holds that figure would let robots and AI PCs run language models on the device instead of in a data center. Samsung said on its Q3 2026 call that its second-generation 2nm process starts mass production of mobile products in Q4 2026, so DX-M2's first silicon is due after that start.
DX-M2 Schedule and Budget Against DeepX Revenue
TheElec lists the schedule: first silicon in March 2027, evaluation samples for customers in the second quarter, initial mass-production shipments in the fourth quarter of 2027, and full-scale production and global supply from 2028. TheElec and Seoul Economic Daily's account of the Bloomberg report both give initial volumes in the fourth quarter of 2027 and full-scale production from 2028; the Korea Herald copy of the Bloomberg report mentions 2028 only. DeepX's DX-M1, built on Samsung's 5 nm process, entered mass production in August 2025. Another Korean NPU developer, Rebellions, is on a similar timeline, as Voltgrid reported.
DeepX plans first silicon in March 2027 and full-scale production in 2028, according to TheElec; Samsung has not commented.
Illustration (AI-generated)
The $100 million is large against DeepX's size. TheElec reports 88 purchase orders worth about $13.3 million across 10 countries as of October 2026, a 2026 revenue forecast of $40 million (up from a $17 million target), and 2025 revenue of 3.3 billion won with a 20.1 billion won operating loss. The more than $100 million compares with a $40 million revenue forecast for 2026 and $13.3 million in purchase orders. These are different kinds of figure, and the reports do not say over what period the $100 million is spent.
What Is Not Confirmed on DX-M2 Yet
Samsung has not commented on the DX-M2 in the reports we read, and the SF2P node name comes from TheElec; Bloomberg says only "2 nanometer". The 5 W figure is a DeepX design target, not a measured result, and no performance numbers, die size or price were given. The reports we read name no DX-M2 customers. DeepX's schedule is a plan: initial shipments in the fourth quarter of 2027 and full-scale production in 2028 per TheElec, with the Korea Herald copy of the Bloomberg report mentioning 2028 only. Silicon in March 2027 is the first point at which the schedule can be checked. Samsung's 2 nm HBM base-die plans at Taylor remain reported, not confirmed.
Sources
- Editor
- Ege Erten
- Published
- Sources
- 2 cited
- Standards
- AI-assisted





