Skip to main content

DeepX completes DX-M2 tape-out on Samsung’s 2nm process

14 NOVEMBER 2026·2 MIN READ·2 SOURCES

DeepX says it has completed the tape-out of the DX-M2, designed for generative AI workloads at around 5 watts. The chip is being manufactured on Samsung Electronics’ second-generation 2nm process, with first silicon expected in March 2027.

DeepX completes DX-M2 tape-out on Samsung’s 2nm process

Key takeaways · 4

  • 01

    The tape-out transferred the finalized design to the foundry and included corrections for lithography-related distortions.

  • 02

    DeepX says the chip is designed for models at around 5 watts, with mixture-of-experts models ranging from 20 billion to 100 billion total parameters.

  • 03

    Dealroom reports DeepX is targeting a price below $50 per chip; this is a target, not a stated selling price.

  • 04

    Treat the cited 80-TOPS and 6.7-times figures as pre-silicon specifications and simulations.

Design moves to the foundry

DeepX said it completed the DX-M2 design on Nov. 10 and completed the mask tape-out in early November[1]. The tape-out transferred the finalized chip design to the foundry and completed corrections for lithography-related distortions[1]. The chip is being manufactured using Samsung Electronics’ second-generation 2nm foundry process, SF2P[1]. DeepX expects its first silicon chips in March 2027[1]. The tape-out is a design and manufacturing milestone; it does not itself establish measured performance, and the cited performance figures include pre-silicon specifications and simulations[1][2].

Capacity and performance claims

DeepX designed the DX-M2 to run generative AI workloads, including large language and vision-language models, at around 5 watts[1]. Dealroom says the chip can hold mixture-of-experts models with 20 billion to 100 billion total parameters, with roughly 3 billion to 20 billion active per token[2]. It describes decoding at 10 to 30 tokens per second within the power budget[2]. DeepX CEO Lokwon Kim told EE Times that memory bandwidth rises from 50 GB/s to about 150 GB/s between generations[2]. A quoted claim says tokens per second could rise up to 6.7 times on Llama 3.1 8B, paired with 2.7 times the power; Dealroom says the 6.7-times and 80-TOPS figures are pre-silicon specifications and simulations[2].

Schedule and price targets

DeepX plans to provide evaluation samples to customers in the second quarter, with initial mass-production shipments scheduled for the fourth quarter[1]. The company plans full-scale DX-M2 production and expanded global supply from 2028[1]. Dealroom reports that DeepX is targeting a price below $50 per chip[2]. Dealroom also says that target would be roughly a quarter of the reported per-chip cost of rival edge accelerators[2]. These are forward-looking plans and a target price, rather than evidence of current availability or a confirmed transaction price[1][2].

Target users and company context

DeepX is targeting robotics and industrial systems; Dealroom also lists robots, drones, AI PCs and industrial devices running models locally[1][2]. The DX-M2 received a CES 2027 Innovation Award in the AI category[1]. DeepX began mass production of its DX-M1 in August 2025[1]. As of October 2026, the company had secured 88 purchase orders worth about $13.3 million across 10 countries, and it forecast 2026 revenue of $40 million, more than twice its prior $17 million target[1]. Those figures provide company context, but do not establish DX-M2 sales or performance[1].

For teams evaluating local AI hardware, the DX-M2’s targets offer a basis for planning which workloads and devices to test. Keep procurement and performance decisions provisional until samples can be evaluated, especially because key performance figures are pre-silicon.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 14 November 2026

    DeepX completes DX-M2 tape-out on Samsung’s 2nm process

Sources

AI fluency, one session a day, built for your work.