A $100 million bet on silicon at the edge
For a company founded only in 2018, a nine-figure chip program is no small undertaking. DeepX said the design of the DX-M2 is finished and the final plans, the step engineers call tape-out, have gone to Samsung’s foundry unit. What follows is the costly, painstaking phase in which the first wafers are produced, tested and refined before a chip can be sold at scale. The company puts the total bill at more than $100 million, a figure that covers performance verification and production engineering.
The timetable is measured but deliberate. First silicon is expected by March 2027, followed by customer evaluations and mass manufacturing in 2028. That schedule is later than the 2027 mass-production target DeepX outlined when it signed its 2nm agreement with Samsung Foundry and design partner Gaonchips in August 2025, a reminder that at the leading edge of semiconductor manufacturing, every milestone must be earned rather than simply promised.
Five watts to run generative AI
The DX-M2’s defining promise is efficiency. DeepX says the chip can run generative-AI models on about 5 watts. Details the company displayed at CES 2026, reported by EE Times Japan, describe a four-core neural processing unit delivering around 80 TOPS, paired with LPDDR5X memory offering roughly 150 GB/s of bandwidth. DeepX says the chip is designed to handle language models with about 20 billion to 100 billion parameters at speeds of roughly 10 to 30 tokens per second.
















