Changes to production plans at Tesla have delayed the mass production of the next-generation neural processing unit (NPU) DX‑M2 by South Korean AI chip developer DeepX by six months.
According to industry sources, the multi-project wafer (MPW) production originally scheduled for April 2026 has been postponed. This shift will impact the quality testing schedule, which is now expected to start no earlier than the third quarter of 2026. DeepX had planned to begin mass production of the DX‑M2 in the second quarter of 2026, but the delay in MPW runs has pushed back the production timeline.
Industry observers noted that the delay is mainly a knock-on effect from production plan changes at Tesla, a major customer of Samsung Electronics.DeepX’s DX‑M2 is the first external customer chip to use Samsung’s 2nm process, while Tesla is also developing its AI6 AI chip using Samsung’s 2nm technology. The delay in Tesla’s MPW runs has directly affected DeepX’s production schedule.
Although Tesla has not publicly disclosed the specific reasons for the MPW delay, industry sources speculate it may be related to adjustments in the mass production timeline for chips used in Tesla’s autonomous vehicles and humanoid robots, as well as its supercomputer investment plans.A Samsung Electronics official declined to comment, stating the company cannot confirm matters related to its customers.
Earlier reports indicated that Tesla recently sent a senior procurement executive to Samsung Electronics to discuss expanding wafer capacity for the AI6 chip.Under the previous agreement, volume was approximately 16,000 wafers per month; the additional capacity under discussion is around 24,000 wafers per month.If finalized, the new agreement could bring total output to about 40,000 wafers per month.
DeepX’s DX‑M2 is a generative AI accelerator designed for AI data center workloads, capable of handling models with up to 100 billion parameters.It is expected to have a maximum power consumption of 5 watts and a computing performance of 80 trillion operations per second (TOPS). The processor supports the low-power DRAM standard LPDDR5X as memory.
According to industry sources, DeepX has not yet received orders for the DX‑M2. Customers typically place orders after testing sample chips, so meaningful revenue contributions are expected to begin only after mass production starts in 2026.DeepX’s customers include Samsung Electronics (for Exynos 2100), Hyundai Motor Group’s robotics lab MobED, and Intel (for certain Core i5 models).
Sources also noted that DeepX recently supplied 40,000 DX‑M1 chips and modules to Baidu.The DX‑M1 chip is priced at approximately $20–$50 per unit, with modules selling for $50–$100.Based on these prices, the shipment is expected to generate revenue of $800,000 to $4 million, accounting for roughly 5% to 24% of DeepX’s full-year revenue forecast of $17 million.
(Reprinted from https://news.eccn.com/）