Groq, the AI chip company indirectly acquired by NVIDIA, has requested Samsung to increase its AI chip production from 9,000 wafers to 15,000 wafers, representing an increase of approximately 70%. Instead of directly assuming managerial control, NVIDIA is collaborating through a non-exclusive technology licensing agreement. Groq's production at Samsung is transitioning from the sampling phase to mass production, entering the early stage of large-scale commercialization this year. This move by NVIDIA aims to extend its ecosystem into the inference domain, with plans to release inference chips based on Groq’s design at GTC 2026.
The rising demand for AI inference chips is reshaping the order landscape of Samsung's foundry business.
On March 10, according to Chosun Ilbo citing informed sources, AI chip startup Groq has requested Samsung Electronics to increase its AI chip production from approximately 9,000 wafers to 15,000 wafers, marking an increase of about 70%.
According to reports, Groq was indirectly acquired for approximately $20 billion in December of last year.$NVIDIA (NVDA.US)$NVIDIA did not directly assume managerial control but indicated that it would collaborate through a non-exclusive technology licensing agreement. The industry widely views this move as an effort to extend NVIDIA's ecosystem into the inference domain, complementing its dominant position in the AI training chip market.
Notably, NVIDIA is expected to unveil a chip optimized for inference based on Groq's design at GTC 2026. According to sources, the chip may adopt SRAM instead of HBM, which is commonly used in traditional AI chips. This architectural choice is expected to significantly reduce inference latency and power consumption.
Meanwhile, according to South Korean tech media The Elec,$Tesla (TSLA.US)$the postponement of multi-project wafer production plans has forced DeepX, a South Korean AI chip company, to delay the mass production schedule of its next-generation NPU by approximately six months.
These developments reflect the complex situation currently faced by Samsung Foundry: on one hand, the rapid expansion of AI inference chip demand brings new customers and incremental orders; on the other hand, changes in the production schedules of major clients like Tesla have substantially impacted other customers sharing the same production line.
Groq's production expansion allows Samsung Foundry to enter the inference chip market.
According to reports citing industry insiders, AI chip startup Groq's production at Samsung Electronics ( $CSOP Samsung Electronics Daily (2x) Leveraged Product (07747.HK)$ Production is transitioning from the prototype stage to mass production. Previously, Groq's production at Samsung mainly involved sample chips, used to evaluate their suitability for AI inference scenarios. This year is considered the early phase of formal large-scale commercialization.
Although the absolute scale of Groq’s orders placed with Samsung remains relatively limited, this order is expected to help Samsung Foundry establish a foothold in the AI inference chip market.
Additionally, the processor developed by HyperAccel, a South Korean inference AI chip startup, is entirely produced using Samsung Foundry’s 4-nanometer process, further highlighting Samsung’s strategic positioning in the niche market of inference chips.
Tesla's delays affect DeepX as scheduling conflicts arise in the 2-nanometer production line.
Samsung Foundry's 2-nanometer production line is facing order disruptions from Tesla. According to The Elec, citing sources familiar with the matter, Tesla has postponed its Multi-Project Wafer (MPW) production plan, directly impacting South Korean AI chip company DeepX, which also relies on the same production line.
MPW is a model that allows customers to share wafers to split prototyping costs. DeepX had originally planned to initiate MPW production of its second-generation NPU chip, DX-M2, in April this year, but it has now been delayed by about six months. Reports indicate that DX-M2 is the first external customer chip specifically designed for Samsung’s 2-nanometer process, while Tesla is also using the same process to manufacture its AI6 chip.
The specific reasons for Tesla's MPW delay have not been disclosed. According to industry insiders cited in the report, adjustments to the production timelines for autonomous vehicles and humanoid robots, along with changes in supercomputer investment plans, may have collectively caused the delay.
DeepX’s clients reportedly include Samsung Electronics, Hyundai Motor,$Intel (INTC.US)$and the company recently delivered chips to$Baidu (BIDU.US)$The chip has been delivered.
Tesla may significantly expand AI6 chip production as attention focuses on Samsung’s 2-nanometer capacity.
Despite recent order disruptions, the partnership between Tesla and Samsung continues to deepen. According to The Elec, Tesla recently dispatched a procurement executive to Samsung Electronics to negotiate an expansion of its 2-nanometer AI6 chip production.
Under the existing contract, the agreed monthly wafer output was 16,000 units. If a new agreement is reached, the monthly output could increase to approximately 40,000 units, representing an increase of over 100%.
This potential expansion plan indicates that Tesla's long-term reliance on Samsung’s 2-nanometer process may significantly deepen, meaning that Samsung Foundry’s 2-nanometer capacity will face even tighter scheduling pressures. Coordinating capacity allocation among multiple clients, including Tesla and DeepX, will become a key operational challenge for Samsung in the near term.
Editor/Rocky