[DigitalToday reporter Chi-gyu Hwang] Nvidia is preparing to launch a dedicated inference chip to speed up AI response times. Nvidia CEO Jensen Huang (젠슨 황) is expected to unveil the new chip at the GTC developer conference starting on March 16 local time, the Financial Times reported on March 13 local time.

The chip is the first output since the company hired Groq founders in a $20 billion deal in December. Nvidia plans to introduce a record-based language processing unit, or LPU, as a product line alongside Vera Rubin, its next flagship GPU. Groq developed an LPU designed to respond quickly to complex AI queries and has produced products in cooperation with Samsung.

Nvidia has argued that a single GPU can handle both training and inference, but it revised that strategy as AI tools such as agentic coding systems became more sophisticated, the FT reported.

Bank of America analysts estimate that inference will account for 75 percent of total spending when the AI data centre market reaches about $1.2 trillion in 2030.

Nvidia's new chip uses SRAM instead of high-bandwidth memory, or HBM. HBM is expensive, and supplies are tight because memory makers such as Samsung Electronics, SK Hynix and Micron are unable to keep up with AI demand. SRAM has smoother supply and is suited to boosting the speed of AI inference work, the FT reported.

· [Tech Inside] What Nvidia's Groq acquisition signals for AI data centres · [Tech Insight] View Nvidia's Groq acquisition through the lens of HBM risk

Nvidia's move is expected to further intensify competition among leading companies over inference AI chips. Amazon Web Services, or AWS, also announced a multi-year partnership with AI semiconductor startup Cerebras and said it would deploy Cerebras inference chips in its data centres. Meta also unveiled four types of custom-designed chips aimed at AI workloads.

· AWS allies with Cerebras, takes aim at the AI inference market · Meta unveils its own AI chips after large Nvidia and AMD deals