“I would like to express my gratitude to Samsung.”

At NVIDIA’s annual developer conference “GTC 2026,” held at the SAP Center in San Jose, California, on the 16th (local time), NVIDIA CEO Jensen Huang took the unusual step of publicly thanking Samsung Electronics while announcing the performance of NVIDIA’s latest artificial intelligence (AI) accelerator, “Vera Rubin,” during his keynote speech.

He said, “Samsung is making the ‘Groq 3’ Language Processing Unit (LPU) for us,” adding, “Shipments are expected to begin around the third quarter (July–September) of this year.” The Groq 3 LPU is an inference AI chip that NVIDIA expects will open up a new market beyond the current AI accelerators centered on graphics processing units (GPUs). With the public confirmation that NVIDIA has entrusted production of its next-generation inference AI chip to Samsung Electronics, observers say the “AI semiconductor alliance” between Samsung Electronics and NVIDIA has expanded from memory supply to include foundry, or contract semiconductor manufacturing.

● NVIDIA entrusts “inference AI” to Samsung Electronics

NVIDIA announced that to boost the inference performance of its new AI accelerator Vera Rubin, it will equip it with the LPU of AI semiconductor startup Groq, which NVIDIA effectively acquired in December last year for USD 20 billion (about KRW 29.8 trillion).

Until now, AI has focused on how quickly it can learn large volumes of information, making GPUs, which excel at large-scale parallel processing, the key component. However, as AI moves beyond the training phase into the actual task-execution phase—often described as the era of “AI agents”—inference-optimized AI chips capable of processing data with relatively low power consumption are drawing attention. The Groq 3 LPU mentioned by CEO Huang is NVIDIA’s answer to secure dominance in the inference AI chip market, in addition to its existing stronghold in GPUs.

The Groq 3 LPU is produced using Samsung Electronics’ 4 nm (nanometer; 1 nm is one billionth of a meter) foundry process. Hwang Sang‑joon, Executive Vice President in charge of memory development at Samsung Electronics, told reporters at the GTC 2026 venue, “The Groq inference‑only chip is being manufactured at the Pyeongtaek campus,” adding, “We have already received more orders than expected.”

The inference AI chip market has already become a “battlefield” for Big Tech companies seeking to secure future leadership. Meta has introduced its own inference‑optimized AI chip series “MTIA.” Microsoft and Amazon are also rolling out and advancing their respective inference chips, “Maia 200” and “Inferentia.” In this environment, NVIDIA has joined hands with Samsung Electronics to push for dominance in the inference AI market.

During his keynote, CEO Huang repeatedly stressed the importance of inference AI. “With the emergence of AI agents, the amount of required inference has increased to 10,000 times what it was when ChatGPT first appeared,” he said, adding, “The inflection point for inference has arrived.” NVIDIA plans to equip its next‑generation AI accelerator Vera Rubin with the Groq 3 LPU and operate it in a configuration where large‑scale computation is handled by GPUs and rapid responses are handled by the Groq 3 LPU. NVIDIA expects that combining GPUs and LPUs can increase performance‑per‑watt efficiency by up to 35 times.

● “Let’s go, Samsung,” “Jensen♡SK hynix”

CEO Huang also showed keen interest in Korean companies after his keynote. Visiting the Samsung Electronics booth at GTC 2026, he remarked, “Samsung is the best in the world,” and said, “Let’s go (GO) Samsung.” At the SK hynix booth, he spoke with SK Group Chairman Chey Tae‑won and said, “You are perfect.” He also signed a Vera Rubin prototype with the phrase “Jensen♡SK hynix.” CEO Huang presented a revenue target of USD 1 trillion (about KRW 1,490 trillion) for NVIDIA’s AI chips next year, which is interpreted as highlighting the need for cooperation with Samsung and SK to achieve this.

Meanwhile, Samsung Electronics unveiled for the first time its next‑generation high bandwidth memory (HBM), the HBM4E (7th generation), at the event. Samsung plans to supply HBM4E samples to customers in the third quarter of this year and start mass production in the fourth quarter (October–December).

Popular News