With Nvidia's annual technology conference "GTC 2026" just a week away, attention is focused on whether Nvidia will unveil a new artificial intelligence (AI) chip specialized for inference. Rivals have released custom in-house chips strong in inference, putting Nvidia—which has led the graphics processing unit (GPU) market—on alert. As Nvidia has begun providing customers with the first test samples of "Vera Rubin," the latest AI accelerator it unveiled early this year, attention is also expected to focus on the competition between Samsung Electronics (005930.KS) and SK hynix (000660.KS), key suppliers of memory chips.
According to Nvidia on Monday, the company will hold GTC 2026 in San Jose, California, from the 16th to the 19th. More than 30,000 developers, researchers and corporate executives from over 190 countries will attend, and more than 1,000 topics will be covered during the event.
CEO Huang will deliver the keynote address at the SAP Center on the 16th. In the speech, he will introduce Nvidia's latest innovative technologies and present its direction, ranging from accelerated computing, AI factories, open models, and agentic (assistant-type) systems to physical AI. Discussions with industry leaders including Andreessen Horowitz, the Allen Institute for AI, Black Forest Labs, Cursor, Reflection AI and Thinking Machines Lab are also planned.
Will an AI Chip for Inference Be Unveiled?
One of the most closely watched events ahead of GTC is whether Nvidia will unveil a new inference chip. The Wall Street Journal (WSJ) reported last month, citing sources, that Nvidia will unveil a new chip specialized for inference at this GTC, calling it a potentially significant business overhaul that could shift the landscape of the AI competition.
Nvidia has dominated the AI data center server market with GPUs strong in AI training. AI startups such as OpenAI have focused on developing foundation models, including large language models (LLMs), through large-scale training on servers equipped with GPUs. But as the AI industry shifted toward agentic (assistant-type) AI last year, the importance of inference chips came to the fore.
GPUs offer versatility capable of performing both training and inference, but they have limitations, being inefficient at the inference stage due to high power consumption and cost burdens. In agentic AI, the ability to act according to circumstances based on what has been learned is important, so inference chips that quickly retrieve data when needed are preferred. While broad training requires GPUs optimized for parallel computing and large-capacity memory, inference depends on speed and efficiency, requiring neural processing units (NPUs) specialized for lightweight computing and lighter memory.
If an inference chip is unveiled this time, it would be the first collaboration chip released since Nvidia made a roundabout acquisition of Groq for a record 20 billion dollars (29.33 trillion won) late last year. Groq is an inference chip developer. Nvidia, leveraging its financial power, absorbed the thorn-in-its-side Groq by buying up key technology and personnel.
With rivals such as Google, Amazon and Meta having already developed or developing their own inference-specialized AI chips, Nvidia's new inference chip could be a reversal card to expand its influence in the inference market as well. With OpenAI—considered a powerful partner—having raised complaints that Nvidia's GPUs are not suitable for inference, and reports that Nvidia's investment in OpenAI had shrunk, an inference chip could provide an opportunity to strengthen the partnership with OpenAI again.
1-Nanometer Class? Hopes for Feynman Hints
Whether an early teaser hinting at the AI accelerator "Feynman" will be released is also of interest. Feynman is a next-generation GPU expected to launch as early as the second half of next year. It is observed to surpass the performance limits of existing accelerators by adopting Taiwan's TSMC 1-nanometer (nm; 1 nm is one billionth of a meter) class process and HBM5, the eighth generation of high-bandwidth memory.
Earlier, at last year's GTC, Nvidia presented a roadmap to release "Vera Rubin" and Feynman in succession following the "Blackwell" series, which was its latest chip at the time. Nvidia has unveiled next-generation AI chips through GTC every year. Some analysts say that the 1-nanometer class process will be applied to Feynman, which is being developed with a target of launching in the second half of next year or in 2028.
Has the China H200 Export Been Shelved?
Whether CEO Huang will continue to attempt to export the H200 semiconductor to China is also of interest. Foreign media reported last May that Nvidia switched the production facilities of Taiwan foundry company TSMC from the H200 to the latest "Vera Rubin." Ahead of the US-China summit to be held in China later this month, whether H200 exports will resume has been regarded as an indicator to confirm changes in the trade conflict.
The H200 is two generations behind Vera Rubin, but it is still rated as a GPU suitable for advanced AI development. Nvidia was previously known to have received orders for 1 million H200 units from Chinese companies. Nvidia CEO Jensen Huang said early this year, "Demand is very high, so we are operating the supply chain, and the H200 is being rapidly supplied to the production line."
But even with the administration of US President Donald Trump allowing the resumption of H200 exports, the US-China standoff has continued, making H200 sales uncertain. President Trump announced in December last year a policy to allow exports of Nvidia's H200 chip to China, but approval has been delayed due to strict conditions from the Commerce Department. China, too, has issued guidelines telling its companies to purchase the H200 only when necessary, while encouraging the use of domestically produced AI chips.
David Peters, Assistant Secretary of Commerce, said at a hearing of the House Foreign Affairs Committee on the 24th of last month that the H200 had not yet been sold to China. Colette Kress, Nvidia's chief financial officer (CFO), also said at a quarterly earnings conference call on the 25th of last month, "We received US government approval for a small quantity of H200 products, but have not yet generated revenue," adding, "It is also unclear whether imports into China will be allowed."
Samsung Electronics and SK hynix Compete for Supply Before Jensen Huang
Interest is also high in the competition between Samsung Electronics and SK hynix, Nvidia's main memory semiconductor suppliers. SK Group Chairman Tae-won Choi is known to be planning to attend GTC 2026, and if he does, a meeting with CEO Huang is also anticipated. It would be a reunion just a month after they held a "chicken and beer meeting" in the United States last month. It is the first time Chairman Choi will visit the GTC site in person. At Samsung Electronics, Song Yong-ho, vice president of Samsung Electronics' DS Division, will give a presentation on the theme of "The Future of Semiconductor Manufacturing Through AI."
At this year's event, fierce competition between Samsung Electronics and SK hynix is expected to seize leadership in the market for HBM4, the next-generation HBM to be mounted on Vera Rubin. Nvidia is known to have allocated about two-thirds of the HBM4 volume for use in Vera Rubin and other products this year to SK hynix, but Samsung Electronics has entered the HBM4 market first.
CEO Huang predicted, "AI is no longer a single application or technological innovation but is establishing itself as essential infrastructure," adding, "Every company will utilize AI, and every country will build AI." He stressed, "Every layer of the AI stack, from energy, chips, infrastructure and models to applications, is advancing simultaneously, and at GTC you will be able to see the scene of this change firsthand."
Will Samsung Ultimately Take the Seat Next to Nvidia's Rubin? [Semiconductor Encyclopedia]
