Chips Worth Tens of Millions of Won Unobtainable Even After a Year... Tech Companies Dream of Escaping Nvidia

NVIDIA Challenges 'CUDA' Monopoly Structure
Requires Careful Strategy and Long Preparation Time

Jensen Huang, CEO of NVIDIA, announced the Blackwell platform at a pre-event of COMPUTEX held in Taipei, Taiwan on June 2. Photo by Reuters and Yonhap News Agency

Movements to break away from the AI chip ecosystem monopolized by Nvidia are spreading globally. Apple, in collaboration with Broadcom, has embarked on developing AI server chips optimized for its own ecosystem, while Canadian fabless startup Tenstorrent is attempting to design customized chips that are not dependent on specific hardware. Interest is focused on efforts to move away from Nvidia, as reducing dependence on Nvidia could lead to a decrease in semiconductor prices. However, experts point out that overcoming Nvidia's powerful 'CUDA' ecosystem and network requires careful strategy and long-term preparation.

According to industry sources on the 16th, efforts to break away from the Nvidia ecosystem are appearing across the industry. Apple recently partnered with Broadcom to develop server chips for AI computation, and Tenstorrent is developing customized design chips using the open semiconductor architecture 'RISC-V' that are not dependent on any single company. Recently, to expand the RISC-V ecosystem, they have started advanced semiconductor design contract services in Japan.

AMD has developed the open-source high-performance computing platform ROCm in response to Nvidia CUDA, and Intel is competing with Nvidia GPUs through its Gaudi AI accelerator. Google and Amazon are also building their own AI accelerators for their data centers to reduce dependence on Nvidia. South Korean AI semiconductor company FuriosaAI is preparing to launch the high-performance AI semiconductor 'Renegade (RNGD)' with enhanced cost efficiency. Park Seong-hyun, CEO of Rebellion, evaluated that "the global AI semiconductor market restructuring has begun."

The reason behind these corporate movements is the 'Nvidia dominance in AI chips' phenomenon. Nvidia has built a powerful ecosystem centered around CUDA, going beyond being a simple hardware supplier. CUDA is a proprietary technology that only works on Nvidia graphics processing units (GPUs). Programs developed with this technology do not run on GPUs from AMD, Intel, or Apple. This monopoly structure limits companies' choices and is cited as a factor driving up chip prices. The next-generation product B (Blackwell) 200, which Nvidia began releasing in the fourth quarter of this year, is known to cost $40,000 (about 57 million KRW) per chip.

Even obtaining it on time is difficult. An industry insider said, "The biggest buyers of Nvidia products are global big tech companies like Google, Microsoft (MS), and Amazon Web Services (AWS). However, even these companies have to prepay and place orders but still wait over a year and cannot receive as many units as they want." He added, "At the same time, Nvidia continues to raise product prices, so any company has increasing motivation to break free from dependence."

The monopoly supply also leads to a lack of technical support. An industry source said, "For companies like Google, hundreds of Nvidia engineers are directly dispatched to complete projects, but from Nvidia headquarters' perspective, Korea is a low priority and is thus excluded from technical support." Recently, Yoo Young-sang, CEO of SK Telecom, mentioned at the 'SK AI Summit,' "While big tech companies are engaged in a GPU acquisition war, it is not easy for us to access the latest GPUs," adding, "Meta and MS own 150,000 H100 units, but in Korea, the total is said to be only about 2,000."

However, the consensus is that attempts to move away from Nvidia are still in the early stages. Tenstorrent has secured customer contracts worth about $150 million so far, but there is still a large gap compared to Nvidia, which earns hundreds of billions of dollars in revenue per quarter. An industry insider said, "Non-Nvidia camps have repeatedly united and separated while discussing ecosystem construction plans, but there is still a long way to go," adding, "It will take several more years to become a real competitor."