The relationship between NVIDIA and TSMC, which has been fruitful for almost two decades, is facing significant tensions in the context of the new Artificial Intelligence (AI) GPUs. According to recent reports, NVIDIA is considering alternatives to reduce its dependence on TSMC due to increasing pressure on wafer prices and difficulties with the production capacity of the latest lithographic nodes of the Taiwanese company. In this scenario, Samsung emerges as a possible option for the American giant.
Prices and delays at TSMC are driving NVIDIA towards Samsung
TSMC, known for its advanced technology and production capacity, has been NVIDIA’s main supplier for years. However, the need to maintain profit margins above 50% and the increases in wafer costs, which could exceed $30,000 for advanced nodes, have led NVIDIA to seek more cost-effective alternatives. AI GPUs, such as those using the Blackwell architecture, require manufacturing capacity that is not only advanced but also competitive in terms of pricing.
TSMC’s N2 node, expected for 2026, is now scheduled to begin mass production in 2025. However, the success rate and performance of this node are causing uncertainty. Even companies like Apple have decided to postpone the use of the N2 node in their future A19 Pro chips. In this context, NVIDIA is considering the possibility of shifting some of the production of its future GPUs, including those based on the Vera Rubin architecture, to Samsung’s factories.
Samsung, an alternative with competitive advantages
Samsung, which recently signed a significant contract with Arm for production on its 2nm SF2 node, is seizing the opportunity to attract NVIDIA. Although the South Korean company has faced financial challenges and lost customers to TSMC in previous nodes, its ability to offer significant discounts, ranging from 20% to 30%, is a factor that NVIDIA cannot ignore.
Furthermore, Samsung promises to surpass TSMC in mass production with its SF2 node, which incorporates GAA (Gate-All-Around) and BSDPN (Backside Power Delivery Network) transistor technology. This technical advancement, along with more competitive prices, is enticing NVIDIA, which could move the production of Vera Rubin GPUs and other future products to South Korea.
Issues with Blackwell and HBM3E memory
Another aspect that has complicated the relationship between NVIDIA and TSMC are the issues with HBM3E memory, which have affected the performance of Blackwell architecture GPUs. These setbacks have caused delays and increased tensions between the two companies, leading NVIDIA to rethink its loyalty to TSMC.
The last time NVIDIA moved part of its production to Samsung, using its 8nm node for its GPUs, TSMC reacted by increasing its fees when the American company decided to return to its usual supplier. This dynamic of tense negotiations and fluctuations in costs has led NVIDIA to once again explore Samsung as a viable option to secure its future in the GPU market.
The future of GPUs and NVIDIA’s decision
NVIDIA’s final decision is still unclear. The company led by Jensen Huang could continue with TSMC, which offers higher transistor density and superior performance in its N2P node, or switch to Samsung, which proposes earlier production and more competitive pricing.
With the expected launch of Vera Rubin GPUs in 2026, the competition between these two tech giants could intensify. NVIDIA is at a crossroads: stay loyal to TSMC, its main partner for years, or diversify its production and take advantage of the benefits offered by Samsung in terms of costs and advanced technology.
The evolution of these negotiations will determine not only the future of NVIDIA but also the development of the semiconductor market, a sector in constant change and essential for the current technological revolution.
via: El Chapuzas Informático