As manufacturing demand from big tech corporations concentrates on the advanced processes of TSMC, the world's largest foundry (semiconductor contract manufacturing) company, the bottleneck is worsening, and even top client Nvidia is reportedly set to adjust its output. The plan is to cut production of the top-tier model of the next-generation artificial intelligence (AI) accelerator "Rubin," called "Rubin Ultra," slated for release next year, while increasing production of the previous-generation Blackwell platform.
According to the industry on the 2nd, Nvidia is said to be scaling back production of Rubin Ultra, which was scheduled to go into mass production on TSMC's 3-nanometer (nm; one-billionth of a meter) node compared with the original plan. Instead, it reportedly will increase production of the Blackwell platform, which is made on 4 nm. TSMC is expanding capacity for its state-of-the-art 3 nm process now in mass production, but demand concentrated among clients such as Apple is seen making on-time delivery difficult.
Bottlenecks have intensified as the foundry division of Samsung Electronics, a competitor to TSMC, struggled to secure clients due to poor Production yield on nodes at 3 nm and below. On top of that, the AI industry has been growing steeply, sending demand for AI chip manufacturing surging. Unlike the past, when smartphone application processor (AP) clients such as Apple and Qualcomm accounted for a significant share of manufacturing volume, more big tech corporations like Google, Amazon, and Meta are seeking to manufacture AI chips.
Amid this, with TSMC's production capacity constrained and on-time delivery appearing difficult, Nvidia also seems to have moved to adjust its output. Starting this year, Nvidia will mass-produce the AI chip Rubin, which introduces TSMC's 3 nm process for the first time. The previous-generation Blackwell platform is mass-produced on the 4 nm process.
Nvidia is TSMC's largest client, and the fact that even Nvidia adjusted output indicates just how severe TSMC's capacity constraints are. Nvidia is seen as aiming to manage the mass-production schedule and volume of the top-end Rubin version while ramping up volumes of the Blackwell platform, which is produced on the relatively less constrained 4 nm lines.
Some say the bottleneck from limited capacity is materializing—TSMC has said its processes are "sold out" through 2028—which could present an opportunity for the foundry division of Samsung Electronics. While the TSMC concentration worsened as Samsung Electronics' foundry division was hampered by low Production yield, the advanced process yield at 3 nm and below is now said to have stabilized.
In fact, Tesla has formalized 2 nm cooperation with the foundry division of Samsung Electronics for AI chip production, and global big tech corporations such as Qualcomm and AMD are reportedly exploring 2 nm cooperation with Samsung Electronics' foundry division. Groq, an AI Semiconductor startup indirectly acquired by Nvidia, also asked the foundry division of Samsung Electronics to increase output.
A semiconductor industry official said, "Concerns were raised 2 to 3 years ago that a bottleneck could worsen as TSMC effectively cornered advanced processes, and that trend is becoming reality in tandem with the steep growth of the AI industry," adding, "At this point, as Samsung Electronics' advanced process Production yield stabilizes, it is becoming an opportunity to win orders for volumes from big tech corporations."