AI Data Center Demand Spills Into Industrial Chip Shortages, Texas Instruments Says Mature Nodes Are the Choke Point

On Nov. 11, at a conference called Embedded Labs 2026 at COEX in Seoul's Gangnam district, Park Joong-seo, CEO of Texas Instruments Korea, told reporters that the squeeze created by AI data centers has moved beyond memory. Power and embedded semiconductors are getting tight too, he said, and building more advanced fabs does nothing to increase the supply of industrial chips.

His explanation is a mismatch in process nodes. Microcontrollers for general-purpose and real-time control, along with communications and radar chips, are mostly made on mature processes between 28nm and 130nm. Global investment, meanwhile, is concentrated on 2nm and 3nm advanced nodes and high-bandwidth memory. Park put numbers on the economics: advanced wafers cost $40,000 to $50,000 each, while legacy wafers run roughly $1,000 to $2,000. From the standpoint of outside foundries, he said, the legacy business is not profitable enough to justify new capacity, so they avoid adding it.

TI's answer is to rely on its own factories. The company, an American analog chipmaker that introduced the first silicon transistor in 1954, runs 15 sites for manufacturing, assembly and testing across eight countries and regions, including the United States, Germany and Japan. In 2010 it began moving to 300mm (12-inch) wafers, the first in the analog industry to do so. A 300mm wafer yields 2.2 to 2.3 times as many usable dies as a 200mm one. Park said that cuts chip production costs by about 30% and more than doubles shipment volumes at the same process and lead time. He added that TI has internalized 95% of its embedded production equipment, covering its mainstay 45nm and 65nm processes and extending to 28nm in the future, which he said guarantees stable lead times even when outside foundry capacity is short.

The demand side is visible in TI's third-quarter results. Total revenue was $6.055 billion, up 28% from a year earlier, with analog accounting for 81.1%. Embedded revenue reached $858 million, up 21%. Data center revenue rose about 95%, industrial revenue grew more than 40%, and automotive posted a mid-teens increase. Only growth rates, not dollar figures, were disclosed for those three segments.

AI servers explain part of that growth. Beyond GPUs and CPUs, they carry large numbers of voltage regulation chips that deliver power and remove heat, plus sensor and motor control chips for cooling equipment. TI has worked with Nvidia on an 800-volt direct current design meant to reduce power loss. Delivery efficiency is 84.6% on the existing 48V-centered structure and 85.2% with 800V DC, and is expected to reach 86.7% after 2027 if transformer and distribution structures also change. Park said every conversion stage between high-voltage power and the sub-1V levels AI processors use adds loss and heat, and TI is working with major global platform companies to cut the steps from as many as five to three.

Park also questioned the industry's focus on TOPS figures. Setting price and power aside, he said, one could build a 10,000-TOPS chipset, but an autonomous driving platform costing 20 million won (about $15,000) inside a 20 million won car would not be accepted by the market. He argued that competitiveness lies in delivering practical functions at affordable prices through software optimization, and that Level 3 functions can be implemented at 400 to 500 TOPS.

On humanoid robots, he said price, not technological maturity, will decide when they go mainstream. A 500 million won (about $370,000) robot would sell only for research, he said, while one in the 10 million won (about $7,500) range doing human-level labor would open the market explosively. The industry expects a large market to form around 2035, starting from 2030. He named GaN-based power chips, motor drive integration for longer battery life, and the integration of functional safety and cybersecurity standards as challenges.

TI is also trying to ease switching costs. Its Code Composer Studio development environment has shortened development from years to months or weeks, and migration tools let customers short of other companies' parts carry over existing firmware expertise. For SimpleLink wireless chips, TI said wireless battery management network availability exceeds 99.999%. The assessment accompanying Park's remarks is that suppliers with their own 300mm capacity hold an edge in lead times and costs, while industrial and automotive customers dependent on mature nodes must weigh supply diversification against the cost of redesigning their products.