# AI chip demand surges as shorter development cycles and higher yields become critical

By Riza Dagoc

Canonical URL: https://www.tokenpost.com/news/insights/34639
Published: 2026-10-19T07:28:05.000Z
Updated: 2026-10-19T07:28:05.000Z
Section: Insights & Views

> AI's rapid development is increasing pressure on chipmakers to shorten research cycles, improve yields and build integrated systems. Existing data centers may also play a larger role in meeting demand.

Keeping pace with advances in AI will require shorter semiconductor development cycles, higher manufacturing yields and integrated systems that make better use of existing data centers, according to an industry analyst.

As AI develops on an annual cycle, the semiconductor industry is struggling to expand its technology and production capacity fast enough. Competition is also shifting from individual chips toward end-to-end data center design.

In an Oct. 17 interview with Schwab Network, Steven Sopko, practice lead of Semiconductor and DeepTech at Hyperframe Research, said AI can quickly outpace the industry's ability to keep up.

Sopko said Applied aims to reduce the time needed to research chip materials and manufacturing equipment from the current 10 to 15 years to six to eight years.

Capacity constraints are another major challenge. Citing ASML's remarks during its earnings announcement, Sopko said equipment suppliers including Applied, ASML and Lam Research will struggle to meet demand.

Improving yields could also increase supply. A modern chip requires 500 to 1,000 process steps, leaving numerous opportunities for defects, and some chips are discarded even when production runs well. Reducing that waste would make more AI chips available, Sopko said.

Broadcom operates across several areas, including custom chips for hyperscalers and photonics that help data center systems communicate more efficiently. Sopko said the breadth of its operations can make it difficult to identify which segment is driving performance, while describing the company's guidance as strong.

AMD's new server racks reflect efforts to organize data center systems more efficiently. In early AI training environments, one CPU managed 8, 12 or 24 GPUs, but agentic AI is bringing the CPU and GPU mix closer to parity as CPUs orchestrate other chips.

The shift is prompting chipmakers to put CPUs back at the center of data center design and build integrated systems spanning networking and data. Sopko said the focus is no longer only on silicon but on the end-to-end system needed for all components to work together.

Intel has improved execution over the past 18 months and has a distinctive U.S. capability for packaging complex chips, Sopko said. Its card for servers in air-cooled data centers could add GPU-like computing power, allowing enterprises to run more AI and agentic AI workloads in existing facilities.

Sopko said the industry must focus not only on building large new data centers but also on adapting its extensive existing infrastructure to the AI surge. Using older facilities more effectively could help address added costs and pressure on the energy grid.