Intel's Falcon Shores AI GPU will use TSMC's 3nm process and CoWoS advanced packaging.

Intel's next-generation Falcon Shores AI GPU will utilize TSMC's 3nm process rather than Intel's own manufacturing capabilities. The decision reflects broader challenges at Intel's Foundry Services, which has struggled with competitive execution. More significantly, Intel has yet to establish itself as a meaningful competitor in the AI accelerator market through its Gaudi product line, prompting the company to pursue this alternative approach.

Falcon Shores represents Intel's first true GPU IP designed specifically for the AI segment, built upon the next-generation Xe graphics architecture. The company has positioned the architecture as substantially superior to its Gaudi offerings, framing it as the product that could bring "redemption" to Intel in the markets. However, Intel's track record offers caution: while the Ponte Vecchio chip was an engineering achievement, it failed to build a significant customer base, and the Aurora Supercomputer deployment was severely delayed. According to Taiwanese media outlet Cnyes, Intel is pursuing an "outsourcing" route to ensure it can compete directly with NVIDIA's Blackwell products.

Intel's product strategy for Falcon Shores is notably ambitious. Rather than focusing on specialized segments, the company plans to develop high, mid-range, and low-end variants to serve diverse customer bases across the AI market, including in heavily regulated regions such as China. The GPUs will employ TSMC's CoWoS-R packaging, the same advanced solution used in NVIDIA's Blackwell chips and recognized as among the best available for HPC and datacenter applications. This complete reliance on TSMC for production offers quality assurance but introduces supply chain risk as demand across AMD, NVIDIA, and Intel all draw from the Taiwan-based manufacturer.

Falcon Shores GPUs are expected to demonstrate substantial power delivery, with reported thermal design powers reaching as high as 1500W, suggesting significant generational performance improvements. The architecture will utilize liquid cooling across its lineup, though some entry-level variants may employ alternative cooling solutions for lower-power configurations. The timeline points toward launch in 2025, though questions remain about TSMC's capacity to fulfill Intel's orders amid surging AI demand.

Intel's push into competitive AI accelerators reflects necessity—its Gaudi generation generated only $500 million in the current year, far below competitor earnings. Despite this revenue lag, Intel's software team continues development work to bring its AI offerings closer to competitive parity. Whether Falcon Shores can successfully establish Intel as a credible alternative in a market dominated by NVIDIA remains an open question for the year ahead.