U.S. chip giant Broadcom has reportedly canceled plans to build a semiconductor packaging and testing plant in Spain, dealing a blow to the country's ambitions of becoming a major player in the European chip industry. The investment, originally announced in 2023, was expected to be worth nearly $1 billion and create around 500 jobs. The facility was to be a key part of Spain's "PERTE Chip" initiative, supported by up to €12 billion in EU pandemic recovery funds.

Talks between Broadcom and the Spanish government broke down in 2024 amid reported political complications in both countries. Spain's unstable internal government dynamics may have slowed progress, while the re-election of U.S. President Donald Trump and his renewed focus on domestic semiconductor manufacturing through tariffs are believed to have contributed to Broadcom's retreat.

The project had faced uncertainty from the beginning. Broadcom never publicly confirmed the $1 billion investment figure—originally provided by Spain's Ministry of Economy—and no final site was selected. Meanwhile, competitors such as Cisco have made tangible progress with local chip design centers, casting further doubt on Broadcom's stalled initiative.

This is not an isolated case. Other major chipmakers, including Intel and Wolfspeed, have delayed or scaled back European projects, citing economic and geopolitical uncertainties.

Despite pulling back from the Spain deal, Broadcom is doubling down on its AI and high-performance computing (HPC) ambitions. On July 15, the company launched its most powerful Ethernet switch to date—Tomahawk Ultra—a cutting-edge networking processor designed to support massive AI workloads and compete directly with Nvidia's NVLink Switch.

The Tomahawk Ultra acts as a high-performance data traffic controller within AI clusters and HPC systems, enabling efficient, low-latency communication among tightly coupled chips. With a staggering 51.2 Tbps throughput and ultra-low latency of 250 nanoseconds, the chip supports up to four times more connections than Nvidia's competing solution—all while leveraging a speed-optimized Ethernet protocol instead of a proprietary standard.

Broadcom engineered the chip to meet the needs of modern AI and scientific computing applications. It introduces features like in-network collectives (INC), which offload collective operations such as AllReduce and Broadcast directly onto the switch chip, boosting computing efficiency and freeing up resources on AI accelerators. Its adaptive Ethernet headers reduce protocol overhead from 46 to 10 bytes while maintaining full Ethernet compatibility, increasing network efficiency and flexibility.

Additionally, the chip implements a truly lossless Ethernet fabric through mechanisms like Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC), ensuring no packet loss during high-volume data transfers—an essential feature for AI training and inference at scale.

Manufactured by TSMC using a 5nm process, the Tomahawk Ultra took three years to develop and represents Broadcom's latest offensive in the battle for AI infrastructure dominance. As Broadcom shifts focus from uncertain overseas investments to core AI technology leadership, the Tomahawk Ultra positions the company as a formidable challenger to Nvidia in the next-generation networking landscape.