Broadcom's 3.5D Chip Tech Ignites New AI Hardware Arms Race

PALO ALTO, CA – February 26, 2026 – Broadcom Inc. today announced a significant escalation in the artificial intelligence hardware arms race, confirming it has begun shipping the industry’s first custom compute System-on-Chip (SoC) built on a cutting-edge 2-nanometer process. The new chip leverages the company’s proprietary 3.5D eXtreme Dimension System in Package (XDSiP) platform, a novel approach that could redefine the performance and efficiency of the massive data centers powering the AI revolution.

The first shipments are destined for Japanese technology giant Fujitsu, which will utilize the advanced silicon as a core component of its next-generation FUJITSU-MONAKA processor initiative. The move signals a broader industry shift, where advanced packaging technology—how multiple chips are integrated together—is becoming a critical battleground for achieving computational breakthroughs as the benefits of traditional chip shrinking, dictated by Moore's Law, begin to wane.

The New Frontier in Chip Packaging

At the heart of Broadcom's announcement is its 3.5D XDSiP platform, a hybrid technology that merges established 2.5D packaging with advanced 3D stacking. While 2.5D integration, which places chips side-by-side on a silicon interposer, is common in today's high-end GPUs, and true 3D stacking has faced commercialization hurdles, Broadcom's 3.5D approach offers a powerful middle ground.

The key innovation is its use of Face-to-Face (F2F) bonding. Instead of stacking silicon dies in a conventional top-to-bottom manner, F2F integration connects them directly top-to-top. This seemingly simple change has profound implications. According to technical documentation, this method achieves a sevenfold increase in signal density between the stacked dies compared to traditional methods. Furthermore, it enables a tenfold reduction in the power consumed by the die-to-die interfaces by utilizing 3D Hybrid Copper Bonding, drastically cutting down on energy-wasting interconnect bottlenecks.

This architecture allows chip designers to build modular systems, referred to as XPUs, where compute, memory, and networking components can be scaled independently and then integrated into a single, compact package. The platform, which relies on TSMC's advanced CoWoS-L (Chip-on-Wafer-on-Substrate) packaging, can integrate over 6000 mm² of silicon and up to twelve High Bandwidth Memory (HBM) stacks, providing the immense memory bandwidth essential for training and running large AI models.

“We’re proud to deliver the first 3.5D custom compute SoC for Fujitsu – a testament to the outstanding execution and innovation by the Broadcom team,” said Frank Ostojic, senior vice president and general manager of Broadcom’s ASIC Products Division. “Since introducing our 3.5D XDSiP platform technology in 2024, Broadcom has expanded its 3.5D platform capabilities to support XPUs for our broader customer base that will ship from 2H '26.”

Custom Silicon's Ascendance in the AI Era

Broadcom’s partnership with Fujitsu highlights a growing trend among technology giants: the move away from off-the-shelf hardware toward custom-designed silicon. As companies like Google, Amazon, and Meta have demonstrated with their own custom AI chips, bespoke hardware offers significant advantages in performance, power efficiency, and cost, providing a crucial competitive edge in the AI-driven economy.

Fujitsu's FUJITSU-MONAKA initiative is a prime example. Set for a 2027 release, the Arm-based processor is designed to deliver unparalleled performance and power efficiency for data centers, high-performance computing, and edge AI applications. The project aims to double both application performance and power efficiency compared to competitors, a goal directly supported by the capabilities of Broadcom's 3.5D SoC.

“The launch of Broadcom’s 3.5D XDSiP technology marks a transformative milestone in advanced semiconductor integration,” stated Naoki Shinjo, SVP and Head of Advanced Technology Development Unit at Fujitsu. “This breakthrough is a key enabler for Fujitsu’s FUJITSU-MONAKA initiative to deliver cutting-edge, high-performance, and low-power processors. We highly value our strategic partnership with Broadcom and believe this technology will help power a more scalable and sustainable AI-driven society.”

By opting for a custom solution from Broadcom, Fujitsu can tailor its processor architecture precisely to its needs, integrating confidential computing for enhanced security and a 3D many-core design optimized for a diverse range of workloads, from AI model training to complex data analysis.

Taming the Power Demands of Gigawatt AI

The insatiable appetite of AI for computational power has created a significant challenge for data center operators and environmental sustainability. The press release explicitly mentions the need to support “gigawatt-scale AI clusters,” a term that underscores the colossal energy footprint of modern AI infrastructure. Broadcom’s technology directly addresses this critical issue.

The use of a 2nm manufacturing process alone provides a substantial efficiency boost, offering up to 30% lower power consumption compared to the previous 3nm generation for the same level of performance. However, the most significant gains come from the 3.5D packaging itself. By shortening the physical distance data must travel between compute and memory components and reducing interconnect power draw, the F2F architecture minimizes latency and energy waste. This is crucial for improving the overall efficiency of a data center, where energy costs and cooling are major operational expenses.

This focus on efficiency aligns perfectly with the goals of initiatives like FUJITSU-MONAKA, which aims to contribute to carbon-neutral green data centers. As AI models continue to grow in complexity, such advancements in hardware efficiency are not merely beneficial—they are essential for the sustainable scaling of artificial intelligence.

Reshaping the Competitive Landscape

With this announcement, Broadcom is carving out a powerful niche in a market overwhelmingly dominated by NVIDIA, whose GPUs hold an estimated 80% to 92% market share for AI training. While NVIDIA and AMD have also pioneered advanced packaging, Broadcom’s 3.5D F2F approach offers a distinct and compelling alternative for the burgeoning custom silicon market.

Industry analysts project the advanced packaging market to soar to over $80 billion by 2033, driven almost entirely by the demands of AI. It has become the new frontier for innovation, offering a path to greater performance as traditional transistor scaling becomes more difficult and expensive. Broadcom's platform allows hyperscalers and other large enterprises to break free from a single-vendor ecosystem and build custom AI accelerators optimized for their specific software and workloads.

Looking ahead, Broadcom is not stopping with Fujitsu. The company has stated that more than five 3.5D products are already in development for other customers, with broad production shipments slated to begin in the second half of 2026. Recent reports have also linked Broadcom to a major collaboration with OpenAI for custom AI accelerators, suggesting that this advanced packaging technology is poised to become a foundational element for some of the world's most advanced AI systems.