On May 22, CryptoCity reported that AMD (Advanced Micro Devices) CEO Lisa Su announced that AMD will invest more than $10 billion in Taiwan’s industrial supply chain, with a focus on expanding AI chip manufacturing and advanced packaging capacity, and will also release a new-generation rack-level AI server platform, Helios, with plans to launch in the second half of 2026.
Confirmed Technical Specifications of the Helios Platform
Helios is AMD’s new-generation complete platform designed for rack-scale AI infrastructure. It uses a vertically integrated architecture and includes the following confirmed components:
MI450X GPU (AMD Instinct MI400 series): each card is equipped with at least 288GB of HBM4 memory, with a maximum configuration of 432GB; FP4 compute performance of about 50 PetaFLOPS; supports high-performance Ethernet and 800GbE NIC connections; mass production is expected in the second half of 2026. The 6th-generation EPYC “Venice” processor: manufactured using TSMC’s 2-nanometer process. ROCm open AI software stack: provides openness at the ecosystem level, directly contrasting with NVIDIA’s closed CUDA ecosystem.
AMD also announced a new Ryzen AI Max+ Pro 495 processor, using the Zen 5 architecture, 16 cores / 32 threads, up to a 5.2GHz clock speed, and supports up to 160GB of display memory; AMD claims this is the first x86 processor capable of running an AI model with 300 billion parameters on a single machine.
Taiwan Advanced Packaging Supply Chain: Confirmed Partner Companies and Technical Roadmap
AMD’s core focus on its $10 billion investment is advanced packaging capacity. The following are the confirmed collaboration directions:
ASE (Amkor Technology), Siliconware: cooperating to develop the next-generation 2.5D EFB cross-chain interconnect technology, improving data transmission bandwidth and power efficiency between AI chips
Powertec: collaborating to validate panel-level 2.5D EFB technology, aiming to reduce the packaging cost of large AI systems and increase packaging scale
Unimicron (欣興), Nanya PCB (Nanya PCB), Kyocera: confirmed as the main substrate/packaging substrate suppliers
TSMC: AMD has increased some server processor production capacity via TSMC’s Arizona wafer plant, and will continue to expand in Taiwan; the 6th-generation EPYC “Venice” uses TSMC’s 2-nanometer process
Lisa Su said that memory supply is still “very tight.” AMD is working closely with all supply-chain partners to ensure CPU, GPU, and memory production capacity expand in sync.
Common Questions
How does ROCm’s open software stack on the AMD Helios platform form a competitive strategy against NVIDIA CUDA?
NVIDIA’s CUDA is a closed, proprietary GPU programming platform. After AI workloads are developed on CUDA, they are typically difficult to port to other hardware. AMD’s ROCm is an open-source software stack that allows developers to build AI workflows without being locked to a specific platform, theoretically reducing the migration cost for enterprises switching from NVIDIA to AMD infrastructure. By positioning ROCm as a core software component, the Helios platform is part of AMD’s long-term strategy to break through CUDA’s moat via ecosystem openness, but the practical friction costs of migration and the existing accumulation of CUDA’s ecosystem are still the main obstacles AMD needs to overcome.
What role does 2.5D EFB packaging technology play in the performance competition of AI chips?
2.5D packaging is an advanced packaging technology that allows GPUs, HBM memory, and other chips to be arranged on the same silicon interposer, achieving data transmission bandwidth and power efficiency far beyond traditional packaging through high-density, short-distance interconnects. EFB (Embedded Face-on-Board) is the cross-chain interconnect technology AMD and its packaging partners are developing, aimed at further improving packaging density and thermal management capability. For large AI model training and inference workloads that require close cooperation between GPUs and high-bandwidth memory (HBM4), advanced packaging efficiency directly determines the upper limit of overall system performance—this is why advanced packaging capacity has become a core bottleneck in AI chip competition.
How does AMD’s $10 billion investment in Taiwan compare in scale to similar Taiwan investment plans by Intel and NVIDIA?
Based on confirmed publicly available information: AMD’s $10 billion commitment is among the largest single-supply-chain-investment scales by global AI chipmakers in Taiwan in recent years. NVIDIA mainly expands capacity indirectly in Taiwan through foundry partners such as TSMC, and there is no record of similar-scale public announcements for direct investments in Taiwan. Intel has customer relationships with TSMC, but its main capital expenditures are concentrated on building its own fabs (in the United States, Europe, and Israel). The special aspect of AMD’s investment is that it covers the entire vertical supply chain at the same time—chip manufacturing (TSMC), advanced packaging (ASE, Siliconware, Powertec), substrates (Unimicron, Nanya PCB, Kyocera), and system integration (Wiynn, Wistron, and Inventec)—making it an ecosystem investment with vertical integration rather than merely expanding by increasing foundry orders.
