Phison Electronics (8299.TT), a global provider of NAND flash controllers and storage solutions, has announced a collaboration with Intel aimed at expanding the capabilities of AI PCs to run larger and more complex artificial intelligence workloads locally. The initiative combines Intel’s Core Ultra Series 3 processors with Phison’s Pascari aiDAPTIV technology, designed to address memory constraints that limit the performance of advanced on-device AI applications.
The collaboration comes at a time when AI PCs are evolving beyond basic assistant tools toward more sophisticated applications, including document analysis, multi-step automation, and agent-based workflows. These workloads increasingly require larger model sizes, sustained memory access, and the ability to maintain long-running sessions—demands that often exceed the limits of standard system memory configurations.
Phison’s aiDAPTIV technology is positioned as a memory extension solution that enables AI workloads to scale beyond conventional hardware limitations. It achieves this by distributing AI working memory across system DRAM and high-performance NAND flash storage through its Pascari aiDAPTIV Cache Memory architecture. This approach reduces reliance on high-capacity DRAM while maintaining performance for AI inference tasks.
In internal testing cited by the company, aiDAPTIV enabled a 26-billion parameter model to run on a system equipped with 16GB of DRAM, compared with a requirement of 32GB of DRAM without the extension layer under identical conditions. The company said the result highlights how memory optimization can expand access to larger AI models on mainstream client hardware.
The collaboration specifically targets Intel AI PC platforms powered by Intel Core Ultra processors and includes integration with the OpenVINO toolkit, Intel’s open-source AI software framework designed to optimize deep learning inference across hardware architectures. Phison and Intel are also working with independent software vendors (ISVs) to support technical evaluations, performance benchmarking, and workload optimization.
“AI PCs are evolving into platforms for more sophisticated local AI workloads, including agentic applications and larger MoE models that place increasing demands on memory capacity and responsiveness,” said KS Pua, CEO and founder of Phison Electronics. He added that the collaboration with Intel Corporation enables broader memory availability for AI workloads, allowing developers and device makers to deploy more capable applications locally while improving efficiency and data privacy.
The companies are showcasing the technology at COMPUTEX, where demonstrations include a local chat interface running a Mixture-of-Experts (MoE) model that would typically exceed system memory limits. The setup illustrates how aiDAPTIV extends effective working memory to enable execution of larger models directly on-device.
Another demonstration highlights a hybrid large language model (LLM) routing system built on OpenClaw, an open-source AI agent framework. The system dynamically routes tasks between local AI models and cloud services, aiming to reduce reliance on cloud compute while preserving scalability for complex queries.
Beyond Phison and Intel’s core demonstration, the ecosystem showcase includes contributions from software partners such as Ollama, LLMWare, TurinTech, Intel AI Superbuilder, and Intel AI Playground. Hardware collaborations with PC manufacturers including ASUS, MSI, and Acer are also part of the initiative, signaling a broader push to integrate AI-optimized memory architectures into consumer and enterprise devices.
Industry software developers highlighted the potential impact of extending usable memory through storage-based architectures. Ollama, which specializes in local AI model deployment, noted that memory constraints are often the primary barrier to running larger models on client devices. LLMWare emphasized that enterprise GenAI workflows, including retrieval-augmented generation (RAG) and domain-specific agents, increasingly depend on efficient local inference capabilities.
TurinTech AI also pointed to the importance of balancing performance, privacy, and cost in on-device AI systems. The company said integrating AI-driven optimization tools with Intel AI PCs and Phison’s memory extension approach could support more complex workloads without requiring constant increases in system DRAM capacity.
From Intel’s perspective, the collaboration reflects a broader industry shift toward decentralized AI computing. “More users and businesses want to run AI locally – faster, more private and without the cost of sending everything to the cloud,” said Jim Johnson, Senior Vice President and General Manager of Client Computing at Intel Corporation. He added that enabling larger AI workloads on simplified memory configurations could help organizations extract greater value from their own data while reducing total cost of ownership.
As AI applications continue to expand from cloud-centric models toward hybrid and edge-based deployments, memory efficiency is emerging as a critical constraint in hardware design. Phison and Intel’s collaboration reflects an industry effort to address this bottleneck by rethinking how system memory is allocated and extended for AI workloads.
The companies have not disclosed commercial timelines for the integration beyond ongoing demonstrations and developer evaluations.
