As SC26 approaches, one message is likely to cut through the noise on the show floor: throwing more GPUs at an AI problem doesn’t automatically solve it. In this pre-show Q&A with VMblog, Michael Wu, GM and President of Phison USA, argues that memory capacity, data movement and storage performance are becoming just as critical as raw compute, particularly as longer context windows, agentic workflows and larger datasets strain budgets already under scrutiny. For a company with more than 25 years in the NAND flash industry, spanning controllers, firmware, embedded storage and enterprise SSDs, the conversation has shifted from faster, bigger storage to how memory, storage and compute work together.
Wu walks us through how Phison is applying that experience through its Pascari enterprise storage portfolio and Pascari aiDAPTIV Cache Memory technology, which adds a flash-based memory tier beyond GPU memory and system DRAM for suitable workloads. He explains what that means for KV cache reuse, agentic AI, and getting more value from hardware organizations already own, while candidly noting that flash is not a DRAM replacement. He also discusses Phison’s collaboration with Intel announced at Computex 2026, and where he sees the industry heading by SC27. Attendees can continue the conversation with Phison at booth #5413.
++
VMblog: If you were giving an SC26 attendee a quick overview of your company, what would you say? How has your company evolved since SC25, and what has been the biggest change in your business over the past 12 months?
Michael Wu: Phison has been in the NAND flash industry for more than 25 years. We started with controller technology, but over time, we’ve expanded our expertise into embedded storage, enterprise SSDs and now systems-level infrastructure.
What has changed most is the complexity of the problems our customers are asking us to solve. It’s no longer about delivering faster storage or greater capacity. With AI, customers are looking at how memory, storage and compute work together, and where infrastructure limitations can affect performance and cost.
That’s where our experience in controllers, firmware and NAND becomes valuable. Through our Pascari enterprise storage portfolio and Pascari aiDAPTIV Cache Memory technology, we’re applying that expertise to real-world infrastructure challenges.
We’re looking forward to discussing these developments with attendees at booth #5413.
VMblog: What key message will SC26 attendees hear from you this year? With AI infrastructure budgets under heavy scrutiny and hardware costs rising, what concrete value proposition can they take back to their CFO and CIO?
Wu: One of the biggest misconceptions about AI infrastructure is that adding more GPUs automatically solves the problem. As AI workloads become more complex, memory capacity, data movement and storage performance can become just as important as compute.
For enterprises, the question is how to get more value from the infrastructure they already have. GPUs and high-performance memory are expensive, and not every part of an AI workload requires the same level of performance.
With Pascari aiDAPTIV, we’re addressing that challenge by introducing an additional memory tier for suitable workloads, helping organizations expand available memory capacity without necessarily moving to larger GPU or DRAM configurations.
The business value is about using the right resources for the right workloads, rather than continuously adding more expensive hardware. That’s an important consideration when CIOs and CFOs are evaluating the long-term economics of AI.
VMblog: This year’s SC theme is “HPC Unites.” Where does your company fit into that idea? Which partnerships, joint reference architectures or open standards are you announcing or demonstrating that make the whole greater than the sum of its parts?
Wu: AI and HPC infrastructure are becoming increasingly interconnected. No single component or company can address every challenge, which makes collaboration across the ecosystem essential.
For Phison, that means working across hardware, software and application environments to understand how storage and memory technologies can support the broader system.
One example is our collaboration with Intel (announced at Computex 2026), to help enable larger local AI workloads on Intel AI PC platforms. By integrating aiDAPTIV with AI runtimes and applications, we’re exploring how memory tiering can support larger models and longer AI sessions.
The important point is that technology becomes more useful when it works within the systems customers already depend on. That’s how collaboration translates into practical value.
VMblog: Nearly every vendor now claims to be “AI-native” and “HPC-ready.” What are your provable differentiators, and which independent benchmarks or production customer results back them up?
Wu: There are a lot of new terms being used across the AI industry, but what matters to customers is whether a technology can solve a specific problem in their environment.
Phison’s differentiation starts with more than 25 years of NAND engineering experience. We design controllers, develop firmware and understand how flash behaves under different workloads. That gives us a strong foundation for addressing challenges beyond traditional storage.
With Pascari, we’re applying that expertise across the AI data lifecycle, from performance-intensive training and checkpointing to environments where large datasets need to be retained efficiently.
With aiDAPTIV, we’re taking a different approach to memory capacity by using high-endurance flash alongside middleware software to enable larger AI workloads.
We’ve demonstrated this technology on Intel AI PC platforms, including support for larger models and longer-running workloads
Ultimately, we believe differentiation needs to be demonstrated through workload behavior and real-world performance, not simply product labels.
VMblog: The memory and accelerator supply crunch has pushed up prices and stretched lead times for DRAM, HBM and NAND. How are you helping customers plan and get more from constrained hardware?
Wu: Memory is becoming one of the most important constraints in AI infrastructure. Larger models, longer context and more concurrent users all increase demand for GPU memory and system DRAM.
But not all data needs to always remain in the fastest and most expensive memory. That’s an important distinction when organizations are trying to manage limited resources.
Pascari aiDAPTIV introduces a flash-based memory tier that complements GPU memory and DRAM. Frequently accessed data remains in faster memory, while suitable less-active or reusable AI data can be retained in aiDAPTIV Cache Memory and retrieved when needed.
Of course, flash does not deliver the same performance as DRAM, and we’re not positioning it as a direct replacement.
The goal is to better match memory resources to workload requirements so customers can make more efficient use of the hardware they have.
VMblog: Agentic AI is now on the SC program, including a dedicated workshop. How are AI agents changing your product?
Wu: Agentic AI introduces a different set of infrastructure demands compared with traditional chatbot applications.
When an AI agent is working through multiple steps, accessing tools and retaining information across a longer workflow, the amount of data that needs to remain available can grow significantly.
Agents also tend to use larger, more capable models for multistep reasoning and tool calling. The challenge isn’t simply running a larger model. It’s managing the information that accumulates as the model continues working.
That’s where memory architecture becomes important. With Pascari aiDAPTIV, suitable data such as saved KV cache can be retained in flash-based aiDAPTIV Cache Memory while active data remains in GPU or system memory.
For example, when many requests share the same long prompt instructions or documents, saved KV cache can be reused instead of recomputed. This can help support longer-running AI sessions, although the benefits depend on the workload and how frequently information needs to be accessed.
As agentic applications become more sophisticated, understanding those memory access patterns will be increasingly important.
VMblog: With memory scarce and expensive, and modern AI leaning heavily on KV cache, checkpointing and massive datasets, what innovations in memory tiering, storage orchestration or interconnect are you unveiling to keep costly accelerators fed?
Wu: AI is changing the way we need to think about storage. It’s not just about how much data you can store or how fast you can access it. It’s about what happens to that data throughout the workload.
During AI training and inference, data is constantly being prepared, moved, accessed, checkpointed and retained. Each stage has different requirements, and trying to address all of them with one type of storage can create unnecessary cost or performance trade-offs.
Through our Pascari portfolio, we’re addressing these different requirements. The X201 is designed for sustained, performance-intensive workloads, while the D201P focuses on environments where capacity is a priority.
Pascari aiDAPTIV adds a flash-based memory tier beyond GPU memory and system DRAM. For suitable workloads, it can retain KV cache so the GPU can reuse it instead of recomputing it.
What we’re seeing is that infrastructure efficiency increasingly depends on how these technologies work together, rather than optimizing each component in isolation.
VMblog: Talent shortages and lock-in fears are real. How are you making HPC and AI infrastructure easier to operate and more portable so lean teams can do more with less?
Wu: One of the challenges enterprises face with AI is that building a working demonstration is very different from operating an application reliably over time.
You need to think about memory capacity, storage performance, software integration, data security and ongoing management. For organizations with limited technical resources, that complexity can become a barrier to adoption.
Phison’s experience across controllers, firmware and enterprise storage gives us a foundation for approaching these challenges at the systems level.
By bringing together Pascari enterprise storage, aiDAPTIV AI memory technology and software integration, we’re working to make infrastructure components function more cohesively.
The objective is to reduce the amount of integration work customers need to manage themselves, while recognizing that every environment has different operational requirements.
VMblog: Looking toward SC27 and beyond, which bets are you making? What will attendees be talking about a year from now that they aren’t talking about today?
Wu: I think one of the conversations we’ll continue to see is how organizations improve AI infrastructure efficiency without relying entirely on additional compute.
Today, there’s a lot of attention on GPU performance, and understandably so. But as AI applications become more complex, memory capacity, data movement and the relationship between compute and storage are becoming equally important considerations.
Longer context, agentic workflows and larger datasets create challenges that cannot always be solved by scaling a single component.
From Phison’s perspective, our controller, firmware and NAND expertise gives us a way to contribute to those discussions across enterprise storage, memory technologies and emerging infrastructure environments.
The bigger question for the industry is how to build systems that can support more demanding AI workloads while remaining practical to deploy and operate.
