How a traditional SSD manufacturer is breaking down the cost barriers that keep AI locked away from most businesses

You know what's fascinating about Silicon Valley? It went from being called the "Valley of Heart's Delight" - all fruit orchards and wheat fields - to becoming the epicenter of technological innovation. And now, we're witnessing another transformation right before our eyes: companies that made their names in traditional hardware are pivoting to solve the most pressing challenges in AI infrastructure.

That's exactly what's happening at Phison Electronics, a company better known for SSD controllers and storage solutions. But during our recent conversation during the 62nd edition of the IT Press Tour with members of the Phison leadership team - Michael Wu (GM & President), Brian Cox (Product Marketing Director), and Heather Davis (Product Marketing Director) - it became clear that Phison isn't just adapting to the AI era. They're actively reshaping how organizations can afford to deploy AI on their own terms.

The Problem That Started It All

Sometimes the best innovations come from solving your own problems. The company shared a telling story about how Phison's CEO faced a classic AI adoption dilemma:

"We want to apply AI to our design and manufacturing supply chain management processes... Can you tell me what the price tag will be? And the answer came back... 'Oh, $2 million.' And he replied, 'I don't have that. That was not in the plan. Go back and figure out a more efficient way.'"

Sound familiar? This scenario plays out in boardrooms across the world every day. Companies see the potential of AI but balk at the astronomical costs of traditional GPU-heavy deployments. The irony? Most organizations don't actually need all those GPUs - they just need the memory that comes with them.

Enter aiDAPTIV+: Making AI Affordable

Phison's answer to this problem is their aiDAPTIV+ platform, a technology that fundamentally changes the economics of AI deployment. Think of it this way, as Michael Wu so aptly put it: if you're hosting a party for 50 people but your house only has seating for 10, you don't go out and buy five houses. You just need creative seating solutions.

The aiDAPTIV+ platform achieves similar results by using flash storage as an extension of GPU memory. Instead of loading entire AI models onto expensive high-bandwidth memory (HBM or GDDR), the system intelligently manages model data between flash storage and GPU memory, feeding the processor exactly what it needs when it needs it.

The Technical Magic Behind the Scenes

The real innovation lies in Phison's middleware - what they call aiDAPTIVLink. This software layer coordinates the constant swapping of data between different types of memory, ensuring the GPU stays fed while dramatically reducing costs.

Here's what makes this approach work:

• Cost Reduction: Organizations can achieve 8-10x cost savings compared to traditional all-GPU deployments
• Flexibility: The system works across various form factors, from IoT devices to data center servers
• Performance: While there's a speed trade-off, most organizations can easily work with slightly longer processing times in exchange for massive cost savings

Brian Cox put it simply: "If I can cut the price by 8 to 10 times and do multiple trainings per day, maybe not every hour, does that work for you? From a budget standpoint, yes. And from a service activity level, yes."

The Pascari Series: Built for the AI Era

Phison also recently expanded their enterprise SSD lineup with the Pascari X200Z, specifically designed for AI workloads. This isn't just another fast drive - it's engineered for the unique demands of AI training and inference.

What Makes Pascari Special

The X200Z delivers some impressive specifications that matter for AI applications:

• Endurance: Up to 100 drive writes per day (compared to the typical 1-2 DWPD for standard drives)
• Performance: PCIe Gen5 interface with near-Storage Class Memory (SCM) latency
• Optimization: Specifically tuned for the sequential write patterns common in AI training

The key insight here is that AI workloads behave differently from traditional enterprise applications. Where typical database operations might involve lots of random reads and writes, AI training tends to involve large sequential data movements. Phison redesigned their drives from the ground up to excel at these patterns.

Beyond Earth: The Lonestar Data Holdings Partnership

Here's where the story gets really interesting. Phison recently partnered with Lonestar Data Holdings on what might be the most ambitious data storage project ever attempted: building data centers on the Moon.

While this might sound like science fiction, it addresses a very real concern about data preservation. Throughout history, we've lost countless libraries and archives to natural disasters, wars, and accidents. The Library of Alexandria, business records lost in 9/11, entire companies that disappeared because they couldn't recover their data - the pattern repeats.

Lonestar's lunar data centers represent the ultimate backup strategy, and Phison's involvement speaks to the reliability and durability of their storage solutions. If your SSD can survive and operate in the harsh environment of space, on the moon, it can probably handle whatever your data center throws at it back here on Earth.

Innovation Across the Spectrum: From Chips to Complete Solutions

Phison's recent announcements at COMPUTEX 2025 show they're not content to just solve the cost problem. They're innovating across the entire storage stack:

The E28 Controller: AI Meets Storage

The E28 represents a industry first - an SSD controller with built-in AI processing capabilities. Built on TSMC's 6nm process, it delivers:

• Performance: Up to 2,600K/3,000K IOPS (random read/write) - over 10% higher than competing products
• Efficiency: 15% lower power consumption versus comparable 6nm-based controllers
• Intelligence: Integrated AI processing for enhanced SSD optimization

Expanding the Ecosystem

The company is also introducing solutions like the E31T DRAM-less PCIe Gen5 controller for mobile platforms and developing next-generation PCIe signal ICs, including the world's first PCIe 6.0 Redriver.

Democratizing AI: From Data Centers to Coffee Shops

Perhaps the most compelling aspect of Phison's approach is how it makes AI accessible to organizations that could never afford traditional deployments. Cox mentioned seeing customers at NVIDIA's GTC conference who had bought multiple GPU cards - not because they needed multiple processors, but because they needed the collective memory capacity.

Real-World Impact

Phison is already using their own technology internally with measurable results:

• Code Documentation: Spent 60,000 on adaptive technology instead of 2 million for traditional solutions, saving work equivalent to 30 engineers
• Knowledge Management: Invested 150,000 versus a potential 6 million, achieving productivity gains equal to 50 engineers
• Software Development: 400,000 investment versus 16 million for traditional approaches, replacing work that would have required 200 additional hires

Making Learning Accessible

The affordability factor extends beyond deployment to education and skill development. Cox noted an interesting trend in their training programs: Half of all the students are university professors. Which is pretty wild.

When AI training hardware becomes affordable enough for individuals and educational institutions to own, it solves the access problem that limits skill development. Students and researchers no longer need to compete for limited time slots on shared university servers - they can have their own AI development environment.

The Performance Reality: Trading Speed for Accessibility

Let's be honest about the trade-offs. Phison's approach doesn't make AI training faster - it makes it possible. In their lab demonstrations, they showed how the aiDAPTIV+ system handles complex, multi-turn conversations much better than traditional setups, but with some speed compromises.

The key insight is that most organizations don't need to retrain models every hour. Cox explained: "Am I having to train models with new data every hour? Probably not. I'm going to train a model maybe once a day, maybe once a week, maybe once a month."

This perspective shift - from "fastest possible" to "fast enough and affordable" - opens AI deployment to thousands of organizations that were previously priced out of the market.

Looking Ahead: The Post-Training Era

NVIDIA's Jensen Huang has spoken about the transition from "pre-training" (building foundational models like ChatGPT) to "post-training" (customizing AI with private, organization-specific data). This shift plays directly into Phison's strengths.

When organizations need to incorporate their unique data into AI models while keeping that information private and compliant with data sovereignty requirements, on-premises solutions become not just attractive, but necessary.

The Bottom Line

Phison's approach represents something important happening in the AI infrastructure space: the democratization of advanced technology. By focusing on the cost barriers rather than just performance metrics, they're enabling a much broader range of organizations to benefit from AI.

Their CEO K.S. Pua captured this philosophy perfectly: "AI adoption must not remain exclusive to tech giants. Real competitiveness lies in empowering SMBs, educational institutions, and public sectors with affordable, fast, and secure AI solutions."

As AI moves from the exclusive domain of big tech into mainstream business operations, companies like Phison are proving that innovation isn't always about building the fastest or most powerful solution. Sometimes, it's about building the solution that the most people can actually use.

The question for IT leaders isn't whether AI will become part of their infrastructure strategy - it's whether they'll wait for costs to come down naturally, or take advantage of solutions like aiDAPTIV+ that are making AI accessible today.

##