Phison Electronics has developed IPs for leading storage solutions across SSD, USB, eMMC, and more. Why are storage technologies so critical in today’s technology landscape?
Just like humans need carbon to live, every electronic system needs storage.
Whether it is IoT, automotive, mobile devices, PCs, servers, aerospace, or defence—none of it works without storage. Think of NAND flash as raw rice—it does not function alone. You need a controller to cook the rice and make it usable. That is what Phison Electronics does: we combine NAND with our controllers to create complete systems, supporting USB, SD card, eMMC, UFS, SSD—every major interface. We cover more interfaces than any company in the world—even for lunar and Mars missions.
IoT is defined by computing, connectivity, and storage. At Phison Electronics, we provide the storage—the rice base—then build different dishes for different needs. It is all about tailoring one essential ingredient into a variety of practical uses.
What is the flagship solution in Phison Electronics’ portfolio today?
In traditional systems like PCs, there are standardised interfaces—PCIe Gen 3 or Gen 4—that simplify development. But AI is different. It relies on expensive HBM (high-bandwidth memory), which creates bottlenecks in GPU compute power and high costs. We innovated by combining NAND flash with our own silicon, IP, UI software, and middleware software to complement HBM. Our solution can pre-store massive amounts of cache data—avoiding redundant computation. That is the benefit of our large, affordable AI SSD - aiDAPTIVCache.
For example, fine-tuning a model like LLaMA 370B usually costs $2 million. With our aiDAPTIV+ system, it can be done for $100,000. That is innovation—creating a large, affordable cache that supports edge AI fine-tuning (post-training) tasks.
What has the industry response been to the launch of aiDAPTIV+?
At first, nobody believed us. Even after proof of concept, they said the market was too small, or it needed too many software developers. But that changed this February. Once we proved you could do on-premise fine-tuning affordably, interest exploded. Now we are overwhelmed—with not enough developers to meet demand. If NVIDIA or AMD endorse us, the floodgates will open. We have already proven aiDAPTIV+ can not only fine-tune AI model on-premises but run inference more efficiently, saving power and improving ESG metrics.
Are there use cases beyond language models?
AI used to be focused on images—like facial or license plate recognition. That is mature now. The real challenge is language, which requires vast memory. A single image has limited parameters. But a page of text is huge. That is where costs balloon.
We focus on LLMs, VLMs, LVMs—anything language-related. aiDAPTIV+ is not tied to any specific model. AI models just need memory. Without enough memory, the system crashes. Our solution lets a single GPU run any model—maybe slower, but it runs. And this is not limited to the cloud. We have already installed systems in small and medium businesses. With just $4,000, they can run AI locally in their office.
How is Edge AI-readiness evolving in public and private organisations?
AI has exploded faster than anything in history—even smartphones. But 99.999% of people only know how to use it, not how it works. Education and implementation are lacking. Ask yourself: Do you believe AI is the future? Yes? Then know this—it is not just cloud. It is going to the edge. Why? Data privacy. But no one is ready. The education system has not caught up.
In the next decade, more than 50% of startups will work in edge AI—just like the app boom after smartphones – but there is still a huge gap in software development. Edge AI is about to become ubiquitous. Fifty years ago, people relied on mainframes. Now everyone has a personal PC. The same will happen with AI. Until recently, you had to use cloud-based tools like ChatGPT. But very soon, every company can run AI locally—on-premise—with our hardware, protecting their data.
Where is edge computing expertise taking off? What regions are most active?
Taiwan and Malaysia are leading. We have teams in both. India is booming—huge population, eager to learn, low costs, and many engineers. The U.S. is slow. Europe is nearly inactive.
A European VC came to debate us before Lunar New Year. He claimed Edge AI is not needed—just use the cloud. I asked him: If you are trying to sell slippers in a new market to no avail, is the problem that people do not wear slippers, or that there is no way to buy them? Europe has not invested in foundational tech for decades.
There is data showing developer activity in Europe is picking up. Does that change your view?
Yes, but they are still focused only on the cloud. Developers there have been conditioned to build on cloud infrastructure. In Asia, we have built for the edge. When we show Europeans our $4,000 solution, they say it is garbage. They are brainwashed by GPU companies who only focus on high cost AI-model-training machines.
The cloud model will be replaced by a hybrid edge-cloud model. Cloud has its place—just like notebooks were not replaced by tablets. Cloud suits students and individuals who do not mind sharing their private data. But many companies cannot use ChatGPT or Copilot due to IP risks. They need on-premise solutions, which require IT infrastructure.
What about all the global investment in cloud infrastructure last year? Was it too much?
Not in infrastructure—but cloud companies poured too much into a certain GPU company. The GPU company’s gross margin is 92%. If the hardware cost is $1, they sell it for $10. That profit does not go back into the cloud—it goes to the GPU company.
Now cloud companies need ROI, but AI as a service is costly. And GPU-based inference is no longer viable. They are shifting to NPUs from companies like Broadcom or Marvell—just 3% of the cost. Microsoft, Amazon, Alibaba—they are all pulling back GPU investments.
And companies like NVIDIA are pivoting to the edge?
Exactly. GPU companies are developing edge devices like Spark, a $3,000 AI PC. But it only supports inference. With just 128GB DRAM, it cannot do fine-tuning. It is more a toy than a tool because AI models cannot be post-trained (fine-tuning) by users for specific edge tasks such as medical, finance, or law.
Our AI-100 solution, by contrast, costs only a few thousand and massively boosts performance effectively. Think of the world’s leading GPU company as selling a Mercedes. We offer a Toyota that flies.
How do you see global AI readiness?
At the Beyond 2025 conference in Greece, I spoke with four ministers. They all agreed: only 3% of AI talent globally is truly great—and most of it is in the U.S., where the tax also goes. Europe has exported its talent because it lacks core tech.
Greece invited me to bring our tech and train engineers there.
We are building aiDAPTIV+ in Europe now. It is not just about cloud or edge—it is about global empowerment. Every country is saying “AI, AI, AI,” but most have no clue what it really involves. We are here to help them understand and build.
