Exclusive | Taiwan's Memory Giant Phison Says AI's Real Barrier Is DRAM, Not GPUs

Walk into any conversation about AI infrastructure and you will hear the same refrain: the bottleneck is GPU computing power. K.S. Pua (潘健成), chairman of Phison Electronics (群聯電子), thinks that consensus is wrong — and he has spent decades in the memory industry building the evidence to prove it.

"Everyone assumes the bottleneck in AI is the GPU. It isn't," Pua told Storm Media in an exclusive interview. "The real problem is DRAM."

The distinction matters enormously. GPU constraints are largely a supply chain problem. DRAM constraints, Pua argues, are structural — a fundamental mismatch between what AI workloads demand and what current server architectures can economically deliver.

Dynamic random-access memory is the working memory that AI models draw on during inference. It is also extraordinarily expensive, and the capacity any single server can house is sharply capped. For small and mid-sized enterprises hoping to deploy large language models, that combination has effectively priced them out.

Many AI applications, Pua said, have stalled not because computing is unavailable but because the hardware cost of memory remains prohibitive. "Sufficient compute, insufficient memory" is how he frames the problem — and it is the gap Phison has moved to close.

The company's answer is the aiDAPTIV hybrid AI SSD solution and the aiDAPTIV Hybrid Claw, both of which deliberately break from conventional server architecture. Through tightly co-designed hardware and software, computational data that would normally need to reside in expensive DRAM is offloaded to solid-state drives. The result: the system operates as though it has nearly unlimited memory, at a fraction of the cost.

The commercial impact has been measurable. According to Pua, the approach has cut token costs — the per-unit cost of AI inference — by roughly 70%.

"When servers become cheaper, AI applications naturally become easier to scale," he said. "Once costs fall and the market expands, the volume of data generated by inference will surge. AI revenue ultimately grows in proportion to storage capacity."

In traditional semiconductor markets, NAND Flash was understood as a cyclical business — prices boom, inventories swell, prices collapse, and the cycle repeats. Pua believes that model no longer describes what is happening.

"In the past it was cyclical. Now it's structural demand," he said. "I can no longer see where supply and demand would equilibrate."

To illustrate the scale of the shift, he offered an analogy: imagine a stable community of one million people whose collective appetite suddenly quintuples overnight, while the food supply — wafer fabrication capacity — remains fixed. The imbalance that follows is not a blip. It is the new normal.

He identified three mechanisms driving AI's insatiable appetite for storage.

The first is explosive data generation. Tasks that once took human workers hours to complete — producing video, audio, written text — can now be generated by AI in minutes. Once data is created, it must be stored somewhere. There is no compression factor that changes that arithmetic.

The second is the direct link between cloud service provider revenue and memory purchases. For CSPs, AI inference is not a cost center — it is the revenue engine. "You need memory to store data in order to make money," Pua said. That logic means cloud providers will continue buying storage even as prices rise, because the alternative is leaving inference capacity idle.

The third mechanism is what Pua describes as a widespread misreading of spot market volatility. When short-term traders liquidate inventory and prices dip, casual observers may interpret the movement as a sign of impending oversupply. Sophisticated players, he said, use those moments as buying opportunities.

"The critical question for the storage market going forward is not whether shortages will occur, but how they will occur," Pua said. AI's demand for memory is rigid, cumulative, and permanent — it does not reverse.

The AI wave has not just changed the market — it has transformed Phison's own business, faster than any external observer might expect.

"At least seven or eight of our top ten customers this year will be different from last year," Pua said. Cloud service providers and AI-native clients have displaced the company's traditional buyers at speed, and now constitute its primary revenue base.

Phison's share of high-value design-in business — deep integration into customer systems rather than retail component sales — has exceeded 70%, giving the company meaningful insulation from the price swings that dominate commodity storage markets.

Despite leading a company whose market capitalization has surpassed NT$100 billion, Pua said he continues to work the front lines personally. He described Phison as having become "a kind of tutoring center," with his teams visiting clients directly to walk them through new AI computing architectures hands-on.

On the broader question of why so many companies announce AI initiatives that produce little, Pua was characteristically direct.

The problem, he said, is delegation. Executives declare AI a priority, hand the project to subordinates, and hope the IT department figures it out. It rarely works.

"If a company wants to adopt AI, the boss absolutely cannot simply hand it off to subordinates and expect the IT department to carry it through alone," he said. "The boss has to personally lead the charge. That's the only way the organization builds genuine momentum."

He drew a pointed parallel to the enterprise-wide introduction of ERP systems a generation ago. At first, changing existing workflows always generates resistance — employees find reasons to push back, just as they did when ERP was introduced. Without visible, committed leadership from the top, most efforts quietly collapse.

"If the boss doesn't step forward with conviction, the effort will almost certainly fail," Pua said. "AI adoption is exactly the same. Real transformation only happens when leadership is personally and visibly committed."

Born in Malaysia and having built Phison into a global leader from Taiwan, Pua has made a deliberate move to invest in the country where he grew up.

Phison has established a wholly owned subsidiary in Kuala Lumpur — MaiStorage (群聯大馬) — with Pua personally leading the initiative. The venture is designed to localize Phison's IC design capabilities and advanced AI technologies in Malaysia while contributing to the development of a domestic semiconductor ecosystem.

The strategy operates on three levels. MaiStorage focuses on local IC design and storage solutions, bringing Phison's most advanced technologies — including the aiDAPTIV+ platform — to Malaysian data center, AI enterprise, and automotive clients. The subsidiary has established operations at Malaysia's Semiconductor Accelerator and IC Design Park in Puchong, Selangor, where dozens of senior engineers and Taiwan-based personnel are conducting substantive technology transfer and cultivating high-end semiconductor R&D talent locally.

Pua has also stated his intention to take MaiStorage public independently on a Malaysian exchange — an ambition he frames as something more than routine business expansion.

"We want to replicate in Malaysia the kind of technology-first company that Phison represents," he said. For Pua, MaiStorage is both a homecoming and a foundation for what he describes as a Taiwan–Malaysia partnership built on mutual benefit.

Pua's closing observation is perhaps the most useful for any company navigating the AI era, regardless of whether it operates in memory, cloud services, or any adjacent field.

"Keep asking yourself: what have I not thought of yet?" That, he said, is Phison's guiding principle for sustaining competitive advantage in a rapidly shifting, highly uncertain environment.

At a moment when AI is redrawing the boundaries of the semiconductor industry, Phison is positioning itself not merely as a component supplier but as an indispensable contributor to AI computing and storage architecture. For Pua, the competitive battles ahead will be won not by scale alone, but by being — as he put it — truly needed.

Original Article in Chinese

（Related： Exclusive | Taiwan Could Lose Organ Transplant Independence to China Within a Decade, Top Scientist Warns ｜ Latest ）