Dubai, UAE – October 22, 2025 – As artificial intelligence reshapes industries from healthcare to finance, a looming crisis in data storage could grind the revolution to a halt. Datuk Pua Khein-Seng, the pioneering inventor of the single-chip USB flash drive and CEO of Phison Electronics Corporation, issued a stark warning on the sidelines of GITEX Global here: AI machines are poised to generate 1,000 times more data than humans, unleashing an unprecedented demand that will trigger severe shortages in NAND flash memory supplies over the next decade.
Speaking to Anadolu Agency amid the bustling halls of the world's largest tech and AI event, Khein-Seng, a Malaysian-born innovator based in Taiwan, reflected on his groundbreaking 2001 invention that revolutionized portable storage. "That little pen drive changed how we carry data," he said, holding up a modern equivalent. "But today, AI isn't just carrying data—it's creating oceans of it." Phison, which he founded in 2000, has since pivoted to NAND flash controllers and systems, powering everything from smartphones to data center servers. The company now supplies memory solutions to giants like Apple and NVIDIA, underscoring Taiwan's pivotal role in the global semiconductor ecosystem.
Khein-Seng's prognosis aligns with mounting industry analyses. AI workloads, particularly in training large language models, devour petabytes of data daily. Industry reports forecast a "memory supercycle," with NAND flash demand surging 50% annually through 2030 due to AI data centers. Phison's own projections predict shortages as early as 2026, potentially lasting until 2035. "Investments in NAND production stalled after years of price crashes," Khein-Seng explained. "Now, with AI reallocating resources to high-bandwidth memory for GPUs, we're facing a perfect storm."
The numbers paint a dire picture. Global data creation hit 181 zettabytes in 2024, with AI contributing just 10%—but that slice is exploding. By 2028, AI-generated data could swell to 97 zettabytes, a 1,000-fold increase over human output, driven by generative models producing synthetic datasets for training. Human-generated content, like social media posts or documents, pales in comparison; AI systems iterate endlessly, simulating scenarios in seconds that would take humans lifetimes. This deluge isn't abstract: the latest GPUs, essential for AI inference, require terabytes of flash storage per cluster to handle the influx, turning data centers into voracious consumers.
Yet, Khein-Seng argues the real villain isn't compute power but memory bottlenecks. "People fixate on GPUs because they're flashy and expensive," he said. "But the true choke point is DRAM and NAND. Limited dynamic random-access memory starves AI during peak loads, idling billion-dollar hardware." In AI training, where models process billions of parameters, insufficient memory forces repeated data fetches from slower storage, slashing efficiency by up to 40%. Phison's aiDAPTIV+ platform, unveiled at GITEX, addresses this with adaptive layering—stacking NAND tiers dynamically to expand capacity without redesigning systems. "Flash memory is the escape hatch," Khein-Seng asserted. "It scales cheaply, unlike HBM, which costs 10 times more per gigabyte."
This innovation couldn't come soon enough for cloud-dependent nations. Khein-Seng revealed that 99% of global AI access flows through cloud systems, dominated by U.S. titans like Amazon Web Services, Microsoft Azure, and Google Cloud, which command 65% of the market. China follows with 20%, via Alibaba Cloud and Tencent. But for the rest of the world—Europe, Southeast Asia, Africa—building sovereign AI clouds is a pipe dream. "It demands trillions in capital, elite talent, and IP that most countries lack," he noted. "Relying solely on U.S. or Chinese clouds risks data sovereignty and backdoor vulnerabilities. Imagine your national health records funneled through Beijing or Silicon Valley."
Geopolitical tensions exacerbate this divide. U.S. export controls, tightened in January 2025, block advanced chips to China, widening the hardware gap: America holds 74% of global AI supercomputer capacity, versus China's 14%. Yet China counters with cost revolutions—its DeepSeek R1 model rivals top Western models at one-tenth the price, boasting 97 million users by April 2025 and exporting open-source frameworks to over 50 developing nations. Beijing's $100 billion AI spend this year funds not just models but governance pushes, like a proposed global AI cooperation body to rival U.S.-led standards.
Khein-Seng's solution? Democratize AI via flash-optimized infrastructure. Phison's edge systems slash cloud costs by 30-50%, enabling "local AI" for mid-tier economies. "Cheaper GPUs process slower, but with ample NAND, they suffice," he said. "We need 'AI blue-collar' workers—engineers customizing open-source like Llama for local apps, from Thai rice farming to Kenyan telemedicine." This echoes Phison's MaiStorage venture in Malaysia, blending Taiwanese hardware with regional talent to bypass U.S.-China duopoly.
Taiwan, Phison's home, exemplifies this hardware prowess but exposes software frailties. As the "silicon shield," the island fabs 90% of advanced chips, fueling 80% of AI GPUs. Yet, as Khein-Seng lamented, "We lead hardware but trail software by miles." Startups thrive in niche AI marketing, but Taiwan lacks large-scale innovators, hampered by venture droughts—$2.5 billion in 2024 versus Silicon Valley's $100 billion—and a manufacturing-first culture. Government initiatives, targeting trillions in AI value by 2040, aim to bridge this via talent pipelines and robotics, but critics call for bolder software subsidies.
Globally, storage leaders cluster in the U.S., Japan, South Korea, and China, holding 85% of NAND capacity. Taiwan's fifth-place rise, via Phison and others, stems from controller IP—the "brains" optimizing raw flash. But as AI edges toward 1,000-layer 3D NAND, cryogenic etching breakthroughs will favor incumbents, potentially locking out newcomers unless collaborations accelerate.
Khein-Seng's GITEX appearance, alongside AWS and IBM execs debating AI ethics, highlighted Phison's aiDAPTIV+ demo: a server cluster handling 10 petabytes of synthetic AI data without latency spikes. "This isn't hype," he urged. "It's survival. Without scalable storage, AI becomes a luxury for superpowers, widening inequalities."
As GITEX wraps—drawing 180,000 attendees and 6,800 exhibitors—the inventor's call resonates. With AI projected to add $15.7 trillion to global GDP by 2030, ignoring storage shortfalls risks stalling progress. Nations must invest in NAND innovation, foster "blue-collar" AI talent, and forge hybrid models blending U.S. compute with Taiwanese flash and Chinese openness. Otherwise, the data deluge drowns us all.
Phison's stock surged 8% post-GITEX on shortage buzz, but Khein-Seng remains pragmatic: "Innovation isn't optional. It's the only way to store our future." In a world where machines outpace humans 1,000-to-1 in data birth, his words aren't just a warning—they're a blueprint.
