Why Memory is the key to Unlocking AI’s future
As AI models grow larger and more complex, they’re rapidly outstripping the capabilities of today’s memory and storage technologies. The widening gap between compute speed and memory capacity is emerging as a critical bottleneck. While traditional high-bandwidth memory (HBM) offers speed, its high cost, limited capacity, and energy demands make it difficult to scale. Without innovation in memory architecture, AI’s momentum could stall. Bridging this gap isn’t just about performance—it’s key to unlocking the next generation of truly transformative AI systems.
As the industry works to develop scalable, efficient memory solutions that can keep pace, Sandisk has introduced its High Bandwidth Flash (HBFTM)—a game-changing memory technology designed to support the future of AI everywhere.
We have hit the ‘memory wall, ’ which means we need to rethink how these systems are built. We need an ecosystem that allows us to run very large and intricate models that were not previously possible.”
Alper Ilkbahar
Chief Technology Officer, Sandisk
As AI models mature, the strain on infrastructure increases. Larger context windows and heavier inference workloads demand far more memory capacity and bandwidth than current systems can efficiently deliver, not only in cloud and training clusters but also at the edge, where smaller devices face even tighter power and memory constraints. The result is a widening performance gap with cascading effects across cost, power, and scalability.
the challenge: AI’S GROWING MEMORY GAP
01
A Mounting Memory Crisis
33%–50%
Amount of data-center Total Cost of Ownership accounted for by memory
Source: Processor Architecture Research Lab, Intel Labs
Source: ABI
Number of data centers expected to be built worldwide between 2025–2030
2,267
Source: International Energy Agency (IEA)
Projected growth in global data center electricity demand by 2030
>2×
The cost of AI is already high. Business leaders are now asking, ‘How are we going to power up all these data centers? What does it mean for the environment?’ All the subsequent problems are things that we are trying to solve.”
Ilkbahar,
Sandisk
HBF is a memory architecture from Sandisk, designed to overcome the limits of existing technologies—redefining what’s possible in AI memory. It builds from Sandisk’s low-cost, high-density CMOS bonded to array (CBA)-based NAND core technology with architectural improvements in silicon technology, design technology, and 3D stack packaging, enabling high bandwidth, high endurance, and durability, as well as energy efficiency.
BREAKING THROUGH THE WALL WITH HBF
02
The performance difference between today’s leading HBM and HBF, the latter of which offers far greater storage capacity and low-power circuits
2.2%
Source: Sandisk
Sandisk
Ilkbahar,
HBF is at the heart of this memory-centric AI architecture. It delivers both high capacity and high bandwidth without increasing the overall system cost.”
HBF Technology:
Designed for the Next
Generation of AI Memory
HIGH BANDWIDTH,
LOW COST
One of the lowest costs per bit in the memory industry, HBF can help maximize data throughput without breaking the budget.
CBA technology delivers ultra-high density, speed, and low-power circuits, redefining performance standards for AI-driven workloads.
CBA
BREAKTHROUGH
Built on NAND technology, HBF retains data without refresh power, cutting energy use and costs.
Nonvolatile power efficiency
Designed to be more resilient to high temperatures present in data center environments, HBF maintains top performance and reliability even in the most demanding conditions.
Thermal stability under pressure
HBF is engineered for high endurance and reliable operation, helping to ensure consistent uptime even in extreme environments.
Smart reliability and redundancy
HIGH BANDWIDTH FLASH
The focus of system design is shifting from compute to memory as AI models grow and diversify. A memory-centric approach enables more flexible architectures that combine technologies such as HBF and HBM, scaling effectively without simply adding more processing power. The approach allows for a personalized, persistent AI experience, in which AI is able to remember previous responses and context to better respond to the user. It also makes it possible to run AI in more places—from hyperscale data centers to edge devices—while improving cost and energy efficiency by reducing unnecessary
data movement across the system.
THE SHIFT TO MEMORY-CENTRIC AI
03
8x–16x
Source: Sandisk
Memory capacity advantage of HBF compared with HBM at similar bandwidth and cost
Sandisk
Ilkbahar,
When memory is no longer the limiting factor, AI can live everywhere—not just in the cloud or massive data centers but on personal devices, in cars, and at the edge. You can bring powerful, personalized intelligence closer to the user instead of moving everything back and forth to a cluster somewhere else.”
UNLOCKING THE FUTURE
The rapid rise of AI has brought the industry face to face with the memory wall, a structural barrier that can’t be overcome by simply adding more computing capacity. Breaking through it requires a different approach, one that puts memory at the center of system design. Memory-centric AI unlocks new ways to scale efficiently, reduce energy demands, and extend AI to more places and applications. As architectures evolve, the future of AI will be defined not just by how fast it computes but also by how well it manages ever deeper and more complex flows of data.
Learn more about how Sandisk’s HBF is enabling the shift to memory-centric AI at Sandisk.
© 2025 Fortune Media IP Limited. All Rights Reserved. Use of this site constitutes acceptance of our Terms of Use and Privacy Policy | CA Notice at Collection and Privacy Notice | Do Not Sell/Share My Personal Information FORTUNE is a trademark of Fortune Media IP Limited, registered in the U.S. and other countries. FORTUNE may receive compensation for some links to products and services on this website. Offers may be subject to change without notice.
04
The performance difference between today’s leading HBM and HBF,which also offers far greater storage and lower electricity use
up to
Memory capacity advantage of HBF compared with HBM at similar bandwidth and cost