Introduction

Over the past two years, the AI conversation has focused almost entirely on GPUs and models, which makes sense given that most of the attention is on outcomes and how we achieve them. While that was happening, I was quietly looking at the AI workload itself and where the friction actually exists. What I found is that the real constraint does not center on compute, but on data movement.

If the problem were simply moving data from storage repository A to repository B, then the focus would likely be on filesystem and network performance. But I tend to think about this a little differently, and it has been interesting to see that what once felt like an isolated line of thinking is now beginning to show up more broadly in industry conversations.

The real constraint shaping the future of AI is memory, specifically advanced memory.

I originally wrote this in October of last year but held it back because the market did not seem ready for this kind of message. This thesis examines why the next strategic battleground in artificial intelligence may be advanced memory manufacturing and why the United States may need to rethink how it approaches this critical layer of infrastructure, particularly as policymakers begin considering what a second phase of the CHIPS Act might need to address. The idea itself is straightforward but increasingly difficult to ignore.

The Great AI Memory Foundry

Artificial intelligence has become the industrial engine of our time, yet the foundation it depends on is dangerously fragile. According to industry analyses, more than 90% of the world’s advanced memory manufacturing occurs outside U.S. borders, concentrated primarily in East Asia, a region that sits amidst growing geopolitical tension and supply-chain risk. It is a strategic vulnerability for any nation betting its future on AI.

We have seen this pattern before. While compute power races ahead, advanced memory capacity and locality lag behind. Billions are poured into GPUs and accelerators, not because the processors themselves are the constraint, but because modern AI workloads demand enormous memory bandwidth and locality. When those requirements are not met, GPUs appear idle, waiting on data movement rather than computation. If AI is the new electricity, memory is the grid. Today, America does not own the grid.

The Memory Gap

In the 1990s, the United States produced nearly 37% of the world’s semiconductors. Today, that number is closer to 12%. Intel and Micron once held a unique advantage with their joint development of 3D XPoint, a technology that bridged the gap between volatile and persistent memory. It represented a fundamental shift in how data could be accessed, moved, and retained.

Both companies ultimately walked away. Micron sold its Utah fab in 2021, and Intel shuttered Optane a year later. From a business perspective, that decision is understandable, as Intel was reportedly losing hundreds of millions of dollars per year on Optane. However, what was lost was not just a product, but America’s last foothold in next-generation memory. The lesson here is about foresight, not necessarily failure. Intel tried to compete in an already competitive space by packaging the 3D XPoint technology as an SSD, which was a lack of foresight. The technology itself, had they focused solely on the persistent memory market (PMEM), was not a failure. That advanced memory was treated as a component when it should have been treated as a capability.

While others stepped back, NVIDIA leaned forward. Through acquisitions such as Mellanox and the Enfabrica “acqui-hire”, it has quietly positioned itself as the architect of AI infrastructure. The strategy, to me, is clear, compute without balance is waste, and performance without data flow is inefficiency. The true constraint on AI today is not processing speed, but the ability to move, stage, and access memory at scale.  This reflects a broader thesis that sovereign AI leadership ultimately depends on control of infrastructure. Every stalled workload and every underutilized GPU points to the same thing, the system is starving for memory.

From Silicon to Sovereignty

The last decade was defined by who controlled compute, but the next will be defined by who controls memory, and we’re not just talking about system memory. If others control the design, production, or distribution of advanced memory, they control the pace and sovereignty of AI itself. You cannot claim leadership in intelligence if you do not control the foundation from which it learns. For too long, the United States has poured investment into processors, accelerators, and training clusters while outsourcing the very layer that makes them usable. We have seen this before with energy dependence and rare-earth minerals, but now it is digital infrastructure. In the age of AI, sovereignty is not measured by borders or bandwidth. It is measured by where your memory lives.

A Blueprint for Renewal

The Great AI Memory Foundry is not a company, it is an initiative. A coordinated effort to rebuild what was lost and accelerate what comes next. It could be supported under the CHIPS Act as a public-private partnership spanning government, academia, and industry, focused explicitly on advanced memory systems. It needs to be part of the next phase of the CHIPS Act.

Companies such as NVIDIA, Intel, Micron, and AMD could collaborate not on products, but on stability. The Department of Energy and the Department of Defense could partner with private enterprise to secure a scalable domestic supply of advanced memory for AI, high-performance computing, and data-driven industries.

The CHIPS Act was a critical first step toward restoring semiconductor manufacturing but it fell short of addressing the most strategic layer, in my opinion. The Foundry completes that effort by focusing on the data and memory systems that give compute its purpose. Compute is inert without memory, which is why I write about data movement, infrastructure flow, and ground it in memory. Sovereignty is hollow without control of the infrastructure that sustains it.

The Rebirth of Industrial Capability

When steel defined national strength, we built mills. When the space race demanded innovation, we built rockets. When data became the renewable energy of the digital age, we built the cloud. As intelligence becomes the next industrial power, we must build the memory that keeps it alive, and this needs to be a priority.

The Great AI Memory Foundry would bring manufacturing, engineering, and research back to American soil. It would ignite an ecosystem of startups, suppliers, and scientific advancement. It would establish the United States as a global center of foundational AI infrastructure.

Every year we delay, leadership slips further into the hands of others.

The time to build that foundation is now.