Originally published on LinkedIn, February 2, 2026.

Introduction

For the last two years, we’ve heard the CHIPS and Science Act positioned as the cornerstone of U.S. AI leadership. The logic is simple, bring advanced chip manufacturing back onshore and the rest will follow, but from my perspective, it won’t. Compute is important and nobody is arguing that fact, but do we really believe compute alone defines AI power? Here’s an “emperor has no clothes” statement that probably isn’t very popular to say, but even if the United States, or any country for that matter, can own the world’s most advanced GPUs it will still lose the AI race if it does not control the layers beneath them.

I saw someone post a link to the America’s AI Action Plan earlier today, and that was the catalyst for finally writing this article.

So I have a few thoughts that were too long for a standard LinkedIn post.

Pushing the Limits of Compute

If you attended CES 2026 this year and heard Jensen Huang‘s keynote on the Vera Rubin GPU, you would agree that modern AI systems are no longer limited by raw compute. And if you’re still thinking about AI performance primarily in terms of number of GPUs, you’re missing what is actually the choke-point for these systems today.

This is one of those things that makes people uncomfortable because it is not solved with better software or a bigger model. The oxygen for AI is memory bandwidth, data movement, packaging density, power delivery, and thermal ceilings, and these are not software problems that can be optimized away with fancy algorithms, this is in the physical realm.

Today’s most advanced AI accelerators are inseparable from their memory. High Bandwidth Memory (HBM) is not an accessory you can swap out or upgrade independently, it is part of the system itself. There is no option to buy HBM as an option on the side, either. For you Apple users, this is very much like ordering a new MacBook – you either get the memory you need or you are stuck. When organizations need more HBM to support their workloads, they do not upgrade memory, they buy more systems, and those systems are GPUs.

And that system, the one that determines whether your AI infrastructure scales cleanly or hits a wall, remains overwhelmingly concentrated outside the United States, and that’s the core issue I see with the path forward in the Action Plan.

If this sounds familiar, it’s because it is, I’ve written before about how AI is becoming less about raw compute and more about memory, data movement, and the cost of moving bits instead of execution. So, really, this is the same issue, just viewed through a national and industrial lens instead of a system architecture one based on that Action Plan I reviewed.

If we preach sovereignty should be a chief consideration for our customer’s AI strategy, then shouldn’t this paradoxically concern us about our own technology sovereignty? We are reshoring logic while leaving the most fragile and irreplaceable components of AI infrastructure offshore, and in geopolitical terms, this is not resilience, it is exposure. The BCDR guy in me is sounding the alarms, anytime you are exposed, there is opportunity for exploitation. It’s like securing your front door while leaving the back door wide open, and then wondering why you’re still vulnerable.

Why CHIPS 1.0 Stops Short

CHIPS 1.0 did what it was designed to do, and we should acknowledge that as a good first step, but what it did not address is equally important, maybe more so when you start thinking about the actual systems that will run the AI models everyone is so excited about. We’re talking about advanced memory supply, 2.5D and 3D packaging, memory to compute fabric control, high density power and cooling, and system level AI integration and validation. Yes, all the unglamorous but absolutely critical infrastructure that sits between the chip and the actual work getting done.

Before I go on, I dropped some terms that some may be familiar with, while others are left scratching their heads.  What is 2.5D and 3D packaging?  It’s worth clarifying because not everyone lives in this world.  One way to think about the difference between 2.5D and 3D packaging is in terms of capability classes. Today’s advanced AI systems are like high end, turbo charged sports cars. Incredibly fast and already operating near physical limits. What comes next is a different class entirely, built on fundamentally different engineering assumptions around power, cooling, and integration. You do not evolve one into the other with incremental upgrades. You have to design the entire system differently from the start. At its core, the shift from 2.5D to 3D packaging is about collapsing the distance between compute and memory, eliminating friction and wasted energy.

One important thing to note is that NVIDIA‘s next generation Vera Rubin GPU still relies on advanced 2.5D packaging, which tells you how close the industry already is to the physical limits of today’s AI system designs.

But, owning the engine without owning the transmission and fuel does not win races, it just creates new failure modes that appear exactly when you need results. And when those failure modes hit, it won’t matter how advanced your logic chips are if the memory can’t keep up, or worse, not even available, if the packaging can’t handle the thermal load, or if the power delivery can’t scale – you’re dead in the water.

By the way, the July 2025 America’s AI Action Plan makes it clear that building American AI infrastructure matters, but turning that priority into real capability requires far more attention to memory, packaging, and system-level reliability than policy alone can provide. Enterprises feel that gap immediately, because they don’t deploy policy, they deploy systems.

The Real Shape of AI Infrastructure

I’ll say it again, AI does not run on chips, it runs on integrated systems. And those integrated systems are not limited by algorithms so much as by how they are engineered, and I’m not talking purely GPUs. This is where AI adoption slows down today, not because the models aren’t good enough, but because the infrastructure underneath them is brittle. The GPU gold rush has definitely cooled and now customers are in need for a forward looking strategy that will help set up the infrastructure investments they have made for success.

To that end, I think the next competitive advantage in AI will not come from another parameter jump or a new architecture that promises 10x improvements, it will come from industrial reliability at scale. Industrial reliability at scale means you can run production workloads day in and day out without worrying whether infrastructure becomes the bottleneck. That turns AI into an execution strategy, not an experimentation strategy.

CHIPS 2.0 Needs More Than Chips

If CHIPS is to evolve beyond its initial scope, and if we’re serious about AI leadership that lasts beyond the current hype cycle, it must expand beyond today’s current logic. At minimum, four pillars are missing from the current approach.

1. Memory sovereignty, a domestic capability for advanced memory and packaging is no longer optional, it is foundational to everything else we’re trying to accomplish.
2. Advanced packaging leadership, because the future of AI performance is stacked, bonded, and integrated, not etched thinner on the same old 2D paradigm.
3. Memory fabric control, and this is critical because data movement now dominates cost, latency, and power in ways that make the actual compute almost secondary.
4. Physical AI systems, where power density and liquid cooling and real world validation must be treated as national infrastructure concerns, not afterthoughts that get addressed once the “important” work is done.

Without these pillars, CHIPS remains a partial solution to a system level problem, and partial solutions tend to fail in expensive and embarrassing ways when you need them most.

Sovereignty Without Isolation

First and foremost, this is not an argument for nationalization or closed markets. When language like this is used, there’s a tendency to conflate sovereignty with isolation, or protectionism to use a more politically charged word. The most effective model is public-private alignment, not control. Private capital builds and operates because they’re better at it and more efficient, while government participates as an anchor customer and strategic guide, not as an owner trying to run industrial operations it has no business running.

Neutrality, trust, and scale are all essential, and none of them are achievable if you try to force a command-economy approach onto what has to remain a market-driven solution with strategic and advisory guidance. An AI infrastructure strategy that is credible has to serve commercial innovation and national security at the same time without distorting either one, which is hard to do, but not impossible if these partnerships are structured properly.

The Ground Beneath the AI Economy

AI leadership will not be decided by who trains the biggest model or who has the most impressive benchmark results on some academic dataset. It will be decided by who owns the integrated systems those models depend on, the infrastructure that makes AI practical, deployable, and reliable at scale. Compute was the first step, an important one that CHIPS 1.0 addressed, but memory, packaging, fabric, and physical integration are the next ones, and they’re just as critical to long-term competitiveness.

Conclusion

In my opinion, if the United States wants durable AI leadership, the kind that persists through economic cycles and geopolitical shifts, it must stop thinking in terms of chips alone and start thinking in terms of infrastructure and systems, integrated systems. This gets back to the car metaphor earlier, if we develop the best engine around but do not have control over the fuel, how far and fast can it really take us? The reality is, the real work is not happening at the top of the stack where everyone is focused on models and algorithms, it’s happening a layer below in the unglamorous world of memory bandwidth, power delivery, and thermal management. And if we don’t get that right, everything built on top of it will be on shaky ground.