The real AI contest will be won by who can best marshal the vast physical infrastructure needed to build the new technology, energy systems, semiconductor fabs, and enormous data centers.
The math is frightening. Demand for AI computing power has grown roughly five times per year since 2020, well above efficiency gains. Rather than fears of a potential bubble, we should fear an inability to keep up with supply.
The supply chain is straining. TSMC, whose advanced chips underpin virtually every frontier AI model, has kept capital expenditure near the top of a $52–56 billion range, while warning that capacity remains tight. OpenAI’s Stargate initiative encompasses nearly seven gigawatts of planned capacity and over $500 billion in projected investment over four years.
The race is now about cornering a physical resource whose production is bottlenecked at every stage: fab capacity, advanced packaging, high-bandwidth memory, power generation, and grid interconnection.
Nowhere is China’s position more constrained than in this tight compute environment. US Commerce Under Secretary for Industry and Security Jeffrey Kessler said in mid-2025 that Huawei would be unable to produce more than 200,000 advanced AI chips in the year — a fraction of domestic demand. Huawei’s Ascend line has improved, with the company aiming to produce 600,000 units of its top-tier Ascend 910C AI chips in 2026.
The gap between Huawei’s Ascend and NVIDIA’s Blackwell-class systems for frontier training remains material. American export controls should continue to focus specifically on preventing that gap from closing at the hardware level. The idea is to continue to compute-constrain China.
Given the immense leverage compute power bestows, the ability to allocate hundreds of thousands of frontier-class chips should be analytically closer to basing rights or nuclear cooperation agreements than to export licensing in the conventional sense. Countries are not merely purchasing technology. They are negotiating for a position for a crucial resource.
When Washington weighed permitting the United Arab Emirates to import 500,000 or more advanced NVIDIA chips annually, the deliberations focused on where the capacity would be physically located, who would operate it, and what obligations would attach to access.
This reframing has consequences for how governments should think about AI policy. The dominant tendency is still focused on debates about model outputs, content liability, and voluntary safety commitments. While addressing those issues remains important, it would be a monumental mistake to misidentify the priority as securing adequate supply.
Compute is the binding constraint on frontier capability; the crucially important policy levers are in energy permitting, transmission infrastructure, advanced semiconductor packaging, high-bandwidth memory supply chains, and the export control architecture that governs who can access the best systems. None of these is glamorous. All of them are, at this moment, more consequential in the strategic competition over frontier AI than anything else.
Two points deserve particular emphasis. First, power. The US possesses the financial capital and the engineering base to lead in frontier AI, but its ability to translate that into deployable compute is increasingly limited by electricity. Grid interconnection queues run to years, and permitting timelines compound the delay. A country that cannot reliably deliver power at gigawatt scale to compute clusters will find its frontier advantages eroding regardless of its chip supply.
Second, the export control architecture requires continuous attention. The history of dual-use technology controls suggests they work best when maintained as a moving target. As such, they must be updated faster than workarounds can be institutionalized. In other words, controls set in 2023 against 2026 hardware are outdated
Computing capacity represents the commanding height industry of the AI era. It will be shaped by electrical grids, fab yields, packaging throughput, and rack-level thermal management. It is high time that we treat compute infrastructure as a strategic asset on the order of energy reserves or military production.
Elly Rostoum is a Senior Resident Fellow with the Center for European Policy Analysis (CEPA).
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
Tech 2030
A Roadmap for Europe-US Tech Cooperation
