The AI silicon arms race moves to chips and grids

Nvidia’s plan to spend about $26 billion building “open‑weight” AI models marks a new phase in the AI arms race, one where control over chips, data centers and electricity grids matters as much as algorithms. The outlay, revealed in recent securities filings and reported by Wired, comes on top of the company’s pledge to help finance up to $500 billion in U.S. AI infrastructure over four years, according to a government-facing briefing from Nvidia. That level of spending is turning AI progress into a contest of physical capital and energy.

At the same time, a growing cast of hyperscalers is racing to secure its own silicon and power footprint. TrendForce estimates that capital expenditures by eight major cloud providers, including Meta, Microsoft and Google, will exceed $420 billion in 2025 alone, largely to buy Nvidia GPU racks, build data centers and design in‑house AI accelerators TrendForce. The scramble is already spawning megaprojects like the multi‑partner “Stargate” super‑data‑center initiative and multi‑billion‑dollar orders for Nvidia GPUs by Oracle and others, as documented by Tom’s Hardware and Wikipedia’s summary of Stargate LLC.

Nvidia’s leverage, Meta’s counter‑move

Nvidia’s dominance of high‑end AI accelerators — one analysis cited by the French think tank IFRI put its GPU share around 90% in late 2024 — gives it outsized influence over who can train frontier models at scale IFRI. By bankrolling open‑weight models itself, the company is positioning as both arms dealer and AI model supplier, aiming to anchor workloads on its own hardware and software stack, as Wired notes.

Big platforms are not standing still. Meta has been rapidly iterating its in‑house Meta Training and Inference Accelerator (MTIA) family, with successive generations designed to run recommendation systems and generative AI inference more efficiently than off‑the‑shelf GPUs for certain tasks, according to company briefings covered by S&P Global and other analysts. Meta is also pouring billions into new data centers, such as a planned $10 billion facility in Louisiana reported by Reuters, underscoring how owning silicon, buildings and power contracts is becoming a strategic hedge against Nvidia’s pricing and supply.

Grids as the new bottleneck

Even if companies secure enough chips, electricity and cooling are emerging as hard constraints. A 2024 Bloomberg analysis found that AI‑optimized chips like Nvidia’s B100 can draw nearly twice the power of their predecessors, with clusters pushing local grids to their limits Bloomberg. A recent report for the Southern Environmental Law Center concluded that current U.S. data‑center power forecasts implicitly assume the country would commandeer more than 90% of the world’s new advanced chips by 2030 — a scenario the economists behind the study called unrealistic given global semiconductor capacity Utility Dive.

The United Kingdom has become a test case for how governments respond when AI build‑out collides with infrastructure limits. Demand for grid connections from energy‑hungry projects more than tripled to 125 GW by mid‑2025, exceeding today’s national peak load, according to data cited by energy analyst Tim Harper and the National Energy System Operator Tim Harper. In response, the UK government set up an AI Energy Council to overhaul the grid connection process and explicitly flagged AI data centers as a priority in its June 2025 reforms UK government.

Concentrated compute, concentrated risk

This convergence of silicon scarcity and grid strain is reshaping where AI power clusters and who benefits. S&P Global notes that requests from data‑center firms in just one U.S. utility territory, Dominion Energy Virginia, jumped to 40.2 GW of power connections in early 2025, nearly double mid‑2024 levels S&P Global. Globally, electricity demand from data centers is on track to roughly double by 2026, putting new stress on transmission networks, according to grid‑planning research summarized on arXiv.

The risk is that a handful of firms — those able to lock up Nvidia supply, build custom chips, and secure multi‑gigawatt power deals — come to control not just AI products but the underlying capacity to compute. That concentration raises familiar concerns about pricing power and competitive barriers, but now fused with questions of carbon emissions, local water use and national energy security. As Nvidia, Meta and their peers escalate spending into the hundreds of billions, decisions about where to site data centers and who gets grid priority will shape which countries lead in AI — and how much physical infrastructure strain societies are willing to absorb to keep training ever‑larger models.