AI Arms Race Escalates Into Full-Stack Infrastructure War

Nvidia is preparing to spend roughly $26 billion building open-weight AI models over the next five years, crystallizing how the AI arms race has become a full‑stack infrastructure war spanning chips, power, and entire data‑center geographies. The sheer scale of that commitment, disclosed in recent financial filings and detailed by WIRED, puts model development on the same capital footing as semiconductor fabs and long‑haul transmission lines.

That level of spending will not just accelerate model capabilities; it will reshape where power plants are built, how grids are managed, and who controls the next generation of machine intelligence.

From GPUs to custom silicon stacks

Nvidia’s $26 billion bet is explicitly aimed at open‑weight models that enterprises can run or fine‑tune on their own infrastructure, part of a push to seed demand for ever‑larger GPU clusters while keeping developers inside its software ecosystem, according to the company’s latest disclosures and reporting by WIRED.

Rivals are racing to control more of the stack. Meta has unveiled multiple generations of its internally designed Meta Training and Inference Accelerator (MTIA) chips and plans to use them in production data centers alongside Nvidia hardware, even as it quietly delayed a next major Llama release, codenamed “Avocado,” in late 2024, as reported by The Information and confirmed in subsequent coverage by Reuters. Broadcom, meanwhile, has become a central behind‑the‑scenes supplier, building custom “XPU” accelerators for hyperscalers like Google, Meta and ByteDance, tailored to their proprietary AI workloads, according to Broadcom’s own disclosures and recent earnings commentary.

This is no longer just a contest over who has the best off‑the‑shelf GPU. The largest platforms are converging on vertically integrated stacks: in‑house models, custom accelerators, optimized networking, and data‑center designs tuned for agentic and always‑on AI workloads. That integration increases performance—and also deepens lock‑in.

Power grids and real‑world infrastructure bend toward AI

The hardware race is slamming into physical limits. McKinsey estimates that data centers already draw 3–4% of total U.S. electricity and could climb to 11–12% by 2030 if AI adoption continues on its current trajectory, effectively tripling sector power demand by decade’s end, according to a recent McKinsey analysis. Utilities and regulators are scrambling to keep up.

In the U.S. Southeast, Georgia’s regulators approved a plan in January 2025 for Georgia Power to boost generation capacity by about 50%, explicitly citing a wave of new data‑center projects, many AI‑driven, as reported by AP News. Across the Midwest, the U.S. Department of Energy has extended a $1.6 billion loan guarantee to upgrade high‑voltage transmission lines, framing the project as necessary to handle surging loads from data centers and AI, according to AP News. A working group advising the Department of Energy warned in 2024 that AI‑driven data‑center growth represents the “leading edge” of U.S. electricity demand and urged coordination between cloud giants, private‑equity‑backed data‑center operators, and grid planners, in recommendations published by the agency’s advisory board and summarized in a DOE report.

Regulators are starting to treat AI campuses as quasi‑industrial loads on par with steel mills. In Virginia and Texas, power demand from data centers is projected to jump by multiple gigawatts year‑over‑year, with utilities warning that failure to meet AI‑related load growth could ripple across the broader economy, according to CNBC’s reporting and grid forecasts compiled by S&P Global.

Cloud, campuses and local spillovers

Cloud providers are reorganizing around this demand. Oracle has signaled plans to sharply raise capital expenditures, with projections of roughly $35 billion in 2026 to expand its cloud and AI infrastructure footprint, according to Data Centre Magazine. The company has also disclosed lease commitments for data centers and cloud infrastructure of about $248 billion as of late 2025—up 148% from just three months earlier—underscoring how long‑dated these AI bets have become, as eWEEK and CNBC have reported.

Smaller infrastructure providers echo the shift. GPU‑rental platforms such as RunPod and others have reported that AI training and inference now dominate their workloads, prompting redesigns of data‑center layouts, cooling, and network fabrics to prioritize accelerator‑dense clusters. At the same time, major operators like Oracle are touting closed‑loop, non‑evaporative cooling systems to reduce pressure on local water supplies, as outlined in a recent Oracle blog post, even as their campuses occupy thousands of acres and tie up gigawatts of potential capacity.

The local consequences—from housing pressure near new campuses to rising land values around transmission corridors—are still emerging. But with data‑center capacity in the U.S. projected to more than double by 2035, from roughly 35 GW in 2024 to 78 GW, and AI loads responsible for a large share of that growth, according to a 2025 synthesis of government and industry forecasts by AIX Energy, this is no passing boom.

Who controls the next intelligence layer?

The result is a geopolitical and economic contest over who owns the infrastructure that will host increasingly capable, agentic AI systems. Nvidia’s $26 billion open‑model program, Meta’s custom accelerators, Broadcom’s private XPUs and Oracle’s quarter‑trillion‑dollar leases are all bets that control of compute, power, and real estate will translate into control of the next intelligence layer.

For policymakers, the challenge is to channel that capital into cleaner grids and more efficient hardware without letting AI data centers crowd out other social priorities. For everyone else, the AI arms race now runs through very physical battlegrounds: substations, substacks and the neighborhoods that grow up around them.