Meta and xAI are channeling staggering sums into data center infrastructure, betting that raw computational power will eventually close the gap their AI models haven’t managed to bridge on their own.

The infrastructure arms race

Meta’s marquee project is Hyperion, a facility in Louisiana that, when completed, will consume more electricity than the entire city of New Orleans. The physical footprint is equally absurd. Hyperion is expected to cover an area comparable to lower Manhattan. Zuckerberg has said Meta plans to invest “hundreds of billions” into AI data centers.

Then there’s Prometheus, Meta’s Ohio-based facility projected to go online in 2026.

xAI’s Colossus supercomputer was constructed in just 122 days inside a repurposed factory. The plan is to scale Colossus to accommodate one million GPUs, backed by strategic acquisitions in the energy sector.

Where the models fall short

For all the infrastructure muscle, the AI models themselves tell a less triumphant story. Meta’s Llama series and xAI’s Grok have been released in successive iterations, but neither has managed to consistently outperform competitors like OpenAI’s GPT series or Google’s Gemini in ways that matter to developers and enterprise customers. The challenges aren’t necessarily about the models being broken. They’re about cost, efficiency, and the increasingly crowded field of foundation model providers.

Community pushback and energy concerns

Both Meta and xAI face growing local opposition in the communities where these data centers are being constructed. Residents are raising concerns about energy consumption, emissions, and the strain on local power grids.

What this means for investors

Companies providing energy solutions, construction services, and specialized data center technology stand to benefit from what could be a multi-year, multi-hundred-billion-dollar buildout cycle. The demand for power, in particular, creates an interesting secondary market for energy producers and grid infrastructure companies.

The key metric to watch isn’t whether Llama or Grok catches up to GPT. It’s whether the sheer scale of infrastructure spending generates returns that justify the cost. The next 18 months, as Prometheus comes online and Colossus scales toward its million-GPU target, will offer the first real data points on that question.