Meta Builds Five Gigawatt Data Centers to Overtake Google in AI Race

A year after the failure of the Llama 4 model, Meta has rebuilt its artificial intelligence program from the ground up, and according to the latest report from analytics firm SemiAnalysis, it could overtake Google in compute power and model quality within the next six months. Analysts describe it as the most aggressive infrastructure buildout the industry has yet seen.

The SemiAnalysis report, titled 'The Future of Meta Superintelligence: A 1 Year Progress Update,' offers one of the most detailed public looks yet at how Mark Zuckerberg has responded to his AI lab's earlier setbacks. After Llama 4 disappointed relative to the competition in quality, Meta underwent a year of restructuring that included creating the Meta Superintelligence Labs unit and a massive injection of capital.

Five Titans at Once

The most striking element of the report is the scale of the data center buildout. Meta is simultaneously erecting five clusters exceeding a gigawatt in power, which analysts say is unprecedented in the tech industry. Prometheus in Ohio is set to eventually exceed 3 gigawatts and is already partially operational, using tent-like structures that speed up construction. Hyperion in Louisiana has 1.5 gigawatts of capacity spread across three buildings of 400 megawatts each.

The Iowa cluster, with 1 gigawatt of capacity, went from lease signing to construction completion in a year, from May 2025 to May 2026. Additional, still-unnamed campuses are being built in El Paso and Indiana. SemiAnalysis estimates that by year's end, Meta will surpass both OpenAI and Anthropic in total compute power.

In-House Data Instead of Vendors

Rather than relying solely on external firms to supply training data, Meta has turned its efforts inward. About 3,000 engineers, mostly fresh graduates but also experienced staff, have been assigned to build an internal factory for reinforcement learning environments. The goal is to give the company access to data that commercial brokers cannot provide.

At the same time, companies specializing in supplying such environments to the industry at large have exploded in the market. According to the report, Mercor, Surge, and Handshake are each already generating more than $1 billion in annual revenue, while Fleet, Mechanize, and Afterquery are approaching $100 million. Mercor alone reported 2.5 million hours of expert work in the second quarter of 2026, roughly equivalent to 4,800 full-time positions, at rates exceeding $100 per hour.

Costly Recruitment

The scale of investment in talent is just as impressive as that in infrastructure. Meta paid $14.3 billion to bring in Alexandr Wang from Scale AI to lead its new superintelligence unit. Some researchers received compensation packages ranging from several hundred million to more than a billion dollars, and top software engineers earn seven-figure salaries annually.

The current set of algorithms is enough to automate office work, provided you have enough data of the right kind - Sholto Douglas, Anthropic researcher, quoted in the report from the Dwarkesh podcast

Models Still Lagging

Despite the massive spending, Meta's model results still fall short of market leaders. The report's authors tested Muse Spark 1.1 and rated it roughly on par with Claude Opus 4.6 or GLM 5.2 in general agentic applications, but noted that the model 'has a bad habit of ignoring warnings instead of fixing them.' In benchmark tests, Meta trails both China's DeepSeek v4 Pro and Kimi K2.6, and has lagged behind Anthropic and OpenAI throughout 2026.

SemiAnalysis stresses, however, that Meta's chances hinge not on its starting point but on its rate of improvement. If Zuckerberg maintains his financial resolve in the coming months, Google could be permanently pushed out of the top tier of global AI infrastructure providers, the report's authors argue.

For Polish companies using AI models and cloud infrastructure, these shifts in the compute race carry practical weight. Greater competition among Meta, Google, OpenAI, and Anthropic typically translates into falling prices for model access and a faster pace of updates, but also growing uncertainty over which provider will hold the technological edge over the coming quarters.

Sources: The Future of Meta Superintelligence: A 1 Year Progress Update (newsletter.semianalysis.com), Meta set to overtake Google's frontier AI models in six months, SemiAnalysis says (uk.finance.yahoo.com)