For years, the global AI race was described as a contest between models. This week’s developments suggest the real competition is moving beneath them—to power, land, chips, cooling and the networks that connect intelligence to businesses.

On 5 September, TCS subsidiary HyperVault announced that it had secured 264 acres in Hyderabad for a phased AI data-centre campus with capacity of up to one gigawatt. The plan targets high-density training and inference workloads. In Europe, the Commission’s AI Gigafactories programme is mobilising €20 billion, with its formal call launched in July after 77 proposals across 16 member states.

The geography is changing, but so is the architecture. Equinix announced a distributed inference programme with NVIDIA and Together AI on 2 September. Scheduled to become available from the first quarter of 2027, it is designed to move model execution closer to the data, applications and users it serves.

These are not isolated construction stories. PwC’s new baseline outlook projects US$31.6 trillion of global AI-infrastructure capital expenditure through 2050, while stressing that the number is a forecast shaped by power access, chip supply, policy and investment conditions. The International Energy Agency reports that data-centre electricity demand grew 17% in 2025; electricity use by AI-focused facilities rose 50%.

The commercial lesson is straightforward. Model capability may be global, but reliable AI delivery is physical and increasingly regional. Latency, energy price, data residency, grid capacity and local regulation can affect where systems run and what they cost.

For smaller businesses, this does not mean buying servers. It means asking better questions of technology providers: Where will our data be processed? What happens if capacity is constrained? Can workloads move between models or regions? How is energy and infrastructure cost reflected in the price? What service level is actually guaranteed?

The countries and companies that secure compute will influence the next phase of AI. But the winners will not be those that build the biggest facilities without discipline. They will be those that turn expensive infrastructure into dependable, affordable business outcomes.

AI’s next global contest is not only about who creates the smartest model. It is about who can run intelligence reliably, close to where real work happens.

Sources

Editorial note: the infrastructure-race framing is SEPIDRA’s synthesis of the cited developments. Planned capacity and investment forecasts are identified as such and are not presented as completed infrastructure.