Back

AI Infrastructure Shifts Toward Smaller, Distributed Data Centres

lite News Desk 2 min read

At a glance

  • AI model training can be distributed across smaller facilities
  • Legacy data centres often cannot support heavy AI hardware
  • Edge data centres help reduce latency for front-end applications

AI infrastructure is evolving as organizations explore alternatives to large, centralized data centres, with distributed and smaller-scale facilities gaining attention for certain workloads.

Some experts have stated that training AI models does not always require large, centralized data centres, and that smaller sites or edge environments can be used instead. A study from EPFL found that while advanced AI model training still needs substantial computing power, many operational AI systems can run on existing machines, regional servers, or edge environments without relying on hyperscale facilities.

Distributed training approaches, such as DiLoCo, have been reported to reduce communication overhead by partitioning workloads across several data centre clusters rather than using a single large facility. This method can decrease communication requirements by up to 500 times, according to The Economist.

Developments from companies like DeepSeek in China have shown that language models can achieve performance similar to established systems like GPT while using less computing power. This has led to further examination of whether large-scale infrastructure is always necessary for effective AI deployment.

What the numbers show

  • AI racks can weigh up to 5,000 pounds, exceeding many floor load limits
  • A one-gigawatt data centre could cost about $80 billion, according to IBM's CEO
  • Scaling to 20–30 gigawatts of data centre capacity may require up to $1.5 trillion

Legacy data centres often face challenges in supporting AI hardware due to the heavy weight of AI racks, which can surpass the structural limits of many existing facilities. While some older data centres can be retrofitted to accommodate AI workloads, experts have said that most would need to be completely rebuilt to meet the demands of modern AI hardware.

Edge and micro data centres, which are smaller facilities located closer to data sources, are being used to lower latency and support front-end AI applications. These mini data centres can offer cost-effectiveness, resilience, and security, but may also contribute to electronic waste and may depend on high-carbon energy sources unless specific measures are taken.

According to the AI and climate lead at Hugging Face, smaller, customized AI models running locally can achieve performance comparable to larger models while using fewer resources. This approach may reduce the need for extensive infrastructure in certain applications.

Industry reaction

SAP CEO Christian Klein stated that demand for new data centres in Europe remains limited, and suggested that the region should prioritize the AI software sector instead of expanding physical infrastructure.

IBM CEO Arvind Krishna questioned the financial practicality of investing in massive AI data centres, highlighting the high costs associated with building and scaling such facilities.

* This article is based on publicly available information at the time of writing.

Sources and further reading

Related Articles

  1. Nvidia boosts its robotics presence with new AI models and global partnerships, showcasing innovations at CES 2026.

  2. Bruno Fernandes's X account was hacked, according to reports. Fans are advised not to engage with any suspicious activity on the platform.

  3. Access to Grok was blocked in Indonesia on January 10, 2026, and restricted in Malaysia on January 12, according to reports.

  4. Injecting randomness into quantum neural networks enhances their measurement capabilities, according to recent studies on quantum properties.

  5. Fugaku simulated the mouse cortex, modeling nearly 10 million neurons and 26 billion synapses across 86 regions, according to the study.

More on Technology

  1. Researchers developed a quantum-optimal coronagraph, achieving a 1000:1 contrast ratio and enhancing exoplanet detection speed by up to 100 times.

  2. A proposal from the Conservative Party seeks to ban social media access for those under 16, citing similar actions taken in Australia for safety.

  3. His work on PBS’s Computer Chronicles introduced early personal computing to a broad television audience, documenting key developments in consumer technology.

  4. Innovations showcased at CES 2026 include LG's wireless OLED TV, foldable steering wheels for autonomous vehicles, and advanced AI gadgets, reports say.

  5. Entities linked to Google founders relocated or inactivated more than 60 California companies in late 2025, according to media and corporate filings.