Intel and SambaNova have unveiled a colossal computing rack that squeezes 36,864 CPU cores into a single 100-kilowatt chassis, designed specifically for the emerging field of 'agentic AI' – artificial intelligence that can plan, reason and act autonomously. The system, built on a disaggregated inference architecture, separates processing from memory to boost efficiency when running large language models. The first commercial customer has already been secured, though neither company has named them.
Disaggregated inference means that instead of forcing every processor to carry its own dedicated memory, the rack pools memory resources and allocates them dynamically. This approach reduces idle time and cuts the energy wasted when traditional servers wait for data. For UK businesses, this could lower the total cost of running AI models on premises, avoiding the need to rent cloud capacity from US hyperscalers.
However, the 100kW power draw per rack raises serious questions about energy consumption and carbon targets. Dr Eleanor Hargreaves, a computing infrastructure researcher at the University of Manchester, commented: 'A single rack consuming as much electricity as a small housing estate is unsustainable without clean energy. UK data centre operators will need to pair these systems with renewable power or face higher costs and regulatory pressure.'
The timing coincides with increased regulatory activity from the UK's Information Commissioner's Office (ICO) and the EU's AI Act, which imposes strict transparency and risk-management rules on high-impact AI systems. UK firms deploying agentic AI must ensure their models comply with data protection law, particularly around automated decision-making. The ICO has already warned that organisations using AI must be able to explain how decisions are reached.
For the UK economy, the technology offers a double-edged sword. On one hand, it could spur domestic AI innovation by giving British startups access to powerful, on-site hardware without cloud lock-in. On the other, the high energy demands may clash with net-zero pledges, potentially requiring grid upgrades or special tariffs. SambaNova claims its architecture is more efficient than rival systems, but independent benchmarks are not yet available.
Looking ahead, industry analysts expect more UK enterprises – particularly in finance, healthcare and defence – to explore on-premise AI infrastructure as data sovereignty concerns grow. If the Intel-SambaNova system proves reliable, it could accelerate the shift away from purely cloud-based AI, giving UK firms greater control over sensitive data and inference costs.