Axelera's power-efficient chips let enterprises build AI infrastructure without liquid cooling
As companies shift from training models to running inference at scale, expensive GPU racks with liquid cooling are keeping many locked out of on-premises AI. Axelera AI is targeting that gap with low-power chips designed to fit standard enterprise data center racks.

Inference hardware that consumes far less power is emerging as an accessible entry point for organizations seeking to deploy AI on their own equipment. The industry's transition from model development to daily inference workloads has created a problem: traditional GPU infrastructure demands both substantial capital investment and specialized cooling systems, forcing many enterprises to lease computing resources instead of owning them. This constraint has opened an opportunity for alternative chipmakers to challenge established players. Axelera AI B.V. began by developing low-power chips and tools for edge AI applications, and is now bringing that expertise into data center deployments, according to Manuel Botija, vice president and head of product at the company.
Enterprises cannot own their own infrastructure. They're locked out because of the cost and the power. We are enabling them now to own it. It's not that we are helping the world save power or save money — which we are — but it's that we are unlocking use cases that were simply not possible before.
Manuel Botija, Axelera AI
Botija shared these insights during an exclusive conversation with theCUBE Research's John Furrier and Rebecca Knight at SC26, where the discussion centered on how power-efficient inference processors are moving into data centers and enabling enterprises to build their own AI infrastructure rather than depending on cloud compute rentals.
Power-efficient inference chips head for the data center
Axelera's Europa chip entered production in September. The company announced Titania, its third-generation architecture, ahead of SC26. This new design scales up to a 400-watt processor, and the resulting six-kilowatt server delivers performance comparable to leading data center GPUs while remaining compatible with the standard 20- to 30-kilowatt air-cooled racks found in typical enterprise data centers, Botija explained.
Having a six-kilowatt server that can perform like a data centre GPU is a pretty big deal. It basically means that they have some storage equipment, some networking equipment, some traditional CPU equipment, and next to all of that, they can also insert some AI inference hardware acceleration without needing to go and rebuild another server room or another data center.
Manuel Botija
For organizations that prefer to purchase AI hardware outright rather than pay ongoing cloud fees, the obstacles have been formidable. A single GPU rack typically costs a minimum of $1 million and frequently exceeds that figure, and most setups require liquid cooling infrastructure, Botija noted. Axelera's approach leverages digital in-memory computing, a methodology rooted in research conducted by its founders that enables the company's power-efficient inference processors to achieve superior performance-per-watt efficiency.
https://www.youtube.com/embed/VkhJAK_r19M?feature=oembed
Now you can own your own infrastructure. You can amortize those investments and … reduce your opex, and therefore be more profitable as a company.
Manuel Botija


