NVIDIA has unveiled a factory-built modular system for AI data centres, aiming to cut deployment time from several months to as little as 14 days as demand for computing capacity surges.

Called DSX POD, the system integrates computing racks, networking, cabling, power distribution and cooling into factory-tested units shipped in transportable containers. Deliveries are scheduled to begin worldwide in early 2027.

“An AI factory creates value when it starts producing tokens, not when construction is complete,” NVIDIA founder and chief executive Jensen Huang said at the company’s GTC event in Berlin on Wednesday.

Tokens are units of text or other data processed and generated by AI models. For data centre operators, faster deployment means an earlier start to commercial operations and revenue generation.

Mr. Huang highlighted the financial implications of delays, saying a data centre expected to generate $40 billion in annual revenue would lose the equivalent of $10 billion if delayed by three months. The figures were illustrative estimates presented by him, not losses reported at a specific facility.

The DSX POD approach shifts much of the installation, integration and testing work from customer sites to factories, reducing the complexity of on-site deployment.

Built around NVIDIA’s Vera Rubin computing platform, the reference design includes seven liquid-cooled computing racks connected by 17 kilometres of Spectrum-X networking cables. The systems are fully cabled, integrated and tested before being shipped to customers.

“Once site utilities are ready, estimated on-site deployment goes from several months to days,” Mr. Huang said.

The 14-day timeline is conditional on site utilities being ready, meaning customers must still complete essential site preparation and infrastructure work before deployment.

NVIDIA said the factory-built approach would help operators bring computing capacity online faster, reduce the time equipment remains idle and limit capital tied up in infrastructure awaiting commissioning.