At GTC Berlin, Nvidia funds supplier capacity and unveils factory-built AI data centres

At its GTC Berlin conference on 21 October, Nvidia made a run of announcements, four of which bear on how AI infrastructure gets built and governed: money paid to suppliers ahead of delivery, a factory-built data-centre design, Google Cloud’s Vera Rubin instances, and a reference architecture for containing AI agents. All figures below come from Nvidia’s own releases, which carry the usual forward-looking-statement caveats.

On supply, Nvidia announced capacity agreements with Amkor, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, SPIL and Unimicron. Jensen Huang, Nvidia’s chief executive, said: “Demand is extraordinary. Capacity is the constraint.” The disclosed money is a $1.1bn capacity-reservation advance to KYEC for new US testing facilities, targeted for high-volume production in April 2027, and a “multibillion-dollar” advance to IBIDEN’s Ono plant in Japan under an agreement announced earlier in the year. Other amounts were not given. Nvidia’s targets include 2.5 times its substrate capacity over three years, more than 300% more interposer wafer capacity at Amkor in South Korea by the end of 2028, and 15 times more co-packaged-optics engine capacity at SPIL by mid-2027. These are plans, not delivered capacity.

DSX POD is a reference design for Vera Rubin systems built into modular containers by partners, including Dell, HPE, Lenovo, Schneider Electric, Vertiv and Bull, and validated before delivery. Nvidia says this cuts deployment at a prepared site from six months to as little as 14 days; the figure applies to a site already prepared with power, cooling and networking. Partners expect to deliver in early 2027.

Nvidia also announced with Google Cloud A5X instances built on Vera Rubin NVL72 systems. The headline says general availability, but the text says availability comes “next month”, with Europe to follow. Google Cloud says it plans to scale to as many as 960,000 Rubin GPUs across sites. The first customer named is Ineffable Intelligence, a London company that Nvidia describes as founded by AlphaGo creator David Silver, which is using it for reinforcement learning.

The OpenShell Reference Architecture, available now, extends the September launch of OpenShell with a full stack: Vera Rubin for inference, Vera CPUs for agent harnesses and BlueField-4 chips running Sentry software for out-of-band monitoring. Nvidia says Sentry quarantines an agent that leaves its boundary “in milliseconds”. Anthropic, OpenAI, Meta and Microsoft are among the companies Nvidia lists as working with it; the release does not say each has deployed it.