Turnkey AI factory emerges

SUSE and Nvidia on April 21 introduced a packaged “AI factory” that lets companies build and run AI systems inside their own controlled infrastructure. (suse.com) SUSE said the product combines SUSE AI with Nvidia AI Enterprise and is aimed at moving models from local development into production across the edge, data center and public cloud. The company said the stack includes pre-validated blueprints and GitOps-based management so customers can deploy and govern systems in a more standardized way. (suse.com) The New Stack reported the pitch is “sovereign” deployment: organizations keep data, models and operations under their own jurisdiction and policy controls instead of handing everything to a public cloud provider. SUSE unveiled the offering at SUSECON 2026 in Prague on April 21. (thenewstack.io) An AI factory is the industry’s shorthand for a repeatable production line for artificial intelligence: chips supply the computing power, software schedules jobs, and management tools handle security, updates and policy. What vendors are selling now is less a single machine than a full operating environment for building, tuning and serving models at scale. (thenewstack.io) That framing is showing up beyond SUSE. Google on April 22 introduced two separate eighth-generation Tensor Processing Units, TPU 8t for training models and TPU 8i for inference, the stage where a trained model answers prompts or processes live requests. (cloud.google.com) Google said the split reflects a change in workloads: pre-training, post-training and real-time serving now have different infrastructure needs. CNBC reported both chips are scheduled to become available later in 2026, underscoring how chip design is being carved up around inference as a distinct business. (cloud.google.com, cnbc.com) Nvidia is also pushing lower in the stack, into the power systems that feed AI data centers. Asiae reported on April 22 that South Korean power-equipment makers were preparing responses to Nvidia’s roadmap for data centers based on 800-volt direct current infrastructure. (asiae.co.kr) TradingPedia, citing that Korean report, said the proposed 800-volt direct-current design is intended to cut current, copper use and cable bulk compared with today’s 54-volt systems. The report also said compatibility with existing equipment remains a central hurdle. (tradingpedia.com) The common thread is control over the full chain: software stack, chips, and even the electrical architecture underneath the racks. In that model, the “factory” is not just where AI is trained; it is the infrastructure for serving models continuously under a company’s own rules. (suse.com, cloud.google.com, asiae.co.kr)