Scientific-model startup ZenteiQ increased batch-inference throughput 4.07 times, reaching 5.97 billion daily tokens per 1,000 accelerators using Google Kubernetes Engine’s Inference Gateway.

ZenteiQ is a startup spun out of the Indian Institute of Science and selected under the IndiaAI Mission to build sovereign scientific foundation models. Its models address physical laws, governing equations, and numerical solvers across fields including thermodynamics, computational fluid dynamics, and electromagnetics.

The Google Cloud case study reports that GKE and its Inference Gateway increased aggregate inference throughput to 4.07 times the previous level. Daily generated output rose from 1.47 billion to 5.97 billion tokens per 1,000 accelerators; this is an accelerator-normalized batch-generation metric, not a GPU inventory or an interactive-response benchmark.

ZenteiQ’s independently designed pipeline separates prompt preparation on worker pods from vLLM model serving across multinode accelerator pools. Its four stages expand scientific seed concepts into prompts, generate candidate reasoning traces, score responses with evaluator models, and apply guardrail checks using another model family before packaging verified training data.

The startup says its fixed compute allocation creates a trade-off between synthetic-data generation and pretraining, with reserved multihost training clusters waiting when datasets fall behind. Its original Kubernetes ClusterIP round-robin routing encountered bottlenecks as it scaled toward 9-billion- and 35-billion-parameter training schedules, including uneven inference-engine queues caused by prompts of varying complexity and response length.