The Pulse
Turba Labs Announces $52 Million Funding Led by Creandum and Cusp Capital
Turba Labs announced $52 million in funding led by Creandum and Cusp Capital, without naming a financing round. The company says its AI performance platform aims to increase useful output from existing infrastructure.
AI.info Team · Written with AI ·
Turba Labs announced $52 million in funding led by Creandum and Cusp Capital. The company’s announcement does not identify a financing round or say whether the amount represents total funding or only this financing.
Funding for a focus on existing hardware
The company is developing an AI performance platform for organizations that operate AI infrastructure. Rather than optimize a model or hardware component in isolation, Turba says it treats workloads and infrastructure as one execution system, linking analytics, a calibrated digital twin and orchestration.
Turba’s founders describe the operating problem as a mismatch between nominal capacity and useful work. Accelerators can wait on memory or communication, while capacity in one pool sits unused as jobs queue elsewhere. High utilization alone, the company argues, does not establish that a system is delivering useful output at the required cost, speed and reliability.
The founders, Dr. Patrick Jahnke and Dr. Hans-Juergen Schmidtke, write:
“The system is the unit of optimization.”
They say system performance depends on the critical path across compute, memory and network, not just local gains in one component.
Software that predicts before it changes workloads
Turba says its platform first uses predictive scenario planning to estimate how a new model, more users or different hardware could affect performance, reliability and cost. In live operation, it can steer routing, admission, replicas, placement, batching and parallelism as conditions change. The company says operators also receive an explanation of the binding constraint, the expected improvement and what observation could show a forecast was wrong.
That approach is meant to account for how a workload’s requirements shift with precision, context length, batch composition and concurrency. A model fitting in device memory does not by itself establish that a service will meet its performance targets at the intended load, the founders write. Turba says its system factors in those interactions before and during deployment.
A 9.4-fold result—with a modeling caveat
Turba cites a modeled mixed-inference example in which blended cost per token fell by a factor of 9.4, or about 89%, while service objectives were maintained. The example spans five workload profiles and six optimization measures, according to the company. It is a modeled result for that stated scenario, not a reported outcome from a customer deployment or a promise of the same savings across other systems.
The distinction matters: Turba says it will separate modeled results from measured deployment outcomes and compare its platform with a well-tuned baseline on representative workloads, including bursts. The announcement does not identify a customer or provide independent measurements validating the modeled example.
More than two years of engineering
Jahnke and Schmidtke say they spent more than two years on the research and engineering behind the platform. The announcement does not provide a valuation or name additional investors.
Turba frames its ambition as doubling useful compute without building a new data center. That is the company’s goal, not a claim that it has already doubled compute globally. Its concrete test is whether infrastructure operators can reproduce gains against a strong baseline while meeting their own service requirements.