AI workload optimization startup Callosum raises $100M

Artificial intelligence company Callosum has secured a significant investment of $100 million to enhance its cloud-based Tailored Inference service. The funding round was led by Atomico, with additional participation from Plural, DCVC, and the U.K. Sovereign AI Fund. This is a follow-up to the company's previous $10.25 million funding round in February.

Tailored Inference is a cloud service designed to optimize AI application performance by efficiently allocating computational tasks to the most suitable models. In tests, Callosum reported that its technology can perform certain inference tasks up to 3.7 times faster than GPT-5.6 Luna while maintaining superior output quality. By breaking down complex tasks into smaller, standalone modules called blocks, Tailored Inference can route each block to the AI model best equipped to handle it.

Simple tasks are directed to less resource-intensive algorithms, while more challenging work is processed by frontier models.

The service also aims to reduce infrastructure costs by intelligently deploying AI models on the most efficient hardware. Upon determining the optimal model for each inference task, Tailored Inference deploys the appropriate model on the chip that can run it most effectively. At launch, the platform supports AI accelerators from over half a dozen companies, including Cerebras Systems Inc.

Cerebras Systems will integrate its WSE series of wafer-size inference accelerators, with the latest WSE-3 Turbo model featuring twice the performance of its predecessor thanks to an increased clock rate. Cerebras' engineers have also upgraded the on-chip network that connects the cores, further enhancing the accelerator's capabilities. The accelerator is delivered as part of a Wafer-Scale Backpack, which includes liquid cooling hardware, power management components, and other auxiliary parts.

Three of these backpacks form the backbone of the CS-4, a rack appliance that Cerebras customers can use to deploy its silicon in their data centers. One CS-4 can handle models with 10 trillion parameters. The prefill and decode phases of the AI model generation process, which comprise the steps through which AI models produce prompt responses, are particularly well-suited for the CS-4's performance capabilities.

Callosum's co-founder and CEO, Danyal Akarca, emphasized the importance of intelligently orchestrating compute resources rather than simply increasing computational power. The company's platform is capable of running customer workloads on the chipsets from Advanced Micro Devices Inc. and Amazon Web Services Inc., showcasing the service's versatility and potential for widespread adoption in the AI landscape.