AI workload optimization startup Callosum raises $100M

Callosum offers a cloud service called Tailored Inference that helps developers increase the efficiency of AI applications. The company says its technology can complete some inference tasks 3.7 times faster than GPT-5.6 Luna while delivering better output quality, and can reduce infrastructure costs.

The service breaks down the multi-step tasks performed by AI applications into standalone software modules called blocks. Each block is then routed to the AI model best suited to run it: simple tasks go to low-cost algorithms, while more difficult work is handled by frontier models. Tailored Inference also optimizes the underlying infrastructure, deploying each model on the chip that can run it most efficiently. The service supports AI accelerators from more than half a dozen vendors at launch.

Cerebras Systems Inc. announced a partnership with Callosum on the same day as the funding round, saying its WSE series of wafer-size inference accelerators will be integrated into Tailored Inference. Cerebras introduced its newest WSE chip, the WSE-3 Turbo, on Wednesday. The chip has the same core count as its predecessor but provides twice the performance through an increased clock rate, and its on-chip network has been upgraded.

Cerebras ships the accelerator as part of a module called the Wafer-Scale Backpack, which includes liquid cooling, power management and other components. Three such modules form the backbone of the CS-4, a rack appliance that customers can use to deploy the silicon in their data centers. According to Cerebras, a single CS-4 can run models with 10 trillion parameters.

The process by which AI models generate responses has two phases: prefill and decode. Cerebras says the CS-4 is particularly well-suited for decode, and can be paired with prefill-optimized chips from Advanced Micro Devices Inc. and Amazon Web Services Inc. Callosum's platform can run customer workloads on both companies' silicon.

Callosum co-founder and CEO Danyal Akarca said in a statement: 'The next-generation of AI will be defined by how intelligently compute is orchestrated, not simply how much compute is available.'