Skip to main content

Callosum

Category: AI Infrastructure

System-level software platform that orchestrates AI workloads across heterogeneous chips and models to enable 'Heterogeneous Intelligence' — diverse AI models running on different hardware types working together as an integrated, co-evolved system. Callosum was founded in 2024. The company is led by Danyal Akarca. Based in London, United Kingdom. Team size: N/A. Total funding raised: $10.25M. Latest round: Pre-Seed. Key investors include ["Plural (Lead investor - European VC firm)","ARIA (Advanced Research and Invention Agency - UK government R&D funding)","Charlie Songhurst (Angel investor)","Stan Boland (Angel investor - FiveAI)","John Lazar (Angel investor - Royal Academy of Engineering)"].

Founded
2024
Headquarters
London, United Kingdom
Team size
N/A
Total funding
$10.25M

Value proposition

Delivers 2x accuracy, 7x faster performance, and 4x lower cost compared to homogeneous single-chip solutions by exploiting diversity across models and hardware. Achieves up to 12x cost reduction on specific workflows while setting new SOTA on benchmarks using only open-source models.

Products and solutions

["Heterogeneous Intelligence Platform (core orchestration software)","Heterogeneous Recursion Engine (deep context reasoning across mixed models/chips)","Heterogeneous Vision-Language-Action System (web automation and multi-modal agents)","Topology-aware Cache Management System (optimal eviction and pre-fetching)","On-die Grammar Enforcement Kernels (custom silicon optimization for AWS Inferentia2)"]

Unique value

First company to co-evolve heterogeneous chips and intelligence together, inspired by brain architecture principles — the human brain achieves intelligence through diverse specialized cell types working together, not by copying one neuron billions of times. Vertically integrated 'Intelligent System' where workflows are hardware-aware, models are task-graph-aware, and kernels are output-constraint-aware, with every layer co-optimized in context of the whole.

Target customer

Companies building multi-agent AI systems requiring superior performance across complex workflows; emerging chip manufacturers seeking to demonstrate hardware capabilities at scale; enterprises with heterogeneous AI workloads (automation, inference, complex reasoning)

Industries served

["AI Infrastructure","Enterprise AI & Automation","Cloud Computing","AI Chip Manufacturing","Data Centers","Inference-Optimized Compute"]

Technology advantage

Combines neuroscience-inspired heterogeneous architecture with full-stack optimization (from kernels to workflows). Achieves: 2x accuracy/7x speed/4x cost improvements on complex workflows; 25% improvement over ICLR 2026 SOTA on Visual WebArena using zero frontier API calls; up to 12x cheaper workflows (SambaNova GPT-OSS-120B vs GPT-5); 2.4x speedup via topology-aware caching; 1,767x faster grammar enforcement at batch-64 through on-die masking; multi-cloud/multi-chip support (AWS, GCP, Azure across Nvidia, AMD, Cerebras, SambaNova, Trainium/Inferentia).

How they differentiate

First company to co-evolve heterogeneous chips and intelligence together with brain-inspired architecture. Achieves 2x accuracy, 7x faster performance, and 4x lower cost compared to homogeneous single-chip solutions by orchestrating AI workloads across diverse hardware (Nvidia, AMD, AWS Trainium/Inferentia, Cerebras, SambaNova) with neuroscience-inspired heterogeneous computing principles.

Main competitors

["Ray/Anyscale","Nvidia","Cloud providers with custom chips (Google TPU, AWS Trainium/Inferentia)"]

Key partnerships

["AWS (custom NKI kernels for Inferentia2 silicon, on-die grammar enforcement)","Cerebras (heterogeneous recursion benchmarking and deployment)","SambaNova (heterogeneous recursion configurations)","Coworker AI (enterprise production partner for GitHub workflow benchmarks)","ARIA/UK Government (£50M Scaling Inference Lab access for chip validation)","Google DeepMind (founder research collaboration history)","Photonics/interconnect companies (next-gen data center connectivity R&D)","Academic: University of Cambridge, Oxford, MIT, ETH Zurich, Imperial College London (founder backgrounds and ongoing research ties)"]

Notable customers

["Coworker AI (enterprise production partner for GitHub workflow benchmarks)"]

Major milestones

["Company founded in November 2024 (UK Companies House registration #16076949)","Emerged from stealth mode in February 2026","Raised $10.25M pre-seed round led by Plural with support from ARIA","Published first technology results demonstrating heterogeneous recursion capabilities","Developed custom NKI kernels for AWS Inferentia2 silicon","Achieved benchmark results beating single-call Claude Opus 4.5 at every context length"]

Growth metrics

Recently emerged from stealth mode in February 2026 with $10.25M pre-seed funding; demonstrated 25% improvement over ICLR 2026 SOTA on Visual WebArena using zero frontier API calls; achieved up to 12x cost reduction on specific workflows

Market positioning

Early-stage UK-based AI infrastructure startup challenging Nvidia's dominance by enabling multi-model, multi-chip AI workloads. Targets two customer segments: companies building multi-agent AI systems requiring superior performance, and emerging chip manufacturers seeking to demonstrate hardware capabilities at scale.

Geographic focus

United Kingdom (London), Europe, with global reach through multi-cloud platform support (AWS, Google Cloud, Microsoft Azure)

Patents and IP

No registered patents publicly disclosed as of February 2026. Core IP resides in proprietary heterogeneous orchestration algorithms, topology-aware runtime, and custom silicon kernels.

About Danyal Akarca

Cambridge-trained computational neuroscientist and medical doctor. PhD candidate at University of Cambridge's MRC Cognition and Brain Sciences Unit, researching generative models of brain networks. Previously Research Fellow at Medicalchain. Co-author of research published in Nature Machine Intelligence on brain-inspired AI systems.

Official website: