London-based Callosum has raised a $100m seed round to build software that distributes artificial-intelligence workloads across different models and processors, as AI economics begin to depend as much on using computing capacity efficiently as obtaining more of it.

Atomico led the financing, with participation from Plural, DCVC, the UK Sovereign AI Fund, and other investors. Callosum is also the first company backed by the government’s £500m Sovereign AI Fund, established to invest in strategically important British AI businesses and infrastructure.

The company is developing what it calls heterogeneous intelligence: instead of assuming that one model running on one type of processor should handle an entire workload, its software separates work into constituent tasks and directs them towards combinations of models and hardware selected around latency, cost, energy consumption, and quality.

Callosum has also announced partnerships with chip and infrastructure companies including Cerebras and South Korea’s Rebellions. The proposition rests on the expectation that AI computing will become more diverse rather than consolidating permanently around one model family, accelerator architecture, or cloud provider.

The AI stack becomes less uniform

Much of the current infrastructure build-out has been organised around obtaining enough high-end accelerators to train and serve increasingly capable models. Shortages encouraged companies and governments to treat computing capacity almost as a single commodity even though workloads differ considerably in what they require.

A complex reasoning task may justify an expensive frontier model and high-performance accelerator, whereas classification, retrieval, summarisation, or a narrowly specialised agent may run economically on a smaller model or different processor.

Sending all of those jobs through the same stack can simplify engineering initially, but it can also mean paying premium inference costs where they produce little additional value. Callosum is trying to turn that mismatch into a software optimisation problem.

Its platform decomposes workloads and orchestrates individual tasks across heterogeneous infrastructure rather than requiring developers to optimise separately for each architecture. The intended result is a better balance between performance, cost, and energy consumption.

The approach resembles earlier layers of abstraction in computing, where applications became able to use increasingly diverse underlying infrastructure without users managing every component directly. AI makes that problem harder because both models and processors are changing quickly, while performance depends heavily on the task being executed.

Callosum’s partnership with Cerebras illustrates the concept. Its low-latency inference systems can sit alongside other architectures inside the same orchestration environment, while Rebellions brings another specialist AI processor rather than another version of the dominant accelerator stack.

Sovereignty without manufacturing every chip

The Sovereign AI Fund made Callosum its first investment earlier in 2026, describing the company as infrastructure for an increasingly heterogeneous computing market. The decision offers a relatively pragmatic interpretation of technological sovereignty.

Britain is unlikely to recreate the entire advanced semiconductor supply chain domestically, while leading AI services already depend on processors, manufacturing equipment, cloud infrastructure, memory, networking, and software supplied across several countries.

Orchestration provides a different form of resilience if it genuinely allows workloads to move between processors and models. A system designed around multiple suppliers can be less exposed to a shortage, export restriction, price increase, or capacity limit affecting one architecture.

The extent of that portability will depend on how applications and models are engineered. Software built around proprietary acceleration libraries or one provider’s unique capabilities can still be difficult to move even when an orchestration platform presents the underlying hardware as interchangeable.

The same technology could influence the economics of publicly funded computing. National infrastructure needs high utilisation if expensive accelerators are to generate enough research and commercial value, while software capable of matching workloads to mixed hardware may help operators use diverse estates more efficiently.

Orchestration layers can also create dependencies of their own. If organisations rely heavily on proprietary scheduling and optimisation, switching the control layer may eventually become difficult even where the models and processors underneath remain portable.

Compute efficiency moves up the investment agenda

The size of the financing reflects how important optimisation has become. A $100m seed round is unusual for a systems-software company, but AI infrastructure businesses are being financed against a market in which customers are already committing billions of dollars to processors and data centres.

That capital intensity creates a powerful incentive to extract more useful work from every installed machine. Moving a workload onto cheaper hardware without materially reducing performance can produce substantial savings once an inference operation is repeated millions of times.

Electricity adds another constraint because different architectures can have markedly different power requirements for particular workloads. Data-centre operators facing grid limits may eventually have to treat energy consumption as a scheduling input rather than merely an operating expense.

Callosum is entering the market as computing diversity expands alongside operational complexity. New processors do not create meaningful competition if customers cannot move software onto them without extensive engineering work.

The $100m round gives the company the resources to test whether an independent layer can make that diversity practical. Its value will depend on whether customers actually want software choosing between models and chips, and whether the resulting optimisation saves more than the additional infrastructure complexity costs.