On August 20, 2026, Bloomberg reported that London-based AI startup Callosum raised $100 million in seed financing, one of the largest seed rounds in European history, to build what it calls a heterogeneous intelligence routing layer: software that breaks each AI workload into its component tasks and sends every one to the model and chip best suited to the job.
The round was led by Atomico, with participation from Plural, DCVC, and a significant investment from the UK’s £500 million Sovereign AI Fund, making this the fund’s second commitment to Callosum after its initial equity investment in April 2026.
The implications for enterprise AI teams are direct. Most agentic workloads today funnel every step through the same general-purpose frontier model on the same Nvidia GPU cluster. Callosum is betting that the next infrastructure layer to matter is the one that decides, at the task level, which model and which chip should handle each unit of work.
What Callosum Builds and Why It Matters Now
Callosum’s core product is an orchestration layer, not a model and not a chip. Its software distributes AI workload across mixed hardware, routing each part of a task to the processor and model that handle it best instead of assuming a homogeneous grid of GPUs.
The company’s founding thesis comes directly from neuroscience. Co-founders Danyal Akarca and Jascha Achterberg met at Cambridge, where they researched the corpus callosum, the bundle of nerve fibers connecting the brain’s two hemispheres. Their argument: intelligence emerges from separate specialized systems coordinating, not from one very large system scaling further. That framing maps onto AI infrastructure in a specific way. Real-world enterprise problems are heterogeneous. A single agentic workflow might include document extraction (suited to a fast, cheap model), multi-step reasoning (suited to a capable frontier model), and a terminal task (suited to a low-latency inference chip). Routing all three steps to the same endpoint wastes money and time.
Callosum’s software handles that routing decision automatically.
Performance Numbers from the Cerebras Partnership
At launch, Callosum’s most concrete production data comes from its partnership with US chip company Cerebras, whose wafer-scale inference systems deliver high-throughput, low-latency compute at economics that differ substantially from Nvidia GPU clusters.
On complex agentic workloads in financial services delivered through that partnership, Callosum reported:
| Metric | Callosum vs. single frontier model on conventional GPU |
|---|---|
| Inference speed | 4x faster |
| Compute cost | 70% lower |
| Task success rate | 10% higher |
For autonomous computer-use tasks specifically, the company reported two times the accuracy, seven times the speed, and one quarter of the cost versus a conventional single-GPU setup.
These numbers come from Callosum’s own reported demonstrations. Independent third-party benchmarks are not yet published. But the directionality matches a pattern visible across the infrastructure layer this year: the inference cost crisis in enterprise AI is real, and heterogeneous compute is one credible structural answer to it.
Andrew Feldman, CEO of Cerebras, called out the opportunity plainly: “By integrating Cerebras into Callosum’s platform, we are making ultra-low-latency inference available exactly where it creates the greatest impact, enabling customers to build AI systems that simply were not practical before.”
Why the Routing Layer Is Becoming Strategic
The fundamental economics of enterprise AI infrastructure have shifted. As Gartner forecast in August 2026, global inference spending will reach $23.3 billion in 2026, surpassing training at $19 billion for the first time. Inference is projected to represent 55% of AI-optimized infrastructure spending in 2026, rising to 59% in 2027, driven by the scale-up of agentic workloads.
Agentic workflows amplify the cost problem. A single user query to an AI agent might drive six to twelve inference calls across document reading, tool invocation, verification, and output formatting steps. At scale, those calls are the dominant operating cost. The IBM and Together AI $240 million deal and NVIDIA’s Nemotron Switchyard routing system both addressed parts of this problem. Callosum’s approach addresses it at the task-routing level rather than the model or infrastructure level.
The AI chip market is also fragmenting in ways that make routing infrastructure more valuable, not less. Callosum’s current hardware partner network includes Cerebras, Rebellions (the Korean inference chip company), Axelera, d-Matrix, Lumai, and Tendrils, alongside infrastructure partners Supermicro and HPE. Each new chip that enters production adds to the value of a routing layer that abstracts over all of them. Without routing software, enterprises have to rebuild applications every time they want to try a new accelerator.
This is the positioning Callosum is claiming: not a bet on any single chip winning, but a bet on the routing layer winning regardless of which chips win.
The UK Government’s Bet
The UK Sovereign AI Fund’s involvement deserves separate attention. The £500 million fund, launched with backing from technology secretary Liz Kendall and chancellor Rachel Reeves, and chaired by James Wise, made Callosum its first equity investment in April 2026. Its participation in this larger seed round confirms and scales that commitment.
For European AI infrastructure, the investment signal is meaningful. The UK government is betting that the valuable layer in the AI stack is not the model and not the chip. It is the software that connects them efficiently. That framing, if correct, implies that infrastructure advantage in AI does not require winning the model race or the chip race: it requires owning the coordination layer above both.
The round’s size, one of the largest seed rounds ever in Europe according to Tech.eu’s reporting, also signals that capital is flowing toward AI infrastructure in Europe at a scale that was not visible twelve months ago.
What Enterprise AI Builders Should Watch
Callosum is still in early production. The company has not published a public pricing model, has not disclosed customer names or revenue, and the performance claims are from its own reported demonstrations rather than third-party benchmarks. The practical challenges of routing production workloads across multiple providers (network reliability, failure attribution, cost predictability) are real engineering problems that seed-stage funding does not automatically solve.
But the directional thesis is sound. Enterprise AI infrastructure teams are already navigating decisions that Callosum’s platform addresses:
- How do we reduce inference costs without sacrificing agent task completion rates?
- How do we incorporate specialized chips (Cerebras, Rebellions, AMD) without rewriting our agent architecture?
- How do we add redundancy and cost optimization across cloud providers without managing three separate integration layers?
- How do we make our agentic AI systems more reliable in production without simply throwing a more expensive frontier model at the problem?
The answer the industry is converging on is not a single better model. It is better routing. NVIDIA’s Switchyard architecture uses difficulty-based routing to escalate agent steps to more capable models only when complexity demands it. Callosum’s approach extends that idea across chip types and cloud providers: not just “which model,” but “which model on which hardware.”
The Shift in Enterprise AI Infrastructure Strategy
There is a structural change visible in how serious enterprise AI teams are approaching infrastructure in 2026. The first wave of enterprise AI deployments assumed the answer was always the best available frontier model. The second wave, now underway, is discovering that deploying frontier models for every step of every agent workflow is prohibitively expensive, slower than necessary, and often less reliable than a well-chosen mid-tier model.
Writer’s Palmyra X6 showed that a purpose-built model can undercut frontier token costs while maintaining quality for enterprise GTM tasks. Solar Pro 4 from Upstage made the case for reliable, cheaper document agents over frontier overkill. Now Callosum is building the layer that lets enterprises mix all of these options dynamically, task by task.
For enterprise leaders and GTM teams thinking about how to build AI systems that scale economically, the routing layer is no longer an optimization. It is becoming a strategic infrastructure decision. The companies that build it well will have lower per-agent operating costs, better task completion rates, and the flexibility to adopt new hardware as the chip landscape evolves.
If your AI agent stack is pushing every step through a single frontier model on a homogeneous GPU cluster, that is an architecture worth revisiting. The infrastructure decisions enterprises make in the next six to twelve months will define their AI cost structures for years.
To explore how to structure your enterprise AI infrastructure and agentic strategy, talk to the Enera team.
Sources
- Bloomberg: AI Startup Callosum Raises $100 Million to Make AI Tasks Cheaper (August 20, 2026)
- Atomico: Our investment in Callosum
- Tech.eu: Callosum raises $100M seed round (August 20, 2026)
- The Next Web: Callosum raises $100m to make AI workloads cheaper (August 20, 2026)
- RuntimeWire: Callosum raises $100M to route AI tasks across models and chips (August 20, 2026)