Two companies moved on the same day to stake claims in what is crystallizing as a distinct product category: the intelligent routing of AI workloads. Callosum, a London startup founded by Cambridge neuroscientists, closed a $100 million seed round — one of Europe's largest ever — to match AI tasks to the right models and chips3,4. Hours later, corporate expense platform Ramp productized an internal tool it has run for three years, launching Router, an API service that lets companies switch between large language models on the fly1. ANALYSIS Taken together, the moves signal that the industry's center of gravity is shifting from who builds the best model to who delivers the right model for a given job — and that a software layer sitting between developers and compute is attracting serious capital and strategic attention.
Why it matters
AI inference costs are becoming the dominant line item as enterprises move from experimentation to production. ◆ A routing layer that can steer queries to the cheapest or fastest model for each task directly attacks that cost structure, and both Callosum and Ramp are betting the market is large enough to support dedicated products rather than ad hoc internal tooling.
The big picture
Callosum raised $100 million in a seed round led by Atomico, with participation from Plural, DCVC, and the UK Sovereign AI Fund2. The UK Sovereign AI Fund's website indicates the Callosum investment was its first, made in April. The round follows a $10.25 million pre-seed led by Plural in February7. Since its founding in 2025 by Danyal Akarca and Jascha Achterberg, Callosum has raised approximately $110 million in total.
Callosum's thesis is rooted in its founders' neuroscience research: just as biological intelligence arises from diverse neurons, AI should not rely on identical chips. The company's software breaks AI tasks into steps and routes each to the most suitable model and chip. It offers a cloud service called Tailored Inference and has announced a flagship partnership with Cerebras focused on low-latency inference at scale. It also has partnerships with chipmaker Rebellions and other infrastructure providers.
"As AI moves from training to inference, the question is no longer just who builds the best models. It's how we deliver intelligence efficiently, across an increasingly diverse landscape of models and chips," said Niklas Zennström, CEO of Atomico. "They're building the software layer that intelligently routes AI workloads to the right compute, making AI faster, more efficient and more accessible without adding complexity."
Ramp's Router, meanwhile, operates at a different layer. It offers access to models from OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI, and Z.ai through a single API. Users can set routing strategies — for example, preferring flex-usage pricing tiers, selecting models based on up to three user-specified benchmarks, or routing only difficult problems to expensive models. The service is free for the remainder of 2026, though users still pay model inference costs, and it launches with a $26 credit offer. Ramp says it has used the router internally for three years5.
ANALYSIS Ramp is following Stripe's playbook of building financial infrastructure around AI usage — TechCrunch described both companies as "setting up toll houses for AI inference" — while Callosum is going deeper into the compute stack, routing not just across models but across chip architectures.
Between the lines
Callosum is explicitly challenging what it calls the "monoculture" assumption that superintelligence will come from a single model running on identical chips. UK AI Minister Kanishka Narayan endorsed the approach, saying that integrating Cerebras into Callosum's platform makes "ultra-low-latency inference available exactly where it creates the greatest impact". ◆ The UK government's involvement — through both the Sovereign AI Fund's first investment and Callosum's inclusion in the government's £1.1 billion AI hardware plan — suggests London views heterogeneous compute routing as a strategic capability, not merely a startup bet.
Ramp's entry from a different direction is equally telling. A fintech company productizing an internal AI routing tool signals that model selection has become operationally complex enough to warrant dedicated infrastructure even inside companies whose core business is not AI. The dashboard exposing token spend, cost, latency, and fallback attempts frames routing as a financial-management problem — natural territory for an expense platform.
The two products occupy different positions on the stack: Ramp routes across models via API; Callosum routes across models and chips. But both rest on the same premise — that a multi-model, multi-hardware world needs an orchestration layer, and that layer is a product, not a feature.
What's next
Ramp has not disclosed Router's pricing for 2027. Callosum said its Tailored Inference product is already delivering performance and cost improvements in cybersecurity and finance. ANALYSIS The competitive question is whether routing remains a standalone category or gets absorbed into cloud platforms and model providers. For now, the $100 million seed and a fintech's same-day product launch suggest the market is voting for standalone.