Callosum 融资1亿美元以在模型和芯片之间路由AI任务

Why it matters

Callosum is betting the valuable layer in AI will be software that selects the cheapest workable model-chip pairing, giving it upside across hardware winners.

Callosum，这家由 Danyal Akarca 和 Jascha Achterberg 创立的伦敦 AI 基础设施初创公司，据 Bloomberg 在 8 月 20 日报道，已完成由 Atomico、Plural 和 DCVC 领投的 1 亿美元种子轮融资。Bloomberg 称英国的 Sovereign AI Fund 也参与了此次融资。政府支持的该基金进行了 Callosum 所称的“重大投资”。Callosum 未披露估值。

Bloomberg 报道的这轮融资此前，Callosum 已披露过一笔早期的 $10.25 million pre-seed。正如 Achterberg 在 Callosum 露面时所解释的那样，创始人基于他们在 Cambridge 的神经科学与计算研究提出了异构 AI 的理论。他们认为 AI 系统应当结合专用模型和计算子层，而不是将每项任务都送入单一的通用模型。Callosum 将这一理念应用到基础设施层面，按照成本、速度和能力，将工作流的不同部分分配给不同的模型、处理器和云实例。

创始人正在把这一观点带入一个主要由同质化 Nvidia GPU 群集构建的市场。Callosum 希望成为位于这些硬件之上的路由与编排层，为企业提供一种在不围绕每个供应商重构应用的前提下，将现有加速器与新芯片结合的方式。Callosum 表示其面向部署多模型 AI 工作流的组织，以及寻求生产负载的硬件公司在构建产品。

Two neuroscientists take on the AI stack

Akarca trained as a medical doctor at the University of Southampton before completing a PhD in computational neuroscience at Cambridge. His doctoral work examined how brain networks develop under physical and metabolic constraints, followed by research at Cambridge's MRC Cognition and Brain Sciences Unit and Imperial College London.

Achterberg completed his Cambridge PhD under neuroscientist John Duncan and Google DeepMind researcher Matthew Botvinick, with research collaborations involving DeepMind and Intel Labs. He later became a research fellow at St John's College, Oxford, studying how specialized neural circuits coordinate to produce flexible cognition.

The pair's academic backgrounds are central to Callosum's architecture. Their work treated cognition as a systems problem involving specialized modules, communication constraints and distinct computational jobs. In a company technical post, Callosum said a "seed model" sets a plan, sub-models handle tasks such as expansion, retrieval and verification, and the software dispatches each operation to hardware suited to its compute profile.

Achterberg wrote when Callosum emerged from stealth in February that intelligence depends on "the diversity of co-optimised mechanisms working together". The founders have carried that view into a product strategy that treats models and computing substrates as components software should coordinate for customers.

Callosum has not publicly disclosed verified headcount, customer count, revenue or commercial deployment figures.

The reported $100 million bet on the routing layer

Callosum previously disclosed a $10.25 million pre-seed led by Plural, with participation from ARIA, the UK's Sovereign AI Fund and unnamed angel investors. The UK government described Callosum as the Sovereign AI initiative's first equity investment. The relationship between that earlier capital and Bloomberg's reported $100 million seed remains unclear, so Callosum's total financing cannot be stated cleanly from the available disclosures.

Callosum describes its product as software for coordinating heterogeneous models, chips and workflows. Callosum says the system assigns different parts of a workload according to the cost, speed and capability of the available components.

That position could benefit Callosum across several hardware suppliers. New accelerators need compatible software and workloads before buyers will deploy them. Enterprises, meanwhile, want lower inference bills and less dependence on a single hardware provider. Callosum is trying to serve both groups by making less-established processors usable inside production AI systems.

The UK's Sovereign AI Fund is described by the government as a fund with $677 million offering equity investment, public-compute access, procurement pathways and fast-track talent visas to help British AI companies start, scale and remain anchored in the UK. Its backing puts Callosum's orchestration software inside the government's effort to build domestic influence over AI infrastructure.

Callosum said in February that it was partnering with Normal Computing, Mixx, Cortical Labs and Great Sky on heterogeneous-compute projects.

Callosum has also worked with CommonAI on a co-located heterogeneous computing project supported by a $2.9 million ARIA grant. The projects are research-heavy, but they define the technical scope of the founders' wager: AI infrastructure will fragment beyond today's mix of GPUs and mainstream cloud accelerators.

The benchmarks now need to become a business

In a company technical post, Callosum claimed that early demonstrations produced up to 12 times lower cost and 5.5 times faster performance on certain deep-context workloads. It also reported higher quality at lower cost than a single-model baseline on a partner's GitHub activity summarization task. Those figures came from Callosum's own tests, and the results vary by model, hardware pairing and workload.

Callosum's commercial test is whether task-level routing can preserve those savings under production requirements such as reliability, observability and predictable pricing. Routing work among several providers creates networking and failure-management problems of its own. Customers also need clear responsibility when one component in a multi-model workflow fails. Callosum has not established a public pricing model or disclosed customers, revenue or the scale of any commercial deployments.

The founders are entering a heavily financed inference market. Baseten raised $300 million at a $5 billion valuation in February. In a July announcement, Fireworks AI said it raised a $1.505 billion Series D at a $17.5 billion valuation. Tensormesh, which emphasizes caching and reuse of previously processed context, raised $20 million in May. Infinity, which adapts inference workloads to new chip architectures, raised $15 million in July.

Callosum's narrower distinction is its attempt to optimize individual tasks across several models, cloud providers and types of hardware, including processors that have yet to achieve broad adoption. Callosum's reported $100 million financing would give Akarca and Achterberg capital to test whether task-level routing can lower inference costs across multiple models, cloud providers and accelerator types. Their wager depends on continued fragmentation in the AI stack. Concentration around a handful of model and hardware providers would leave Callosum with fewer differences to exploit.