Callosum
Infrastructure for routing AI workloads across heterogeneous models and chip architectures to optimize performance and cost.
Website: https://www.callosum.com/
Cover Block
From the public record
| Field | Value |
|---|---|
| Name | Callosum |
| Tagline | Infrastructure for routing AI workloads across heterogeneous models and chip architectures to optimize performance and cost [Callosum] |
| Headquarters | London, UK [Callosum] |
| Founded | 2025 [Tech.eu, February 2026] |
| Stage | Seed [Tech.eu, February 2026] |
| Business model | API / Developer Platform |
| Industry | Deeptech |
| Technology | AI / Machine Learning |
| Geography | Western Europe |
| Growth profile | Venture Scale |
| Founding team | Co-Founders (2), Danyal Akarca and Jascha Achterberg [Fortune, February 2026] |
| Funding label | $100M+ total disclosed [Tech.eu, February 2026] [SiliconANGLE, August 2026] |
| Total disclosed | $110,250,000 [Tech.eu, February 2026] [SiliconANGLE, August 2026] |
Links
From the public record
- Website: https://www.callosum.com/
- Product Blog: https://www.callosum.com/blog/welcome-heterogeneous-intelligence
- Team: https://www.callosum.com/team
- Jobs: https://jobs.ashbyhq.com/callosum/c342204a-6bb8-47ae-a926-049a50e950ae
The Short Version
PUBLIC Callosum builds infrastructure for routing AI workloads across multiple models and chip architectures, and it merits investor attention now because it has moved from stealth to more than $110 million in disclosed seed financing within roughly six months, a pace that suggests unusually strong conviction around the heterogeneous-compute thesis [Tech.eu, February 2026] [SiliconANGLE, August 2026]. Founded in 2025 and based in London, the company was started by Danyal Akarca and Jascha Achterberg, two Cambridge-trained computational neuroscientists who reportedly met during their doctoral studies and are positioning the company around a simple claim: AI applications should be optimized across the full workflow, not tied to a single model family or chip stack [Fortune, February 2026] [TechTimes, August 2026].
The core product, Tailored Inference, is described as a cloud service and API layer that breaks workflows into smaller tasks and assigns them to the model and hardware configuration best suited to cost, speed, capability, or power requirements, with launch coverage spanning Nvidia and AMD processors, AWS Trainium and Inferentia, and newer architectures such as Cerebras and SambaNova [SiliconANGLE, August 2026] [Fortune, February 2026]. That differentiation is credible at the level of product narrative and partner breadth, particularly because Callosum is also reported to have relationships with Cerebras, Rebellions, Axelera AI, Tendrils, and Supermicro, although the operational depth of those relationships is still unevenly documented in public sources [SiliconANGLE, August 2026] [en.sedaily.com, August 2026] [Quiver Quantitative, August 2026].
The founding team reads as technically serious rather than commercially overextended. Akarca is identified as a medical doctor and computational neuroscience PhD student at Cambridge's MRC Cognition and Brain Sciences Unit, while Achterberg's public record shows a Cambridge PhD completed in collaboration with Google DeepMind and Intel Labs, which maps well to a company trying to sit between model behavior, systems orchestration, and silicon heterogeneity [neuroscience.cam.ac.uk, 2026] [jachterberg.com, 2026].
On capitalization, the company announced a $10.25 million seed led by Plural in February 2026, followed by a $100 million seed led by Atomico in August 2026 with participation from Plural, DCVC, and the UK Sovereign AI Fund, bringing disclosed funding to about $110.25 million; the business model is presented as API and developer infrastructure rather than end-user software [Tech.eu, February 2026] [SiliconANGLE, August 2026]. Public traction signals are promising but still early: Callosum has cited use cases in cybersecurity, finance, and enterprise applications, and ai.engineer reported that Round Treasury reduced mean document-extraction time from 15.9 seconds to 2.1 seconds per document using the company's APIs [eu.vc] [ai.engineer, 2026].
Over the next 12 to 18 months, the key questions are whether Callosum can translate technical claims into repeatable production deployments, show that cross-chip orchestration creates durable developer lock-in rather than a feature advantage, and convert headline funding into a visible ecosystem position among both enterprise AI builders and emerging chip vendors [Fortune, February 2026] [SiliconANGLE, August 2026].
Single-source, plausible -- Section relies on a mix of independent reporting and company-linked materials, with funding and founder background corroborated across multiple sources but product performance and customer breadth only partially verified.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Seed |
| Business Model | API / Developer Platform |
| Industry / Vertical | Deeptech |
| Technology Type | AI / Machine Learning |
| Geography | Western Europe |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (2) |
| Funding | $100M+ total disclosed, approximately $110.25M [Tech.eu, February 2026] [SiliconANGLE, August 2026] |
The Company in Brief
PUBLIC
Callosum presents as a London-based AI infrastructure company formed around a specific systems view: AI workloads should not be tied to one model family or one chip stack when routing across heterogeneous compute may improve speed, cost, and efficiency [Callosum]. The company says it is building "the infrastructure where heterogeneous chips & intelligence co-evolve," and its public materials frame the business around routing workloads across different models and chip architectures rather than around a single proprietary model [Callosum].
The public chronology is short but unusually compressed. Callosum was founded in 2025 and is headquartered in London, according to the structured company record and its public-facing site positioning [Callosum]. By February 2026, the company had emerged from stealth with a $10.25 million seed round led by Plural, and by August 2026 it had announced a further $100 million seed financing led by Atomico, bringing disclosed funding to about $110.25 million across two rounds [Callosum].
The founding team consists of two co-founders, Danyal Akarca and Jascha Achterberg, both identified in public materials as part of the company's leadership bench [Callosum]. The website also describes Callosum as "The Intelligent Systems Company," which is better read as a positioning statement than as a legal descriptor, since no legal entity name is established in the cited company materials used for this section [Callosum].
Unconfirmed -- This section relies primarily on company website materials, with chronology limited to public company disclosures used in the structured record.
What They Have Built
Mixed sourcing
Callosum is making a fairly specific infrastructure claim, not a general AI-platform one. The company says it routes AI workloads across different models and chip architectures, with the stated goal of improving speed, cost, and energy efficiency through heterogeneous compute rather than relying on one model family or one hardware stack [Callosum] [Callosum] [Fortune, February 2026]. In public coverage, its principal product is Tailored Inference, described as a cloud service and API layer that breaks workflows into smaller tasks and assigns each task to the model and chip configuration best suited to cost, speed, capability, or power requirements [SiliconANGLE, August 2026].
The technical wedge appears to be orchestration across both the model layer and the hardware layer. Callosum has said the system is cloud-agnostic and designed to execute across major cloud providers [PUBLIC][Callosum], while third-party reporting says the launch hardware scope included Nvidia and AMD processors, AWS Trainium and Inferentia, and newer architectures such as Cerebras and SambaNova [Fortune, February 2026]. SiliconANGLE also reported support for accelerators from more than half a dozen vendors and said developers can use competing chip suppliers without rewriting existing application code [SiliconANGLE, August 2026]. Taken together, that suggests the product is trying to abstract away infrastructure fragmentation at the application level, though most of the deeper implementation detail still comes from company statements rather than independent technical documentation [Callosum] [SiliconANGLE, August 2026].
Public performance evidence is directionally encouraging but still early and unevenly sourced. In August 2026 coverage, SiliconANGLE reported that Tailored Inference could complete some inference tasks 3.7 times faster than GPT-5.6 Luna, while Plural said Callosum's system delivered 2x accuracy, 7x speed, and 4x lower cost than homogeneous compute for autonomous computer use [SiliconANGLE, August 2026] [Plural]. A named deployment offers a more concrete data point: ai.engineer reported that Round Treasury reduced mean document-extraction time from 15.9 seconds to 2.1 seconds per document using Callosum for invoice and receipt workflows [ai.engineer, 2026]. The same section of the record also includes stronger company-only claims, including average 7.5x faster response and up to 20x faster in complex document-ingestion cases, plus an 11x speedup and 43x cost reduction using an 8B model in one workflow [PUBLIC][Callosum]. Those results may be real, but on the public record they should be read as case-specific benchmarks rather than broadly validated system performance [Callosum] [ai.engineer, 2026].
Inferred, not confirmed -- Material product details are supported by third-party coverage from Fortune and SiliconANGLE, but several core capability and performance claims remain company-sourced or only partially corroborated.
Market Size and Demand
From the public record
This market matters now because model performance is no longer the only constraint in applied AI, compute routing, hardware choice, and inference economics are increasingly part of the product decision for any team deploying AI at scale [Fortune, February 2026] [Tech.eu, February 2026].
The difficulty in sizing Callosum's addressable market is that the company sits across several existing budgets rather than inside one clean software category. No cited third-party market report in the source set provides a direct TAM, SAM, or SOM for heterogeneous AI workload orchestration, so the public record supports a narrower conclusion: Callosum is selling into the intersection of inference infrastructure, model orchestration, and accelerator utilization [SiliconANGLE, August 2026] [Fortune, February 2026]. That matters because buyers appear to include both application developers building multi-agent systems and chip vendors looking for production pathways, which broadens the commercial surface beyond a single developer-tool budget [Fortune, February 2026] [TechTimes, August 2026].
The demand drivers described in the reporting are consistent across sources even where the exact performance claims are not equally verified. Tech.eu and Fortune both frame the company's wedge as a response to dependence on uniform Nvidia-centric stacks, with Callosum arguing that AI workflows should be split across different models and hardware according to speed, cost, capability, or power requirements [Tech.eu, February 2026] [Fortune, February 2026]. SiliconANGLE adds that Tailored Inference is positioned as a cloud service and API layer that lets developers use multiple accelerator vendors without rewriting existing application code, which points to a practical adoption tailwind: buyers want optionality across chips, but they do not want to rebuild their applications to get it [SiliconANGLE, August 2026].
Adjacent markets are at least as important as the core one. On one side sits inference infrastructure, where the economic pressure comes from rising usage and the need to reduce latency and cost per task; on the other sits the emerging market for non-Nvidia accelerators that need software layers capable of abstracting hardware differences for end users [Fortune, February 2026] [SiliconANGLE, August 2026]. Callosum's disclosed hardware scope, including Nvidia, AMD, AWS Trainium, Inferentia, Cerebras, and SambaNova, suggests it is not trying to win by replacing any single provider, but by becoming an orchestration layer across them [Fortune, February 2026]. That creates substitute risk as well, because cloud-native model routing, inference gateways, and accelerator-specific software stacks could each absorb part of the same budget over time [SiliconANGLE, August 2026].
Macro and policy conditions also appear supportive, if still early. The presence of the UK Sovereign AI Fund in the August 2026 round suggests that public-sector interest in domestic AI infrastructure and compute sovereignty is part of the backdrop, not merely venture enthusiasm [SiliconANGLE, August 2026]. At the same time, the broader semiconductor supply picture and the push to diversify away from a single hardware ecosystem make heterogeneous deployment more compelling in theory, especially for enterprises that care about resilience, cost control, and energy use [Fortune, February 2026] [Callosum]. The caveat is that most hard evidence in the public set remains company framing or financing-era reporting, so the market case is stronger as a directional thesis than as a fully quantified category today.
| Cited market lens | What the source supports | Relevance to Callosum |
|---|---|---|
| AI workload orchestration | AI workflows can be split across models and chips to optimize for speed, cost, capability, or power [SiliconANGLE, August 2026] | Describes the category logic behind Tailored Inference |
| Multi-agent AI systems | Reported target buyers include companies building multi-agent AI systems [Fortune, February 2026] | Expands demand beyond infrastructure teams alone |
| Emerging accelerator ecosystem | The product is intended to let developers use competing chip suppliers without rewriting application code [SiliconANGLE, August 2026] [TechTimes, August 2026] | Ties Callosum to growth in alternative AI hardware |
| Compute sovereignty and infrastructure policy | The UK Sovereign AI Fund participated in the $100 million seed round [SiliconANGLE, August 2026] | Suggests policy-aligned demand for domestic or diversified AI infrastructure |
The table shows why this is best read as a convergence market rather than a neat standalone software segment. The opportunity grows if inference spending, multi-model application design, and non-Nvidia accelerator adoption all continue to expand together.
Single-source, plausible -- Section relies on named publisher coverage for market framing, but no third-party market sizing report is present in the source set and several category claims remain partially company-framed.
Who Else Is Fighting for This
MIXED
Callosum is positioning itself one layer above single-model inference tooling, with the bet that orchestration across models and chip types can become a control point that point solutions and hardware-specific platforms do not fully own [SiliconANGLE, August 2026] [Tech.eu, February 2026].
| Company | Positioning | Stage / Funding | Notable Differentiator | Source |
|---|---|---|---|---|
| Callosum | API and cloud service for routing AI workloads across heterogeneous models and chip architectures | Seed, disclosed funding of about $110.25M across two 2026 seed financings | Cross-model and cross-chip orchestration, with launch hardware scope spanning Nvidia, AMD, AWS Trainium and Inferentia, Cerebras, and SambaNova | [Tech.eu, February 2026] [SiliconANGLE, August 2026] [Fortune, February 2026] |
| Inference.ai | Named competitor in AI inference infrastructure | Not established in the provided public sources | Public materials provided here do not establish a product-level differentiator relative to Callosum | [Structured Facts] |
| Baseten | Named competitor in AI inference infrastructure | Not established in the provided public sources | Public materials provided here do not establish a product-level differentiator relative to Callosum | [Structured Facts] |
| FriendliAI | Named competitor in AI inference infrastructure | Not established in the provided public sources | Public materials provided here do not establish a product-level differentiator relative to Callosum | [Structured Facts] |
The table is directional rather than complete. It shows that Callosum is the only company in this set for which the source pack substantiates a specific wedge, namely routing tasks across both models and hardware rather than serving a single deployment surface or accelerator family [SiliconANGLE, August 2026] [Fortune, February 2026].
The competitive map breaks into three segments. First are incumbents in AI compute, especially the Nvidia-centered stack that Callosum was explicitly framed against in early coverage, where the status quo advantage is ecosystem depth, developer familiarity, and procurement inertia rather than workflow-level optimization [Fortune, February 2026] [Tech.eu, February 2026]. Second are inference and deployment challengers such as Inference.ai, Third are adjacent substitutes, including cloud-native hardware options and chip vendors themselves, because Callosum's own pitch is that developers should be able to use competing chip suppliers without rewriting application code, which implies that the substitute is often building directly on a preferred hardware stack or cloud offering instead of adopting an orchestration layer [SiliconANGLE, August 2026].
Where Callosum appears strongest today is at the intersection of talent, capital, and partner alignment. The founders' Cambridge neuroscience backgrounds are unusual for infrastructure, but the more relevant point is that the company has raised roughly $110.25 million in disclosed seed funding within months of emerging from stealth, from backers including Plural, Atomico, DCVC, ARIA, and the UK Sovereign AI Fund, which gives it room to integrate broadly across chips and absorb the cost of working with a fragmented supply base [Tech.eu, February 2026] [SiliconANGLE, August 2026]. The partnership and relationship set also matters: public reporting ties the company to Cerebras, Rebellions, Axelera AI, Tendrils, and Supermicro, and launch coverage said the product spans Nvidia, AMD, AWS Trainium and Inferentia, Cerebras, and SambaNova [SiliconANGLE, August 2026] [en.sedaily.com, August 2026] [Fortune, February 2026]. That edge is real, but probably perishable if larger platforms make heterogeneous routing a standard feature or if chip vendors decide to expose similar orchestration directly to developers.
The main exposure is that Callosum does not appear to own the endpoint customer relationship in the way a full-stack model provider, hyperscaler, or deeply embedded inference platform can. Its reported use cases span cybersecurity, finance, and enterprise applications, and one cited deployment with Round Treasury points to clear workflow gains in document extraction, but the public record still describes more of an enabling layer than a destination platform with obvious lock-in [eu.vc] [ai.engineer, 2026]. That leaves an opening for a named rival such as Baseten, if it can pair deployment simplicity with enough model and hardware flexibility to satisfy teams that do not want another orchestration layer. It also leaves Callosum exposed to the adjacent substitute of direct cloud procurement, particularly for buyers who optimize around vendor concentration, support terms, or existing discounts rather than absolute inference efficiency.
The most plausible 18-month scenario is a sorting market rather than a winner-take-all one. Callosum is the likely winner if heterogeneous inference becomes operationally painful enough that enterprises and chip vendors both want a neutral routing layer, especially as newer accelerators look for workload share alongside Nvidia and AMD [Fortune, February 2026] [SiliconANGLE, August 2026]. The inverse also holds: if buyers continue to prefer vertically integrated stacks from clouds, model vendors, or hardware partners, then Callosum's differentiation compresses fastest because its value sits in coordination, which is useful but easier for larger platforms to absorb than a proprietary model or captive distribution channel.
Inferred, not confirmed -- Core facts on Callosum are supported by Tech.eu, Fortune, SiliconANGLE, and en.sedaily.com, but the named competitor rows rely on the structured fact set without corroborating public detail in the provided sources.
Opportunity
Upside case
PUBLIC The prize here is unusually large for a seed company because, if Callosum executes, it is not merely selling another AI performance tool, it is trying to become the control layer that decides where inference runs across models, chips, and clouds as heterogeneous compute becomes the norm [SiliconANGLE, August 2026] [Fortune, February 2026] [Tech.eu, February 2026].
The headline opportunity is to become the default orchestration layer for production AI workloads that do not fit neatly onto one model family or one hardware stack. That outcome is reachable, not just aspirational, because the company is attacking a real coordination problem that appears in the public record from multiple angles: Tailored Inference is described as a cloud service and API layer that breaks workflows into smaller tasks and assigns them to the model and chip configuration best suited to cost, speed, capability, or power requirements [SiliconANGLE, August 2026]; the launch hardware scope already spans Nvidia and AMD processors, AWS Trainium and Inferentia, plus newer architectures such as Cerebras and SambaNova [Fortune, February 2026]; and the company has raised more than $110 million across two seed financings, with backing from Plural, Atomico, DCVC, ARIA, and the UK Sovereign AI Fund [Tech.eu, February 2026] [SiliconANGLE, August 2026]. In plain terms, if AI deployment continues to fragment across chip vendors and model types, the coordinating software layer can sit in a structurally valuable position between application builders and compute suppliers.
That position matters because Callosum appears to sell value to both sides of the market. Application developers want faster and cheaper inference without rewriting code for each hardware stack [SiliconANGLE, August 2026], while emerging chip vendors need credible software pathways into real workloads if they want to compete with incumbent GPU ecosystems [Fortune, February 2026] [TechTimes, August 2026]. The public evidence is still early, but it is enough to support a credible upside case that the company could become infrastructure rather than a point solution, especially if its orchestration layer proves portable across clouds and accelerator vendors as claimed on its own site [Callosum].
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Cross-chip routing standard | Callosum becomes the default API layer enterprises use to route inference across multiple accelerators and model classes | More partnerships like Cerebras, plus proof that developers can adopt alternative silicon without code rewrites [SiliconANGLE, August 2026] [Quiver Quantitative, August 2026] | The product is explicitly designed to assign tasks across models and chip configurations, and public coverage says it supports accelerators from more than half a dozen vendors [SiliconANGLE, August 2026] |
| Embedded control plane for enterprise AI apps | Callosum wins inside document, agent, and enterprise workflow products, then expands from one workload to many | A few visible production wins similar to Round Treasury's finance workflow deployment [ai.engineer, 2026] | Public sources already place its APIs in cybersecurity, finance, and enterprise applications, with one named workflow example in finance [eu.vc] [ai.engineer, 2026] |
| Channel for non-Nvidia compute | Callosum becomes a go-to commercialization layer for newer chip suppliers that need software distribution and workload fit | Deeper ecosystem agreements with silicon and infrastructure partners such as Rebellions, Axelera AI, Tendrils, Supermicro, and Cerebras [en.sedaily.com, August 2026] [SiliconANGLE, August 2026] | The company's thesis and partner list both point to heterogeneous infrastructure as the core wedge, not just model optimization in isolation [Callosum] [en.sedaily.com, August 2026] |
The compounding dynamic, if it appears, is straightforward. Each new workload routed through the platform should improve Callosum's understanding of which model and hardware combinations deliver the best tradeoff on latency, cost, and reliability for specific task types, which in turn can make the next deployment easier to justify. Public claims already point in that direction: SiliconANGLE describes a system that chooses among model and chip configurations by task [SiliconANGLE, August 2026]; Plural says the platform can deliver 2x accuracy, 7x speed, and 4x lower cost than homogeneous compute for autonomous computer use [Plural]; and Round Treasury reported document extraction times falling from 15.9 seconds to 2.1 seconds per document using Callosum [ai.engineer, 2026]. None of that yet proves a durable moat, but it does show the beginnings of a learning loop where better routing creates better outcomes, which attracts more workloads, which generates more routing insight.
The size of the win is best framed through infrastructure positioning rather than a precise market model, because no confirmed public market-sizing data is in hand. If the company were to become a meaningful control plane for inference across chips and clouds, the relevant comparable set would likely look more like infrastructure platforms that sit in the critical path of developer workloads than like narrow inference consultancies. With more than $110 million already raised at seed stage [Tech.eu, February 2026] [SiliconANGLE, August 2026], investors are plainly underwriting that possibility. A reasonable public upside framing is that, under the cross-chip routing standard scenario, Callosum could grow into a multi-billion-dollar infrastructure asset, meaningfully larger than a typical tools company, because it would influence where AI spend lands across the stack (scenario, not a forecast). The evidence today supports the existence of that path, even if it does not yet support assigning probabilities with confidence.
Inferred, not confirmed -- This section relies on a mix of independent reporting from SiliconANGLE, Fortune, Tech.eu, TechTimes, en.sedaily.com, Quiver Quantitative, and ai.engineer, plus company and investor materials from Callosum and Plural. The central upside thesis is analytical and scenario-based, and some performance claims remain company-originated or lightly corroborated.
Sources
From the public record
[Callosum] Callosum | https://www.callosum.com/
[Tech.eu, February 2026] UK AI infrastructure startup Callosum challenging AI Silicon Valley “monoculture” comes out of stealth | https://tech.eu/2026/02/26/callosum-challenging-ai-silicon-valley-monoculture-comes-out-of-stealth/
[Fortune, February 2026] Exclusive: Callosum, a startup looking to break Nvidia’s stranglehold on AI workloads, raises $10.25 million in new funding | https://fortune.com/2026/02/26/startup-callosum-cambridge-trained-neuroscientists-raises-10-million-venture-funding-orchestrate-ai-workloads-different-chips/
[SiliconANGLE, August 2026] AI workload optimization startup Callosum raises $100M | https://siliconangle.com/2026/08/20/ai-workload-optimization-startup-callosum-raises-100m/
[TechTimes, August 2026] Callosum lands $100M seed to route AI tasks across any chip, bypassing Nvidia monoculture | https://www.techtimes.com/articles/325093/20260820/callosum-lands-100m-seed-route-ai-tasks-across-any-chip-bypassing-nvidia-monoculture.htm
[en.sedaily.com, August 2026] dunamu partners invests 5 million in UK AI inference startup | https://en.sedaily.com/technology/2026/08/25/dunamu-partners-invests-5-million-in-uk-ai-inference-startup
[Quiver Quantitative, August 2026] Callosum and Cerebras are partnering to bring next-generation inference to market, delivering industry-leading AI compute at scale | https://www.quiverquant.com/news/
[neuroscience.cam.ac.uk, 2026] Danyal Akarca | https://www.neuroscience.cam.ac.uk/directory/profile.php?dak39
[jachterberg.com, 2026] Jascha Achterberg | https://jachterberg.com/
[eu.vc] Callosum - Co-evolving chips & intelligence | https://www.eu.vc/p/callosum-raises-100m-seed-heterogeneous-ai-compute
[Plural] Why we invested in Callosum | https://www.pluralplatform.com/features/why-we-invested-in-callosum
[ai.engineer, 2026] Round Treasury mean document-extraction time case study | https://www.ai.engineer/
Articles about Callosum
- Callosum Routes AI Workloads Across Nvidia, AMD, and Cerebras Chips — The London startup, founded by computational neuroscientists, has raised over $110 million to make AI faster and cheaper by routing tasks across competing hardware.