Networking & Interconnect Systems - Member of Technical Staff
About the job
- Deep understanding of networking and communication systems, including reasoning across protocols, hardware, memory systems, and software layers.
- Experience with high-performance data movement such as RDMA, Ethernet fabrics, accelerator interconnects, programmable networking, or similar latency-sensitive systems.
- Strong systems performance instincts, including tracing bottlenecks across abstraction layers and distinguishing fundamental constraints from accidental ones.
- Ability to question existing abstractions, generalise across unfamiliar technologies, and simplify systems around workload requirements.
- Characterise and simplify communication paths between accelerators, tracing latency and unnecessary overhead across networking, runtime, memory, and device layers.
- Develop faster data movement across heterogeneous hardware, exploring RDMA, programmable NICs, DPUs, switches, and software approaches to cross-device memory management.
- Build flexible mechanisms for transferring model state across accelerator types and execution strategies, including KV cache connectors across different forms of parallelism.
- Evaluate existing and emerging scale-up and scale-out fabrics, informing cluster architecture, infrastructure integration, and how heterogeneous accelerator communication should evolve.
About Us
We’re living through a Cambrian explosion of intelligence: new models and new chips, each specialised for different tasks, are arriving all at once. The result is a new era for AI, one of radical heterogeneity.
Callosum is the Intelligent Systems Company. We believe the next generation of AI won't be defined by any single model or chip, but by intelligent systems in which hardware and intelligence co-evolve. We are building the infrastructure that unifies heterogeneous compute across the full stack. This opens a new axis of scaling intelligence: a dynamic system that tailors itself to what each workload actually needs, whether that's speed, cost, precision, or whatever unit comes next.
The last era scaled on a different bet: one bigger model, more of the same chip, more data. That bet is running into structural limits. Frontier models offer extraordinary capability at unsustainable cost, one that today's monolithic infrastructure was never designed to serve.
Our founding principle is that intelligence comes from many specialised systems working together, not from any single component. We build the software orchestration layer that co-evolves models, workflows and silicon into one system, delivering inference tailored to every workload, and demonstrating orders-of-magnitude leaps in capability and cost.
Because our software spans the full stack, our engineering team works directly with heterogeneous accelerators and frontier silicon, including Cerebras, d-Matrix, Intel, NVIDIA, AMD, Normal Computing, Tenstorrent, GreatSky, and Mixx. We are not stopping at today's chips: each new generation of silicon unlocks algorithms that couldn't run before, and we intend to be first to them, every time. If we get it right, it will belong to everyone building on it - not to any single vendor.
In our latest funding round, we raised $100M, led by Atomico with participation from Plural, DCVC and the UK Sovereign AI Fund’s first investment. With this, we are building the infrastructure for the next era of intelligence.
We are engineers and scientists based in London, working across the full depth of the stack. We are curious, intellectually honest, and building what doesn't exist yet. If you thrive on uncharted territory and are energised by the scale of the challenge, we'd love to hear from you.
About the Role
Callosum believes that orders of magnitude improvements in AI systems will come through application-aware orchestration across heterogeneous hardware. As workloads become increasingly disaggregated, the performance of the system is determined not only by where computation happens, but by how efficiently state and data can move between the hardware best suited to each part of the workload. Communication latency sets the boundary on how far that disaggregation can go.
This role owns that boundary. You will develop the communication systems that allow heterogeneous accelerators to operate as parts of one system, working across protocols, memory movement, networking hardware, and the software layers above them. That means understanding where overhead actually comes from, removing assumptions inherited from more general-purpose workloads, and finding new ways to move data directly and efficiently across vendor boundaries. The remit is intentionally broad: from RDMA paths and programmable NICs to cache transfer formats and emerging interconnect technologies, you will own the technical strategy for making communication less of a constraint on the systems we can build.
What You'll Build
Characterise and simplify communication paths between accelerators, tracing latency and unnecessary overhead across networking, runtime, memory, and device layers
Develop faster data movement across heterogeneous hardware, exploring RDMA, programmable NICs, DPUs, switches, and software approaches to cross-device memory management
Build flexible mechanisms for transferring model state across accelerator types and execution strategies, including KV cache connectors across different forms of parallelism
Evaluate existing and emerging scale-up and scale-out fabrics, informing cluster architecture, infrastructure integration, and how heterogeneous accelerator communication should evolve
What Sets You Apart
Deep understanding of networking and communication systems, with the ability to reason across protocols, hardware, memory systems, and software layers
Experience with high-performance data movement such as RDMA, Ethernet fabrics, accelerator interconnects, programmable networking, or similar latency-sensitive systems
Strong systems performance instincts, with the ability to trace bottlenecks across abstraction layers and distinguish fundamental constraints from accidental ones
A demonstrable tendency to question existing abstractions, generalise across unfamiliar technologies, and simplify systems around the requirements of the workload
What We Offer
Competitive Salary, determined by skills and experience
Equity & Ownership
Private healthcare
We offer Visa sponsorship and relocation benefits to hire the best in the world
We work in person at our London office. You'll have the tools, space and setup to do your best work, and if you have specific needs, just tell us
We're committed to building an inclusive workplace where everyone feels welcome, and believe in equal opportunities for all.
About the company
Callosum builds AI infrastructure that unifies different AI models and hardware chips to run workloads on the best combination of model and chip for each task. Its Tailored Inference APIs route tasks to the optimal model–chip pair while honoring cost, energy, and latency constraints, using partners like Cerebras and Rebellions for specialized silicon. Unlike approaches that rely on one model and uniform hardware, Callosum orchestrates multiple models and chips to maximize efficiency. The company's goal is to enable scalable, cost-efficient AI inference across heterogeneous hardware by intelligently matching tasks to the right combination of model and accelerator.
Company Size
11-50
Company Stage
Seed
Total Funding
$110.3M
Headquarters
London, United Kingdom
Founded
2025
Get referred to Callosum
See people who can refer or advise you
Simplify's Take
What believers are saying
- Atomico led Callosum’s $100 million seed on 20 August 2026, signaling strong demand.
- The UK Sovereign AI Fund made its first investment in Callosum, validating national relevance.
- Reported 4x speed and 70% cost savings with Cerebras strengthen enterprise sales conversations.
What critics are saying
- Callosum’s September 2026 cap table shows 24.1% preferred dilution and investor board access.
- The product depends on third-party silicon vendors, so supply friction hits deployment timelines immediately.
- Nvidia’s CUDA ecosystem and entrenched inference tooling threaten Callosum’s interoperability pitch.
What makes Callosum unique
- Callosum’s Tailored Inference APIs route workloads across chips and models, not one GPU stack.
- Founders Danyal Akarca and Jascha Achterberg bring Cambridge neuroscience, DeepMind, and Intel credibility.
- ARIA-backed partnerships with Cerebras and Rebellions anchor its heterogeneous compute positioning.
Help us improve and share your feedback! Did you find this helpful?
Benefits
- Health Insurance
- Company Equity
- Relocation Assistance
Growth & Insights and Company News
6 month growth
-23%1 year growth
-26%2 year growth
-26%Callosum's $100m seed put 24.1% of the holdco into a new preferred class. Callosum's $100m seed maps to a filed £57.20m Series Seed 2 issue for 24.1% of its holdco, while Atomico and Plural-linked directors joined the board. By Hagen Hoferichter · Published 23 August 2026 at 10:02 CEST Callosum's $100 million seed round looks different when the financing company's filings are put beside the announcement. The latest share allotment at Callosum Technologies Ltd records £57.20 million of cash for a new preferred class equal to 24.1% of the resulting issued shares. The same financing period changed the board. Two people publicly associated with Atomico and Plural joined the two founders on a four-seat board. That gives the new capital board-level access, but the filings do not establish investor control or disclose which institution owns which part of the new class. The finding is not that Callosum's $100 million headline is wrong. A round can close in tranches, combine instruments or include commitments that are not yet allotted. The narrower point is more useful for investors: the public record shows a £57.20 million preferred-share close, a 24.1% issued-share block and a new board configuration at the UK holdco that sits above the operating company. The filed close is smaller than the headline, but large enough to reset the denominator. Callosum announced the seed on 20 August 2026, describing a $100 million round led by Atomico with significant participation from Plural, DCVC and the UK Sovereign AI Unit. The company said the capital would support software that matches artificial-intelligence workloads to computing hardware. Bloomberg Law independently reported the round and its headline size. Companies House records for Callosum Technologies Ltd tell a more specific capital story. On 10 August, the holdco reported 4,329,819 Series Seed 2 Preferred shares allotted between 24 June and 8 July at £13.21 cash per share. The arithmetic is straightforward: | Filed position | Shares or cash | What it changes | | Shares before Series Seed 2 | 13,669,818 | Existing ordinary and Series Seed holders' base | | New Series Seed 2 Preferred | 4,329,819 | New class added in the June-July closing window | | Shares after the allotment | 17,999,637 | Post-issue issued-share denominator | | Registered cash for the new class | £57,196,908.99 | Filed-to-date cash, not a claimed full-round total | | New class as a share of issued capital | 24.055% | Simple issued-share dilution for earlier holders | Every earlier holder therefore owns a smaller percentage of the issued-share count after this allotment. The 24.1% figure is a share-count calculation, not a fully diluted ownership percentage. The preferred class has distribution-priority mechanics and one vote per share, while the amended articles contain the complete rights package. That distinction keeps the comparison honest. The filed amount is in pounds and the announcement is in dollars. The public record does not say whether the difference reflects exchange rates, a later tranche, commitments, grants, another instrument or a different measurement date. It does show exactly what was allotted and paid for in the reported close. Callosum built the financing in a new UK parent. The holdco was incorporated on 13 August 2025, shortly before the first major capital steps. A 26 August 2025 allotment issued 8,999,998 ordinary shares for non-cash consideration described as the legal and beneficial ownership of shares in SERNN, Inc. That structure placed the operating business under Callosum Technologies Ltd before the later cash financings. The 9 September 2025 allotment then added 3,043,301 Series Seed shares for £1.36 each and 803,556 Series Seed shares for non-cash consideration described as SAFE conversion. Further cash allotments followed in February and March 2026. By the time the Series Seed 2 class arrived, the holdco had 13,669,818 issued shares. Callosum's operating company record is therefore not the whole financing perimeter. The new preferred class sits in the parent that became the operating company's corporate person with significant control in August 2025. For diligence, the legal-entity choice matters: the round's dilution and governance entry are visible at the holdco, while the product and team are described at the Callosum operating-company level. This is the same analytical problem that appears in Amber's concentrated Series A issuance. The funding announcement supplies the event and participants. The registered allotment shows which legal entity issued the equity and how far the denominator moved. Those are different pieces of the financing map. Two investor-linked directors joined the founders. The capital reset was accompanied by a board change. Companies House filings dated 6 August record the appointments of Ian Hogarth and Alexandr Vidiborschii. Hogarth is publicly listed by Plural, while Vidiborschii appears on Atomico's investment team. Their appointments put two investor-affiliated individuals beside Callosum's two founders, Jascha Achterberg and Danyal Akarca. The board is therefore evenly split by role on the public record: two founders and two people associated with the lead and participating venture firms. That creates board-level access for the new financing group. It does not, on its own, prove a majority of voting directors, reserved-matter rights, a casting vote or unilateral control. The filings identify the people and their appointment dates, not the shareholder agreement behind the appointments. The timing is commercially relevant. The director appointments were filed shortly after the allotment window for the Series Seed 2 shares. The documents do not state whether the board rights were a condition of the round, part of a broader governance reset or an existing contractual right being exercised. They do show that capital and board composition changed in the same financing cycle. For a venture investor, that is a more concrete diligence point than the headline round size. A quarter of issued shares with a preferred distribution position can influence future financings even when the investor allocations are private. A four-person board can make information flow and strategic consent more immediate, while still leaving control questions unresolved. The ownership question is narrower than the headline question. The public announcement names Atomico, Plural, DCVC and the UK Sovereign AI Unit, but the Series Seed 2 SH01 does not allocate the 4,329,819 shares among them. It is therefore safe to describe the class as a 24.1% issued-share block and to connect the board appointments to named people with public ties to Atomico and Plural. It is not safe to assign a percentage to any one investor or to say the UK Sovereign AI Unit owns a particular part of the class. The same boundary applies to the $100 million figure. The registered £57.20 million is a filed amount for one class and closing window. It is not evidence of a funding shortfall, and adding the earlier public $10.25 million announcement to it would manufacture a total without knowing the instruments and dates that each announcement covers. The next decision-changing records are specific: a post-round shareholder list, the amended articles in readable form, and any shareholder or subscription agreement that sets allocation, preference and board rights. Those documents would answer what the 24.1% class means economically, whether further shares were committed but not allotted, and how the two investor-linked directors operate within the four-seat board. Callosum's financing is already legible at one level. A new UK holdco issued a preferred class for £57.20 million, the class represents 24.1% of the resulting issued shares, and two investor-linked directors joined the founders. The unresolved question is no longer whether the seed changed the cap table. It is how the private rights package turns that visible block and board access into economic and governance power.
Callosum raises $100M seed. Callosum raises $100M in a seed round led by Atomico to advance its platform for unifying heterogeneous computing resources for AI. Updated August 20, 2026 Callosum, a company building a global platform to unify heterogeneous compute for AI, has announced the closing of a $100M seed round. The company focuses on the orchestration of AI workloads across diverse silicon and models to optimize cost, energy, and performance. Investors. This round was led by Atomico, with participation from Plural, DCVC, and the UK Sovereign AI Fund. Callosum use of funds. Callosum plans to use the capital to scale its heterogeneous compute platform, expand its global compute partnerships, and further develop its programmable heterogeneity and Tailored Inference technology to deliver performance and cost-efficiency in AI systems. About Callosum. Founded by Danyal Akarca and Jascha Achterberg, Callosum is an AI infrastructure company building a platform for heterogeneous intelligence. Its technology enables AI workloads to run across different models and computing architectures, allowing organizations to optimize inference for specific requirements. Funding details. Company: Callosum Raised: $100M Round: Seed Funding Date: August 20, 2026 Lead Investor: Atomico Additional Investors: Plural, DCVC, UK Sovereign AI Fund Software Category: Artificial Intelligence Source: https://www.callosum.com/blog/seed-round Updated August 20, 2026
Callosum raises a $100 million seed led by Atomico, UK Sovereign AI on the cap table: HedgeCo.Net - Callosum, a London AI-infrastructure startup, raised a $100 million seed round led by Atomico, with participation from Plural, DCVC, and the UK's Sovereign AI fund. Atomico announced the lead on August 20. Bloomberg independently reported the same $100 million seed, the Atomico lead, Plural and DCVC in the round, and a "significant" investment from the UK's £500 million Sovereign AI vehicle. Callosum did not disclose a valuation. It did not disclose how much of the $100 million came from the public fund versus private capital. The company builds software that splits AI workloads and routes each piece to a model and chip suited to the job, rather than assuming a homogeneous GPU grid. Atomico said a first family of tailored-inference APIs is live. The lead's own note, on complex agentic workloads in financial services run with Cerebras, claimed four times the speed, 70% lower compute cost, and a 10% lift in task success versus a single frontier model on conventional infrastructure. Those operating stats are Atomico's and the company's. They are not a third-party benchmark. Bloomberg did not independently verify them. This is a seed, not a growth round. "One of the largest ever raised in Europe" is Atomico's ranking language, not a league-table print. Bloomberg put the UK vehicle at £500 million and called its cheque "significant." It did not print a dollar amount for that ticket, and neither did Atomico. Cheque size from the state remains unpublished. The valuation remains unpublished. For venture allocators the diligence is the cap table, not the orchestration pitch. A £500 million public vehicle taking a "significant" but unquantified slice of a $100 million seed is industrial policy sitting next to Atomico. It is not a priced round. Anyone marking a European AI-infra unicorn off this close is inventing a post-money the company refused to print. The $100 million is the raise. The valuation is blank. The second-order read is how sovereign AI money is attaching to software that sits between models and silicon rather than to a national frontier lab. Allocators should mark $100 million as the seed, Atomico as lead, and Sovereign AI as a disclosed but unquantified participant. Leave the 4x / 70% / 10% claims on the sponsor's page. The scarce object is a valuation and a public-fund ticket size, not another heterogeneous-compute manifesto.
Callosum raises $100M to cut AI costs by routing workloads to specialized chips. 20 August 2026 Key Takeaways * Callosum raises $100 million led by Atomico, one of the largest early-stage rounds in UK history. * The startup routes AI tasks to cheap, specialized chips instead of defaulting to expensive Nvidia GPUs. * Founded by Cambridge neuroscientists, the platform assigns each workload to the most efficient model and chip. The human brain doesn't use one giant processor for everything. It uses small circuits built for specific jobs. AI companies do the opposite, and it's killing their margins. To change that, Callosum raises $100M to build software that routes AI tasks to the cheapest chip that can do the work. The round was led by Atomico, with participation from Plural and DCVC. It also marks the first investment from the UK's new Sovereign AI Fund. At £73.5M, it ranks among the largest seed rounds in UK history. The money comes at a critical time. As AI moves from training massive models to deploying them at scale, companies are discovering that running those models is where the real money burns. The problem Callosum solves. Most companies today run every AI task on expensive, general-purpose Nvidia GPUs. It doesn't matter if the job needs massive compute or just basic processing; it all gets the same premium chip. That brute-force approach is draining budgets fast. For many AI-native companies, inference takes up over 50% of revenue. Training gets the headlines, but inference is what determines if a business is profitable. Callosum routes compute like the human brain. Callosum's answer is to stop treating all compute the same. Founded by Cambridge neuroscientists Danyal Akarca and Jascha Achterberg, the company took inspiration from biology. The human brain doesn't use one giant processor for everything. It uses small, specialized neural circuits tailored for specific actions. Callosum built software that does the same for AI. It acts as a smart routing layer between applications and hardware. It breaks down complex AI tasks and automatically assigns each piece to the cheapest, most efficient model and chip available. Instead of defaulting to Nvidia, Callosum partners with specialized chipmakers including Axelera, Cerebras, and Rebellions. A simple task gets a simple chip. A complex task gets the heavy hardware. 70% savings and 4X speed. The results are already showing up in early deployments. In partnership with chipmaker Cerebras on financial agentic workloads, Callosum reports 70 percent reductions in cost. By matching tasks to specialized silicon instead of general GPUs, it's also delivering 4 times speed gains. The pitch to investors is simple: the next phase of AI isn't about buying more chips. It's about using the hardware you already have as efficiently as possible. "Running generative AI models is wildly expensive," the company notes. Callosum wants to change how software talks to physical microchips so companies stop overpaying. The AI market is shifting. The race to train foundational models is slowing, but the race to deploy them cheaply is accelerating. That makes workload orchestration the next major hardware battleground. With backing from Atomico and the UK government, Callosum is positioning itself at the centre of it. The future of AI compute, the founders argue, isn't running every task on the biggest GPU available. It's picking the right tool for the job, just like the brain does. I'm a crypto writer with 4+ years of experience passionate about turning big, technical ideas into content anyone can understand. From blockchain to stablecoins to everything in between, I enjoy helping readers stay informed in a space that never stops moving. Disclaimer VentureBurn is a media platform covering the latest in cryptocurrency, artificial intelligence, venture capital, and the startup ecosystem. Opinions expressed on VentureBurn are for informational purposes only and do not constitute investment advice. Before making any high-risk investments in digital assets or emerging technologies, readers should conduct their own due diligence. All transactions and financial decisions are made at your own risk, and any losses incurred are solely your responsibility. VentureBurn does not endorse or recommend the buying or selling of any digital assets and is not a licensed investment advisor. Please note that VentureBurn may participate in affiliate marketing programs.