Callosum

“A few years ago, we had a handful of unicorns. Today we have 700. In a few years we’ll have 1200.” Bernard Liautaud has had a front row seat to European tech for decades - first as a founder, then as one of our most influential investors. He joined Katie Prescott to discuss the recent shift. What’s changed? Belief, for one thing. “Entrepreneurs now feel that Europe is a great base. They want to be global leaders, but with European cultural values, responsibility and headquarters.” Danyal Akarca, co-founder of Callosum, sees another European advantage emerging: heterogeneity. “The market has moved past training massive models. Now we want AI to do actual work”, he tells Maggie Buggie of InThisTogether.vc. That means more emphasis on specialisation. As one of the first companies backed by the UK’s Sovereign AI fund, Callosum is showing that faster, more efficient, purpose-built AI can be based on a diverse mix of models and chips, rather than dependence on a single technology or provider. “That is an unbelievably hopeful future if we get it right. It’s an inherently democratised, decentralised way of building our AI infrastructure.” And the proof of Europe’s emerging strengths is in the room this morning. Evaro’s Thuria Wenbar and BioOrbit’s Dr Katie King, PhD joined Karim Jalbout to explore the advantages of building in Europe: from access to the entire healthcare value chain to scaling across multiple markets. Dr. Ben Maruthappu shared how Cera has saved the NHS and government more than £1bn to date, by scaling with patience and discipline over the past decade - a far cry from the “move fast and break things” mode of growth. Different companies and markets, but a common thread: Europe’s diverse, regulated ecosystem is becoming its superpower. #BoardwaveLive2026 #EuropesMoment

Today, we are thrilled to launch Europa, our next-generation AI Processing Unit (AIPU) designed specifically for physical AI and enterprise AI workloads. Being supported by over 30 customers and partners globally is truly humbling - thank you to all of them (full list in comments). Enterprise AI compute demands are shifting rapidly. Rising cloud costs, strict power budgets, and stringent data sovereignty needs mean organizations need a practical path to run persistent AI on infrastructure they control. Europa delivers up to 6x higher energy efficiency (tokens/s/watt) compared to GPU-based alternatives, enabling high-performance inference within real-world power and thermal limits. Europa is available as standalone silicon, as a half-height, half-length PCIe card (Axelera Edge 232p), and as a full-height, full-length PCIe card (Axelera Server 250p). Through our partner ecosystem, Europa is shipping in fully validated systems from leaders like Dell Technologies (XE5) and Supermicro (111AD), and works in existing infrastructure from Hewlett Packard Enterprise, Lenovo, 2CRSi, Advantech, Axiomtek, and SECO and more!   Alongside Europa, we are introducing AxeleraScript (AxScript), our novel Python-based domain-specific language that solves the model support bottleneck. Most AI accelerators limit you to a fixed menu of supported models. AxScript gives developers direct, kernel-level programmability over our silicon at full performance-enabling teams to compile custom networks, proprietary operators, or new architectures across CNNs, LLMs, VLMs, and beyond. Using AxScript with AI assistance, a single engineer ported the complex Qwen3.5 model in just one day.   Learn more about the Europa architecture and AxScript: https://lnkd.in/eraf_Jxf #EdgeAI #AIInference #ComputerArchitecture #AIHardware #Europa

After co-inventing ChatGPT, I spent 2 years in stealth building a new type of AI model. Today, TypeSafe AI is launching Jev, the first System One Model, and announcing a $40 million seed round led by DCVC. I left OpenAI to answer a question: if these models are so intelligent, why have they automated so little of the world’s work? After 2 years in stealth and a $40M seed round, we’re excited to introduce Jev, our first public System One model 🚀 📈 Jev is built for the decisions inside software. Unlike AI chatbots, Jev can’t hallucinate and won’t just tell you what you want to hear. It returns typed decisions with probabilities, so your code can decide when to act and when to ask a person. You can break a workflow into individual judgments, run them in parallel, and combine them however you want. Intelligence is now something you can reliably build with. And at $42 per billion input tokens and free output, intelligence becomes a resource accessible to all. We can't wait to see what this unlocks for developers 👀 A huge thank you to the team and everyone who supported us when TypeSafe was just an idea 🙌 ❤️ Try it today: https://typesafe.ai

“We're in the early stages now of this Cambrian explosion. What are the boundary cases where we want certain tasks on certain models on certain silicon? We’re making that programmable. This is where all of the frontier performance is going to come from.” This is one of the core problems we are solving at Callosum & I'm very happy to have sat down to discuss this with Andy Hock of Cerebras. Looked at closely, most tasks we ask AI to solve are compositional: a bundle of distinct sub-problems assembled as one larger problem. For example, A single workflow has multiple steps, and each step can call for different models and different kinds of compute. The number of possible combinations grows fast and finding the optimal match means searching and navigating all of it. That search is exactly what we built Callosum to navigate, to make AI orders of magnitude faster, cheaper and more performant. It's why I'm excited about our new partnership with Cerebras! Grateful to Andy and the Cerebras team! Andrew Feldman, Angela Yeung, Jason W., Ben Duggan

"You now have new surface area to expose the benefits of the silicon in new ways. When doing this over a huge amount of different silicon, Cerebras was unbelievably good. We could actually not only improve the speed by a significant amount in these long-context reasoning problems, we also found that we could also make them very cost-effective because we're being efficient with the tokens." — Danyal Akarca, co-founder of Callosum, in conversation with Andy Hock, our SVP & Chief Strategy Officer, on the new Cerebras-Callosum partnership. Long-context reasoning means reasoning over a huge amount of data. Callosum's approach chunks that context into subparts and runs each on a model specialized for it — heterogeneous recursive models. Run across Cerebras hardware, that opens new surface area to expose the benefits of the silicon, making the work both faster and more efficient with tokens. This is the idea at the center of our new partnership with Callosum: match each part of a workload to the model and compute that handle it best. Where would per-step routing make the biggest difference in your stack? Watch the full conversation: https://lnkd.in/gYVGVC2C

Having the most money and the biggest data centers used to mean you'd end up with the smartest AI model. Danyal Akarca from Callosum says that's not really true anymore, now that so many AI models are already good. What's left is a tiny gap between the best and good enough, and some people will still pay billions to be on the right side of it. Check out the full interview by searching "Danyal Akarca - Scaling Europe" wherever you get your podcasts. The Scaling Europe show is presented by Deel, and sponsored by SurrealDB, Chargebee, Airwallex and Lovable.