Today, Cohere announced the initial close of its Series E funding round targeting up to CAD $7 Billion, led by Schwarz Group with participation from new and returning investors including Bell Ventures, Bullhound Capital, Inovia Capital, NVIDIA, and Prosperity7 Ventures, and the Government of Canada. We will use the funds to build something increasingly essential in AI: a frontier company outside the U.S. and China. Our ambition is to build frontier models that compete with the best in the world, a secure platform that lets institutions own the AI they run on, and cyber defenses democracies can rely on. We don’t have to accept a future of AI run by just a few, and we shouldn’t want one.
Within a few years, every hospital in Bavaria, every bank in New York and every agency in Ottawa will run a significant part of its operations on AI. They should be able to choose world-class systems built under laws they trust, using technology they can inspect and adapt, deploy inside their own environments, and on platforms they can depend on to remain available when it matters most. Today, many organizations face an uncomfortable choice: cheap models from China or powerful ones you can’t fully see into or control. The market is already choosing: more than 70 percent of the open-model tokens served through OpenRouter came from Chinese models earlier this year – up from 3 percent fourteen months earlier.
We’re focused on changing that.
The need for choice and control is no longer theoretical. Recent history has shown how technologies behind communications, energy, and national security infrastructure can end up controlled by decisions made elsewhere. Access can change overnight because of commercial decisions, export controls, or geopolitical events, often with few alternatives. Each situation was different, but the lesson is the same: countries need guaranteed access to the technologies that keep them secure, resilient, and competitive. As allied governments take on greater responsibility for their own defense, infrastructure, and technological capabilities, AI belongs on that list.
The case against frontier fatalism
When Canada and Germany launched their Sovereign Technology Alliance in Munich earlier this year, they said they were "choosing to build," and that’s exactly what we intend to embody. Yet a growing consensus holds that the frontier of AI belongs to two countries and a handful of companies, and that everyone else should make peace with renting intelligence from whoever wins. You hear versions of this argument in boardrooms, government agencies, and investment committees across the democratic world, usually from people who mean well. Its advocates call it realism. We call it frontier fatalism.
Taken to its logical conclusion, frontier fatalism is a plan for managed decline. It asks countries to abandon capabilities they spent decades building. After investing in research, talent and innovation, they would wire their economies, public services, and defense infrastructure around technologies they don’t control, can’t meaningfully influence, and may not even be able to access indefinitely. And they would be expected to be grateful for API access.
The fatalists are right about one thing: building at the frontier is difficult. It requires capital, compute and research talent at a scale few organizations can assemble. Where they are wrong is in assuming there is only one way to compete.
There are many frontiers
Much of the confusion in discussing frontier AI models comes from talking about “the frontier” as if it were one thing. But there are the myriad domain frontiers: the best achievable performance on a specific class of valuable problem, from finding software vulnerabilities, to analyzing intelligence across dozens of languages, to running a bank's compliance work. Researchers at Harvard Business School referred to this unevenness as the "jagged technological frontier." In many domains, success depends not only on the model itself, but on the system around it: how context is assembled, how state is tracked, how reasoning is verified, and how outputs are evaluated. That reality creates room for more builders. And there are dimensions of the jagged technological frontier where owning, rather than buying, is crucial to one's sovereignty.
The proof: cybersecurity
Cybersecurity is one of the most important domain frontiers today. Anthropic's Claude Mythos Preview demonstrated the potential of advanced models to discover previously unknown vulnerabilities, while Project Glasswing extended that work to operators of critical infrastructure in more than 15 countries. Those advances matter. But they have largely occurred behind closed doors, with access limited to a small group of partners on terms set by a single provider, reflecting the same concentration that defines the broader frontier AI landscape. More importantly, they do not prove that leadership in cyber defense will belong exclusively to whichever company builds the strongest general-purpose model.
AISLE, which will soon join us as Cohere Cyber, offers a different perspective. The company describes vulnerability discovery as a production function with two inputs: intelligence per token and tokens per dollar. In practice, that means model capability is only part of the equation. A well-designed system can use efficient models to explore more possibilities, run more tests, and validate more findings. Raw intelligence matters, but so does the ability to turn that intelligence into results.
The evidence is compelling. In August, curl founder Daniel Stenberg reported that Mythos, OpenAI's Codex Security, and ZeroPath had found nothing further in his codebase. AISLE's system subsequently identified 18 of the 24 vulnerabilities fixed in curl's October release. A specialized cyber platform and a general-purpose model are not directly comparable. But that is precisely the point. Organizations do not deploy models in isolation. They deploy systems, products, and workflows. What ultimately matters is the performance of the complete system.
Cybersecurity also offers something uniquely valuable for AI development: ground truth. Suspected vulnerabilities can be reproduced, validated by maintainers, and tested against proposed fixes. That makes cybersecurity not only a critical domain in its own right, but also an ideal environment for improving the models that power it.
How the frontiers feed each other
It may be expected that a company focused on specific domains would fall behind on general models, but in practice the domain work is one of the best ways to make general models better. Working at a domain frontier means pushing for the best possible results on one kind of valuable problem, whether that takes a specialized model, a general model with the right tools and instructions, or a full system built around several models.
Some of the biggest recent advances in model reasoning have come from reinforcement learning on specific domains where a computer can check the answer, such as whether a proof holds, a test passes or a bug that reproduces. The skills models learn inside these domains transfer beyond the domain itself and into related settings, driving general model improvements alongside domain specific ones.
Working with customers on real-world tasks also shows us where today's models fall short in ways public benchmarks rarely capture and each failure tells us what the next model needs to do better. Customers also tell us which capabilities matter most. Public sector customers usually ask first for strong analysis across many languages and the ability to read images and documents alongside text, the same capabilities a bank, a telecom operator or a hospital needs from an AI agent.
All of those improvements flow back into our general models, and better general models make everything we build on top of them more capable, so a customer in one domain benefits from what we learned in another. That loop is why we build on all of these frontiers at once, and this funding round is designed to speed it up.
Europe has done this before
Whether a challenger can take on entrenched giants this way, with a different approach and allies building alongside it, is a question Europe has answered before. Half a century ago, the answer came from commercial aviation.
In the late 1960s, American manufacturers dominated the jet airliner market. Europe’s national aircraft makers were individually too small to challenge them. Britain’s government pulled its funding from the project that became Airbus in 1969. France and Germany pressed ahead. Airbus did not begin by matching Boeing aircraft for aircraft. It competed through a different design: the A300, the world’s first twin-engine widebody. Commercial success took years. A breakthrough came when Eastern Air Lines, led by former Apollo 8 commander Frank Borman, agreed in 1977 to trial four A300s and subsequently ordered more. A decade later the A320 brought fly-by-wire controls to airliners. In 2025, Airbus delivered 793 aircraft to Boeing's 600.
The lesson is that an apparently settled market can be moved by a serious competitor with patient investment, a differentiated approach, and partners willing to build. Airbus builds jets in Mobile, Alabama, and spends more than $15 billion a year with over 2,000 American suppliers. Its A220 began life as Bombardier's C Series, designed in Quebec. Many of its engines come from CFM International, a fifty-fifty venture between GE and France's Safran whose LEAP engine also powers Boeing's 737 MAX. None of that makes Airbus less European. Competition on the airframe and partnership on the engine made both sides of the Atlantic stronger.
That is the model we have in mind for AI. What we are building is a second option for the democratic world, and the United States benefits from it as much as Canada, Germany, and other allies do. The US frontier should not have to carry the competition with China alone. Building more frontier capability across the democratic world makes the entire alliance stronger.
Two decades of building
The foundation for this ambition already exists. AI has always been an international project, and Canada and Germany have been important contributors. In 2004, when neural networks were out of fashion almost everywhere, the Canadian Institute for Advanced Research backed an AI program led by Geoffrey Hinton that brought him together with Yoshua Bengio and Yann LeCun, the three researchers later honored with the Turing Award for deep learning. The Transformer paper, published at Google in 2017, was written by eight researchers with roots in seven countries. One was a University of Toronto undergraduate who went on to co-found Cohere. That same year, Canada became the first country with a national AI strategy, anchored by the Vector Institute, Mila and Amii.
Public investment has continued. Ottawa has committed more than C$5 billion to AI since 2017: about C$742 million for the Pan-Canadian strategy, a C$2.4 billion package in Budget 2024 that included the C$2 billion Sovereign AI Compute Strategy, and at least C$2 billion more in this June's AI for All strategy. It has paid off in innovation and talent. With about half of one percent of the world's population, Canada is home to roughly 10 percent of the world's top AI researchers. Germany has made a parallel bet, committing €5 billion to its national AI strategy since 2018 and growing Cyber Valley in Tübingen into one of Europe's largest AI research collaborations.
Research and talent are the foundation. The next step is building companies that can turn them into frontier capabilities, create opportunities for researchers to build at home, and deliver systems institutions can use. That is what this round is for. Investors, research institutions, infrastructure partners, and customers on both sides of the Atlantic are building with us because they believe the future of AI should be more competitive, not ceded to fatalism. Our Series E is the largest funding round ever raised by a private Canadian company, and it carries two decades of public and private investment into a new chapter.
What we will build
Models at every price point, up to the frontier. The frontier is jagged, and the cost of useful capability is falling fast. We will continue investing in frontier models while also building smaller, more efficient models for tasks that don't require the largest systems available. Customers need both access to advancing capabilities and the flexibility to choose the right model for each job based on performance, cost, latency, and deployment requirements. The goal is not one model for every task. It is a portfolio of capabilities that lets customers deploy AI on their own terms. We will continue working with partners across research, sovereign infrastructure, enterprise deployment, and cloud services to meet the full range of customer needs.
A platform customers control. If models are swappable, the durable asset is the layer that governs an institution's data and workflows, and that layer should belong to the customer. North, our agentic AI platform, is built to be that layer. It runs in a customer's own cloud, data center or air-gapped environment and remains model-agnostic, allowing organizations to use Cohere models alongside others of their choosing. The workflows, data, and controls remain with the customer, and no external provider can switch them off. Reliant AI, which joined us earlier this year, extends North’s capabilities into biopharma and healthcare. We will apply the same approach across additional industries, pairing world-class models with purpose-built products designed for real-world deployment.
Frontier-class cyber defense. Cyber is where the domain-frontier thesis has already been proven, and where dependable access matters most. Through our planned acquisition of AISLE, which identified more than 500 CVEs in its first year, we are building Cohere Cyber: a frontier-class system designed to help defenders find, validate, and remediate vulnerabilities before attackers can exploit them. It runs on-premises, behind the firewall, or in fully air-gapped environments, ensuring sensitive source code never leaves the building. Because it operates on models hosted by the customer, no external provider can throttle access or turn it off when it matters most. Most importantly, it is built for defense. Its purpose is to identify and fix weaknesses before they become incidents, putting advanced cyber capabilities in the hands of defenders without meaningfully increasing risk to everyone else.
The next twelve months. We will complete our combination with Aleph Alpha and acquisition of AISLE, launch Cohere Cyber for air-gapped and classified environments, release the next generation of our models, and begin frontier-scale training with allied partners. Each step advances the same strategy: stronger models, better systems around them, and more control for the institutions that rely on both.
Building it together
AI will become part of the operating infrastructure of economies, public services, and national defense. Who builds it, under what law, and with what accountability will shape the choices available to every country that adopts it. No country needs to build every part of that infrastructure alone. But the democratic world needs more countries and companies capable of contributing to it.
Canada and Germany have chosen to build, and so have we. If you are a government, a critical infrastructure operator, a researcher, or an American partner who believes the alliance is stronger when more of its members can build, we want to work with you. The frontier is not reserved for a fixed set of countries or companies. It belongs to those willing to advance it.
