Cohere's AI Sovereignty: A New Era in Enterprise AI?

Cohere VP of Product Engineering Rachad Alao argued at VB Transform 2026 that AI sovereignty—tight control over data and infrastructure—is critical for enterprises like banks and hospitals, even as token prices fall. He highlighted that rising agent workload complexity drives token usage up, and advocated using smaller, efficient models for most tasks to balance cost and security.

Cohere's AI Sovereignty: A New Era in Enterprise AI? Cohere's VP of Product Engineering, Rachad Alao, debates AI sovereignty at VB Transform 2026. He emphasizes tighter control over AI systems and data amidst falling token prices. At the VB Transform 2026 conference in Menlo Park, the focus was clear: AI sovereignty is more than just a buzzword. Rachad Alao, Cohere's VP of Product Engineering, spoke candidly about the complexities of maintaining control over AI systems without compromising on data security or vendor flexibility. AI sovereignty, Alao explained, is essential for organizations operating mission-critical systems, such as banks, hospitals, and governmental bodies. 'it's important to have very tight control on where the data resides,' Alao emphasized. This means having jurisdictional authority over AI operations, from GPUs to private-cloud infrastructure. Rising Demands on Agent Workloads During the discussion, VentureBeat's CEO, Matt Marshall, offered a critical viewpoint on the economics of locally deployed models. With inference /glossary/inference prices dropping, why optimize every token /glossary/token ? Alao countered with a key insight: the complexity of agentic use cases is increasing, driving token usage up exponentially. Enterprises are transitioning from basic chatbots to complex agents that perform multi-step processes. This leads to a pointed question: Are enterprises truly prepared for the surge in token utilization? Cohere's approach diverges from models that bill per token. Instead, they focus on solving enterprise challenges securely while minimizing model usage. Alao's advice was clear: 'Use the right model for the task at hand.' Efficiency with Smaller Models Cohere's open-source North Mini Code model, running on a single Nvidia /glossary/nvidia H100 GPU, targets agentic software engineering. Alao acknowledged that while larger models might perform better on the most demanding tasks, they aren't always necessary. For 80% of use cases, smaller models offer effectiveness and cost-efficiency. Cohere's Command A+, a 218-billion- parameter /glossary/parameter mixture-of-experts model, further reduces the hardware needs for private deployments. Its Apache 2.0 license grants enterprises flexibility in operation and modification. So, are larger models the future of AI, or is there substantial merit in these smaller, more efficient solutions? Integrating Search in Agent Workflows Cohere's ongoing work in embeddings and enterprise search is evolving. Alao discussed the shift to multimodal /glossary/multimodal search, indicating a move beyond mere text retrieval. Search is now an integral part of agentic workflows, with models determining when and how to use retrieval methods. Ultimately, Cohere's focus on data control and portability addresses a significant concern for enterprises wary of vendor lock-in. With a governance layer that routes traffic to suitable models, they break free from dependency on cloud provider bundles. The question remains: Will other enterprises follow suit in prioritizing sovereignty and control? Get AI news in your inbox Daily digest of what matters in AI.