The Rise of Agentic AI and Openchip's Vision

The European tech scene is witnessing the emergence of new players, including the Spanish startup Openchip, which has announced its intention to launch its solutions by 2028. The company's focus is on the growing wave of agentic AI, a paradigm that promises to redefine the interaction between artificial intelligence systems and operational environments. Agentic AI refers to systems capable of perceiving their environment, making autonomous decisions, and acting to achieve specific goals, often through the orchestration of multiple Large Language Models (LLMs) and the use of external tools.

This long-term vision, projected to 2028, underscores the complexity and transformative potential of such technologies. For businesses and technical decision-makers, the advent of agentic systems implies a profound reconsideration of deployment strategies and underlying infrastructures. The ability of these agents to operate semi-autonomously requires not only advanced models but also a robust and responsive architecture capable of supporting their continuous operations.

Infrastructural Implications of Agentic AI

The implementation of agentic AI systems introduces specific infrastructural requirements that differ from those of individual LLMs. An AI agent, by its nature, often performs a sequence of operations: planning, executing queries on LLMs, using external tools, analyzing results, and subsequent iteration. Each step in this pipeline can generate latency, and the sum of these latencies can compromise the agent's effectiveness and responsiveness. For this reason, the physical proximity of computing resources to data and users becomes a critical factor.

Throughput management is another relevant challenge. If a company intends to deploy numerous AI agents in parallel, perhaps to automate business processes or support real-time decisions, the infrastructure must be capable of sustaining a high volume of requests and responses. This translates into the need for GPUs with high VRAM and compute capability, optimized for LLM inference, and a low-latency network for communication between the various components of the agentic system. The choice between on-premise, cloud, or hybrid deployment thus becomes strategic, directly influencing performance and Total Cost of Ownership (TCO).

On-Premise vs. Cloud for Agentic Workloads

The decision to host agentic AI workloads on-premise or in the cloud involves a series of significant trade-offs. On-premise deployment offers granular control over hardware and the environment, allowing for specific optimizations for agentic pipelines and ensuring data sovereignty, a crucial aspect for regulated industries or sensitive data. Air-gapped environments can be more easily realized, providing a higher level of security and compliance. Although the initial investment (CapEx) may be higher, the long-term TCO can be lower, especially for predictable, high-volume workloads.

On the other hand, cloud solutions offer scalability and flexibility, converting costs into an OpEx model. However, they can introduce higher network latencies and raise concerns about data sovereignty, depending on the jurisdiction and provider policies. A hybrid approach, combining the advantages of both, might represent the ideal solution for many organizations, allowing them to keep the most sensitive or critical workloads on-premise while leveraging the cloud for scalability or less constrained tasks. For those evaluating these complex deployment decisions, AI-RADAR offers analytical frameworks on /llm-onpremise to support the assessment of trade-offs.

Future Prospects and Strategic Planning for 2028

The 2028 horizon set by Openchip suggests a period of maturation for agentic AI and the infrastructures that will support it. By that date, LLM inference hardware is expected to become even more efficient, with GPUs offering greater VRAM and throughput at potentially lower costs. The challenge for companies will be to anticipate these evolutions and plan an infrastructure that is not only powerful but also flexible and sustainable.

The ability to integrate AI agents into existing business processes will depend heavily on the robustness and resilience of the underlying infrastructure. Choosing a self-hosted or bare metal deployment for critical agentic AI components could offer competitive advantages in terms of performance, security, and control. Openchip's vision highlights how the future of AI is not just a matter of algorithms but also of architecture and infrastructural strategy.