The Next AI Trade May Not Be What the Market Thinks
| The first wave of AI investing was relatively straightforward. More model training meant more GPUs. More GPUs meant more NVIDIA.The investment framework largely revolved around a single question: who benefits from larger models and more compute? That framework worked well.But the next phase of AI looks structurally different. We are moving from AI that generates to AI that acts. This distinction matters more than most investors appreciate. Agentic AI systems are designed not simply to answer prompts but to plan, reason, remember context, call tools, coordinate with other agents, and execute multi-step tasks autonomously. Instead of a single inference request, these systems create workflows.The shift sounds subtle. The infrastructure implications are not.In this week’s Tech Edge, we explore why agentic AI may create a much broader infrastructure cycle than most investors currently appreciate, especially across networking, memory, orchestration, and system architecture. |
| From Intelligence to OrchestrationTraditional generative AI was largely a GPU problem. A request enters. The model processes it. Tokens come out.Agentic systems create a different architecture entirely: Plan → Search → Retrieve → Reason → Act → Repeat.Each step creates coordination overhead. The model now has to access memory, call external systems, orchestrate workflows, and manage state across tasks. CPU-side orchestration can account for a surprisingly large portion of workload latency in agentic systems. That changes where value accrues. |
| The CPU increasingly becomes the control plane. Memory becomes an active system component. And infrastructure shifts from isolated chips toward integrated systems.Importantly, this does not reduce the importance of GPUs. If anything, GPUs become even more valuable. But they increasingly function as specialized inference engines inside a much larger orchestration stack.As GPUs become more expensive, hyperscalers have a growing incentive to reserve them for token generation while offloading planning, routing, memory coordination, retrieval, and tool execution elsewhere in the system.The first AI cycle was primarily compute-constrained. The next phase may increasingly become coordination-constrained, where latency, memory movement, orchestration, networking, and system architecture determine real-world performance.The market still talks about compute. Increasingly, AI is becoming a coordination problem.The AI Stack Is ExpandingOne of the defining characteristics of the first AI cycle was concentration. Capital flowed toward a narrow set of winners.Agentic AI potentially broadens the opportunity set considerably. As AI systems become more orchestration-heavy, infrastructure increasingly requires additional CPU-heavy layers dedicated to planning, routing, scheduling, memory coordination, and tool execution. The result is a meaningful rise in system complexity at the cluster level. Agentic workloads could create:$32–60B of incremental CPU TAM by 203015–45 exabytes of incremental DRAM demandmeaningful growth across storage, packaging, networking, and system infrastructureBut the broader takeaway may be more important than the specific numbers: AI is becoming a full-stack systems story.Owning the model may not be enough. Owning the architecture increasingly matters. |
| Why This Matters for InvestorsMarkets often extrapolate linearly. But major technology transitions rarely evolve linearly. The internet did not just create website winners. It created routers, data centers, fiber, cloud infrastructure, mobile ecosystems, and entirely new layers of value creation. AI may be entering a similar phase. As intelligence shifts from generating outputs toward coordinating actions, the second-order effects become increasingly important.The conversation may gradually move beyond: “Who builds the best model?”. Toward: “Who enables intelligence at scale?”That distinction could define the next leg of the AI cycle.The Bottom LineThe first phase of AI rewarded scale. The next phase may reward coordination.Markets are still largely pricing AI as a GPU expansion cycle, while the underlying architecture is quietly evolving into something much broader: a full-stack systems buildout centered around orchestration, memory, networking, and infrastructure efficiency.The important shift is not that GPUs matter less. It is that intelligence at scale increasingly depends on everything surrounding them.As agentic AI moves from experimentation into production, the winners may not simply be the companies building the most powerful models, but the companies enabling autonomous systems to operate reliably, continuously, and efficiently at scale.In many ways, AI infrastructure is beginning to resemble the evolution of the internet itself: what initially looked like a narrow compute story gradually expanded into a much larger ecosystem spanning connectivity, storage, coordination layers, and entirely new system architectures.The market may still be focused on the brain. But increasingly, the bottleneck is becoming the nervous system. |
