The agentic era broadens AI’s tailwinds

Since the late-2022 launch of OpenAI’s ChatGPT 3.5, milestones have been frequent as competing AI platforms have continuously expanded their respective models’ capabilities. Among these advancements, last December’s release of Anthropic’s most powerful model to date – Claude Opus 4.5 – will likely be remembered as particularly impactful. Much of the excitement was focused on Claude’s ability to accelerate the deployment and power of agentic AI.

No longer theoretical, the era of agentic artificial intelligence has arrived, and its reverberations are being felt across the technology sector. As with other watershed developments within AI, the dawn of agentics has caused both corporate managers and investors to recalibrate their expectations for what will be required to bring the full capabilities of these novel applications to light.

Similar to how AI’s inference stage surprised the market with its sustained demand for graphics processing units (GPUs), the adoption of agentic AI is making potential winners out of legacy technologies that had previously not been considered as direct beneficiaries in the productivity revolution. Chief among these are memory chips, optics, and even central processing units (CPUs).

To be sure, demand for AI’s foundational infrastructure – GPUs, data center capacity, electricity generation –remains robust and, in most cases, exceeds near- to mid-term supply forecasts. But each incremental advancement has further revealed the magnitude of AI’s computational intensity, and meeting those needs will command a far greater breadth and depth of inputs than initially anticipated.

Agentics: The power of many

Most users’ early AI interactions were premised on prompts, not unlike how they had historically interfaced with legacy search engines. While getting prompts to leverage large language models to deliver a concise, bespoke output is no small feat, agentics is a whole new level. Rather than requiring the user to guide the conversation or task, agents – since they can reason – are able to take action, thus exercising a degree of autonomy or agency.

For example, a worker could create a range of agents that act as virtual assistants carrying out a host of tasks, with each unit of output iteratively guiding the next course of action. And unlike the employee or her subordinates, this agent can work around the clock, always processing new data and executing tasks.

While GPUs largely do the thinking of AI agents, the scheduling and carrying out of duties typically fall to CPUs. Those chips – largely an afterthought in the AI era – are well equipped, and indeed essential, to helping AI agents maximize their capabilities. The upshot is that erstwhile unloved CPU producers have seen their near- to mid-term prospects improve.

Agents also require a significant amount of memory – the putatively dull and commoditized corner of the chip industry. In the past, we’ve spoken of the large volumes of incremental data produced by GPUs as they perform test time inference. That data must be stored somewhere and be readily accessible for future reasoning.

Optics, meanwhile, have always been utilized to connect data centers, but as the volume of information flowing within racks and between racks increases on the back of ever more powerful GPUs, hyperscalers must rely on more efficient optical connectivity. Historically, copper often sufficed within racks and data centers. Going forward, these connections will increasingly rely upon optics. This trend to scale up, scale out and scale across stands to be a secular tailwind for optics providers in the era of a GPU-augmented human economy.

A lens on industry structure

As has been the case with GPUs and data centers, the breadth and pace of the AI buildout has resulted in demand for memory, optics, and CPUs outpacing current supply. In many instances, specific industry structures intensify these AI tailwinds. Memory, for example, has become a highly rationalized industry. Low competitive intensity, coupled with bottlenecks, is leading to significant pricing power that could last over the near- to mid-term.

With respect to CPUs, while certain players build their own chips, third-party foundries dominate semiconductor manufacturing. When given the choice between allocating capacity toward more lucrative GPUs or lower-margin CPUs, foundries will opt for the former. In the instances where CPU producers secure manufacturing capacity, they too should enjoy pricing power given production bottlenecks.

On an even more elemental level, there is only so much material that goes into chips and optics being produced. Production constraints in those industries – as well as trade policy and geopolitics – represent frictions in getting these materials into factories.

These segment-specific bottlenecks represent a governor on the AI infrastructure CapEx cycle that, in our view, refutes the bear case against the magnitude of this investment. Limited supply, along with the time it takes to construct a data center and procure sufficient GPUs and electricity to power it, is likely to keep the pace of incremental supply below that of incremental demand for the foreseeable future.

Secular by nature

The shortfall in supply of CPUs, memory, and optics is yet another indication of the durability of the AI theme. Much of this strength is owed to the rationale underpinning the historic level of capital expenditure. Scaling laws are holding. Incremental investment continues to deliver step-function advancements in AI models’ capabilities. Importantly, competing models are leapfrogging each other with each new version. This represents a healthy dynamic. As long as this dynamic continues – and we see no reason to think they won’t – AI companies are likely to generate the returns necessary to justify their investments.

We see agentics as a compounder for this theme. The reason is networking laws. We see a future where workers – and people outside of the office – leverage four or more agents to prioritize, juggle, and execute myriad tasks. In such an environment, a person’s agents will interact not only with each other, but also with agents created by other individuals and entities across the AI ecosystem. With each agent added, the data and associated linkages increase exponentially. As we are only beginning to see the adoption of agentics (which, again, operate around the clock), we expect the power of networking laws to keep demand strong for nearly all of AI infrastructure’s inputs.

An opportunity for software?

Lastly, a word on how agentic AI could affect the beaten-down software industry. We recognize that the ability of models like Claude’s Opus and OpenAI’s Codex to replicate the functionality of existing software suites represents a competitive threat to many companies. But executive teams are also aware of this, and many are taking action. In a notable example, software that historically acts a system of record could leverage agentics to transition to a system of action. These systems already possess volumes of company- and industry-specific data that – when properly parsed – could add significant value to an organization.

IMPORTANT INFORMATION

Artificial intelligence (“AI”) focused companies, including those that develop or utilize AI technologies, may face rapid product obsolescence, intense competition, and increased regulatory scrutiny. These companies often rely heavily on intellectual property, invest significantly in research and development, and depend on maintaining and growing consumer demand. Their securities may be more volatile than those of companies offering more established technologies and may be affected by risks tied to the use of AI in business operations, including legal liability or reputational harm.

Technology industries can be significantly affected by obsolescence of existing technology, short product cycles, falling prices and profits, competition from new market entrants, and general economic conditions. A concentrated investment in a single industry could be more volatile than the performance of less concentrated investments and the market as a whole.