As the AI story moves from training large language models to delivering myriad “agentic” applications, we believe investors have an opportunity to capitalize on this crucial shift.
On April 22, Google marked a fundamental turning point in the AI era: It split its flagship, custom-build processor into two chips: the TPU 8t for training large language models (like Google’s own Gemini model), and the TPU 8i for powering “agentic” AI applications that rely on inferencing to complete multistep tasks.1
Subtle as it may seem, we believe Google’s decision may mark a crucial inflection point for investors: As the “Mag 7” basket has started to splinter, more granular themes within the AI story have begun to take shape (as my colleague Jeff Blazek noted in a recent CIO Weekly column).
More specifically, we believe the shift from model training to agentic AI is already driving a fundamental shift within AI’s underlying infrastructure and ushering a significant expansion of the addressable market across the hardware supply chain.
Under the Hood: Model Training vs. Agentic AI
Training AI models requires superfast chips, called general processing units (GPUs). These chips can run multiple tasks in parallel, which accelerates the training process. (GPUs were initially developed to render graphics for gaming, which is how they got their name.)
Agentic AI requires a different kind of horsepower: When an AI agent responds to a query, it doesn’t simply crank out a bunch of computations; instead, it reasons through a problem, reaches across databases and coordinates with other agents. GPUs weren’t designed for such orchestration. Those decision-heavy steps—which account for up to 90% of an agent’s total response time2—are the realm of central processing units (CPUs). And demand for those is rising.
For example, Nvidia’s latest Rubin data center “platform”—an integrated set of components with various processors, memory chips and switches—pairs one CPU with every two GPUs.3 (In a model-training setup, one CPU for every four to eight GPUs was considered sufficient.) AMD CEO Lisa Su recently doubled the company’s estimate of the total addressable market for server CPUs to over $120 billion by 2030, a 35% compound annual growth rate, driven predominantly by agentic AI.4 Yet CPU supply has not kept pace with ambition, and a renaissance moment for CPUs may have arrived: Since March 2026, Intel and AMD have raised server CPU prices by as much as 20%,5 with delivery windows stretching to six months.6
Up, Out and Across: The Challenges of Scaling Agentic AI
In the model-training era, networking was a relatively straightforward affair. In a typical network, clusters of training chips are concentrated and connected by the fastest possible short-range links. Within a cluster, data traffic was as orderly and predictable as a pack of Ferraris zipping around a closed track: fast and smooth.
Agentic AI is much messier. A single query can travel across data centers and cities, each step governed by a different system—one searching for information, another running calculations, yet a third assembling the response. Traffic is random, bursty and time-sensitive: A delay of milliseconds (acceptable during a model-training session) can break an agent’s entire chain of reasoning. With agentic AI, the quality of the racetrack has become as critical as the engine inside the Ferrari.
That track also has multiple, increasingly challenging levels: The “scale-up” level involves connections between chips within a single server, where bandwidth is already moving from 800 gigabits per second toward 1.6 terabits per second; the “scale-out” level includes connections between servers within a data center managing surging volumes of unpredictable agent traffic; hardest of all, the “scale-across” level governs connections between data centers across buildings, cities and even continents.
That’s where optical connectivity, including the fiber cables that carry data between data centers, comes in. We believe optical fiber is one of the most consequential infrastructure bottlenecks—and target for investment—in the AI era. Corning, a leading optical-fiber maker, inked a $6 billion multiyear supply agreement with Meta in January 2026, and earlier this month it struck an additional $500 million strategic partnership with Nvidia to build three new plants, expanding Corning’s U.S.-based manufacturing of optical-connectivity solutions by tenfold.7,8
But Corning won’t be the only beneficiary, in our view: As AI traffic explodes and connections multiply, we believe the optical cable market could be entering a growth phase broad enough to pack demand pipelines for the next couple of years, even for tier-2 vendors such as Japan’s Furukawa, India's Sterlite and China’s YOFC.
Longer, denser optical cables also demand more sophisticated signal management. Hence the rising demand for next-generation digital signal processing (DSP) chips from companies like Broadcom and Marvell. DSPs decode, clean and re-encode optical signals that can distort over longer distances. As networks grow wider, the connections among them will require smarter handling of the light that carries the data, providing a potentially strong tailwind for the DSP market.
The Investment Opportunity: Rebuilding the Racetrack
We believe delivering agentic AI at scale will require a new and improved information racetrack—an overhaul that could drive a multiyear hardware-upgrade cycle, including CPUs, DSPs, optical fiber and a variety of networking gear.
For investors, we believe Google’s decision to build two chips where one sufficed was not a footnote in the AI story. It was the starting gun to a new hardware race.