Listen to the audio version of this article (generated by AI).
In 1943, Alan Turing wrote to his brother that machines could think.
He was wrong about the timeline, but he was right about the destination.
Today, we stand at that inflection point again: a fundamental shift is underway that will reshape who wins in the AI buildout, and which companies are positioned to capture the value.
The difference is that Turing never saw Marvell Technology Inc. (MRVL), ARM Holdings Limited (ARM) or Nebius Group Ltd. (NBIS).
These three names occupy different layers of the infrastructure stack, yet each benefits from one critical transformation: the shift toward agentic AI.
I talk about this in the latest episode of Being Exponential with Luke Lango, including why that transition could create an entirely new wave of demand for custom silicon, CPUs, high-speed networking, and AI cloud infrastructure.
The Shift That Is Already Happening
Query-based AI is query and stop. You ask ChatGPT to recap an earnings report. It thinks. It answers. Then it goes idle. This pattern repeats… start, stop, start, stop… Every time you issue a command, the compute starts. Once the answer is ready, the compute stops. The result? A series of discrete, sporadic compute demands.
Agentic AI is constant and recurring, meaning you can tell an agent to monitor Nvidia (NVDA) for daily news that could impact earnings and it does so without waiting for your permission. When earnings arrive, it recaps them automatically. It provides you with a daily briefing on stock movements, insider activity and Jensen Huang’s statements. Then it goes back to monitoring. The compute runs continuously. Every day, it fetches data. Every day, it completes tasks when instructed.
This pattern creates substantially higher bandwidth requirements for AI infrastructure. You cannot run constant inference on Nvidia GPUs alone at scale. You need custom silicon, upgraded interconnects and more orchestration capacity… or you choke under the traffic.
Marvell Technology (MRVL): Building the Chips and Highways for Agentic AI
Marvell Technology operates two businesses that benefit massively from this agentic AI shift.
First, the custom silicon business. As agents perform more tasks, cost per task becomes consequential at scale. Hyperscalers design accelerators around their workloads, memory requirements, and power constraints.
Marvell helps customers implement those designs. They have wins with Amazon (AMZN) for Tranium. They have wins with Meta (META). They have wins with Microsoft (MSFT). They have wins with Alphabet (GOOG, GOOGL) and the TPU program. Marvell is embedded in custom silicon world, probably second only to Broadcom (AVGO), and some may argue MRVL is first.
Second, the networking business. Inference systems need to move model data and intermediate results among processors and memory while applications access external data and services. Marvell supplies interconnect and switching technologies across racks and clusters. They have optical connectivity. They support emerging scale-up architectures. These are the highways for data centers. If you run constant agentic traffic, you flood those highways. You need more capacity. You upgrade from copper to fiber optics. That is where Marvell is a massive player.
The Numbers: Marvell reported $8.2 billion in fiscal 2026 revenue, which is calendar 2025. By fiscal 2031, they projected a 10x increase. That guidance update was way above estimates. Their long-term total addressable market expanded 4x beyond prior expectations. Stock surged to $300, breaking through the resistance shelf that held before momentum took over.
Valuation sits at 48x forward earnings and 40x forward EBIT. These are big multiples — but you have 45-51% revenue growth each year through 2029. You see EBITDA margins expand from 10% to 44%. EPS grows from $1.50 to $10.69. When you combine that growth with a multiple that already accounts for acceleration, this stock is technically attractive in the short term and fundamentally compelling over 12 months.
Arm Holdings (ARM): Agentic AI Powers the CPU Comeback
ARM Holdings is not just another chip designer. It licenses CPU technology to other companies, but now it is launching its first in-house AGI CPU with Meta as a lead partner. Sales ramp in 2027, then accelerate through 2028 and 2029. This launch coincides precisely with the surge in CPU demand inside data centers from agentic AI.
Query-based AI focuses on model capability. In other words, how well was it trained? Can it answer your questions?
Agentic AI focuses on inferencing and orchestration across clusters. You need CPUs to direct traffic between GPU clusters, show where data should go, and economically service models efficiently.
The CPU is having a massive resurgence because the infrastructure needs have changed.
The Growth Trajectory: ARM expects 24% revenue growth this year. Next year: 36%. The following year: 30%. Then 64%, then 57%. This is an accelerating story unlike most AI infrastructure equities, which decelerate into single digits by decade’s end. ARM accelerates through the mid-2020s and beyond.
Gross margins will drop from high 90s to low-mid 70s because building CPUs increases cost of goods. Yet you still see 30-40-50% EPS growth annually. The short-term technical chart shows a classic turnaround: after the big rally, sell-off and reinsertion into uptrend, ARM has retaken all moving averages, held the $200 support level and turned it back into support now at $100. Long-term fundamentals look even more attractive than the short-term technical picture.
Nebius (NBIS): Supplying the Compute AI Agents Will Consume
Agentic AI reinforces the neo cloud bull thesis: we remain compute-short long-term, so companies that supply marginal compute can charge high prices and margins. If everyone uses agentic AI, compute consumption growth is enormous. That favors neb clouds like Nebius.
The investment case for Nebius starts with a simple idea: AI demand is growing faster than the infrastructure needed to support it.
Agentic AI could make that shortage much more pronounced.
Think about how you use a chatbot. You ask a question. It generates an answer. Then you go back to whatever you were doing.
Now imagine an agent managing customer inquiries, monitoring inventory, or coordinating a company’s sales outreach. It repeatedly accesses data, runs models, checks results, and takes additional steps. Each completed task can require many separate rounds of computation.
Multiply that activity across millions of businesses and consumers, and you start to see the opportunity.
We’re moving from occasional requests to recurring workloads.
That reinforces the bull thesis for neoclouds: specialized cloud providers supplying the additional AI computing capacity the market needs. As long as that capacity remains scarce, providers that can bring it online reliably have an opportunity to command attractive pricing and grow rapidly.
And among those providers, Nebius stands out for the sheer ambition of its growth trajectory.
The Growth Opportunity: In the projections discussed on the podcast, annual revenue rises from roughly $530 million to about $50 billion by 2030.
That is an extraordinary expansion. It also sets an extraordinarily high bar.
Nebius needs to secure power, deploy hardware, bring capacity online, and convert customer demand into revenue. Strong demand creates the opportunity; execution determines how much of it reaches shareholders.
But the shift toward agents strengthens the demand side of that equation. If businesses increasingly rely on AI to perform ongoing work, they will need substantially more infrastructure to keep those systems running.
That is why I think agentic AI makes Nebius’s ambitious growth outlook more plausible.
The Next AI Winners Will Help Agents Get to Work
The AI investment story keeps evolving.
First, the race was to build models that could answer our questions. Now, the opportunity is expanding to systems that can perform work on our behalf.
That transition changes what the infrastructure must deliver.
More specialized chips to make each task economical. More networking capacity to move data. More CPU resources to coordinate workloads. More computing capacity to run it all.
Marvell, Arm, and Nebius each address a critical piece of that buildout.
There are plenty of other potential winners. But when I weigh the opportunity against the risks, these three sit near the top of my list for the next 12 months.
As AI moves from answering questions to completing tasks, the companies supplying its infrastructure could see a much larger market open up.
For the full discussion of these three stocks — including the growth expectations and valuation assumptions behind the thesis — watch our .
