In our previous report, we outlined the $5.5 trillion AI capex supercycle and highlighted three critical bottlenecks that could ultimately determine how much value the ecosystem captures.
The capital is already flowing. The key question now is which supply constraint becomes the limiting factor first. As the industry shifts from training large models to deploying autonomous AI agents, all three bottlenecks are being fundamentally reshaped.
Memory is emerging as the first major pressure point, and Micron’s latest earnings delivered a clear signal of just how rapidly demand is accelerating.
This Week’s TechEdge
- CPUs: the most overlooked growth inflection
- Memory: the new battleground and what Micron’s results reveal
- Networking: our highest-conviction opportunity
- The Bottom Line: key implications for investors
The traditional hardware equation is changing. AI training was primarily a GPU-driven story, with large clusters of GPUs supported by relatively few CPUs.
Agentic AI introduces a different dynamic. Like human workers, AI agents require dedicated compute resources, their own memory footprint, and continuous communication with other systems. This reverses the hardware ratios that defined the training era and channels significantly greater demand toward infrastructure categories that were previously secondary considerations.
We believe the agentic AI era could generate roughly three times the hardware spending of the training era over the next two to three years.
CPUs: The Most Overlooked Inflection Point
As AI agents become more prevalent, infrastructure requirements move closer to a one-to-one CPU-to-GPU ratio, since each agent requires orchestration, task management, and system coordination.
We project the CPU market could expand from approximately $35–40 billion today to more than $200 billion by 2030. That outlook exceeds AMD’s estimate of roughly $120 billion and is well above the current Wall Street consensus of around $170 billion.
To put the scale into perspective, Cloudflare estimates that supporting 100 million knowledge workers in the United States would require roughly 10 million CPUs, with global demand potentially approaching one billion CPUs. AMD, Arm, and Intel have all pointed to this emerging trend, and 2025 appears to mark the beginning of the transition. While related stocks have already responded, we believe current valuations reflect only the early stages of a much larger growth cycle.
Memory: The Battleground — What Micron’s Results Revealed
This quarter, the memory story moved beyond theory and into reality.
Micron delivered strong results across every major segment, with the upside driven primarily by pricing rather than volume growth. The most striking figure was adjusted gross margin, which surged to nearly 80%—an extraordinary level for a business that has historically generated margins of 30–50% during favorable cycles and often slipped into negative cash flow during downturns.
Memory pricing has climbed roughly sevenfold from the cycle trough, creating a powerful earnings tailwind for Micron. Those higher prices are flowing directly into the company’s profits while simultaneously increasing infrastructure costs for hyperscalers and AI leaders such as NVIDIA, Alphabet, and Microsoft.
The takeaway is clear: memory is no longer just a supporting component in the AI stack. It is becoming one of the most critical constraints in the industry, with pricing power increasingly concentrated among the companies capable of supplying it.
The forces driving this trend are structural rather than cyclical.
Agentic AI is creating a second, distinct source of memory demand alongside high-bandwidth memory (HBM). While AI training workloads require HBM, autonomous agents also rely heavily on conventional DRAM and NAND—the same memory technologies found in everyday PCs and enterprise systems.
At the same time, manufacturing capacity is becoming increasingly constrained. Producing HBM consumes significantly more resources, with each HBM wafer requiring the equivalent of three to four conventional DRAM wafers. As memory manufacturers shift capacity toward HBM production, the supply available for traditional memory products tightens just as demand for those products is accelerating.
The result is a convergence of two demand streams that historically moved independently, now competing for the same limited supply base. According to Micron’s management, there is still no clear timeline for when supply—particularly in HBM—will fully catch up with demand. We believe these constraints could persist through 2028, extending the favorable pricing environment for memory suppliers and reinforcing memory’s position as one of the most critical bottlenecks in the AI infrastructure stack.
Perhaps the most significant development is not technological, but contractual.
Micron announced 16 strategic customer agreements with fixed-price structures, each spanning roughly three years and representing a combined minimum value of approximately $100 billion through 2030.
Historically, memory has been sold largely on the spot market, a key reason the sector has long traded at a valuation discount due to its cyclical nature. The emergence of multi-year, fixed-price agreements has the potential to reduce earnings volatility, soften future downturns, and support the case for a structural re-rating of memory companies.
That said, we are not prepared to fully underwrite that thesis yet. The real test will come when these businesses generate strong cash flows through an entire down cycle—something that remains unproven. These agreements also carry execution risk. If AI infrastructure investment slows, customers could scale back commitments, while increasing competition from Chinese memory producers remains a meaningful long-term consideration.
For that reason, we continue to favor the picks-and-shovels approach. Companies such as KLA, Lam Research, and Applied Materials supply the critical tools required to expand manufacturing capacity, allowing them to benefit from industry growth without taking direct exposure to memory price cycles.
One emerging catalyst deserves close attention. Micron has highlighted humanoid robotics as a potential second wave of demand growth. These systems could require roughly ten times the memory capacity of today’s AI deployments, creating an additional demand engine that may extend the current cycle well into the next decade.
Networking: Our Highest-Conviction Opportunity
As AI agents become more prevalent, the volume of machine-to-machine communication is set to increase dramatically. Every interaction, task delegation, and data exchange generates network traffic that must be moved efficiently across increasingly complex infrastructure.
The debate is often framed as a choice between copper and optical connectivity, but we believe that view is overly simplistic. Both technologies will play important roles at different performance, distance, and cost thresholds. Rather than betting on a single transmission medium, we prefer technology-agnostic infrastructure providers such as Astera Labs and Credo Technology Group, whose products enable faster and more efficient data movement regardless of the underlying architecture.
Signs of strain are already emerging across the ecosystem. Lead times for certain optical networking products have reportedly extended to as much as 12 months, while fiber pricing has risen approximately 50% since the start of the year. These pressures reflect a market struggling to keep pace with accelerating AI infrastructure demand.
Industry forecasts suggest the optical networking market could ultimately exceed $150 billion in value—roughly nine times its current size. While networking has already outperformed both memory suppliers and hyperscale cloud providers during this cycle, we continue to see the greatest potential for upward earnings revisions in this segment, alongside semiconductor capital-equipment companies.
As AI workloads evolve from model training to large-scale deployment of autonomous agents, networking is increasingly becoming a mission-critical layer of the infrastructure stack. In our view, this remains one of the most compelling opportunities across the AI value chain.
The Bottom Line: What This Means for Investors
The AI opportunity is no longer defined by who has the most capital to deploy—it is increasingly determined by who controls the most constrained resources.
CPUs are entering a major growth inflection and, in our view, remain underappreciated by the market. As AI agents proliferate, demand for orchestration, coordination, and general-purpose compute is set to rise sharply, creating a powerful tailwind for the CPU ecosystem.
Memory has moved from a future thesis to a present reality. Pricing power is already reshaping industry economics, and the market has begun to reflect that shift. However, much of the near-term optimism is now embedded in valuations. Whether memory suppliers deserve a sustained re-rating will ultimately depend on their ability to generate durable cash flows through the next downturn. Until that is proven, we prefer exposure through semiconductor equipment providers rather than the memory manufacturers themselves.
Networking remains our highest-conviction investment theme. As AI systems become increasingly agent-driven, the volume of data moving between machines will grow exponentially, making connectivity infrastructure one of the most critical—and scarce—components of the AI stack. We continue to see the strongest potential for earnings upside in this segment.
The broader takeaway is straightforward: in the agentic era, value accrues to the owners of scarcity. For years, GPUs were the primary bottleneck. Today, the constraints are shifting toward compute orchestration, memory capacity, and network infrastructure. Investors who identify these emerging choke points early will be best positioned to capture the next phase of AI-driven growth.
THE ETERNAL SOVEREIGN 
Leave a comment