Related Articles
Stay Up To Date
Something has gone wrong, check that all fields have been filled in correctly. If you have adblock, disable it.
The form was sent successfully
Since the release of ChatGPT in late 2022, the dominant narrative around artificial intelligence has centered on compute. More specifically, on GPUs. As increasingly powerful models required exponentially greater training capacity, accelerator chips became the foundational scarce resource of the AI era and the primary beneficiaries of the initial infrastructure buildout.
Yet technological shifts rarely remain concentrated in a single layer of the stack for long. As AI systems evolve from passive prediction engines into autonomous agents capable of reasoning, planning, and executing complex tasks, the nature of infrastructure demand is beginning to change. Increasingly, the limiting factor is no longer compute alone, but the infrastructure required to coordinate, feed, connect, and contextualize that compute.
In many ways, this marks a transition from the first phase of AI infrastructure toward the second. The first phase required immense computational power to train large models. The next phase increasingly depends on persistent memory, orchestration, and high-speed networking capable of supporting millions of continuously operating, context-aware AI agents. As a result, bottlenecks across the AI stack are beginning to migrate outward from accelerators alone and toward the systems that enable those accelerators to scale efficiently.
This shift is already becoming visible in the economics of the semiconductor ecosystem. While NVIDIA and Broadcom defined the first chapter of the AI era through their exposure to accelerator compute, leadership is broadening toward a wider set of infrastructure providers across memory, CPUs, and networking. Historically, many of these markets were viewed as cyclical and commoditized, with demand tied largely to enterprise hardware or telecom spending cycles. Today, however, AI appears to be transforming several of these industries into strategic bottlenecks within a rapidly expanding ecosystem.
From Compute Scaling to Context Scaling
The emergence of reasoning and agentic AI systems may represent one of the most important architectural shifts in modern computing. Traditional AI systems largely responded to prompts. Agentic systems increasingly operate autonomously, capable of managing workflows, retrieving information, interacting with software tools, and completing tasks end-to-end with limited human intervention.
The effectiveness of these systems depends heavily on context. The more of the relevant world an AI agent can actively access, including codebases, documents, prior decisions, workflows, and user preferences, the more capable and economically valuable it becomes. In practice, context increasingly functions as a form of working memory for AI systems, allowing them to retain information, maintain continuity, and operate with greater autonomy over time.
This dynamic is especially powerful in coding workloads, where the economic returns are already compelling. An AI coding agent that saves even modest amounts of time for highly compensated engineers can generate extraordinary productivity gains. But the difference between a basic assistant and something closer to a true software engineer increasingly comes down to memory and context depth. Agents capable of maintaining awareness across repositories, testing environments, style guides, and architectural history can perform substantially more valuable work than systems operating with fragmented information.
As AI systems become more persistent and autonomous, memory increasingly shifts from a supporting component to a gating factor in system capability. If the value of an AI agent scales with the amount of active context it can maintain, then demand for memory scales not only with the number of deployed agents, but also with the depth of information each agent continuously accesses.
At the same time, the supply side of the memory industry is becoming structurally more constrained. Advanced forms of memory such as High Bandwidth Memory (HBM) and Dynamic Random-Access Memory (DRAM) are becoming increasingly difficult and capital intensive to manufacture, while the lead times required to build cutting-edge fabrication capacity continue to lengthen. The transition toward HBM, which remains essential to advanced AI accelerators, further tightens supply because it requires materially greater manufacturing complexity and capital investment than conventional memory technologies.
The result is a market where memory demand may accelerate materially faster than supply can realistically respond. In our view, this dynamic may support a more durable period of favorable industry economics than has historically characterized the memory sector.
The Return of the CPU
While GPUs remain central to AI computation, the rise of autonomous systems is also driving a meaningful resurgence in the importance of CPUs.
Traditional AI workloads were heavily concentrated within the model inference process itself, with GPUs performing the overwhelming majority of economically valuable work. Increasingly, however, orchestration is becoming part of the workload. Autonomous AI systems require continuous scheduling, memory management, retrieval, planning, tool usage, and coordination across increasingly complex workflows.
Many of these functions occur outside the core inference layer and are significantly more CPU intensive. As a result, CPUs are becoming more deeply integrated into AI inference architectures than many initially expected.
This transition is already visible in changing deployment patterns across data centers. Traditional AI systems often paired one CPU with multiple GPUs, as CPUs primarily acted as lightweight coordinators for accelerator clusters. Emerging AI architectures require materially greater CPU resources per accelerator as orchestration and workflow management workloads expand.
In many ways, this reflects a broader shift in AI infrastructure. Historically, compute itself was the primary scarce resource. Increasingly, however, the scarce resource may become the ability to coordinate increasingly complex systems efficiently. That shift has important implications not only for CPUs, but for the broader architecture of AI infrastructure.
Networking as the Next Bottleneck
If the first AI bottleneck was compute and the second may be memory, the next increasingly appears to be networking.
As AI clusters scale, moving data efficiently between accelerators is becoming as important as the accelerators themselves. Modern AI systems rely heavily on parallel computation, requiring continuous communication between thousands of chips operating simultaneously across increasingly large data center environments.
This creates an important dynamic: networking demand scales faster than compute. Every additional accelerator added to a cluster must communicate with an increasing number of other accelerators across the system, driving exponential growth in bandwidth and interconnect requirements.
Historically, networking infrastructure was often viewed as a secondary consideration relative to compute. Increasingly, however, accelerator utilization rates themselves are becoming constrained by communication overhead and data movement limitations. In effect, AI data centers are transitioning from compute-bound architectures toward networking-bound architectures.
This shift is reshaping the economics of the broader networking ecosystem. Optical interconnects, advanced circuit boards, lasers, optical engines, and testing equipment are becoming increasingly strategic components of AI infrastructure. Technologies such as co-packaged optics, which integrate optical networking closer to compute, may ultimately become essential to scaling future AI systems efficiently.
Importantly, the significance of these technologies extends beyond simple hardware demand. As bottlenecks migrate outward through the stack, the systems responsible for enabling compute increasingly become as strategically important as compute itself.
A Broader Infrastructure Cycle
Taken together, these developments suggest the AI investment cycle may be entering a new phase. The first chapter of AI infrastructure was overwhelmingly defined by accelerators and training compute. The next chapter increasingly appears likely to be defined by the broader ecosystem required to support persistent, context-rich, autonomous systems operating at scale.
That does not imply all infrastructure markets will evolve equally. Many of these industries remain cyclical and historically prone to periods of oversupply. In our view, industry structure matters as much as end-market growth. The most attractive opportunities are likely to emerge in segments where accelerating demand intersects with structurally constrained supply, high barriers to entry, and increasingly consolidated competitive dynamics.
Memory may represent one such example. Connectivity may represent another, though portions of the optical ecosystem could ultimately prove more vulnerable to aggressive future capacity expansion. More broadly, however, we believe the key takeaway is that AI infrastructure is no longer scaling linearly with compute alone. Increasingly, it scales with memory, orchestration, and data movement.
Ultimately, the rise of autonomous AI systems represents more than simply another increase in computing demand. It reflects the emergence of a fundamentally different compute architecture, one that requires a broader, more interconnected, and more sophisticated infrastructure stack than the first phase of the AI era. In our view, that shift may create a durable and expanding opportunity set across select areas of the semiconductor and data center ecosystem in the years ahead.
Disclosures:
The views expressed are the opinion of Sands Capital and are not intended as a forecast, a guarantee of future results, investment recommendations or an offer to buy or sell any securities. The views expressed were current as of the date indicated and are subject to change. This material may contain forward-looking statements, which are subject to uncertainty and contingencies outside of Sands Capital’s control. Readers should not place undue reliance upon these forward-looking statements. There is no guarantee that Sands Capital will meet its stated goals.
All investments are subject to market risk, including the possible loss of principal. Recent tariff announcements may add to this risk, creating additional economic uncertainty and potentially affecting the value of certain investments. Tariffs can impact various sectors differently, leading to changes in market dynamics and investment performance. Companies that manufacture AI infrastructure are subject to risks including rapid technological change, competitive pressures, supply chain disruptions, regulatory developments, customer concentration, and fluctuations in demand for AI-related products and services, which may adversely affect financial performance and valuations.
As of May 15, 2026, NVIDIA and Broadcom were held across Sands Capital strategies. The specific securities identified and described do not represent all of the securities purchased, sold, or recommended for advisory clients. There is no assurance that any securities discussed will remain in the portfolio or that securities sold have not been repurchased. You should not assume that any investment is or will be profitable. The companies were chosen based on an objective criterion and represent the two primary providers of AI semiconductor chip design, as measured by market share.