Artificial intelligence is reshaping industries at a pace that demands a disciplined approach—both to capture meaningful opportunities and to manage risks before they become visible in financial results. Over the past several quarters, the team has developed three analytical frameworks that guide how we think about AI across the portfolio. Together, they inform where we lean in, where we stay cautious, and where we have made proactive changes to reduce risk.
At the highest level, we see three fundamental shifts (see chart below) affecting the broader AI landscape. These shifts—the shift from expensive general-purpose chips to cost-efficient custom silicon, the shift from a single dominant AI model to a genuinely competitive multi-model landscape, and the shift from growth-at-any-cost to demonstrated return on investment—shape where we see opportunity and manage risk.
Our positioning in Alphabet (with its TPU AI chips and Gemini LLM model), Amazon (with its low-cost Trainium and Inferentia chips, which supports Anthropic Claude model development), and Broadcom (with its leadership position in custom AI chip development) reflects our conviction in the first two shifts, while our preference for businesses with clear AI monetization reflects the third.
The infrastructure underpinning AI—the vast data centers, specialized hardware, and power systems that make modern AI possible—requires staggering levels of capex investment. Individual facilities can cost tens of billions of dollars (cost estimates for an advanced 1GW AI data center range from $35-50B), and the largest technology companies are collectively committing hundreds of billions to build this out. Within that buildout, we have identified a few structural bottlenecks where demand is both durable and difficult to satisfy. Companies providing solutions for these bottlenecks occupy positions we find highly attractive, as demand may be more durable in a variety of capex growth scenarios.
The first bottleneck is cost. The scale of AI spending creates relentless pressure to reduce the cost of both training and, most importantly, inference (cost-per-query or token). This is driving a broad shift away from general-purpose processors toward custom-designed chips built specifically for AI workloads. Google, Amazon, and Broadcom—all portfolio holdings—are leaders in this transition. Custom chips also tend to be substantially more energy-efficient than their general-purpose counterparts, making them doubly attractive as data center electricity costs climb.
The second bottleneck is memory. Today’s most advanced AI models require a specialized type of memory—known as high-bandwidth memory, or HBM—that can move data to and from the processor at very high speeds. As AI models grow more sophisticated, handling longer conversations and larger context windows, more complex reasoning, and multi-step tasks, the demand for HBM is growing faster than the industry can supply it. This imbalance has transformed what was once a commodity-like memory market into one where the leading producers have meaningful pricing power and strong margins. We own Micron to capture this opportunity.
The third bottleneck is power. Data centers are extraordinarily energy-intensive, and the pace of construction is outstripping the capacity of existing electrical grids. There is simply not enough power available to support all of the facilities currently planned. This has created acute demand for reliable power generation—particularly technologies that can be brought online quickly. Siemens Energy, a leader in gas turbines, is owned in the global portfolio for exactly this reason.
The fourth bottleneck is return on investment and a path to AI monetization. Unlike the first three, this one is economic rather than physical. Companies investing heavily in AI are facing growing pressure from shareholders and boards to demonstrate that these investments will generate real, measurable returns. The market’s patience for speculative AI spending is narrowing. We position the portfolio toward companies that can draw a clear and credible line between their AI investments and revenue generation—through existing products, established customer relationships, and proven distribution channels. We avoid businesses where the path from AI spending to meaningful profit remains speculative or dependent on assumptions that have yet to be tested.
Our third framework addresses a fundamental question for every company in our portfolio: is AI a benefit, a risk, or a neutral factor for this business? We developed a four-bucket framework to assess this systematically, and we have evaluated every holding (and new idea) based on this framework (see chart below).
The first category includes clear AI beneficiaries, companies where AI creates meaningful opportunity with limited potential downside to the business. This includes infrastructure suppliers—providers of compute, memory, and power that are direct inputs into the AI buildout—as well as companies demonstrating real revenue acceleration from AI. Meta is a strong example: its application of AI to content recommendations and advertising targeting has produced measurable gains in both user engagement and revenue.
The second category comprises net beneficiaries with some uncertainty—companies that stand to benefit from AI but where those benefits could be partially offset by competitive pressure. These businesses tend to succeed when they have strong competitive moats, whether through loyal customer bases, network effects, or proprietary data, that make displacement by AI difficult. Google's integration of Gemini into its search product is a clear example of leveraging AI to strengthen an existing business rather than being disrupted by it.
The third category includes AI targets or vulnerable businesses—those with limited competitive differentiation whose products or services could be replicated by AI at lower cost. In our assessment, our portfolios avoid companies in this bucket. In some cases, positions were exited proactively when the risk became visible. Monday.com is an example: its software features, while strong, were assessed as potentially replicable by AI tools over time. The position was sold proactively as our assessment of the long-term competitive risk evolved.
The fourth category includes businesses that we believe are inherently insulated from AI. Some industries simply cannot be replicated digitally—in-person travel and hospitality being a clear example. Viking Holdings is a portfolio position in this category: no AI system can replicate the experience of a river cruise voyage, and AI is far more likely to benefit Viking's operations through cost reduction and improved operational productivity than to threaten its core business. Healthcare products and medical devices are tied to physical goods and FDA-approved clinical treatments, which represent another insulated category, as does defense technology. We are actively looking for these types of AI-insulated industries.
These three frameworks—the structural shifts, the infrastructure bottlenecks, and the company-level assessment—work in concert to inform how each portfolio navigates both opportunity and risk presented by AI development. In each of our strategies, more than half of the portfolio, in our view, is positioned in companies we believe to be clear or net AI beneficiaries. We have deliberately avoided companies in the vulnerable category, and in several cases took action well before market prices reflected those risks. The frameworks are not static or rigid, as we continue to revisit, refine, and add to them as the AI landscape evolves. What has not changed is our objective: to position the portfolios to capture the long-term opportunity AI represents.