Q1 2025 Review
The first quarter of 2025 was defined by a compression in risk appetite across technology equities, driven by macro uncertainty, renewed rate sensitivity, and a market recalibration following the emergence of more capital-efficient AI model architectures. The DeepSeek moment in late January introduced genuine uncertainty about the near-term demand trajectory for the highest-cost AI hardware — a development that warranted careful analysis rather than reflexive response.
Our Bottleneck-to-Ticker framework proved its analytical utility in this environment. The key question the framework forces is: which spending is discretionary, and which is forced? When that question was applied to the portfolio following the January volatility, the answer produced differentiated conclusions across our positions.
GPU compute at the frontier — the very highest-end accelerators from NVIDIA's Blackwell line — represents the category most sensitive to the efficiency argument. If models can be trained and served at equivalent capability for lower cost, the demand case for the most expensive hardware compresses at the margin. This is a legitimate concern worth taking seriously.
Power delivery, thermal management, and memory infrastructure are different. A data center that exists requires power infrastructure to operate. A facility under construction requires electrical equipment to be commissioned. A GPU cluster already deployed requires cooling to run safely. These are not discretionary line items — they are operational requirements. The spending happens regardless of which AI model architecture is preferred. Our concentrated exposure to these subsectors provided meaningful protection through the quarter.
Portfolio Performance and Changes
The Technology pool outperformed the broader technology sector through Q1, driven primarily by our infrastructure-weighted positioning. Power infrastructure names held up well relative to the AI hardware trade, as investors recalibrated toward capital that is required regardless of the AI efficiency dynamic. Our thermal management exposure was a particular source of relative strength.
We made three meaningful changes to positioning during the quarter. First, we reduced our single largest position by approximately 15% following a review that concluded our entry price discipline had been insufficiently applied when we added to the position in Q4. The thesis remains intact, but position sizing relative to current valuation warranted trimming. Second, we initiated a new position in a power electronics company that had pulled back to a level our technical framework identified as a high-conviction entry point. Third, we exited one smaller position entirely — a company where our research identified execution risk in their manufacturing transition that we had initially underweighted.
The Commodities pool performed well through a period of crude oil volatility and precious metals strength. Our energy positioning benefited from tightening physical markets in late February and March. Gold's significant move higher through the quarter was supportive of our metals allocation.
The DeepSeek Framework
The January announcement of DeepSeek's R1 model — a model that matched frontier performance at a fraction of training cost — deserves careful analysis beyond the immediate market reaction.
The efficiency improvement in model training has real implications for some hardware demand. If training compute requirements for a given capability level decline materially, the near-term demand path for the highest-end training accelerators is less linear than the prior consensus assumed.
However, the history of computing is unambiguous on one point: efficiency improvements in underlying technology have never led to less total computing. They have led to more applications, more users, and ultimately more total compute spending. This principle — Jevons Paradox applied to computing — is the framework we apply to the efficiency argument. Cheaper inference does not reduce the installed base of AI infrastructure. It expands the addressable use case for AI, which over time produces more demand for compute, and therefore more demand for the infrastructure that delivers and cools that compute.
We are not dismissing the near-term implications for training compute demand, which are real. But we are highly skeptical that infrastructure spending — power, cooling, memory, networking — declines as a consequence of training efficiency gains. The inference layer is the growth layer, and inference at scale has exactly the same infrastructure requirements as training.
Looking Ahead to Q2 2025
Our primary research focus for Q2 is the grid modernization theme, where we are building out new coverage in the Commodities pool. We expect to publish initial research on utility and IPP positions in May. We are also watching the copper supply picture closely — physical copper markets have tightened materially, and our framework is generating a high-conviction signal in the sector that we are working to translate into specific position ideas.
The macro environment for our strategy — forced-spend infrastructure, structural bottlenecks in power and memory, and commodity supply constraints — remains constructive. We enter Q2 with high conviction in our positioning and a research agenda that is advancing across all our major themes.
Ahijah Ireland Founder & CIO Green Zone Capital
