During the second quarter and first half of 2026, Delphi generated returns of 21.54% and 17.15% respectively. Most of the return in this period came from investments in the semiconductor sector. Our selection within this area has been highly selective and continues to be so, for reasons we outline below.
This year’s strong return has not been trivial. Part of the challenge stemmed from the war with Iran, but the more meaningful part came from market dynamics that drove share-price declines in a number of companies we invest in. We believe much of this price movement, both up and down, is not economically grounded, which led us to sell certain positions at a profit while increasing our investment in other companies whose low share prices now represent opportunities.
The two best-performing sectors this year have been technology and energy. Energy benefited from the rise in oil prices following the war with Iran. We have no way to estimate how or when the conflict will end, and with a large number of possible scenarios, forecasting oil prices over the coming months or years is very difficult. For that reason we hold no investments in that sector. When we cannot assess future earnings with high confidence, we avoid investing.
In technology we do find companies whose future earning power we have conviction in, and we invest in a select few. Although the sector performed well overall, dispersion within it has been — and continues to be, wide: chip-company share prices soared, while software-company share prices fell.
The Semiconductor Sector and Artificial Intelligence
The massive demand for chips is being driven by the enormous capital expenditure being made by the cloud companies, which are building data centers at an unprecedented pace to meet ever-growing demand for compute. A meaningful share of that demand stems from increasing use of AI tools.
This phenomenon presents several challenges. The first concerns the efficiency with which AI tools are actually used. On one hand they genuinely contribute in many cases, especially in technology and content creation. On the other, there is significant inefficient usage. For example, converting a Word file to PDF is a simple button-click that requires minimal compute, yet many users ask an AI model to perform the conversion, a task that consumes far more compute and takes longer than the “old, simple” way. Another example: people asking an AI model for the weather, while every mobile phone already has a dedicated, efficient app.
Given this inefficiency, we are starting to see the beginning of a trend toward selectivity about which tasks are worth assigning to AI tools, since the cost of these operations is often not trivial. This raises a question: is there economic justification for such widespread use of these tools, and by extension, for building out data-center capacity on such a massive scale?
A further challenge lies in the dynamics of the memory-chip market. This applies both to RAM chips, which serve as high-speed working memory for compute, and to NAND chips, used for long-term storage of large data volumes.
The point can be illustrated through the RAM market, dominated by Samsung, Hynix and Micron. Given the sharp rise in chip demand, these companies did not have time to build enough new fabs, a fab typically takes three years to come online. The result has been demand well above supply, driving memory prices up. Volumes sold are at record highs, prices are at record highs, and so are the profits at those three companies.
In response, memory manufacturers have announced unprecedented capacity-expansion plans. Samsung and Hynix have announced roughly half a trillion dollars of spending in the coming years, while Micron expects to spend another quarter-trillion over the coming decade. New fabs are expected to begin production as early as 2027 and will increase supply. If that supply growth meets a slowdown in demand, or even a temporary pause, chip prices could collapse at a time when the companies are committed to vast spending. Even a temporary slowdown of this kind could materially affect their profits.
To participate in the growing demand for chips without taking the memory-market risk described above, we added Broadcom (AVGO) and Marvell (MRVL) to our portfolio. Both help Google, Microsoft and others design and produce chips, but without the memory-market exposure.
The hype around AI tools has led to sharp run-ups in chip-stock prices, in some cases without regard for risks like those we’ve described. Alongside that, the prevailing view holds that AI will cause software companies to disappear. We see this as a mistaken framework.
Some software companies are more vulnerable, but others are using AI tools to streamline internal workflows and build new customer-facing capabilities, growing their sales and earnings in the process. It seems the investing public has not bothered to look at the data showing a reality quite different from the general narrative.
As a result, we have seen companies grow earnings while their share prices decline. In some cases this has created very attractive opportunities, which we are happy to use.
The Economics of AI Compute
Negative sentiment also surrounds the companies building out data-center capacity. The fear is that their massive investment will not be met by sufficient compute demand. We think that scenario is highly unlikely. Yes, there is inefficient use of AI tools, but even if an efficiency wave takes place, long-term demand for compute will continue to grow. And even if demand softens somewhat, these companies can slow their investment pace to protect cash flow.
Investors’ concerns stem from the fact that today’s heavy capex is weighing on these companies’ cash flows, and in some cases they are even borrowing to accelerate the build-out. Three points are worth keeping in mind here. First, as noted, these companies control their own investment pace. Second, they are not building on hope, demand already exists, and every unit of compute capacity is sold almost immediately. Third, these companies can deliver AI compute to the market at substantially lower cost than peers without physical infrastructure.
We estimate that Google’s cost per unit of AI compute is roughly 80% lower than that of companies buying compute externally. Google designs its own chips, operates its own infrastructure, and builds its own models and AI tooling. Because it does not have to purchase processors from a third party or rent compute from a third party, its costs are materially below those of companies that do not own this stack, such as OpenAI.
We expect that as AI tool use, and compute demand more broadly, continues to grow, more attention will be paid to the economic efficiency of running these tools. When attention turns to cost, certain companies have clear advantages, and we believe that over time they will capture them.
Summary
As always, we remain focused on the earning power of our companies and on the level of conviction we are able to build around their future ability to generate profits. We continue to invest in a select group of companies in the areas described above (chips, software and compute infrastructure), alongside a number of companies in other areas, profitable, growing companies that we believe are mispriced relative to their future earning power.
In keeping with every year of our activity, we also continue to avoid investing in companies where there are too many question marks about their ability to earn growing amounts of money in the coming years, even when current data looks very impressive. We do not let tall towers on shaky foundations dazzle us; we stay focused on a high degree of certainty about future profitability.