The Minotaur Global Opportunities Fund returned +4.0% in April vs. the MSCI ACWI (Net, AUD) +5.0%. April steadied the portfolio after March’s drawdown, but the more important point was reporting season. The print cycle validated rather than challenged the work we did during the selloff: memory is tighter than the market believes, AI infrastructure capex is broadening beyond GPUs, software is bifurcating, healthcare is doing more than its defensive quality implies, and European franchises are quietly working.

Prints, not promises

Reporting seasons turn stories into scorecards. April’s scorecard was clear: AI demand remains strong, but the market is becoming more discriminating about who captures the economics.

A year ago, “we are doing AI” could carry a stock. This quarter, the questions were harder: who is getting paid, who has pricing power, who is supply-constrained, and who is simply spending heavily in the hope that returns arrive later?

The hyperscalers are still spending at extraordinary scale – Google, Microsoft, Meta and Amazon collectively spent over US$130 billion on capex in the quarter alone – but the bottleneck is moving down the stack. Google said it would have generated more Cloud revenue if compute had been available. Microsoft talked about land, power, networking and servers. Meta pointed to memory pricing. Amazon said memory and storage costs had “skyrocketed.”

This is no longer a GPU-only story. Prysmian’s result showed how far the AI build can extend, with fibre spot prices roughly doubling in six months and the company finalising a >€1 billion hyperscaler supply agreement representing around 40% of existing global optical capacity. Saint-Gobain’s data-centre pipeline doubled year-on-year, NextEra reported 21 GW of large-load interest, and Qualcomm signed its first hyperscaler custom-silicon engagement. AI infrastructure is becoming an economy, not a trade.

Our exposure is spread across that widening profit pool: memory through SK hynix and Micron; networking and custom silicon through Broadcom and Qualcomm; power and grid through NextEra, Prysmian and Saint-Gobain; and the platform layer through Alphabet.

Software was also more nuanced than the market shorthand suggests. The “SaaSpocalypse” framing appears too blunt. April’s prints separated tools where AI increases the value of the seat from tools where AI erodes the value of human typing time. Atlassian’s 29% cloud growth and the rally in communications and contact-centre infrastructure names, set against Coursera’s enterprise stagnation, point to that split. The more salient question is: does an agent use your software, or replace it?

Healthcare was another bright spot. Eli Lilly delivered 56% revenue growth, 156% EPS growth and raised guidance 6% above consensus. Chugai grew core EBIT 17%, with an accelerating royalty stream. HCA reaffirmed guidance and continued buying back stock. These are not sleepy defensives. Lilly compounding earnings at GLP-1 quality plus pipeline optionality, at a PEG below 0.5x, is indicative of the opportunity.

Europe is similarly working. Barclays is delivering a 13.5% RoTE with NII growth and capital returns. UniCredit remains consistent in shape. Saint-Gobain is holding 15.5% EBITDA margins through a difficult macro. Indra’s defence backlog grew 279% year-on-year. The European narrative is better than the consensus headline suggests, and our positioning across banks, defence and select industrials reflects that.

Memory – the profitability of the bottleneck

In last month’s letter, we wrote that the market had mistaken inference efficiency for demand destruction in the post-TurboQuant memory selloff. Our view was that lower inference cost increases total usage, supply was sold out through 2026 with new capacity not arriving until late 2027, and Micron and SK hynix had become cheaper on better fundamentals. We added into the March pain.

SK hynix’s print provided evidence that the debate is moving in our favour. DRAM ASPs rose mid-60% quarter-on-quarter and NAND ASPs rose mid-70%. Operating margin reached 72%. Customers are signing multi-year volume agreements, the company is moving from net debt to KRW 35 trillion of net cash, and management is reviewing additional shareholder returns. It also cited agentic AI as broadening demand across all memory types.

The buyers are corroborating the demand side. Meta lifted its 2026 capex guide to US$125-145 billion and cited memory pricing as a key driver. When customers themselves flag memory cost inflation as a binding constraint, the supply-tightness story has moved from analyst thesis to operating reality.

The market still treats memory as an inventory cycle. That may matter at the edges, but AI is changing the economics. Memory is becoming a scarce, essential input into the most important capex cycle of the decade.

On current forecasts, Samsung and SK hynix are on track next year to be the number one and number two most profitable companies in the world, and the big three memory players have the potential to out-earn the Mag 7. That is not yet reflected in how the sector is discussed. Memory remains comfortably our largest portfolio exposure.

Talos at the gate

The next layer we have built inside Taurient is sector-specialist agents. The first is “Talos”, our specialist AI infrastructure and semiconductor analyst. Where Taurient gives us breadth, Talos gives us depth.

Talos has an opinionated view on the AI infrastructure stack, weighs competing data sources, tracks leading indicators across HBM, hyperscaler capex, packaging, memory, networking, power and custom silicon, and is designed to push back when our priors drift away from the data.

That last point matters. We are not trying to build a chatbot that agrees with us faster. We are trying to build a research system that argues with us better.

We built Talos because AI infrastructure is now too large, too technical and too fast-moving for a generalist process to cover well alone, and because we wanted to test whether sector-specialist agents could materially lift our research output. Early evidence suggests they can. The plan is to extend the architecture across healthcare, consumer, financials and beyond. We are building out a research team, except the headcount is measured in tokens rather than salaries.

Our own understanding of AI’s role in the investment process has changed dramatically. Two years ago, we thought AI would mostly help with idea generation, screening and summarisation. It now touches monitoring, modelling, thesis validation, transcript analysis, scenario work and risk review.

Decision authority still does not leave the humans. We make every trading decision and AI does not get a vote. Taurient outputs cite primary sources, and we have codified fact-checking into the system. But the surface area we can cover, the rigour of our scenario analysis, and the speed with which we can respond are materially different from a year ago.

That lived experience is another reason we remain constructive on AI infrastructure. Our own AI spend is up close to 4x year-on-year in 2026. Put another way, by the end of April, we had already spent more on AI than in the whole of 2025. We are not extrapolating demand from a slide deck. We are living it.

No idle hands

April helped separate mark-to-market pain from thesis damage. Reporting season validated much of the work we did during the selloff: memory looks structurally tighter, AI infrastructure demand is broadening, software is splitting rather than disappearing, healthcare quality is compounding, and Europe is producing better operational evidence than its market narrative implies.

Hesiod said, “Work is no disgrace; idleness is the disgrace.” With Taurient and now Talos, we can take the maxim almost literally. There is no reason for idle coverage or untested priors when the tools exist to keep working while we sleep. The agents do not replace judgement, but they make idleness harder to excuse.

April’s lesson was simple: the facts are still moving, the opportunity set is still broad, and the work is compounding.