The Minotaur Global Opportunities Fund returned +0.2% in June, underperforming the MSCI ACWI (Net, AUD), which returned +3.0%. For FY26, the Fund returned +20.0% vs. +17.0% for the benchmark.

June was not a month where our view on AI infrastructure weakened. If anything, the fundamental evidence improved. But it was a month where our view on portfolio balance changed. The strongest theses can still become too large and too correlated for our book. The right response is not always to abandon them but rather to own less of them.

The coverage engine accelerates

Taurient, our proprietary AI research system, is now more than 600,000 lines of code. June alone saw 2,438 commits to Labyrinth, our main research repository, roughly 195 new company directories scaffolded, and 96 companies taken through finalised initiations.

We also added four new AI agents. Spectra covers global telecom operators. Vesta covers regulated and concession-based infrastructure. Tempo focuses on market structure, positioning, liquidity and timing. Vigil, our portfolio risk analyst, was the most consequential addition of the month.

That takes the team to 22 investment agents and five operations agents. More importantly, the agents can now collaborate across desks, flagging work to each other when coverage overlaps. Our aim isn’t to build isolated chatbots – we’re aiming to build something closer to an investment team: specialised, argumentative, cross-checking, and always at work.

Vigil and the decision to cut risk

The defining portfolio decision in June came from Vigil.

Vigil, on watch.

Vigil’s job is to read every holding’s thesis against the live portfolio and our risk outputs, then surface hidden common drivers, concentration, and sizing-versus-edge mismatches that a covariance matrix may not see. It is designed to ask uncomfortable portfolio-level questions before the market does.

The question that mattered this month was simple: what single event hurts the most positions at once?

Vigil’s answer was a simultaneous de-rating of the AI-capex compound trade and a violent unwind of the leveraged Korean carry structure sitting on top of it. His variance decomposition showed that the memory pair of SK hynix and Micron alone contributed roughly 30% of book variance, or 35% including NVIDIA. AI-hardware crowding was above the 90th percentile. At the same time, BofA’s Bull & Bear indicator had sat at 8.8, an extreme bullish/sell signal, for four consecutive weeks, while Korean equities were up 97% year-to-date and retail margin debt had risen to ₩38 trillion, up 140% since January 2025.

That was enough. We halved our AI infrastructure exposure from roughly 31% of the portfolio to around 15%. This was not a demand call. It was a concentration call.

We exited Hut 8, Broadcom, Infineon and Alphabet, and reduced Micron from an 8% position to 2%. SK hynix remains our largest position. It still trades at a discount to Micron, and we think a potential US exchange listing could become a meaningful flow catalyst.

The irony is that our conviction in the memory thesis rose even as our exposure fell. Micron’s FQ3 result was exceptional, with revenue, earnings and margins all well ahead of expectations, and HBM effectively sold out through 2027. The cycle still has room to run. But one reason to build risk systems is to listen to them when they tell you that being right and being oversized are not the same thing.

Where the capital went – Minotaur’s “Chicken and Chips” portfolio

The proceeds were redeployed into less correlated return streams.

Sterling, our financials agent, initiated Lloyds, Citigroup and JPMorgan, prompting us to start positions in all three. They join UniCredit, Barclays and Commerzbank. Banks are one of the few sectors where a hawkish rate backdrop can be a tailwind rather than a headwind, and UK banks in particular carry pre-hedged net-interest-income support into 2026-27. Capital came out of one of the most crowded trades in the market and moved into one of the least loved.

We also added to Korean gaming through Shift Up, Pearl Abyss and Neowiz, alongside existing positions in Krafton and CD Projekt. These are idiosyncratic content businesses, with upside tied more to releases, franchises and execution than to AI capex.

Elsewhere, we initiated Comcast, doubled Talen Energy – Joule’s top energy pick – and increased Spotify.

And then there were chickens.

We initiated Pilgrim’s Pride, one of the largest chicken producers in the US and Europe. It is cheap, cash-generative, and benefits from favourable protein economics as feed costs ease and demand holds up. After several months of violent moves in semis and memory, there is something attractive about an unglamorous business selling an uncorrelated product at a reasonable price.

Eli Lilly keeps compounding

Outside AI infrastructure, Eli Lilly continued to deliver.

Two developments mattered. First, orforglipron, Lilly’s oral GLP-1, produced strong Phase 3 data in type-2 diabetes, strengthening the case for Lilly’s oral obesity and diabetes franchise. Second, Medicare obesity-drug coverage begins from 1 July under the Trump administration’s pricing arrangement with Lilly and Novo. That is a major expansion of the addressable market, shifting obesity treatment further from cash-pay luxury to reimbursed medical care.

Lilly remains one of the clearest examples of what we like in healthcare: a company with the ability to grow the market, defend the franchise, and disrupt itself before competitors do.

The cost of intelligence is falling

One of our favourite charts this month came from Coinbase’s Brian Armstrong. The bars show Coinbase’s AI spend; the black line shows token usage. Token usage keeps climbing to new highs, while the dollar cost has decoupled and come down.

Coinbase AI spend (bars) vs token usage (black line).

The lesson is simple. You do not control AI costs with friction, usage caps and spend alerts alone. You control them with engineering: cheaper defaults, better model routing, aggressive caching, leaner context, and picking the right model for each task.

That is exactly what we are doing at Minotaur. We have spent a meaningful amount of time benchmarking models and building routing logic so that frontier models do the hard reasoning, while cheaper models handle extraction, summarisation and routine analysis.

Our own benchmarks show the same pattern. Claude Opus 4.8 scored highest on the first phase of our initiation reports, but cost roughly $33 per run. Minimax-m3 scored only slightly lower and cost $4.12.

Initiation-report model benchmark (score vs cost).

The signal for investors is simple: the cost of intelligence is falling. For all the debate about an AI capex bubble, usage can keep rising dramatically without cost rising in lockstep. Increasingly, the edge is the unglamorous plumbing around the model, not access to the model itself.

μηδὲν ἄγαν

The Delphic maxim μηδὲν ἄγαν means “nothing in excess.” It is inscribed at the Temple of Apollo at Delphi and is one of the foundational ethical principles of ancient Greece.

It is a useful phrase for a month like June. We remain constructive on AI infrastructure. We remain constructive on memory. We remain constructive on the companies building the physical backbone of the AI economy.

But a good thesis can still become too large. A winning trade can still become crowded. And conviction is not a substitute for balance.

June was a month where our research engine got stronger, our risk system became more useful, and our portfolio became less dependent on a single outcome. The work continues, but with a little less excess.