The Minotaur Global Opportunities Fund returned 2.0% in August, 1.4ppts ahead of the MSCI ACWI (Net, AUD). The parts of the portfolio that hurt us most in July rebounded, with Freee, Micron and NVIDIA the strongest individual contributors. The short book that cushioned July gave some of those gains back.

Turnover was low, largely because the decisions that drove August had already been made: cutting concentration without abandoning the memory thesis in June, then adding to Micron and rebuilding NVIDIA exposure into July’s sell-off. In several places August was just the market catching up with evidence we thought was already sitting there.

So the more useful question this month isn’t what went up, it’s where price and evidence moved together and where the evidence actually changed our view.

When price and evidence agree

Start with NVIDIA. Its August result again showed demand for AI infrastructure running well ahead of supply. Revenue beat, guidance moved higher, and management still describes itself as supply-constrained. The part that matters more for us is that higher memory costs are now showing up directly in NVIDIA’s gross-margin outlook. A thesis that started as a view on memory suppliers’ economics is turning up in their customers’ P&Ls.

Freee is the better story, though. We bought the Japanese accounting-software company in July, and its August result delivered the operating leverage we had been waiting for. Sales and marketing fell materially as a percentage of revenue, management guided to a sharp increase in adjusted operating profit, and they announced a ¥5 billion buyback. The shares jumped on the result and kept going through the month.

Sales and marketing fell to 50% of revenue in FY2026, with management guiding to 45–48% this year.

Scoring memory against our tests

We’ve written enough about memory this year so rather than retell the thesis, August is a chance to mark it. In June we wrote down four things that would make us reconsider the position: a meaningful cut to hyperscaler capex, evidence of supplier inventory building, a normalisation in AI and server demand, or a material deterioration in SK hynix’s gross margin.

None of them showed up. NVIDIA’s result argued against both capex retrenchment and demand normalisation and memory pricing stayed extremely strong. In short, we saw nothing to suggest suppliers are accumulating unwanted inventory.

We also started a small position in Kokusai Electric, the Japanese semiconductor-equipment maker, which reported record revenue, strong orders and upgraded guidance. It’s another way into the same memory cycle, through the equipment needed to expand production rather than the producers themselves.

Our diversifiers are producing

August was also a useful check on the work we did in June to broaden the portfolio. Our June monthly introduced Vigil, our portfolio risk and analytics agent, whose job is to find hidden common drivers and alert us to concentration risk. His work fed into our decision to cut the portfolio’s dependence on AI infrastructure and move capital into less-correlated return streams.

One of those was Pilgrim’s Pride, one of the largest poultry producers in the US. At the time we wrote that there was something attractive about an unglamorous business selling an uncorrelated product at a reasonable price. JBS evidently agreed, proposing an all-stock take-private in August. Pilgrim’s rose around 10% and was one of the Fund’s stronger contributors.

Our Colombian oil producers also did well. Brent has rallied around 15% from its late-July low as renewed conflict with Iran disrupted traffic through the Strait of Hormuz. Colombian E&P independents are still amongst the cheapest oil equities anywhere, trading on roughly 2.6–3.2x EV/EBITDA even on conservative oil-price assumptions. They also have one advantage that suits this particular environment: crude exported from Coveñas avoids Hormuz, Suez and most of the tanker-war risk priced into Middle Eastern barrels. In a month driven by geopolitical supply risk, that made Colombian production look a lot like a “safe barrel”.

Crude from Coveñas reaches the US Gulf, and Asia via Panama, without passing Hormuz, Bab el-Mandeb or Suez. Most crude loaded inside the Persian Gulf has to clear Hormuz first.

Key Market Question 1: What does the AI build actually earn?

One of the bigger pieces of work we did in August asked a question we couldn’t answer off the top of our heads: who is actually earning the return on the hundreds of billions of dollars of AI infrastructure investment?

The answer is less obvious than the headlines suggest.

At the physical data-centre layer, contracted developments seem to earn fairly ordinary returns. Our analysis of a cohort of large projects points to low-double-digit yields, very close to what traditional hyperscale data centres earn. The AI premium, for the physical shell at least, is hard to find.

Accounting muddies things further. Builders can understate their true economic returns because rapid growth means much of the capital deployed hasn’t yet had time to generate revenue. Hyperscalers can temporarily do the opposite because the recent wave of capex hasn’t fully flowed through depreciation and so the numerator arrives before the denominator. This can give you a counterintuitive result: in a fast-growing capital cycle, rising reported returns can be a signal that growth has slowed rather than that the economics have improved.

The biggest open question is what today’s GPUs will be worth in five years. Nobody really knows and that residual value matters enormously to returns right across the stack. So NVIDIA’s willingness to provide very large residual-value support on some infrastructure arrangements is worth watching. In places it is helping underwrite the financing assumptions behind the build, not simply selling into it.

Our provisional conclusion is that the best economics may sit in merchant compute bought at the right vintage and price and in the financing layer funding the build.

Key Market Question 2: Has the SaaSpocalypse become a SaaSy comeback?

There’s a connection here to the SaaSpocalypse debate from earlier this year. Two reporting seasons on, we still have very little primary evidence of broad seat destruction across horizontal software. What we see instead is a widening gap between strong enterprise platforms and weaker mid-market and SMB vendors.

We came out of this reporting season more constructive on software but selectively so. Among the stronger companies, the value of revenue already contracted for the year ahead continued to grow at double-digit rates, existing customers were spending more, and cybersecurity companies raised their expectations for new business. But the weaker end stayed soft showing that this isn’t a uniform sector recovery. The durable demand is where enterprises are consolidating onto platforms with obvious strategic value.

The commercial model is also changing, which may matter more. GitLab, CrowdStrike and others are moving away from rigid per-seat pricing towards consumption commitments where customers commit to a dollar spend and can switch on additional products or modules as they need them. That is a good deal less threatening than the seat-compression narrative implied a few months ago. It makes revenue less dependent on customer headcount and gives customers an easy path to consuming more over time. Add early adoption of AI features, and meaningful AI monetisation in the near term starts to look credible.

The market feared AI would damage GitLab and CrowdStrike. For GitLab, the worry was that AI writing more code would make developer tools less valuable. In practice AI is letting more people build software, and that code still has to be stored, governed and managed. For CrowdStrike, agentic AI is expanding the attack surface, which increases the need for security rather than reducing it. Their August results backed both theses up. CrowdStrike raised its full-year net-new ARR growth outlook materially and GitLab delivered its first improvement in net revenue retention in years alongside growth above 20%.

We’re not buying software indiscriminately and we’re still cautious on the mid-market tier. But in the strongest franchises the evidence increasingly says AI is changing the software model rather than killing it. The AI value chain isn’t disappearing; the rents are moving around inside it.

The scoreboard has some preliminary numbers

In the June quarterly we said we had started keeping market-neutral conviction books for our specialist AI analysts but that there wasn’t enough history to show anything. We now have some numbers, with a large health warning attached.

Twenty-one analyst books have a track record and twelve are positive. The average return is around +1% and the median +1.2%, with outcomes ranging from roughly +21% to -10%. Start dates vary and two months is obviously far too short to conclude anything.

But that’s partly the point. A scoreboard is only useful if it can make its participants look bad and if every analyst is always right, the scoring system isn’t doing enough work. The dispersion is already forcing us to ask whether the differences come from sector conditions, stock selection, timing, or the quality of the underlying research.

The volume underneath all this keeps growing. During August Taurient completed more than 500 quarterly earnings reviews and initiated research on more than 200 additional companies. Most new ideas still end in a pass, which is fine, because more research only helps if we can get to a quick yes or no.

Gathering atoms

August was a month where most of the performance came from decisions made before it started. But a good August doesn’t prove every July decision was right, any more than July’s losses proved the opposite. The job is to keep separating what the market is saying from what the businesses are saying, write the tests down in advance, change our minds when the evidence demands it and continue to use Taurient to assist with all of that.

Democritus said: “Nothing exists except atoms and empty space; everything else is opinion.” Markets produce a great deal of opinion and our job is to keep looking for the atoms with Taurient helping us with our search.