The Minotaur Global Opportunities Fund returned -3.3% in July, underperforming the MSCI ACWI (Net, AUD), which fell -1.3%. The result was unusually concentrated: SK hynix, Artrya and Micron cost ~5.1ppts between them, more than the Fund’s entire decline, while all other security-level contributions added 2.2ppts.

That distinction is not a way of dismissing the loss but it does change the question we need to answer. What mattered in July was not how many things went wrong but whether the evidence changed in the few places that did.

There were meaningful offsets. Europe added ~1.1ppts and the short book added 0.6ppts. CD Projekt, Parex, UniCredit, Crocs and Freee were among the stronger individual contributors, while the SpaceX and Tesla shorts did most of the work on the short side. The portfolio outside the three largest detractors was not standing still.

Three positions explain more than the whole loss

The attribution is useful because it separates a broad portfolio failure from three live research questions. In Artrya, the operating evidence changed and exposed a real commercial bottleneck. In memory, the stocks fell sharply but the strongest new fundamental evidence came from a customer confirming that pricing pressure was still intensifying. In SpaceX, a known technical catalyst brought billions of dollars of forced demand and still failed to establish a floor. The share-price moves were all violent but when we did the analysis of them we got comfortable.

A billing bottleneck is better than a demand problem

Artrya was July’s second-largest detractor. Its June-quarter update contained genuine progress with Northeast Georgia Health System live as the second revenue-generating customer, all five Tanner hospitals live, and Cone Health still moving through integration. Management continued to expect all three foundation systems to be generating revenue from both the core anatomy platform and the plaque module by the end of the year.

The market’s concern was more immediate. Cash receipts had not yet caught up with the commercial ramp implied by the company’s milestones and some brokers had far too optimistic expectations around this. It turns out that the clinical workflow and the payment workflow had moved at different speeds. Artrya’s software can read a scan in minutes, but a US private insurer’s pre-authorisation can take two to seven days. The immediate bottleneck was therefore not whether physicians wanted to use the product; it was whether the hospitals’ revenue-cycle systems could obtain approval, document the service and bill for it without introducing too much friction. Management plans to introduce semi-automated and then automated workflows through August and September. That explanation is better than discovering that clinicians do not value the product and in fact, our primary research has confirmed quite the opposite.

The scan takes minutes; getting paid for it takes days.

We had trimmed the position earlier in July, but Artrya remains a top-ten holding. The evidence required from here is straightforward: rising scan volumes, actual plaque billings, a measurable attach rate, Cone Health going live, and progress on the Flow submission. We have confidence that all of these come through.

The evidence is unchanged in memory, only the price

SK hynix and Micron together accounted for the whole of the Fund’s decline. This was the concentration risk we acted on in June, when we cut AI-infrastructure exposure from roughly 31% of the portfolio to ~15% thanks to Vigil’s, well, vigilance. This materially reduced the amount of the portfolio exposed to the same crowded unwind. (See our June quarterly for the full memory thesis.)

The important thing is that June was a risk-management decision, not a change in our view of memory demand. We reduced the size of the position even as the operating evidence improved. During July’s sell-off, we added to Micron on three occasions and rebuilt some NVIDIA exposure. NVIDIA’s relative resilience was also notable: it rose while the memory producers fell sharply.

On Apple’s 30 July earnings call, Tim Cook described the market as a “hundred-year flood” in memory pricing. Apple said it paid significantly more for memory in June than in March, expects to pay even more in September, and sees pricing continuing to rise beyond that. Its CFO said memory costs more than explained the sequential decline in adjusted gross margin from March to June and the further decline implied by September guidance.

Apple's gross margin, adjusted for tariff refunds: memory costs more than explain each step down.

Buyer-side corroboration resonates highly. Memory producers have an obvious incentive to describe a tight market whereas Apple has an equally obvious incentive to resist higher component costs. When one of the world’s most powerful purchasers says memory is more than accounting for its margin pressure, the supply-demand imbalance has moved beyond the producers’ narrative and into the customer’s P&L.

Apple’s evidence supports the duration of memory pricing but it does not tell us that positioning risk has disappeared, that every producer has equal upside, or that the drawdown has ended. SK hynix rebounded by roughly 30% on 31 July alone. That is evidence of how crowded and volatile the trade has become, not a one-day change in intrinsic value. We continue to balance fundamental evidence and technical flows as a result.

Forced buyers could not hold the price

Our SpaceX short was one of July’s largest positive contributors. SpaceX entered the Nasdaq-100 before the market opened on 7 July and J.P. Morgan estimated that index inclusion could attract approximately US$4.3 billion of passive buying as funds tracking the benchmark adjusted their holdings. Yet the shares fell by roughly one-third.

SpaceX in July: index inclusion on 7 July, an estimated US$4.3bn of passive demand, and a one-third decline anyway.

Index funds do not buy because a price is attractive; they buy because the rules require it. That demand can be large but it is also finite and once the rebalance is complete, the marginal buyer can and does matter again. In this case, mechanical demand provided liquidity but not a durable clearing price.

It is a useful contrast with memory. In one case, a violent share-price fall occurred while the underlying company evidence improved. In the other, billions of dollars of forced buying could not overcome the market’s willingness to sell.

When reasoning gets cheap, truth gets expensive

A new investment idea in July was a basket we call “The Source”. Most discussion of AI economics begins with the models. We think an important part of the value chain may sit one layer earlier. As access to reasoning becomes broader and the cost of generating an answer falls, the scarce input is increasingly not the ability to write a fluent response but the right to use trustworthy, licensed, point-in-time facts.

An AI agent can draft a credit memo in seconds but cannot recreate decades of ratings history, a legally defensible index methodology, a point-in-time record of corporate actions, or an authoritative legal, scientific or tax corpus. In high-stakes workflows, provenance is part of the product.

Companies like Moody’s and S&P Global own credit data, ratings histories and workflow positions that are difficult to reproduce legally and institutionally. MSCI owns indices, risk models and datasets embedded in capital-allocation processes. RELX and Thomson Reuters own specialist legal, scientific, tax and financial corpora where the source, timestamp and rights attached to the information matter as much as the words themselves. All of these are sources of truth so to speak.

AI agents create a new distribution surface for those assets. A data owner may be able to charge not only for human seats and traditional feeds but also for the machine consumption: more frequent queries, more granular retrieval and more workflows that require an auditable source.

The basket, however, is not a generic bet on every financial-data vendor and there are ones that we’ve excluded while we verify how much of their value rests on truly proprietary data rather than aggregation and delivery. That selectivity matters because the central risk to the thesis is substitution rather than expansion. Agents could reduce human seat revenue faster than machine-consumption revenue grows. Furthermore, rights leakage or direct deals between frontier model companies and original data owners could weaken aggregators.

The thesis will be tested by whether these businesses can price machine use, preserve their rights and expand consumption without cannibalising the economics of their existing products. But the direction of travel seems clear: AI may commoditise answers while increasing the value of authorised facts.

What appears is a vision of the unseen

Though the visible result in July was a loss, underneath it were three different mechanisms: an operating bottleneck at Artrya that became more understandable but not yet resolved; a crowded memory trade where important new company evidence strengthened; and a technical catalyst in SpaceX that failed to support the shares. We know that losses are not made better by being explainable but explanations matter because they determine what we do next. July proved out Anaxagoras’ notion that “What appears is a vision of the unseen.” This month, as we do every month, we gather evidence on how we should treat the future and act accordingly.