SaltLight Capital's commentary for the third quarter ended September 30, 2025.
Dear Co-Investor
Our letter to you is arriving somewhat later in the quarter than usual. Admittedly, it has been a busy earnings season, but the more honest answer is that we have been working on some very intriguing potential investments for next year. As ever, we are unrepentant opportunists. We end up pulling on a thread of an idea and suddenly find ourselves tumbling down the rabbit hole.
Over the years, we have trained our noses to follow the subtle scent of opportunity – with the periodic risk that we discover we’ve been chasing a “value trap” wearing cheap perfume. In those moments, we occasionally envy our peers who keep to the reassuring paths of benchmark constituents to avoid going off-piste. Such is the unusual path that we have chosen.
In our last letter, we pondered the risk of an “AI Air Pocket” as market participants digest a growing gap between AI infrastructure capex and the demonstrable return on those investments. That scepticism has intensified in the last couple of weeks, alongside a general apprehension about the US economy and the government shutdown. For perspective, at the end of the quarter, we had 41% exposure to North America — dominant enough to make macro matter, even if we spend most of our time thinking in micro.
We have been “AI bulls” for a long time, investing in NVIDIA before ChatGPT entered the zeitgeist, and, five years on, we remain firmly convinced this is a multi-decade technology epoch. That said, our enthusiasm has become more discriminating about where we place our bets.
What has really sharpened our discernment is OpenAI’s recent deal-making 2 . The size, structure, and implied capacity behind these announcements have blown past even the upper bounds of our internal assumptions. The market, quite reasonably, has been forced to confront the obvious questions about over-capacity, irresponsible spending and a bubble frenzy.
For us, the question is more nuanced and, to be clear, remains unsettled at this early stage. The debate is not whether there will be an eventual return on investment, but who will capture the value. It is striking that, at this stage, aggregate industry returns on capital are very respectable; however, the profit pool is not evenly distributed across the industry between hardware, model providers and the upper software layer. For now, the fabs (TSMC and its downstream supply chain) and the infrastructure designer, NVIDIA, are capturing most of the value.
We believe that the “press release war” being waged over the last few weeks — via increasingly zealous capex announcements — is a strategic attempt to reshape industry dynamics and unlock the profit pool.
It is too early to know how this will play out, and we see heightened risk in the infrastructure layer precisely because it is absorbing so much capital, with a common thread that generally leads back to OpenAI.
Of course, one genuine breakthrough in token economics or model technology could change our calculus. For now, we keep doing what we have always done: watching closely, updating our priors, and trying to pick the best risk/reward spots in what is, undeniably, a tremendous opportunity set.
Our consistent investment approach is not to make predictions about exact outcomes, but rather to be humble, admit that we don’t know exactly what will happen, and structure our portfolio to absorb multiple outcomes.
As things stand today, we continue to find attractive risk–reward opportunities higher up the stack from businesses that should benefit from falling inference costs and improving model capability.
Our portfolio companies Tencent, AppLovin, and Google are already experiencing significant, tangible benefits from AI. Interestingly, they are not generating incremental profits from new generative AI use cases, but rather by replacing their CPU “recommender” engines with GPUs to enable hyper-targeting of content and ads across a broad user base. For these companies, the result is tangible cash-on-cash investment returns.
It is a more appealing bet on their unique distribution abilities and product ecosystems to implement AI technology at scale. Conversely, we have deliberately avoided the so-called “Neo Clouds” and former crypto miners, where we see, at best, only a narrow and fragile competitive edge.
We should note that despite the growing scepticism among equity investors, credit investors appear far more sanguine. In our previous letter, we argued that debt would have to step in to fund this build-out. The tech platforms’ own cash flows, while prodigious, were unlikely to be sufficient to both run the existing businesses and finance the entire compute build-out. On that score, we were not disappointed. This quarter, every major hyper-scaler has now tapped the debt markets. Most relevant for us, the timing and pricing suggest we are still early in the cycle: the issues were heavily oversubscribed and have traded at tight spreads.
Perhaps, we should step back and recognise that if the incremental supplier of capital is unfazed, this should guide us to the stage we are in. From the debtholder’s perspective, this sanguine risk pricing is logical. These companies are among the greatest cash generators in history. Some are still executing repurchase programmes and can raise prices if they ever face genuine debt stress. Moreover, although the absolute amounts of debt seem large, they are only a small fraction of these companies’ market capitalisations.
Going forward, equity investors will have to factor in credit markets to take the temperature of the AI infrastructure trade. There are a few things to consider when one stares into the dangerous eyes of the credit market. Firstly, credit (and specifically private credit) is likely to be the incremental supplier of capital for the next 10 GW of the AI built out. Secondly, the risk capital has been tapped out. Credit has a defined repayment tenor, and 50% of the underlying assets have relatively short lifespans. Lastly, credit funds are often leveraged themselves, which magnifies stress risks. At some point, the “AI Infrastructure” opportunity could start behaving like the real estate sector, which is heavily defined by banking, liquidity, and credit cycles.
*****
ZAR Strength
We should also note that a strong rand is a headwind to our reported returns in the upcoming quarter. We are a ZAR-denominated fund with roughly 80% of our assets offshore. A strong Rand is our kryptonite: one can pick the right businesses, at the right prices, in the right structural trend — and disappointingly still watch a bout of Rand strength quietly shave off a meaningful slice of those hard-won foreign returns when translated back into local currency.
New Fund Administrator
Finally, a brief note on some housekeeping. Our fund administrator has been acquired as part of the industry's ongoing consolidation. As a result, BCI has amalgamated our fund with effect from 21 November 2025.
Our investment strategy remains entirely unchanged: same philosophy, same process… same tendency to fall into rabbit holes. The changes are confined to the plumbing. In the back office, our custodial and bank accounts will move to new partners, and the investment application forms will be updated to reflect the new arrangements. Beyond that, you should notice very little difference — which, in the world of administration, is precisely the outcome one hopes for.
*****
As always, we'd like to remind our co-investors that your manager’s liquid wealth is invested in the same fund as yours. We share the inevitable ups and downs with you.
If you have any questions, please don't hesitate to get in touch.
David Eborall - Portfolio Manager
Read more hedge fund letters here
