SaltLight Worldwide Flexible Fund – Investor Letter 4Q23

                                                                                                23 February 2023

PDF Version

Dear Co-Investor

The SaltLight Worldwide Flexible Fund returned 30.8% in 2023 – the final quarter turned out to be a very strong finish for the year with a +15.4% return.

What went in our favour? We were fortunate to have some exposure to some of the ‘Magnificent Seven’ – Amazon, NVIDIA, Meta Platforms and Google (although in aggregate, we still hold a smaller weighting than the S&P500).  The return wasn’t all due to the so-called ‘AI Winners’. Our investments in MercadoLibre, Roblox, and the Brookfield entities contributed significantly to the return.

Whilst we are pleased with the fund’s result, the extreme variance highlights to us, once again, that predicting short-term stock market gyrations is practically impossible.

Throughout 2023, we had the belief that the fund was undertaking a year of ‘sowing’ – laying the groundwork for our 2028 investment horizon with the possibility of ‘green shoots’ appearing in 2024 (we’ll talk about these later in the letter).

Our internal metrics indicate that, when accurate, our capital deployed into a business is typically early. However, by the 18 to 24-month mark, our investment theses start to manifest in tangible metrics and are subsequently recognised by the stock market.

We’re left with a quandary about attributing our skill or luck to the fund’s performance in 2023. It’s a tricky business to try to divine whether the market is finally catching a glimpse of what we’ve been seeing all along or if we’re just riding a wave of favourable risk conditions.

Turning to 2024

At SaltLight, we’ve been somewhat candid about our positive view of AI and its transformative potential across various industries. In our 1Q23 letter, we showed this chart, an extract from Carlota Perez’s book “Technological Revolutions and Financial Capital”.

Figure 1 – Technology Epochs (Technological Revolutions and Financial Capital)

Summed up, all technology epochs take a long time, but follow a broadly similar diffusion framework over a multi-decade period. Where are we today? We’ll tell you with perfect hindsight in ten years. But the price-inferred odds and some broad unknown, unknown questions make us cautious in the AI Infrastructure space.

How will all this AI infrastructure capex generate a return on capital over the next five years whilst technology diffusion is still very low?

History tells us that if a technology incumbent does not adapt in the early phase of an epoch change, it can fall behind. Therefore, the rationale for the Large Tech spend is somewhat logical – in the hope that if spaghetti is thrown against the wall, some will stick.

The issue is that the real AI opportunity in the enterprise is unlikely to happen within the next 2-3 years as most are projecting. Not because of AI technology but because firstly it’s still too difficult for even the most skilled enterprise to create anything of substantial value for a corporate.  To get to a place where enterprises can use AI on their domain data, they’ll need to substantially rearchitect how they store and utilise data. This will take time.

Secondly, AI compute is just too expensive and therefore hinders every-day adoption.

How does this current batch of capex not become rapidly obsolete?

We are acutely aware of the rapid pace at which AI technology, particularly large language models (LLMs), is advancing. The heuristic is: LLM performance doubles with a tenfold increase in data and compute emphasises a marked departure from traditional Moore’s Law scaling (2x every 18 months). Today’s cutting-edge AI infrastructure could swiftly become outdated.

Our general unease is that what we started with: the market can quickly overestimate what can be done in the short term and underestimate what will happen in the long term.

We believe that the market has discounted ambitious adoption rates in AI infrastructure businesses by a wide margin. There is a high risk of disappointment and so we have pared some of our AI infrastructure investments. It’s a tough thing to sell a broad opportunity set that has incredible potential but at these prices, we think the positive expected value in AI opportunities now rests in the hands of certain AI software applications (more in this letter). While there’s always the risk that our assessments may miss the mark, these decisions are made to balance our portfolio’s risk in terms of our resilience and optionality framework.


In this letter, we’re going to talk about three portfolio companies: Meta Platforms, AppLovin and Transaction Capital.

While we are cautious in AI infrastructure, we do think there are mispriced opportunities in areas of application software where AI can be infused to make a step-change improvement.

Posted in our office is this chart that ASML provides at each of its investor days. This chart is a little outdated from 2021, but we think illustrates how value (in operating profit) was distributed across semiconductors, hardware, and then software & services.

Figure 2 – Operating Margin Distribution in Technology (Source: ASML)

It’s very clear that most of the economic value in the past has accrued to the software services (in grey) built on the backs of highly technical companies run by extremely smart people.

Why is this? We think it is due to a combination of distribution and network effects. Our working hypothesis right now is that this will likely remain a similar outcome in the AI epoch. One outlier right now is Nvidia which is capturing 80% margins.

To put it in perspective, if current estimates are correct, $100bn has been spent on AI data centre expenditure with Nvidia. At 80% margins, NVIDIA is keeping $80bn and their suppliers – the fabs, designers, and equipment manufacturers are only sharing $20bn.

We can offer no forecast of how long this competitive state will last. The arguments for their unique capabilities and ecosystem that justify their margin capture are very good. We will comment that their current price depends on the situation lasting for a very long time.

Whilst we cannot predict this important factor, we are focusing our attention (and capital) on AI and software beneficiaries.

Our investment strategy hinges on identifying and capitalising on opportunities within sectors where we discern a domain advantage, be it through deep industry knowledge or a unique insight into a business model.

The goal for us is to continuously fortify our competitive edge by immersing ourselves further into these niche areas. This process is laborious and, to some, might appear excessive. But sometimes a few ideas intersect to yield a ‘Eureka!’ moment.

A prime example of this approach is the convergence of three areas of our research: AI, network economies and digital advertising.  This trifecta has emerged as a fertile ground for opportunity and these insights have been repeatable and profitable across our portfolio.

Advertising and AI are an Excellent Solution to Generate Profitability

With rising interest rates over the last two years, investors have demanded that fast-growing technology companies ‘get fit’ and move towards free cash flow (FCF) profitability. An unmistakable dispersion now exists in the software sector. Based on research by Meritech, a cashflow-positive company trades at a 33% premium multiple relative to a cash-burning company[1]. The message is now: grow at a high rate but do it profitably!

Advertising is one of the easier levers for management to pull. AI is a key enabler to facilitate targeting and efficiency in the advertising offering.

Advertising Is a Switch That Can ‘Flipped On’ To Generate Margins.

In many cases, a business with a sufficiently large customer base can have a low-margin core product and then monetise it with high-margin advertising. Digital advertising[2] generates high incremental margins (as high as the ~80% level). The low-margin product can almost become the customer acquisition cost, and the ‘real’ business becomes advertising. This strategy has proven effective across a broad spectrum of industries, from e-commerce at Amazon, MercadoLibre, and Shopee (SEA Ltd) to food delivery services such as Uber and Just Eat Takeaway, and even in traditional retail through Walmart’s advertising platform.

The key insight for us was that the high-margin nature of advertising justifies the heavy investment in expensive AI compute for better outcomes.

Why is advertising so attractive? Advertising networks have strong natural network effects – connecting ad buyers with ad consumers. The largest ad networks generate a disproportionate share of value for both sides of the network. There are few variable costs to digital advertising, so scale generates high operating leverage (profit growth exceeding revenue growth).

Now, add a dose of AI to enhance the advertising proposition for the network and a magical (win, win, win) feedback loop starts to take place (see Meta case study later).

The Nature of the Advertising Data to Train AI Models

The signal data required to serve performant ads are generally well-structured, high volume and high value (only if proprietary data from an aggregator’s owned distribution channels).

Importantly, AI enables real-time optimisation as new signal inputs come in. More recently, programmatic advertising (specifically header-bidding[3]) helps assess and leverage the economic value of users.


We thought it worth taking two portfolio companies (Meta Platforms and AppLovin) to illustrate our points. Both are demonstrating tangible economic benefits in recent earnings results.

Meta Platforms

Meta’s primary mission is all about capitalising on user engagement and maintaining its network effects. AI is augmenting their objectives in two ways:

1) Improving engagement time per daily active user (AI Job One)

2) Matching ad buyers (advertisers) with ad consumers (AI Job Two)

Improving Engagement Time per Daily Active User

Meta is in the business of making sure that when you’re scrolling through your feed or watching videos, you’re glued to the screen as long as possible. Why? Because the longer you watch, the more ads they can slip into your viewing experience (think of digital billboards). But these ‘digital billboards’ are a finite resource – only more engagement time creates them.

And here’s where the magic of ‘AI job one’ comes in – finding that perfect video that keeps you hooked – thereby increasing the time you spend on the platform. It’s a cycle that feeds itself: more engagement means more opportunities to serve ads, which in turn means more revenue.

Meta must earn the highest possible yield on each unit of ad inventory – or it is lost. Ad impressions are auctioned off to the highest bidder.

Matching Ad Buyers (Advertisers) with Ad Consumers

Advertising networks play a crucial role in connecting advertisers with their target audience, aiming to optimise advertisers’ marketing objectives such as audience building, brand awareness, and sales conversion. In the realm of performance marketing, the goal is to maximize the return on ad spend (ROAS) through efficient targeting. AI technology enhances this process by identifying the best match between the advertiser, the advertisement, and the consumer, ensuring that ads are delivered to the most suitable audience at the ideal time and place through a sophisticated ranking algorithm.

Figure 3 – Meta Platforms Advertising Metrics (Company, SaltLight Estimates)

The evolution of digital advertising over the past decades has made the process more analytical. Unlike the broad, time-specific reach of traditional TV or radio ads, digital advertising initially adopted a more targeted approach. Advertisers would create and test various ad versions to align with different user personas, a method more precise than traditional broadcasting but still plagued by the issue of irrelevant ad delivery. This misalignment represents a missed opportunity for platforms like Meta, leading to wasted ad inventory and a diminished user experience.

Pre-2022, digital advertisers relied on deterministic targeting methods[4] to track ad interactions and conversions. The landscape shifted dramatically in 2021 when Apple released iOS 14.5 disrupting deterministic targeting by limiting the ability to track users (called IDFA[5] by the industry), significantly impacting Meta’s ad targeting capabilities and consequently its share price.

It took Meta eighteen months to adapt and re-tool by incorporating AI to leverage contextual and first-party data to improve ad targeting. This adaptation has not only helped Meta recover lost ad spend but also positioned it strongly against competitors.

Generative AI promises a profound era of personalised and automated advertising, where ads are dynamically customised for each user. This innovation will allow advertisers to save on service providers creating multiple ad variants and reallocate savings to further advertising spend. Personalisation will likely lead to more relevant ads for users, enhancing engagement rates and the propensity of conversion – a win for Meta’s revenue capture.

Putting this all together, AI’s role in advertising has led to increased user engagement and improved targeting accuracy, creating a beneficial scenario for all parties involved.


AppLovin (APP) operates at the intersection of game advertisers, publishers, and over one billion game players, functioning as a pivotal monetisation enabler in the free-to-play gaming ecosystem.

One of AppLovin’s strengths lies in its primary use of contextual data for ad matching. While this type of data may not offer the high precision of first-party data like Meta’s, it remains invaluable, especially in environments where traditional data signals are weaker. Contextual targeting becomes increasingly relevant in areas like connected TV (CTV), where direct user tracking is more challenging.

Connected TV, which includes devices (smart TVs, consoles, or sticks) that stream TV content, represents a flourishing opportunity set for performance-based digital advertising. As more households move away from traditional satellite or terrestrial TV in favour of internet-connected devices, the potential for monetising this viewership grows. However, advertising on CTV is still operating in the same seventy-year-old way as linear TV – with brand advertising as the dominant part of the funnel.

AppLovin sees considerable opportunity in using its capabilities in contextual game targeting to offer performance-based advertising on CTV. They are targeting a type of household rather than a specific user. This space is shifting quickly – particularly in the US – with tech companies investing in sports content and even retailers such as Walmart buying TV operating systems companies. Connected TV advertising has the potential to be a significantly larger market than the $100bn gaming market.

Transaction Capital (TCP)

During 2023, we discussed Transaction Capital (TCP) frequently. As we suspected, in November, management announced that they were considering spinning off WeBuyCars (WBC). This was confirmed in February 2024. We anticipate that our entry price will exceed the value of the spun-off company and we’ll still have the remainder of the ‘rump’ consisting of the troubled SA Taxi and the resilient Nutun.

WBC is the ‘hidden diamond’ that we’ve been waiting for. Management’s IPO presentation has a projection that they will double their second-hand vehicle market share over the next five years from 10%-12% today – a grand target that we’ll celebrate if they hit. The more alluring part for us is this growth, according to management, will require marginal capex to get there.

Since Transaction Capital acquired WBC in 2020, the company has significantly increased its operational capacity, expanding parking bays from 4,000 to 10,500 and witnessing a 2.5-fold increase in vehicle sales. The profit-driving lever for this business is inventory turnover. This business model is more akin to a grocer that needs to turn fresh produce over quickly at thin margins rather than a tech SaaS business that earns high gross margins. One positive is that cars don’t expire in a few days. Should WBC double its market share within the capital expenditure parameters provided, we anticipate growth in free cash flow growth exceeding sales growth by a large measure.

The interesting question post-unbundling that we ponder is: what becomes of the remaining TCP business? Lenders no longer have leverage over the holding company through HoldCo debt cross-default clauses. Theoretically, if SA Taxi lenders choose not to support debt-rollover relief, TCP could simply put SA Taxi into business rescue in an orderly process rather than the voluntary workout process they’re following at the date of writing. We’ll know in two months how things will look post unbundling.


As we always remind investors, most of our liquid wealth is in the same fund as yours. We share in the inevitable ups and downs right with you. Please free to reach out if you have any questions.

David Eborall
Portfolio Manager


Collective investment schemes are generally medium to long-term investments. The value of participatory interest (units) or the investment may go down as well as up. Past performance is not necessarily a guide to future performance. Collective investment schemes are traded at ruling prices and can engage in borrowing and scrip lending. A schedule of fees, charges, minimum fees, and maximum commissions, as well as a detailed description of how performance fees are calculated and applied, is available on request from FundRock Management Company (RF) (Pty) Ltd (“Manager”). The Manager does not provide any guarantee with respect to the capital or the return of the portfolio. The Manager may close the portfolio to new investors to manage it efficiently according to its mandate. The Manager ensures fair treatment of investors by not offering preferential fees or liquidity terms to any investor within the same strategy. The Manager is registered and approved by the Financial Sector Conduct Authority under CISCA. The Manager retains full legal responsibility for the portfolio. FirstRand Bank Limited is the appointed trustee. SaltLight Capital Management (Pty) Ltd, FSP No. 48286, is authorised under the Financial Advisory and Intermediary Services Act 37 of 2002 to render investment management services.

[1] Meritech Software Pulse, 22 February 2024 for the cohort of companies with 20%-40% NTM revenue growth. Today’s medium NTM revenue multiple is 11.6x for cash positive cohort’s vs 8.7x ‘burning cash’ cohorts. Source

[2] If there are no distribution or network costs

[3] Header bidding is a programmatic advertising technique that allows publishers to offer their ad inventory to multiple ad exchanges simultaneously before making calls to their ad servers. Each party can bid on the same inventory in real time, generative high yield for an inventory seller.

[4] Deterministic targeting methods in ad technology involve using precise and identifiable user data, such as email addresses or device IDs, to target advertisements directly to specific individuals. This approach ensures high accuracy in ad delivery, targeting users based on verified personal information.

[5] The change forced a user to ‘opt-in’ to share data for tracking. It is also referred to as ATT (App Transparency Tracking)