Weekend Reading #373

This is the three-hundredth-and-seventy-third weekly edition of our newsletter, Weekend Reading, sent out on Saturday 18th July 2026.

To receive a copy each week directly into your inbox, sign up here.

*****

What we're thinking.

The selloff that threatened over the past weeks has turned into a major rout in anything related to AI capex over the past week. Led by memory names, the damage is now severe. There are so many things to think about now in terms of the near-term path ahead that it helps to unpack them one by one. Firstly, is this a real selloff that heralds the end of the capex cycle or is it driven by overleverage and latecomers to these names? You'll be unsurprised to learn that we don't have the answer. There are no certainties in markets and that's why we love them! It's not an exam with a set of answers that can be learned and recited at the test.  We know there has been crazy amounts of leverage in South Korea where something like 3% of the adult population apparently got margin called this week. We know that there has also been leverage and intense retail participation in the US names too. We also are hearing as we always do when this type of move happens that the pod shops are reducing risk. All of these are valid. While there is zero indication fundamentally that capex plans are affected in any way, the market will always work it out before it happens so we cannot rely on the so-called fundamental state of things in real time. There is a balance to be had between price and fundamentals to compensate for the uncertainty of simply not knowing. At 40-50% lower in some of these names, at what point is that balance back to appealing? The way we look at it is that if this is a non-contagious event and fundamentals are still sound, at these prices for longer term buyers there is much upside. If, however, this begins to spill over into other sectors of the market, all bets are off. We don't know which way it will go. But we warned of the potential for a sharp move in the USD and this is very much still on the table. Mining stocks are also selling off materially. The only 2 parts of the market that have been somewhat spared are the hyperscalers themselves which have not sold off (some have moved up a small bit) and software names for which the same applies. Are we moving to a phase where the market believes LLMs will be commoditised, and value will shift to the application layer? This week's latest Chinese model, Kimi-K3 has shot to the top of the capabilities list competing with the US frontier labs at a fraction of a cost! There are so many possibilities and so many variables that we really are not looking too far ahead. Price is king and right now it is telling us to continue to be very careful.

What we're reading.
As the AI edge starts moving up the curve from GPUs to memory to applications, and the edge of having the best model gets worn down in favour of the “how” of using the best models, I recently learnt a new metric that I didn’t know had a name: prefill. This article demonstrates it well – it’s the time before the model spits out the first word of its reply when you plug in a prompt, during which it ingests all of the context it needs to perform the task at hand. In a comparison between three consoles, the DGX Spark, the AMD-powered Strix Halo and the now-famous mac mini, the ability to work through huge context seemed to be the missing piece in my regular AI use (weapon of choice personally being Perplexity), especially when prompting with huge context, the ability to digest a full set of prior information before launching into generation becomes critical. And because “garbage in, garbage out” still holds, the only way to get good quality output is to provide it with good quality contextual inputs, on a model and a device that can handle it. EL

What we're watching.

I think that over the past few weeks there has been an inflection point in terms of using AI tools to make film and music. I have been blown away by some of the stuff I've watched. Last week I shared this brilliant music video called Sweet Land by Heavy Pulp. And two more clips this week. First up is this from an X account called Samyaza Speaks. This creator has taken a series of old master paintings and brought them to life with AI. He has added a track to it as background music and it is simply breathtakingly good. We are only getting a small taster of what it coming. For a bit of fun there is also this clip from my favourite new creator of theme all, Charles Curran. It is a reimagined episode of Seinfeld with peptides. Superb. DC

What we're listening to.

Gio Valiante is a performance psychologist for elite golfers and had a stint with Steve Cohen at Point72 as well. This is a conversation with Shane Parish on his Knowledge Project podcast where he goes into some of the things that he does and the way he thinks about his clients. And some good stories too. Really cool listen. 
 
A banger of a listen this week on Rick Rubin's Tetragrammaton podcast with a professor named Angus Fletcher who has a title of Professor of Story Science. It's a long one as he unpacks how he believes our brains operate in stories. His work was so compelling that he was hired by the US army to help the special ops train graduates to become more creative. This has so much good stuff it's worth every minute. A lot of what he says goes contrary to common wisdom. I learned a lot. DC

I came across this Joe Rogan podcast featuring Aravind Srinivas, the co-founder of Perplexity, and thought I’d venture a listen. Where it started with a rather strange but fascinating discussion about how ancient Hindu classics describe weapons that sound like autonomous armaments and how ancient civilisations seem to show unbelievable sophisticated levels of technology possibly beyond what we’re currently able to muster, it quickly moved to the core topic of what’s left for humans when the value of knowledge recall is tending towards zero. Aravind’s answer of “curiosity” as being the last remaining human trait that sits squarely outside the realm of Ais is something worth thinking about. It makes sense that it is ultimately the curious human mind, inquisitive and sometimes crazy enough to explore the previously unexplored, that yields the most interesting outcomes from AI use. EL