What I've Been Reading (7/5-7/12)
Software in no-man's land, agentic AI use at Citadel, and more
7/5-7/12
News, filings, etc.
CMCSA
Peacock to be profitable in 2Q (link)
Versant buying golf simulator company Full Swing for $530m
“GFL Environmental (GFL) has held discussions with Apollo Global Management (APO) regarding a potential investment or takeover, Canadian newspaper The Globe and Mail reported Sunday, citing a source familiar with the talks.” (MT Newswires)
“Morgan Stanley analyst Sean Diffley and his team now expect Starlink to reach about 16 million U.S. broadband subscribers by 2030, up from roughly three million last year. That forecast forced Morgan Stanley to lower its long-term outlook for cable and fiber growth across the industry.” (link)
New COO at FUN (link)
Gary selling a bit of RH stock (link)
John Malone 1989 Congressional testimony (link)
Meditation Capital letter on software (link)
WK, TEAM, BRZE, MNDY, KVYO, DT
“Cheap vs. expensive software – this is related to the point above. Core business systems which can cost in the multiple tens of millions paid to a single vendor per year are potentially more at risk… Small (but still #1 in its category) vs. large (Salesforce, ServiceNow, Workday, etc.) in size – we prefer the former. The former can hold strong positions in niche markets too small for venture-scale outcomes even if major success is reached; the latter have more attention and a potential revenue prize that is worth new startups going after. We also have our general thesis that small companies can grow faster and for longer than large ones (hence our usual focus on smaller, lower-liquidity companies). From a stock perspective, the former also have fewer investor eyeballs, especially in the current environment where many have abandoned the sector, making them more likely to be mispriced.”
“The market is making clear that it wants more real margin, and management teams are already starting to deliver (e.g., Workiva GAAP operating margins +1800 bps y/y in 1Q26, Atlassian 10% layoff and hiring freeze while ARR continues to grow 20%+, Dynatrace new commitment to “rule of 50” of margins plus growth, etc.).”
“This also gets into the stock-based compensation debate. Part of why software stocks have declined so much is not just the fundamental risks but also a major growth-to-value transition in the investor base, or lack thereof due to stock-based compensation being such a large cost item that different groups of investors treat differently. Growth investors were happy to value software companies on fake non-GAAP earnings before subtracting stock-based compensation (we care a lot about real earnings, and were not in this category), but they have fled as the category has fallen out of favor amongst growth investors; value investors newly looking at these beaten-down stocks were shocked at the current low level of real economic earnings and concluded, wrongly but understandably in our view, that these companies are still quite expensive. So software stocks are in no-man’s land… In the meantime, we see this as an opportunity – when “Non-GAAP” becomes a bad word and “EV/Sales” is met with revulsion, only said by those who don’t “get it” – we see it as the best time to lean in, confident in our framework around mature GAAP margins for high-quality software.”
“We do however prefer companies that don’t charge on seats, and Workiva, Braze, Klaviyo, and Dynatrace are all in this category.”
“Braze is a deep domain expert in sending emails, texts, and app push notifications at scale and perfectly (a much harder problem than it sounds, both in engineering and also because deliverability of messages is adversarial against recipient systems that try to filter messages into promotions and spam); the customer is not. Workiva is a deep domain expert in the nuances of SEC reporting and internal controls compliance; the customer is not.”
“For example, Braze and Klaviyo, as customer engagement platforms that help companies send emails and texts and app notifications, historically mainly provided the execution layer, with a human deciding what and when to send; AI gives these companies a new opportunity to additionally sell the “brain” behind marketing – messages personalized for each customer rather than by broad category, near-fully automated campaigns, intelligent agent responses when customers reply, etc. Braze, now with AI, is accelerating its disruption of its legacy competitors, Salesforce Marketing Cloud and Adobe Journey Optimizer; and Braze’s organic revenue growth accelerated to 27% y/y in its most recent quarter ending April 2026, up from a low of 20% in the year-ago quarter ending April 2025.”
“Workiva targets $1.8-2.0 billion in revenue in 2030, up from $1.05 billion in 2026; we model $1.85 billion (not because we think they’ll miss, but because our MO is to be conservative, so we can have full confidence in what we underwrite). Assuming an exit at 4x NTM revenue at year-end 2029 (so ~14x mature EBIT for a business that should still be growing in the mid-teens), and incorporating assumptions for share dilution as well as FCF generation, we get an IRR of 31%. If the multiple doesn’t improve much from current levels, a 3x revenue exit gets an IRR of 23%; a 5x revenue exit pushes the IRR to 39%. We see a plausible downside case of 2.0x 2027 revenue, which would be -20% vs. the current price.”
Ben Thompson’s script for Zuck (link)
“We’ve taken our arrows through the years for lots of things that frankly aren’t our fault, but are rather the reality of being the primary communications platform for all of humanity, and humanity is flawed.”
“Second, AI makes our business better — and by “our business”, I mean ads. AI is more than LLMs: it is machine learning, and we have been using machine learning to improve our ads business for years. More recently, we have developed GPU-dependent algorithms that have significantly improved our ability to not just target ads but also recommend content, which keeps people entertained longer, which lets us serve them more ads. And, looking forward, LLMs themselves will transform advertising, not just by generating copy and images, but by predicting the ads and content that people want to see. Every single one of these improvements goes directly to our top line — and remember, because advertising enables us to offer our products for free, the capacity to increase our top line is unbounded by price elasticity.”
“We are not out here to make chatbots or compete with OpenAI and Anthropic; they can fight for work and productivity and charging subscriptions and replacing humans. Our goal is to celebrate humans, to connect them, to entertain them, and to enable commerce among them. We need compute to do this at scale, and I know it will pay off. My commitment to you is that we will structure our business so we have no choice but to do just that. We’ve done it before, and we will do it again.”
Kevin Mak writeup on CMPS (link)
“3) Career risk is asymmetric. No portfolio manager wants to explain a loss on “the magic mushroom stock.”"
Valuation: “Compass trades at roughly $13, a market capitalization of about $1.8 billion… The honest starting point is that nobody can forecast this. I am not going to build a discounted cash flow for a drug that has not launched, and anything beyond a rough frame would be false precision. What can be said is about shape rather than level. Spravato has proven the market exists at scale, roughly $1.7 billion a year, in the same delivery model COMP360 would use. The overwhelming majority of the four million Americans with treatment-resistant depression have never been treated with it or anything like it. And COMP360, on the clinical record, is at least competitive at a fraction of the treatment burden. Those facts do not locate the outcome. They produce a long right tail. The only question that matters is whether $13 reflects any of it… To test that, I use a crude frame and state it plainly: an enterprise value of roughly four times peak annual sales, which is the multiple observed in specialty pharmaceutical acquisitions and which reconciles to approximately the undiscounted lifetime gross margin of a drug with a normal exclusivity runway; a cost of capital of about 10%; peak sales several years after launch; and the roughly 138 million shares implied by the current market capitalization… Point that frame at today’s price and it implies the market is underwriting around $600 million in eventual peak sales. The analysts who cover the stock model something closer to $1.5 billion. I underwrite $2.5 billion, which is what I believe treatment-resistant depression alone supports for a materially better product, and which counts nothing for PTSD, MDD, or anxiety. My revenue estimate maps to a price in the mid $40s. A conservative case, where COMP360 is approved and moderately successful but adoption is slow, lands in the low-to-mid $20s. None of these three numbers is a forecast anyone should trust to a decimal, mine least of all… Failure is possible, and there is no true downside floor. Every figure above is conditional on approval. In the approval scenario, which I believe is highly likely for the reasons set out above, the equity would be worth some multiple of today’s price along the range described. In the scenario where approval does not come, it would be worth a small fraction of it. The reason I find the trade compelling is not that the downside is protected. It is that a fat right tail is available at a price that reflects almost none of it. That asymmetry, rather than confidence in any single number, is why I own the position and why I have sized it as I have.”
Ken Griffin at GS Exchanges (link)
“One of our team members built an agentic system to recreate academic papers in finance. So, academia publishes a plethora of papers in finance. We read these papers thinking about the hypothesis, the quality of the work done, do we think what they have observed will have persistence out of sample? Do stock buybacks cause stocks to outperform? Simple example. And you know, you have a legion of young Masters and PhDs doing this work. It takes roughly six to eight weeks to reproduce a paper. It’s interesting work. We find a few ideas when you’re doing this. But for us, a few ideas could be worth quite a bit of money. My colleague built an agentic AI system that would read a paper, reproduce it, verify the results that were published in the paper, produce the results out of sample, and do all of this work in about, on average, two to three hours per paper.”
Disclosure: Views are my own; none of this is investment advice; I/we may own positions in securities mentioned.