What We’re Reading (Week Ending 07 September 2025) - 07 Sep 2025
Reading helps us learn about the world and it is a really important aspect of investing. The late Charlie Munger even went so far as to say that “I don’t think you can get to be a really good investor over a broad range without doing a massive amount of reading.” We (the co-founders of Compounder Fund) read widely across a range of topics, including investing, business, technology, and the world in general. We want to regularly share the best articles we’ve come across recently. Here they are (for the week ending 07 September 2025):
1. The ROI Question – Abdullah Al-Rezwan
A friend recently DM-ed me to highlight one of the quotes from Nvidia’s CFO in their recent earnings: “New NVFP4 4-bit precision and NVLink 72 on the GB300 platform delivers a 50x increase in energy efficiency per token compared to Hopper, enabling companies to monetize their compute at unprecedented scale. For instance, a $3 million investment in GB200 infrastructure can generate $30 million in token revenue, a 10x return.”
10x return? That’s a bit eye-popping number. My friend was understandably a bit skeptical of this claim, so he asked ChatGPT to show the math and some reasonable assumptions behind this claim…
…Clearly hyperscalers aren’t realizing such revenue from their investments in Nvidia chips yet…
…Batch size, which is the number of concurrent user requests processed simultaneously, is the single most important operational factor for maximizing throughput. The highest throughput numbers are always achieved with the largest possible batch sizes.
However, large batch sizes increase latency,..
…To maintain low latency, providers must deliberately use smaller batch sizes. This inherently sacrifices aggregate throughput to ensure a good user experience. The 1M tokens/sec benchmark mentioned in the ChatGPT screenshot above is likely achieved at latencies that would be unacceptable for real-time use…
…While 20% utilization may seem conservative, achieving this average utilization consistently (24/7/365) with monetized workloads may not be super easy in inference. AI inference demand is often “peaky.” Infrastructure built for peak load sits idle during off-hours…
…Hyperscalers do not only run the most expensive models. A likely material portion of their workload involves smaller, cheaper models (often <$1 per 1M tokens), reducing the actual blended revenue shown in my screenshot above. If the average realized price drops to $2/M tokens, the idealized revenue drops from $31.5M to $12.6M in my example (ceteris paribus).
2. AI Agents and the Future of Grocery Delivery – Thomas Reiner
Whether it’s OpenAI, Gemini, Siri, or some other tool, every consumer will have a personal agent in their pocket. It will book travel for you, make dinner reservations, manage your schedule, and for purposes of this discussion it will order your groceries for you.
For the average American family that gets groceries 1x per week it’ll know what you often order, it’ll recommend recipes, it’ll monitor past usage and wastage of food products, and it’ll know your consumer preferences around store loyalty. If you change your plans and tell your agent that you’re hosting a dinner party for 8 people serving it can assist by recommending what to serve and automatically ordering it from the grocery store…
…Looking across these models there’s two key areas where there are middlemen to be disrupted: 1) Grocery Delivery Marketplaces and 2) White Label Solutions. The Age of AI is going to be the age of efficiency and wringing out middlemen from the equation. Grocery Delivery Marketplaces are definitely middlemen and I’d argue that white label solutions from providers like Instacart Storefront sort of are…
…AI agents mean that the importance of brand goes down, while the importance of service goes up, and that’s where all the incremental dollars from both players are going. They can best position themselves to win in a world where AI agents make decisions based on best outcome (cost, quality, speed).
DoorDash with their DashMart concept is trying to take themselves out of the middleman equation and focus on being the 1P provider which is a lot more defensive in an AI agentic world.
The biggest challenge will be on crossing the trust chasm. While consumers have high trust in Amazon for shelf-stable goods, it’s non-existent for fresh goods. Early returns from Amazon same-day perishable trial showed 75% of consumers were first-time perishables shoppers at Amazon but only 20% reordered multiple times within the first month. “U.S. shoppers have shown they prefer to buy fresh goods from retailers that run brick-and-mortar stores, as evidenced by the struggles of online-only grocers like Peapod and FreshDirect.”…
…Generally the rise of AI Agents is a lot more mixed picture for the delivery marketplaces. On one hand, an AI that can spontaneously order anything might increase demand for delivery, on the other hand these services might be commoditized as the consumer UX slowly fades away and the importance is put on the underlying speed, convenience, and price.
3. An Interview with Cloudflare Founder and CEO Matthew Prince About Internet History and Pay-per-crawl – Ben Thompson and Matthew Prince
The reason to talk now, and we’ve talked offline about this a few times, both this year and last year, is your push for this pay-per-crawl concept. Why don’t you give me the high level overview, the pitch from your perspective, which I think has evolved? I would like to think partially based on some of my feedback, but what’s the pitch in September 2025?
MP: Let’s take Cloudflare out for a second and just talk about—
Talk about Matthew, the English student? The student newspaper editor.
MP: This is me channeling inner law professor. Let me give you the history of the Internet and why the Internet exists the way that it does and what’s changing.
This is usually my job, but go ahead.
MP: And you can tell me where I’m wrong, but this is my quick history of the Internet, and apologies to Michelle who hates history lessons.
For the last 25 years, the interface of the Internet has been search, and Google has dominated that space, and Google, their incentives as a company were to have the Internet grow as much as possible because if you have chaos, then the search becomes the organizer of the chaos. But you need incentives for people to actually create content and so Google not only had to create the thing that organized the Internet, but they then had to take the thing that took the traffic of where people went and then helped people monetize that, largely through advertising, although they also helped with subscriptions, and Google was the great patron of the Internet for the last 25 years. The web would not exist the way it does if there were not something like Google out there to create the incentives around.
There were a lot of problems with incentivizing around traffic, we created systems where people would just literally try and create rage-baity headlines to get people to click on things so that they could put ads against them and so not perfect, but we don’t have the Internet that we have today unless we have Google and search funding that.
That is changing. The world is shifting where the interface of the web is shifting from search engines and search engines give you a treasure map and say, “Hey, go figure out what your answer is by clicking on these 10 blue links”, to what are effectively answer engines. So if you look at OpenAI, if you look at Anthropic, if you look at Perplexity, even if you look at modern Google, they are not a search engine, they don’t give you a treasure map. Instead, they give you an answer right at the top of that page. That answer, for most users, 95% of the users, 95% of the time, it’s a better user interface. I’m not anti-answer engines, I’m not anti-AI, I think it’s better in every possible way for that to be what the interface is that we all interact with.
But the problem is that if you get the answer and you don’t get a treasure map, then you don’t generate traffic and if you don’t generate traffic, then the entire business model of the web, which has been based on traffic starts to break down and you can see that, not so much in e-commerce sites, not so much in things that actually sell you the physical thing because if you asked what’s the best camera to buy, even if you get an answer, you’ve still got to go buy it from somewhere. It’s going to take the e-commerce and the people who are selling things that’s going to work but the person who wrote the review—
The great thing about physical products is by definition they are scarce and the problem with text on the Internet is it is not scarce.
MP: It’s not scarce, that’s exactly right, and Google set this expectation that everybody can scrape the Internet for free, but it was never free. The Internet has never been free. Google paid for it for a really long time and the quid pro quo with the content creators was, “We get a copy of your content and in exchange we’ll send you traffic and help you monetize that traffic”.
That quid pro quo breaks down as we shift from search engines to answer engines and so something is going to change. I see three possible outcomes for that. And again, none of this involves — if Cloudflare disappeared tomorrow, this is still happening, one of these three things will happen. One, all of the journalists, academics, and researchers in the world will starve to death and die. And it’s crazy, like when you post this stuff on Twitter, how many people were like, “Well, we don’t really need journalists anymore, we have drones”, and I’m like, “I think we still need journalists”…
…If it’s inevitable though, then why does Cloudflare need to be so aggressive? You’re instituting these policies of doing your best to block bots, putting together protocols for recognizing what it’s worth, payments, etc., all very nascent to be sure, a lot to be figured out. But you are not taking the posture of a company that this is inevitable and it’s going to be great, you are being pretty forceful in trying to make something happen.
MP: Well, I think if we weren’t doing it, someone else would. But what I think we have a unique ability to do is we’re really good at stopping things like bots because we do it every day.
So again, it wasn’t like we were sitting around being like, “Hey, what should we do next? Let’s go change the business model of the web”, it was our customers who were publishers were coming to us being like, “We’re dying and we don’t have the technical wherewithal to step in front of it, but we need to stop this, please help”. And honestly, when Neil [Vogel] at Dotdash Meredith was telling me this, I rolled my eyes and I was like, “Publishers, they’re such Luddites, they’re always complaining about the new technology, they’re always complaining about the next thing, this isn’t a big deal”. And Neil and a bunch of others finally said, “Just go pull the data”, and it was only when we actually saw the data, when we saw that over the course of the last 10 years, it’s become 10 times harder to get a click from Google for the same amount of content on that same kind of basis, it’s now 750 times harder with OpenAI, it’s 30,000 times harder with Anthropic.
The business of traffic on the Internet as being the currency is going away and so something either again, either content creation is going to die, it’s going to become futile, or we’ve got to create a new business model. Again, if our mission is to help build a better Internet, this seems squarely in the line with what we should be working on.
So why does Garry Tan say that you are an axis of evil with Browserbase and you should legalize AI agents?
MP: I really don’t understand. I mean, I’m confused by Garry, I think part of it might be that he’s an investor in Perplexity.
Every story needs four characters, you need to have a victim, you need to have a villain, you need to have a hero, and you need to have the village idiot or the stooge. And if you think about it, any news story has those four characters. Right now, the people who have most been the villains have been Perplexity, where they’re doing just actively nefarious things in order to try and get around content company.
I’ll give you an example of something that we’ve seen them do, which is that if they’re blocked from getting the content of an article, they’ll actually, they’ll query against services like Trade Desk, which is an ad serving service and Trade Desk will provide them the headline of the article and they’ll provide them a rough description of what the article is about. They will take those two things and they will then make up the content of the article and publish it as if it was fact for, “This was published by this author at this time”.
So you can imagine if Perplexity couldn’t get to Stratechery content, they would say, “Oh, Ben Thompson wrote about this”, and then they would just make something up about it and they put your name along it. Forget copyright, that’s fraud, just straight up and that’s the sort of bad behavior of some tech companies that again, I think needs to be called out and punished.
4. Bitcoin TreasuryCos: Lessons From The 1929 Crash – Be Water
The explosive proliferation of Bitcoin treasury companies mirrors that of the 1920s investment trusts, and both gold rushes stem from a perfect storm of greed: intense investor demand for exposure to a scarce asset creates mNAV premiums that promoters rush to monetize. If Goldman Sachs could extract enormous profits from its trust in the 1920s, why couldn’t everyone else? If MicroStrategy can monetize its mNAV premium, why shouldn’t every other company follow suit?
Galbraith documented the explosive growth of trusts in the 1920s:
During 1928, an estimated 186 investment trusts were organized. By the early months of 1929, they were being promoted at the rate of approximately one each business day, and a total of 265 made their appearance during the course of the year…
…The renowned Yale economist Irving Fisher famously declared that stock prices had reached a “permanently high plateau” just prior the 1929 Crash. Fisher’s declaration exemplified the kind of euphoric confidence that typically marks a market top…
… Fisher’s plateau quote is now infamous, but the lesser-known context that gave rise to it tells a more revealing story. He was actually defending investment trusts as a key support for stock valuations, much as Bitcoiners cite built-in demand from Bitcoin treasuries today. The New York Times reported at the time:
Professor Fisher spoke on the subject of investment trusts and presented a defense for them against recent attacks in which they have been charged with responsibility for many present evils.
Fisher defended trusts on the grounds that these vehicles were awakening people to the superiority of stocks over bonds and providing investors with a superior structure for gaining equity exposure—much as Bitcoin treasury advocates today claim MicroStrategy offers turbocharged “torque” over direct Bitcoin ownership, and Bitcoin itself offers superiority over TradFi assets like fiat currency, stocks, bonds, and real estate:
I believe the principle of the investment trusts is sound, and the public is justified in participating in them, with due regard to the character and reputation of those conducting them. Largely through the influence of the investment trust movement, the public has been waking up to the superior attraction of stocks over bonds. And I believe the operation of the investment trusts, as a whole, has acted to stabilize the stock market rather than to make its fluctuations more violent…
…Saylor’s confidence in monetizing NAV discounts—which is perhaps reasonable for MicroStrategy in isolation—mirrors the same logic 1920s trust managers used to justify buybacks—only to find that such support strategies are ineffective when liquidity across the ecosystem vanishes and selling pressure dominates.
The trusts discovered that buying back shares when investors are selling and credit is tightening is vastly different from issuing shares when investors are buying. Desperate to prop up their stock prices, the trusts began buying back shares at a discount to NAV—a strategy Bitcoin treasury companies will likely adopt with equally disappointing results for most:
The stabilizing effects of the huge cash resources of the investment trusts had also proved a mirage. In the early autumn the cash and liquid resources of the investment trusts were large…But now, as reverse leverage did its work, investment trust managements were much more concerned over the collapse in the value of their own stock than in the adverse movements in the stock list as a whole…
Under these circumstances, many of the trusts used their available cash in a desperate effort to support their own stock. However, there was a vast difference between buying one’s stock now when the public wanted to sell and buying during the previous spring—as Goldman Sachs Trading Corporation had done—when the public wanted to buy and the resulting competition had sent prices higher and higher. Now the cash went out and the stock came in, and prices were either not perceptibly affected or not for long. What six months before had been a brilliant financial maneuver was now a form of fiscal self-immolation. In the last analysis, the purchase by a firm of its own stock is the exact opposite of the sale of stocks. It is by the sale of stock that firms ordinarily grow.
As the crisis deepened and the mNAV continued to trade at a discount, trusts depleted their remaining cash reserves in a desperate—and ultimately self-defeating—effort to support collapsing share prices:
However, none of this was immediately apparent. If one has been a financial genius, faith in one’s genius does not dissolve at once. To the battered but unbowed genius, support of the stock of one’s own company still seemed a bold, imaginative, and effective course. Indeed, it seemed the only alternative to slow but certain death. So to the extent that their cash resources allowed, the managements of the trusts chose faster, though equally certain death. They bought their own worthless stock. Men have been swindled by other men on many occasions. The autumn of 1929 was, perhaps, the first occasion when men succeeded on a large scale in swindling themselves.
5. Technology vs Platform Shift, Portfolio Change – Abdullah Al-Rezwan
Casey Winters made this point almost a couple of years ago which I think still holds up pretty well:
What I realized having gone through the internet and mobile platform shifts is that the technological and distribution shifts did not happen at the same time. Platform shifts that create both technological and distribution opportunities happen in a sequence, not all at once…AI has come out and definitely created a technological shift that enables new ways to solve problems that couldn’t be done before. But AI lacks a new distribution channel. ChatGPT is “not it”, as the kids would say. At least not yet…
… Sameer also points out that in a technology shift, users may not even be aware about the tech (it just works) whereas in a platform shift, the change is front and center for the user:
In a technology shift, form factor does not and should not matter. For example, scaling Snapchat’s picture messaging functionality would not have been possible without the shift to cloud computing. While Snapchat’s cloud hosting costs were significant, it would not have been possible to scale it as quickly if it relied on large, operationally complex investments into server infrastructure. The most important part — Snapchat’s end users did not know or care about this in any way. The user interface did not change to call out Snapchat’s “Cloud powered” technology. The biggest changes happened in the backend, not the frontend.
Disclaimer: None of the information or analysis presented is intended to form the basis for any offer or recommendation. We currently have a vested interest in Alphabet (parent of Google) and Amazon. Holdings are subject to change at any time.
