What We’re Reading (Week Ending 05 July 2026) - 05 Jul 2026
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 05 July 2026):
1. As AI Companies Race for Power, Amazon and Google Have the Lead – Lee Jinjoo
Amazon has an incumbent advantage. It is the world’s largest cloud provider and has been building a lot of data centers over the past two decades. The company’s operating, self-built data centers in the U.S. consume up to roughly 9 gigawatts of power, according to Aterio, a data provider. That is comparable with the generation capacity of North Dakota.
By comparison, Microsoft and Alphabet’s Google each have self-built data centers that use up to about 5 gigawatts of power, while Meta Platforms’ data centers have a roughly 4 gigawatt capacity. So far, most of the hyperscalers’ data center capacity is self-built, rather than rented from data-center operators…
…Based on estimates from Aterio, which tracks company announcements, utility filings, building permits and satellite data, Amazon is expected to add the most data center and power capacity in the U.S. through 2030. But Google is expected to add capacity at the fastest rate. In fact, including leased capacity from third-party data center owners, Google will have significantly closed its gap with Amazon by 2030, according to Aterio…
…Amazon plans to build out most of its own capacity, while Google is expected to rely more heavily on leases. Based on Aterio’s data, about a quarter of Google’s expected data center capacity in 2030 is expected to come from leases. Self-built can take longer but is the cheaper option over the long term…
…Google has proved that it can get speed with clean energy. At least three of its planned Texas data centers will be able to skip the long queue to connect to the grid because they are being built next to solar and wind projects, according to a report from Cleanview. In two of these data centers, Google will build the solar and wind capacity through its Intersect Power subsidiary. Texas’ power market rules allow faster grid connection if the data center is co-located with a new source of power. In all three cases, the solar and wind capacity well exceeds that of the data center…
…Microsoft, for example, struck a 20-year agreement with Chevron to power its AI data center in Texas with an off-grid, natural-gas-fired power plant. Both Meta and Amazon have plans for such projects, according to data compiled by Cleanview. Amazon hasn’t publicly confirmed its involvement, but its data center in Fayette County, Ohio, is the only planned source of large power demand near a permitted off-grid, natural-gas-power project, according to Cleanview’s Thomas. In most cases, hyperscalers eventually want these data centers to connect to the grid.
2. What’s the Real Depreciation Curve of a GPU? It Depends on What It Actually Did – Pietro Sette
Identical GPU hardware can age very differently depending on how it’s used:
- A GPU running steady inference at, 60–70% utilization, under moderate thermals, day in and day out
- vs. a GPU running irregular training workloads that repeatedly spike to 95–100% utilization and push thermal limits every afternoon
On paper these two might be the exact same model of GPU. In practice, their aging is radically different. One might still be going strong and profitable after 5+ years (indeed, some 2016-era GPUs are still in active cloud service), while the other might effectively be worn out – or at least no longer economically viable in 3 years or less…
…Consider a mid-market lender financing several GPU deployments in the 0–50MW range (hundreds of high-end GPUs across multiple customers).
Their original underwriting assumed:
- ~80% steady utilization on each GPU (a consistent workload level)
- ~5.5 year useful economic life for the GPUs (before resale or obsolescence)
- No meaningful variance across different customers or workload types (every GPU in the fleet treated uniformly)
But when real telemetry data was collected at the GPU level, here’s what was actually observed:…
…Result: The fleet’s effective depreciation curve varied by 30–45% across different end-customers, even though the GPUs were identical models. In other words, certain customer workloads drove their hardware to lose value almost half again faster than others…
…Different operational events and stressors affect how fast a GPU “ages” or loses reliable performance. Thermal stress, power stress, and workload intensity are chief among them.
3. Rory Johnston on Why His $200 Oil Prediction Didn’t Turn Out Right (Transcript here) – Joe Weisenthal, Tracy Alloway, and Rory Johnston
[Joe Weisenthal]: Let’s start back in early March. Remind listeners what your take and the general wisdom was in the first couple of weeks of March about how long this could persist. And remind people why the Strait of Hormuz was seen as the choke point among choke points when it comes to oil.
[Rory Johnston]: Yeah, let’s transport ourselves back to our last conversation.
[Joe Weisenthal]: We need to get the time-travel-machine music, right?
[Rory Johnston]: They can add that. So, the reason it was such a massive deal — and still remains, I’d say. While we’ve avoided the doomsday prophecies, it is still by far the largest supply disruption in the market’s history. And for the numbers, for the barrel counting, for Tracy: the total flow through Hormuz prior to the war was roughly 20 million barrels a day. We knew we weren’t going to lose all of that, because we had some offsets — the Saudi East-West pipeline, the Emirati pipeline to Fujairah on the Gulf of Oman. But netting out all of those known rerouting options — which, again, at the time we didn’t know if they would fully work, because they’d never been fully tested, though they did work, thankfully — even after netting those off, we were still down roughly 13 million barrels a day of Gulf oil production, excluding Iran, that had been forcibly shut in for the duration of this crisis and is only now beginning to pick back up. That’s a lot of oil. That’s more than 13% of global supply. The reason we thought prices were going to hit $150 or even $200 a barrel is that when you have a supply shock that large without any more offsets, you end up at demand-destructive pricing really, really fast. And to destroy that level, the depth of that demand — we had never seen that before, but $200 a barrel seemed like the reasonable price at which it would happen. Now, thankfully, we did not have to destroy that demand. And what we’ll talk about shortly, I’m sure, is all the ways the system adapted and flexed. I think we saw this most notably, above all, in China.
[Tracy Alloway]: Okay, why don’t we just dive into it? Give us your overview on what happened and why we didn’t actually hit $200 a barrel.
[Rory Johnston]: The two biggest things — one on the fundamentals, the barrel-counting side — was China. We always knew China had huge stockpiles of oil, but we didn’t know how it was going to react to this crisis. What we’ve seen is that Chinese crude oil imports — into the world’s largest crude oil importer — fell by upwards of 5 million barrels a day between the three-month average prior to the war and June. We’re not quite done this month, but that’s roughly where we’re trending for June so far. That 5 million barrels a day was upwards of half of the total spot-market supply hit to Asia, and it allowed a lot of those other Asian importers to not have the competition they would otherwise have had for the barrels they were importing. So the countries that were hit hardest, and the governments that were most panicked — South Korea, Australia, Japan, Taiwan, and so on — there was a period where the Prime Minister of Australia was coming out daily and announcing the government’s successful acquisition of a cargo of diesel. It felt very COVID-y. Those importers saw imports collapse through March and April, but through May and into June they actually recovered basically to pre-war levels. And the largest facilitator of that was the fact that China was not competing for any of the other barrels — it absorbed so much of the shock itself.
[Tracy Alloway]: Just on China specifically — I have so many questions already — do you have any sense of how much of this was genuine demand destruction or substitution in China versus just releasing from stockpiles?
[Rory Johnston]: It’s a good question, and the firm answer is we don’t have 100% certainty as to the exact composition of that swing. We know the oil going in fell by 5 to 6 million barrels a day, products included. But in terms of actual demand destruction — and you guys were actually in China very recently — all the mobility indicators showed no notable decline. The level of implied demand destruction we see is striking, and importantly, China does not publish official demand data, and, very importantly, it doesn’t publish official inventory data either. So we’re left feeling around in the shadows. The implied demand destruction through this crisis was on par with the steepest in history, and on par in volume with the COVID-zero demand shock in 2022. But you guys were in China; I have not seen any reporting that indicates that level of lockdown. So we start asking, what’s going on in the middle? Typically, you’d assume that level of demand destruction without COVID-zero lockdowns would have to be driven by massive price increases. But part of what happened here is that China basically throttled the ability of domestic retail prices to rise through their normal regulatory procedures. Petrol prices in Beijing only rose maybe 30%, versus the doubling we saw globally. So again, it just doesn’t track for me that all of that, or even most of it, was demand destruction.
So then we go to substitution or outright releases of strategic petroleum reserves. The one thing we can say is that the inventories we can see — the floating-roof crude oil storage tanks — are still very, very high, roughly where they stood at the beginning of this crisis. As far as we can tell, they’re not drawing down aggressively on those stocks, at least not yet. The caveat is that with satellite analysis we can’t see underground storage caverns and their proper SPR. They have at least six storage caverns that we know of, about 131 million barrels. The likelihood is that they’ve been drawing those down, because crude oil imports fell far faster than refining run rates. So again, it had to be made up somewhere. The Occam’s Razor here is that they’ve been silently releasing additional crude inventories. But above and beyond that, crude refining run rates also fell dramatically, by 3 to 3.5 million barrels a day. So where’s the implied demand destruction? This is where we get one of two things. Either a very large release of refined product stocks — we know China has large stocks of refined products like gasoline, diesel, and jet fuel. We have virtually no firm information on those levels, and we can’t track them closely day-to-day or week-to-week, because unlike crude they don’t have floating roofs, so we have to infer. Those stockpiles are upwards of a billion barrels, but the implication is that they’re drawing them down very rapidly. The other thing we could have seen — and your colleague Javier Blas was on this very early — is the potential to switch some petrochemical feedstocks from oil-derived products like naphtha and LPG toward more gas-based products, natural gas, or even, in the extreme, coal-based chemical products.
4. How to Buy Cheap Claude Tokens in China – Qian Zilan
Underneath the handful of labs sits a much larger market, one that has been operating in public on GitHub, Taobao, Twitter, and Telegram. It is a grey economy of API proxies (commonly called “transfer stations,” 中转站) that lets Chinese developers access Anthropic’s models at as low as 10% of the official price. The participants extend far beyond selective experienced AI researchers, and the motivations are much broader than building a frontier model to catch up. Everyone who wants to use more advanced AI models or tools, be they university professors and students, tech workers, individual developers, or hobbyists, uses API proxies.1 The logs they generate may have become a commodity, traded for purposes ranging from model training to targeted fraud.
Meanwhile, every layer of control frontier US AI companies have added (geoblocking, phone verification, credit card requirements, and now live biometric KYC checks) has produced a corresponding layer of evasion infrastructure. These new SMS farms and biometric harvesting operations have implications that extend beyond geopolitics into how frontier AI safety frameworks are designed…
…A transfer station (中转站) is what the Chinese developer ecosystem calls an API proxy–an overseas server that sits between a developer and Anthropic’s infrastructure. It accepts API requests, forwards them as if they originated from the transfer station’s location, and passes the response back.2 The user redirects their software to the proxy’s server instead of Anthropic’s, and pays the API proxy RMB via WeChat or Alipay.3 This sidesteps both the VPN and the overseas credit card needed for direct access. Prominent transfer stations are catalogued in community repositories and ranked by real-time price and uptime. Below them, a longer tail of small and individual projects comes and goes.
While this setup sounds functionally identical to legitimate Western API aggregators like OpenRouter, transfer stations operate in an entirely different universe of legality and trust. Legitimate aggregators exist to simplify developer workflows, charging standard rates based on transparent enterprise agreements. Transfer stations, conversely, are built explicitly for evasion, routing data through unaccountable middlemen…
…A transfer station is not a sole entity. It sits in the middle of a layered supply chain, with most participants never interacting with each other directly.
Upstream are the resource providers: account merchants who bulk-register or acquire Anthropic accounts at scale; SMS verification platforms that supply the foreign phone numbers needed to pass sign-up checks; and, at the more technical end, reverse engineers who analyze Anthropic’s client code to find authentication shortcuts or detect when detection logic has changed. The payment infrastructure with card merchants and proxy networks also enables overseas billing from inside China…
…Almost no one operates the full chain. Most participants own one or two links and monetise those well, resulting in a resilient, modular system. AI model providers can suspend individual operators, but the upstream account pools and downstream customer base remain intact. So long as there are developers who want access to Claude and identity black markets willing to supply the credentials, which are both durable features, a replacement can be stood up quickly…
…The most curious thing, however, is not how to get access to Claude or Claude Code in China, but how to get it at a ridiculously low price–usually priced at 1 RMB per $1 of tokens — 70–90% below official prices. According to public discussions, there are at least three ways a transfer station makes this possible–often described as “one fish, three meals (一鱼三吃):
Meal 1: The markup on access. This is possible because of the upstream resource providers who can stack proxies using at least five relatively “innocent” tactics:
- bulk-registering API accounts to farm Anthropic’s $5 free credit
- reselling unused quota from others’ accounts
- corporate/educational discount arbitrage
- “APImaxxing” — one $200 Max plan carved up among multiple users via tokens-per-hour quotas, exploiting the gap between Anthropic’s flat subscription price and the far higher cost of equivalent pay-per-token API access…
…Meal 2: Swapping models and inflating tokens. Because users’ inputs and model outputs are mediated through a proxy, users cannot verify which model their request was actually routed to. A user selects Opus 4.7, but the proxy can silently route to Sonnet, Haiku, or, in the worst case, GLM or Qwen, and fraudulently relabel the output…
…Meal 3: The logs are the product. This is perhaps the most important part as it intersects with data privacy and distillation. Every request that passes through a proxy — full prompt, full response, tool calls, iterations — is sitting on the proxy operator’s server. For AI coding agents, those logs contain long reasoning chains, real engineering decisions, repository context, and human-verified correct outputs. This makes them an ideal dataset for post-training: for supervised fine-tuning on real engineering tasks, and, where full reasoning traces are captured, for distilling Claude’s reasoning patterns into smaller models.
5. How funerals keep Africa poor – David Oks
A modest, mid-level funeral in Ghana costs about $5,000 U.S. dollars; a “befitting” one can easily cost $15,000 or $20,000. And all this in a country with a median income of about $1,500 per year. Ghana is known for its particularly ornate funeral culture; but it’s not the only place in sub-Saharan Africa with a culture of exorbitantly expensive funerals. The average household in KwaZulu-Natal in eastern South Africa, for example, spends the equivalent of an adult’s annual income on a single funeral. We see the same tendency for ultra-expensive funerals in a striking number of places: the Democratic Republic of the Congo, Kenya, Nigeria, Benin, Cameroon, Mozambique, the Ivory Coast. It’s often observed, in fact, that families will spend more money on burying the dead than on keeping the sick alive: indeed, in the Kagera region of northern Tanzania, families spend 50 percent more money on funerals than on medical care…
…The answer, I think, is that the funeral isn’t really about the deceased. Funerals function as a costly signal of kinship group loyalty: and in that context, the expense of the funeral is the point. And, in turn, funerals tell us quite a lot about why so many societies across Africa have had so much trouble achieving economic “takeoff.” Kinship societies are actively hostile to economic growth, because economic growth undermines the basis of kinship: that is why kinship societies demand constant, visible sacrifices of wealth—funerals being the most spectacular—that make it extraordinarily difficult for any individual to accumulate capital, reinvest their assets, and pull ahead. The funeral is a window into a system of wealth destruction that serves, above all else, to keep people poor…
…African societies, by and large, are kinship societies.
So what are kinship societies?
You can think of modern societies as large collections of individuals, their lives structured by impersonal institutions like states and corporations. Kinship societies are much older: they are, in fact, the oldest and most durable type of human society. In a kinship society, life is centered on the extended family: the “clan,” the lineage, the tribe—a group that often includes many people who aren’t actually related. These kinship networks don’t act anything like nuclear families in modern societies. They are highly functional organisms: most of the functions provided by states in the modern world—protection from harm, credit, dispute resolution, eldercare, social insurance—are instead provided by the kinship network. If you fall sick, the kinship group will care for you; if you need cash, the kinship group will lend you money; if a stranger wrongs you, the kinship group will avenge you.
Of course, a kinship network isn’t a charity. It’s more like a mutual aid society that you’re born into and can’t leave: what the kinship group gives, the kinship group must also take. A huge amount of life in kinship societies is structured by the obligations that people owe to their kin…
…In a kinship society, nothing that you earn is truly yours. If you make money beyond the point of subsistence, you’ll be expected to share it with your less-fortunate relatives; if you start a business, you’ll be expected to hire your cousins or nephews or in-laws, even if they’re not the best possible employees; if you buy a car, you’ll be expected to lend it out to relatives who need it.
The result is a constant process of redistribution from the most productive members of a kinship group to the least productive. This informal redistribution is a constant feature of life in African societies: 93 percent of Kenyan entrepreneurs agree that success in business leads to financial demands from family and friends. South Africans even have a name for the sharing obligations that define African kinship groups: “the black tax.”…
…If the productive members of the group can defect—removing their resources from the common pool—then the whole system of mutual obligation begins to unravel. If a productive individual can simply withdraw from sharing obligations, then the network must demand more from those who remain, increasing the incentive to defect: so the entire delicate machinery of mutual obligation collapses in a slow cascade. This is the death spiral for kinship networks.
So from the perspective of the kinship network, wealth is a threat…
…You can think of funerals as another wealth destruction ritual. The genius of it is that it can’t be evaded: it is a public ceremony virtually dedicated to the immolation of wealth. In private, you might be able to evade your sharing obligations by hiding your earnings or your savings; but in public, at the funeral, the claims that your kin make on your wealth are at their most visible and least avoidable. You can’t simply not show up to your uncle’s funeral; and, if you show up, you will obviously be expected to contribute a handsome sum.
And this logic is even more powerful for those who are suspected of shirking their kinship obligations. It’s at the funeral where you must signal your willingness to honor sharing obligations most loudly. The lavishness of the funeral is a costly signal of continued commitment to the system of mutual obligation that holds the kinship group together. The point is that it’s expensive and incommensurate with your means.
This is why Ghanaian funerals, for example, have tended to grow only more lavish with time…
…And so the lavish funeral, in the end, is not a strange cultural quirk of African life, but the most visible manifestation of a social order oriented toward the destruction of accumulated surplus. And until the grip of that social order loosens, much of the wealth that Africa produces will continue to go, quite literally, into the ground.
Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Alphabet (parent of Google), Amazon, Mastercard, Meta Platforms, Microsoft, and Visa. Holdings are subject to change at any time.
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), Amazon, Meta Platforms, and Microsoft. Holdings are subject to change at any time.
