Author: admin

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 02 August 2026:

1. Cyprus – “weird” opportunities galore – Swen Lorenz

In 1999, the main index of the Cyprus Stock Exchange rose nearly tenfold, making it the world’s best-performing stock market.

It was one of history’s greatest investment manias:

• Brokers had such a backlog of orders that it took them up to three weeks to clear a trade.
• Some Cypriot farmers reportedly sold their sheep to invest the proceeds in the stock market.
• Hastily arranged IPOs were oversubscribed by a factor of 50-100x.
• One IPO by a small cruise company attracted subscriptions equivalent to 10% of Cyprus’ GDP.
• A small shopping mall suddenly reached a valuation of EUR 2bn, roughly equal to the entire market cap of the Cyprus stock market before the boom began.

The business section of Britain’s The Guardian reported at the time:

“Everyone, housewives, cleaning ladies, ministers and businessmen are in on the act … The mania has reached such heights that from Cyprus’ taverna-terraced beaches to its remote mountain villages few now talk of anything else.”

It ended much like almost every investment mania ends: in tears.

After breaking records on the way up, the Cypriot stock market broke records on the way down, too. The market index ultimately lost 99%, which, according to some, remains the largest decline ever suffered by a national stock market index. By comparison, Greece “only” lost 92.5% during its sovereign debt crisis, while Germany’s hyperinflation between 1918-1922 saw the market fall by 97%.

2. You just hired a million bad employees – George Sivulka

For the first time in history, humans are cheaper than software.

And AI is creating more jobs than it eliminates…

…Managing AI is harder than managing people, because AI scales dysfunction instantly. Fortunately, we can learn from the past:

Agent workforces and human workforces fail in the same way.

Understanding the 7 major parallels between the two will unlock the next trillion dollars of AI value creation…

…1. Tokenmaxxing is throwing bodies at the problem…

…People are spending so much on tokens because they don’t know how to use them.

Maybe 1 in 100 employees knows how to give AI context. It’s a rare breed of person that can articulate a process clearly, that has the patience to empathize with a polluted context window, or even understands what that means.

Give an agent harness to the other 99 people and they will produce “loops”.

2. Loops are meetings about meetings…

…3. Wasted tokens are the new headcount bloat…

…Just like 80% of employees do nothing, 80% of tokens today do nothing.

People create more people. Tokens create more tokens. Looping is the new empire building.

4. 100X tokens are the new 10X engineers…

…In the same way a handful of employees make others 10X as productive, for any given job some amount of token context can cut AI effort down by orders of magnitude. There exist tokens that will give you 100X as much leverage.

Humans are cheaper than tokens on average, but good tokens are cheaper at scale…

…5. Context hoarding is the latest job security tactic.

There’s a massive political problem with AI inside the firm, and it will only get worse.

Employees don’t want to teach AI systems their secret sauce….

…6. Evals are the new OKRs.

The best way to manage a token workforce is the same as the best way to manage humans: by defining what good looks like.

The one breakout AI use case that escaped politics is coding. It expanded the pie and made every engineer better.

The mechanism is evals. 99% of AI revenue today is coding because coding has built-in evals. Code runs or it doesn’t…

…7. The next trillion-dollar opportunity is the transformation company.

Enterprises have been buying foundation model commits, the application layer, and internal builds for years now. All of it conceals a brutal truth about the economics:

Nobody has AI working reliably yet…

…In fact, the next biggest businesses won’t be eating existing services spend. They will sell a net-new type of service to existing players:

“AI transformation companies” will be 10X larger than any neofirm.

Transformation sounds like a one-off project. But there’s a Jevons paradox at work: every use case an organization adopts surfaces ten more. The more AI-enabled a firm becomes, the more transformation it consumes, while the frontier of what’s possible advances daily. Ongoing AI transformation efforts will become the only way to compete.

3. Data-Center Builders Are Racing to Offload Stakes Worth Billions – Anissa Gardizy

Data-center builders and operators across the U.S. are working with bankers to sell majority equity stakes worth tens of billions of dollars in their companies this summer, according to people familiar with the efforts…

…Sales of data-center operators are on the rise as owners of these firms seek exits and investor interest in owning the physical infrastructure behind advanced artificial intelligence grows. A massive backlog of demand for server capacity pushed companies to pursue novel strategies to secure more of it, from renting chips from direct competitors to launching data centers into orbit…

…Whether there are enough buyers with the means to absorb so many large deals at once this summer remains an open question. Ravi Purohit, the co-head of infrastructure at Paul Weiss, said that while plenty of investors want exposure to data centers, only a handful of firms can afford multibillion-dollar deals…

…But as these megadeals scale up, developers are increasingly running into fierce local opposition, driven by residents’ anxieties over rising utility bills, noise and advanced AI services in general. In some cases, it has pushed developers to pause or walk away from projects. This all makes investing in data centers riskier than it was in the past, investors said.

“I think people are downplaying the significance of Nimbyism in this country right now,” Purohit said, adding that companies looking to sell this summer will be scrutinized for how they plan to handle it.

“The more they can demonstrate to buyers that they have a constructive relationship with these communities…that actually goes a really long way,” he said.

4. How Terrorist Groups Are Using A.I. to Gain an Edge in Battle – Dustin Volz and Eric Schmitt

When a gang of motorcycle-riding members of Boko Haram attacked a military base in eastern Nigeria a couple of years ago, they were stymied by a defensive trench surrounding the complex.The extremists regrouped. Before launching another assault, they asked A.I. for help.

“We saw in a movie how motorcycles can jump over bridges,” a former Boko Haram commander told Antonia Juelich, a terrorism and technology researcher at Cambridge University. “We used A.I. to learn how to do this. We gave it information, like what motorcycles we use and the distance we need to jump and so on, and it gave us steps on what we have to do.”

Using tips from chatbots, mechanics modified the motorcycles to allow for faster acceleration and top speed. The riders dug their own holes, filled them with broken glass and fire, and practiced jumps — sometimes with fatal outcomes — until they achieved enough aerial liftoff to mount a successful attack, defectors said…

…Until recently, the Islamic State, Al Qaeda and other extremists primarily used A.I. in the information-operations realm — propaganda production, translation, recruitment and security tradecraft. But that has evolved as jihadists have turned to A.I. for tactical on-the-ground advantages, according to current and former U.S. military and counterterrorism officials and independent researchers…

…Daniel Byman, a terrorism expert at Georgetown University and co-author of a report about A.I. and the future of terrorism released on Friday by the Center for Strategic and International Studies, said terrorist groups were “mixing and matching” from different A.I. systems, seeking to avoid technical guardrails established by the A.I. companies. Dr. Juelich’s research also found that Boko Haram was platform agnostic, interchangeably working with OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and xAI’s Grok, as well as the Chinese firm DeepSeek…

…Defectors recounted attending organized training sessions focused on how to best leverage the powers of generative A.I. models to inform or enhance their uses of the technology.

The trainings, in which laptops were equipped with virtual private networks and encryption software, were delivered via transnational jihadist networks often led by members of the Islamic State, interviewees said. Common topics included managing an account on an A.I. platform, suggestions on generating useful answers and tips on evading safety restrictions.

5. Are We Repeating the Biggest Mistake of 1873? (Transcript here) – Merryn Somerset Webb and Liaquat Ahamed

Liaquat Ahamed: What started it was the gold rush. Europe, which was the centre of the world, had gone through a terrible depression in the 1840s, fuelled by bad harvests and by revolution. They actually thought every government in Europe was going to fall and that we’d have the equivalent of the Russian Revolution across Europe. That didn’t happen — but at the same time, they discovered gold in the United States, and that provided the fuel to get the global economy going.

You had bankers like the Rothschilds, who had made a ton of money in the early part of the 19th century lending to governments, jumping on the bandwagon and starting to lend for infrastructure — particularly the railroads — but to the private sector. You got a massive boom in lending, and it was a boom based on the bond market. Everyone thinks the bond market is a sleepy place where people who don’t want to take risk put their money, but it was essentially the bond market that provided the impetus to growth. It grew by five times in the two decades from 1850 to 1870…

…Merryn Somerset Webb: That’s interesting, and we’ll come on to it — whether economies are incredibly resilient and whether the best thing to do is just leave them alone. Because when we get to 1873, we find out that everyone decided not to leave it alone. They succumbed to what we’d call “something must be done”-ism, and that’s when things started to get nasty.

But the pivotal moment is this war between Prussia and France that starts in 1870. It’s a short war, but it has enormous repercussions, because France is obliged to pay enormous reparations to Germany — to the tune of how much?

Liaquat Ahamed: Around a billion dollars.

Merryn Somerset Webb: A billion dollars, which would be the equivalent of $1.2 to $1.3 trillion today. And Prussia thought this was fine, because so much money would take France forever to pay off. This would keep them down and out of the way — the enemy dealt with indefinitely. Instead, France somehow managed to go out and get the bond markets to give them a billion bucks over two years. And it was the Rothschilds who were at the centre of that, because they had incredible reach.

Liaquat Ahamed: They raised two bond issues which together raised a billion dollars. But the most important thing was that in one case the issue was three times oversubscribed, and in the other, twelve times…

…Merryn Somerset Webb: And the London Stock Exchange went crazy.

Liaquat Ahamed: Because everyone who didn’t get into that bond issue still had their money.

Merryn Somerset Webb: Exactly.

Liaquat Ahamed: And in the US, the railroads had been a perfectly rational boom until then — and then suddenly railroad bond issues doubled. We went up to $500 million a year in an economy where that was roughly 5% of GDP.

Merryn Somerset Webb: Which today would be $1.5 trillion in the US. So that $1.5 trillion — the sort of money that goes into a SpaceX today — is what went into the railroads then. Fascinating. And then of course that billion dollars went to Germany over a two-year period.

Liaquat Ahamed: Yes, and they then had 25% of GDP to play with in cash.

Merryn Somerset Webb: And that was totally mismanaged.

Liaquat Ahamed: They could have fed it into the economy at a slower pace. But injecting 25% of GDP in liquid cash into an economy that was relatively unsophisticated — where everyone who owned government bonds suddenly found their bonds paid off —

Merryn Somerset Webb: Paid off meaning redeemed.

Liaquat Ahamed: Yes. They looked around and said, “What do I do with this money?” and started hunting for opportunities. And lo and behold, a whole lot of charlatans appeared on the scene to relieve them of it. It was not only a stock market boom but a giant IPO boom. The stock market, which had maybe 30 or 40 companies listed, suddenly ballooned to 500 or 600. A lot of them were banks, a lot were real estate, a lot were railroad companies — but there were also all sorts of things, like companies set up to explore the northern regions of Europe, or to look for banana plantations in West Africa. It was a crazy time…

…Merryn Somerset Webb: So the crash starts in Vienna. But because there had been a global boom and a global bubble, it spreads. There’s contagion across the world.

Liaquat Ahamed: Yes. You actually get a false period of calm. It crashes in Vienna, everyone says there’s going to be a global disaster — and then nothing happens for three or four months. So people said, maybe Vienna was just overpriced. It was a local incident. Everywhere else, earnings will rise to match prices and it’ll be fine.

But meanwhile, Jay Cooke — the premier investment banker, who had raised $2 billion for the Union government during the Civil War — suddenly found that because of the disruption in Europe he couldn’t raise capital. He ran out of money in the middle of constructing the second transcontinental railroad. It was a little like the Lehman Brothers story: he started injecting his own bank’s money into the project, and still couldn’t complete it. When he announced that he couldn’t raise the capital, there was total panic — the same sort of psychological panic that happened after Lehman. People said, “If Jay Cooke, a friend of the president and the most well-connected banker in the United States, can’t raise $100 million, what hope do we have?”

Every railroad stopped construction. There were 500 railroad companies in the United States; by the end of the year a third of them had stopped paying dividends, and within five years half of them had defaulted.

Merryn Somerset Webb: Everything comes crashing down.

Liaquat Ahamed: Everything comes crashing down.

Merryn Somerset Webb: And at the same time everything’s still going horribly in Europe — 70% of the banks in Vienna have gone bust. It’s global carnage…

…Liaquat Ahamed: What caused the boom to end? I think two things. One is that everyone tried to build railroads at the same time, so they started competing against each other and railroad profitability began to decline. The equivalent today is all of the hyperscalers trying to build AI infrastructure at once, and the price of tokens starting to collapse — which, by the way, has already started.

Merryn Somerset Webb: It’s already happening.

Liaquat Ahamed: It changes the economics of their investment. And at the same time, because of the disruption in Europe and the war, the price of capital started rising — and there are some signs of that happening now. Until recently we reassured ourselves that we could finance this boom out of the profits of the giant technology companies. But even that has proved inadequate, and they’re now going out and borrowing.

So the combination of declining profitability and a rising cost of capital at some point causes a crunch. In the US case, when Jay Cooke announced he couldn’t complete his railroad — the equivalent today would be OpenAI declaring, “Actually, we think we’ve miscalculated, we’re not going to be able to complete a whole model, and we’re going to have to sell to Microsoft.” Can you imagine the panic that would occur in the market for AI infrastructure?

---

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). Holdings are subject to change at any time. 

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 26 July 2026:

1. Keynote speech by Chinese President Xi Jinping at opening ceremony of 2026 World AI Conference

We human beings must answer the questions posed by our times: How to get along with thinking machines? How to ensure security when algorithm is part of decision making? How to tackle ethical challenges by technologies through adaptive governance? How to realize AI for all when the divide keeps widening? These questions demand serious consideration and real answers from the whole international community.

In China’s view, all countries should take a people-centered approach and develop AI for the positive and for good. We should ensure that AI is an important driver for shared prosperity and common security. We should join hands to build a just and equitable system for global AI governance. To this end, I wish to share four observations.

First, we should adhere to the principle of openness and win-win and boost innovation-driven development. As a new engine of world economic growth and an accelerator for the shift of growth drivers, AI is moving from the digital world into the physical world. We should seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing. We should facilitate technological innovation, industrial development and scenario-based application of AI. We should make coordinated advances in the transformation and upgrade of traditional industries, the cultivation and growth of emerging industries and forward-looking planning for future industries, so that all sectors and businesses can benefit from AI.

Second, we should strengthen risk-awareness and ensure that AI is secure and controllable. AI should be a trusted tool for humanity. We should take seriously the various types of inherent and secondary risks that AI may trigger. We should put in place laws and regulations, technological monitoring, early warning and emergency response systems in order to strengthen the line of security, prevent abuses and malicious use, and ensure that AI is always under human control. In the meantime, we should jointly oppose overstretching the national security concept in the field of AI and placing one country’s security over that of others.

2. The AI Bubble? – Nothing Linear?

Five very different numbers get quoted in this debate, and they blur together constantly:

• $725B — 2026 AI capex from the big four hyperscalers alone (Microsoft, Google, Amazon, Meta).
• ~$1.5T — estimated total AI infrastructure spend across the whole industry in 2026.
• ~$3T — Sequoia’s David Cahn’s estimate of the annual revenue AI must eventually earn to justify that build-out.
• ~$50–100B — roughly what the largest AI firms actually earn today: OpenAI ~$25B run-rate, Anthropic ~$30–47B ARR, plus Google/Microsoft AI revenue.
• The gap — the distance between the last two is the entire debate. It is the widest it has ever been, and widening…

…The three genuine choke points:

Advanced AI chips. Nvidia holds roughly 80% of the AI-accelerator market, and TSMC is effectively the sole advanced-node foundry able to fabricate the leading chips — virtually every leading chip, Nvidia’s, AMD’s, and hyperscalers’ custom silicon alike, passes through it. TSMC’s advanced-node capacity and CoWoS packaging are reportedly booked out for years, with Nvidia taking a large share. A single geopolitical or manufacturing disruption in Taiwan would immediately choke the entire industry’s ability to train or serve frontier models. Lesson: sovereign AI ultimately needs sovereign or allied access to advanced chip fabrication — software alone cannot route around this bottleneck.

Electricity & data-centre capacity. The IEA projects data-centre electricity demand roughly doubling from ~415 TWh (2024) to ~945 TWh by 2030, with AI the dominant driver. In some US states and countries (Ireland, Virginia) data centres already draw a fifth or more of total electricity. Power availability, not chip supply, is increasingly the binding constraint on how fast new capacity can be built. AI’s physical footprint — power, water, land — is becoming as strategically important as its digital one, and draws the same local political resistance as any heavy industry.

Frontier model training. Roughly six labs — OpenAI, Google, Anthropic, xAI, DeepSeek, Zhipu — can credibly compete at the very top, because training a genuinely frontier model costs tens to hundreds of millions in compute alone. But this choke point compresses faster than the physical ones: open-weight releases (Llama, DeepSeek, Qwen, GLM) have shown near-frontier capability at a fraction of the presumed cost. Algorithmic efficiency keeps lowering the compute needed to reach yesterday’s frontier.

A fourth layer, high-quality training data, is a contested rather than settled choke point: the freely-scrapeable internet text that trained the first LLMs is finite and non-renewing, and a fast-growing share of new web content is itself AI-generated — raising the risk of models training on their own degraded output (”model collapse”). Licensing disputes and lawsuits are actively reshaping what data can legally be used, pushing labs toward paid licensing, synthetic data and proprietary enterprise data. Data is shifting from an abundant free input to a scarce, litigated, paid-for one — a structural change in the industry’s cost base most forecasts do not yet price in…

…Part 5 · The bull case — why this may not be a bubble

The technology genuinely works. Unlike some past manias, the product is real and used daily by hundreds of millions. Generative AI reached an estimated 53% adoption among the general population within about three years — faster than the PC or the internet.

Revenue is growing very fast. The leading labs are scaling revenue at rates rarely seen: Anthropic went from roughly $1B to a ~$30–47B run-rate inside 18 months, and OpenAI reached a ~$25B run-rate. Fast growth can, in principle, close even a very wide gap.

The spenders are the strongest companies in history. Unlike the debt-laden telecom builders of 2000, the biggest AI spenders are among the most profitable firms ever. Nvidia earned roughly $120B in net income last year; Microsoft, Google and Amazon are cash machines. They will not simply collapse if returns are slow.

Valuations are rich, but not insane. The NASDAQ-100’s forward P/E is around 26x today, versus roughly 60x at the 2000 dot-com peak. Expensive, and concentrated in a few names — but not the pure fantasy of the last great tech bubble.

Individuals are extracting real value now. Even where enterprise ROI is unclear, the estimated value of generative-AI tools to US consumers reached about $172B annually by early 2026, with median value per user roughly tripling year over year.

Part 6 · The bear case — why it may be a bubble…

…Big-four AI capex has gone from ~$90B (2020) to $147B (2023) to $410B (2025) to $725B (2026). The gap itself has widened from a $200B question (2023) to $600B (2024) to an estimated ~$800B+ annual gap in 2026 — Allianz pegs the capex-vs-revenue divergence at ~46%, exceeding the ~32% seen in the 2001 telecom bubble.The cash-flow squeeze. PIMCO estimates Big Tech capex will consume ~94% of operating cash flow over the next two years, up from ~40% in 2023 — meaning for every $100 earned, only ~$6 is left for dividends, buybacks, salaries and everything else. For the first time, the hyperscalers are also leaning on debt: the big five raised roughly $108B in new debt in 2025.

Enterprise ROI — the crack in the revenue side. The core bull rebuttal is “enterprises will pay, because AI makes them radically more efficient.” The data from the industry’s own consultants is, so far, harsher. MIT Project NANDA found 95% of enterprise GenAI pilots showed zero measurable P&L impact (2025, 300 deployments). BCG’s “AI at Scale” survey of 1,800 executives found only ~26% of companies are generating meaningful financial value from AI. S&P Global found 42% of companies abandoned most AI projects in 2025 — more than double the prior year. McKinsey’s 2026 State of AI found fewer than 20% of pilots reach enterprise-scale production.

Most failures are not the technology’s fault — MIT found ~70% of the work to make AI pay off is process and workflow redesign, not the model. But from a market standpoint the cause barely matters: if the ROI is not showing up in the P&L, the revenue needed to justify the capex is not showing up either…

…The technology can be completely real and the bubble can still burst. That is not a contradiction — it is the normal pattern of every industrial build-out in history. The pattern runs four steps: high returns attract capital; capital keeps flowing until overcapacity is built; overcapacity triggers collapse; survivors inherit the wreckage cheaply and, when demand finally catches up, make fortunes on assets others paid to build.

3.Maybe Intelligence Ain’t All That – Clifford Sosin

Superintelligence arrived. You probably didn’t notice, because it turned out to be kind of incremental…

…We hold a thin scatter of facts about the world, and intelligence or reasoning is whatever fills the space between them. It extrapolates between things we already know, and LLMs are incredible at that. In studying our words, they learned the structure of our thoughts. We will soon have it in unlimited supply, at 200 IQ, for close to nothing.

Where the space between the facts behaves well, this is close to godlike. Real estate law isn’t hard. We made it up, it’s internally consistent, and a model holding every statute and ruling should beat any lawyer for free. Coding, math and most administrative work are similarly benign. What makes them easy is that they have relatively smooth solution spaces and are tractably verifiable.

Most of what matters doesn’t behave like that. The universe is mostly the emergent behavior of complex systems. Stir cream into coffee. Watch a storm build out of nothing but atmospheric temperature differences and water vapor. The local rules are simple and the tornado is not. In systems like these, no amount of reasoning delivers the answer, because there’s no shortcut hiding in the gaps. You have to run the thing. And running it has a ceiling of its own, since small errors compound, which is why no supercomputer will ever give you a clean two-week forecast. The smallest object that can perfectly simulate the universe is the universe. Human systems are similarly complex…

…An AI will hand you a genuinely clever design for a jet turbine blade. It might be far more likely to work than anything your engineers came up with. It’ll still probably fail, because that’s the base rate at the edge of what anyone knows. The only way to find out is to build the blade and try to break it.

That’s the real limit on learning, and it doesn’t care how smart you are. Coming up with ideas was never the hard part. The hard part is how fast reality answers them.

4. A Stock Certificate From 1941 Taught Me More About AI Than Anyone from OpenAI – Francis Huang

But I think the railroad story, if you take it seriously, tells you six things about what happens next.

First, the technology will work. That part isn’t in doubt. Railroads worked. AI works. The question was never “will trains move faster than horses?” or “will language models generate useful output?” The answer to both was always yes. The technology question is settled. Everything that follows is a finance question and an ownership question, and those are much harder.

Second, the buildout will be larger than anyone currently projects. In 1850, America had 8,879 miles of track. By 1860, it had 30,626. Nobody in 1850 would have believed that number. AI infrastructure spending has quadrupled since 2022 and shows no sign of slowing. McKinsey’s $6.7 trillion projection for 2030 might end up being low. When a technology changes the cost structure of everything, the capital required to build it out has a way of exceeding every estimate, including the ones that already seemed crazy.

Third, the financial system will change in ways we can’t anticipate. Railroads didn’t just use the existing capital markets; they created new ones. Bond markets, equity markets, underwriting syndicates, credit analysis, bankruptcy law, corporate governance, even the concept of the limited liability corporation, all got reshaped or invented to handle railroad finance. AI will do the same. We don’t know what the new financial instruments will look like yet. Maybe compute futures. Maybe revenue-sharing tokens tied to model performance. Maybe something nobody has named yet. The railroad precedent says the instruments themselves will be part of the story.

Fourth, a crisis will come, and its trigger will be something nobody is watching right now. In 1873, it was a Viennese real estate bubble. In 1893, it was the cumulative effect of rate wars nobody thought would last. Whatever hits AI won’t be “AI doesn’t work.” It’ll come from some adjacent system that’s quietly entangled with the AI buildout in ways nobody has mapped. A chip supply disruption. A sovereign debt crisis in a country that’s heavily invested in AI infrastructure. An energy bottleneck. Something nobody is talking about at Davos or on the All-In Podcast. The black swan, by definition, is the one you’re not looking for.

Fifth, the people who build it and the people who own it long-term will be different. The merchants of Boston who funded the Boston & Worcester Railroad in 1831 were not the Vanderbilt family that controlled it in 1900, and neither of them were CSX, which runs it today. OpenAI’s current investors, Anthropic’s current investors, the hyperscalers currently spending $700 billion a year: these may or may not be the entities that own and profit from AI infrastructure in 2050. Morgan got rich not by building railroads but by picking up the pieces after other people’s railroads collapsed. Someone will play that role in AI. We don’t know their name yet.

Sixth, and this is the one I keep coming back to: the infrastructure will outlast everyone’s financial projections. The tracks the Boston & Albany laid in the 1840s are still carrying freight in 2026. Data centers being built today will run computation, in some form, for decades after the companies that built them have been restructured, merged, acquired, or dissolved. The physical layer endures. The capital layer above it churns.

5. Why Are Berries Everywhere, in Every Season? Driscoll’s – Julia Moskin

In just the last decade, berries have completed the journey from fragile, local, seasonal treat to worldwide refrigerator staple and marketing juggernaut. Global production has tripled since 2000, according to research from the U.N.’s Food and Agriculture Organization, and still cannot keep up with demand. In sales and volume, berries are the fastest-growing category in American produce, according to data from the U.S. Department of Agriculture.

Most of that growth has been driven by Driscoll’s, a $7 billion California company that began as a multifamily farm in 1904, patented its first strain of strawberries in 1958 and is still controlled by family members. In 1989, its board made what the company calls the Meadowood Declaration, a resolution that seemed preposterous at the time: to make all four berries available, in every season, in every part of the world.

Today the company is the undisputed global market leader, shipping four billion containers of highly perishable fruit across 60 countries each year…

…According to Circana, a market research firm, Driscoll’s is now the second-highest-earning brand in American supermarkets, behind only Coca-Cola…

…Instead of owning land, the company owns the genetic material of its berries and the knowledge of how best to plant, pick and transport them. It subcontracts with farmers around the world to grow those breeds according to its specifications, then handles sales and distribution after harvest.

But global access to berries has a cost, measured in metrics like water consumption, pollution, pesticides and labor practices. Driscoll’s has come under fire on all four fronts…

…Inside a nearby laboratory, where two full-time sensory scientists make their assessments, 210 raspberry varieties were laid out in a grid of plastic pints. Some had been bred for visual appeal, with more shapely shoulders, uniform drupelets and less “hair” (the thin red styles that sprout where the berry is pollinated). Others were developed to maximize yield, with fewer thorns and better “plant architecture” — tall, fluffy stalks that make the berries easy to pick. Each cultivar is tested for qualities like P.S.I., the interior pressure that determines whether a berry will yield to the teeth with an explosive, juicy pop.

Out of those 210 strains, said Kyle Rak, the company’s chief raspberry scientist, perhaps two will make it to market…

…Many Driscoll’s berries are no longer planted in soil, but grown in pots filled with carefully balanced mixtures of organic materials like coconut fiber and moss. This system of substrate farming was developed over centuries in the Netherlands to produce maximum yields from minimal land.

It requires a substantial start-up investment by Driscoll’s growers, who also absorb the costs of ever-shifting factors like labor, weather, equipment and rent. The company provides seed plants and “inputs” like soil treatments, along with technical support and marketing dollars. After harvest, the company retrieves the filled clamshells, then compensates the grower according to the price those berries command. According to Driscoll’s, growers receive 75 to 80 percent of the revenue…

…Demand for berries has exploded in the United States because of overlapping recent trends: more snacking and the rise of “functional” foods that promise specific health benefits…

…Berries are shifting entire economies.

In 2023, they became Mexico’s most lucrative agricultural export, surpassing avocados, beer and tequila. On Moldovan plantations and in Andean highlands, growers of low-margin crops like sugar cane and corn have switched to berries, which command premium prices.

In 2025, China overtook the United States as the world’s largest blueberry producer. Driscoll’s, the first foreign berry company allowed to operate there, now has about 8,000 acres under cultivation. 

---

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, Meta Platforms, Microsoft, and TSMC. Holdings are subject to change at any time. 

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 19 July 2026:

1. The Great Wave Has Arrived – Tang Jie

If there is one thing we have learned over the past twenty years, it is this:

The greatest commercial opportunities never lie in minor adjustments to products or business models. They arise when the ceiling of intelligence itself makes a leap…

…Whoever can push that ceiling even one inch higher will be able to redefine the boundaries of what thousands of industries are capable of achieving. That single inch is precisely what the new generation of AI companies grounded in first principles are competing to secure…

…We have a simple but demanding definition of AGI:

AGI is not the intelligence of a single genius. It is the aggregate of all human intelligence.

It should be capable of creating original knowledge on the level of the theory of relativity. That is the only standard by which we measure whether the true summit has been reached.

On the road toward that destination, several mountains must be crossed. They are also where today’s technological wave is surging most powerfully.

The First Mountain: Long-Horizon Task Capability

The most exciting breakthrough today is teaching models to complete extremely long tasks—not merely answering questions immediately, but planning and executing over weeks, months, or even years…

…The Second Mountain: Fully Autonomous Agent Systems

Building on long-horizon capabilities, groups of agents that can operate independently, collaborate with one another, and work around the clock will become a new form of productivity…

…The Third Mountain: Self-Evolution

This is the most difficult—and also the most compelling—mountain of all.

AI training AI is already taking shape. Models are beginning to write their own code, clean and synthesize their own data, and train themselves…

…What will happen after these three mountains have been crossed?

AI will begin to learn what the “self” is and what self-awareness means. Beyond that, it may begin to touch human emotion. Farther still lies consciousness itself.

From perception to cognition, from cognition to general intelligence, and from general intelligence toward artificial superintelligence, or ASI—the road has already been laid…

…When AGI arrives, today’s applications may all need to be rebuilt as AI-native systems—or may no longer be needed at all.

Operating systems themselves may be rewritten. In the future, when you turn on a computer, what you see may be an “LLM OS,” with every function generated on demand.

Going deeper still, this represents a challenge to the von Neumann architecture that has underpinned computing for the past eighty years…

…As the supply of high-quality human-generated data approaches exhaustion, we will turn computing power into fuel for evolution.

This means building factories for high-quality synthetic data, using AI-versus-AI competition through self-play to generate knowledge from scratch, and giving systems the ability to reconstruct their own code within secure sandboxes.

The goal is to free the pace of evolution from the physical limitations of human engineers…

…The more powerful AI becomes, the more robust its safety constraints must be.

From the very beginning, Zhipu established a guiding principle:

AI must serve human well-being and national strategic priorities.

The company rejects bolt-on safety patches. Instead, it seeks to encode human ethics, social norms, and national laws and regulations into the model’s value function as foundational axioms.

We plan to commit resources on the scale of tens of billions to advancing mechanistic interpretability—clarifying the neural logic behind model decisions and transforming black-box systems into transparent, explainable ones.

2. The Reverse Information Paradox – Satya Nadella

In the AI age, the buyer risks giving away knowledge, just in order to use what they bought.

You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it!

Over time, the information asymmetry becomes increasingly skewed. The seller learns more and more about you as you use what you purchased, while you learn very little about what the seller is learning in return.

That is what I think of as the Reverse Information Paradox…

… That is why enterprises need a real trust boundary for their human capital and token capital to compound. It is where an organization’s data, traces, evals, adapted weights, and memory accumulate and improve together. And it is a hard boundary across which nothing crosses, not even the intelligence exhaust, without consent. Enterprises will demand the rights to use model outputs to fine tune and/or train their own models.  I think of this as every firm’s right to align models to their enterprise accountability obligations.

3. Away From the Casino Tables: The Wildest Value Gap Since 1999 – Sam Ziff

The chart below splits global equities into four valuation buckets, running from deep value to extreme growth, with each cohort’s current valuation shown as a premium or discount to its own 40-year median…

…The cheapest bucket globally (“deep value”) trades at a 45% discount to its historical average, while extreme growth sits at a 40% premium. Shallow value outside the US remains meaningfully cheap. The further you move from US large-cap growth, the more likely you are to find something trading below its long-run valuation. This is the backdrop against which we invest…

…The last time markets were this concentrated and this expensive, the next decade belonged to the parts of the market the crowd had forgotten. From the end of 1999 to the end of 2009, the MSCI World and S&P 500 delivered close to a zero total return, while emerging markets doubled and value outperformed growth.

4. The 221-year-old company that reinvented itself — 4 times – Eric Markowitz

Meet Jean-Joseph D’Ieteren.

In 1805, the 14-year-old orphan took over a small workshop in Brussels and started making wheels. Over the next 220 years, the company D’Ieteren founded became a carriage maker, then a builder of custom car bodies, then an importer of American automobiles, then the exclusive Belgian distributor of Volkswagen, then the world leader in automotive glass repair, and then — in a move that raised eyebrows across Europe — the owner of Moleskine, the Italian notebook brand.

Today, the D’Ieteren Group operates in more than 40 countries, employs over 32,000 people, and generates annual revenues exceeding $8.2 billion…

…The D’Ieteren family maintains a private museum tucked inside an ordinary working building on the Rue du Mail — roughly 5,000 square feet that most people in Brussels don’t know exists. There’s no sign outside. You call ahead, state your reason for visiting, and wait to hear if you’ve been accepted. If you are, you walk up a ramp, down a long corridor, and through a large sliding door. Then the city disappears…

…The company had walked away from its core business four times, yet it had preserved, with extraordinary care, a museum’s worth of carriages, photographs, and tools. It wasn’t the behavior of a company that had escaped its past. It was the behavior of a company that understood its past so precisely that it knew exactly what to keep and what to release…

…For most of the 19th century, D’Ieteren was a carriage maker of growing renown. It won medals at international exhibitions, earned the title of Supplier to the Royal Court, and built some of the most celebrated horse-drawn vehicles in Europe…

…In 1898, while still producing horse-drawn carriages, they built bodywork for 12 electric vehicles commissioned by Camille Jenatzy, the Belgian race car driver who would soon become the first person to break the 62 mph (100 km/h) land speed record…

…After a fire destroyed the old workshops around 1903, the family rebuilt with modernized facilities capable of handling the new work on a greater scale. For the next two decades, horse-drawn and automotive production coexisted — in many cases with the same craftsmen serving the same clients, who often owned both kinds of vehicles…

…On the eve of the 1929 stock market crash, D’Ieteren employed nearly 500 craftsmen and was exporting 65% of its production to Argentina, Egypt, Spain, and the United States. A single car would cost approximately $468,700 in today’s dollars.

But Lucien had also spent 15 years sitting with an uncomfortable truth.

During World War I, while assigned to a military vehicle depot in Le Havre, he had watched American cars roll through by the thousands: Studebakers, Packards, Fords. Standardized, efficiently built, priced for ordinary people. He came home understanding that the future was volume, not bespoke cars…

…What the crash did for Lucien — what crisis so often does — was to remove the social and psychological costs of changing. The reputation, the pride, the loyalty to craftsmen: all of it was real, and all of it had kept him from acting on what he already knew. The crash stripped that away…

…Crisis has a way of burning off the unnecessary, leaving behind only what has always been true.

One of the most commonly asked questions is, “What do you do?”…

…The better question — the harder question — is not what we do, but who we are.

Consider D’Ieteren.

In a narrow technical sense, they were a carriage maker, then a carmaker, then a glass repair company. If you’d asked them in 1850 what they did, they’d have said carriages. But that wasn’t really who they were. Who they were was a tight-knit family obsessed with quality, committed to the long view, and devoted to finding better ways for people to move through the world. The carriage was just the expression of that deeper question at the time.

5. A Framework for Frontier AI and the Dawning of a New Age – Demis Hassabis

AGI cannot be compared to standard technological breakthroughs, not even ones as consequential as the internet or mobile – it is much more akin to the discovery of electricity or fire…

…The magnitude of this technology’s impact will be unprecedented, perhaps 10x of the Industrial Revolution at 10x the speed. It will help us solve some of the biggest problems society faces from accelerating drug discovery to developing new clean energy sources to creating novel advanced materials. We could even reach a point where resources are no longer the limiting factor for human progress, leading to an amazing new era of abundance…

…Urgent action is needed to address risks that might arise as we get closer to AGI. We’ve already seen the challenges frontier models pose for cybersecurity, and other threats including nuclear and bio risks may soon emerge as capabilities continue to advance. On the horizon, we will need robust safeguards to maintain control of increasingly agentic, recursively self-improving systems – and tackle unknown issues that will only become clearer over time…

…The rapid progress we’re seeing in AI requires a new approach to testing frontier AI model capabilities that is dynamic, adaptable, and rigorous. The US is well positioned, given its economic and technical standing, to take the first step in developing such a framework. It could establish a new Standards Body modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority (FINRA), with a board that includes independent leading technical experts and open-source representatives…

…A model would qualify as ‘Frontier-class’ if it meets certain thresholds on a set of benchmarks determined by the Standards Body and regularly updated to keep pace with evolving AI capabilities. Organisations with ‘Frontier Models’ as defined by those benchmarks would be deemed ‘Frontier Labs’, and be encouraged to adopt best practices, such as publishing model cards with technical details, maintaining strong internal cybersecurity, vetting key personnel, and providing sufficient resourcing for safety and security research, and more.

Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release. Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market…

…Model assessments should include rigorous scientific evaluations of capabilities in cybersecurity, biological threats and other high-risk domains…

…Even if we solve these hard technical challenges, there will be further complex economic and philosophical questions to tackle: what sorts of new economic models will be needed to help everyone thrive in a post-scarcity world? What values do we want to live by, what will meaning and purpose be, and how might even the human condition itself change? Resolving these questions obviously cannot and should not be left to technologists alone. It requires every part of society to come together to help define this new chapter.

---

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 (Demis Hassabis is the CEO of Google Deepmind, an Alphabet company) and Microsoft (Satya Nadella is the CEO of the company). Holdings are subject to change at any time.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 12 July 2026:

1. AI’s Value Capture problem – Jaya Gupta

Imagine State Farm, Progressive, Allstate, Travelers, Chubb, AIG, Liberty Mutual, and 100-plus smaller carriers all running claims through the same model. Every carrier feeds it the same stream of context: the accident description, photos, repair estimate, adjuster’s note, borderline approval, fraud flag, override, payout, appeal, recovery outcome.

At first this is obviously useful. The model moves claims faster, flags suspicious cases, learns which repair estimates run inflated, which medical patterns look strange, and which overrides later become losses.

But if the same model learns from every carrier, is your claims judgment still your advantage? The underwriting exception that protected your loss ratio becomes a benchmark. The fraud pattern your team caught early becomes a feature sold back to the market…

…If everyone gets the same edge, customers keep it. Imagine an auto manufacturer using a model to negotiate semiconductors, resin, freight, contract manufacturing capacity, and substitute parts. The edge is buying better than the next manufacturer: knowing which supplier shortage is real, which quote embeds excess margin, and when preserving supply matters more than squeezing price. If every manufacturer runs procurement through the same model, the model does not just lower costs. It makes buying more “similar”. The best buyer loses the spread between its process and everyone else’s…

…The model also captures what compounds. Imagine 1,000 resource-constrained biotechs using Claude for Life Sciences because they do not have the internal platform of massive pharma company. Each company owns its compound, lab cost, failed program, and regulatory trail. But the workbench can see the pattern across all of them: which tox signal killed the program, which assay gave false confidence, which endpoint was weak, and which patient subgroup wasn’t the right one. If it sits across enough biotechs and pharmas, it can see failure patterns no single company can see. While data advantage is in exclusivity, a shared workbench breaks exclusivity by aggregation. And because Anthropic intends to develop drugs of its own, the tool you adopt for efficiency is built by the entity whose endgame may be to do what you do, using what it learned by watching the field do it…

…Data rights are not learning rights. Companies know how to negotiate retention, confidentiality, security, access controls, and training opt-outs. But the more important question is who owns the derived judgment: tasks, feedback loops, evals, workflow traces, corrections, failure modes, decision patterns, agent skills, and product insights. Once the model company knows the hard problem, it can acquire the job logic another way…

…The gain is front-loaded; the dependency compounds. The first adoption creates a real productivity jump. But once competitors run the same model, that jump becomes the baseline, and what remains is not your edge, it is your dependency on the next upgrade. Everyone will capture the first uplift but the vendor captures the recurring learning curve. Year one, the factory model reduces downtime, but then every rival has the same predictive maintenance workflow and the vendor owns the process intuition you now depend on.

2. High Bandwidth Flash: The Full Report – Austin Lyons

What’s interesting is that it has the same read bandwidth as an HBM4 stack, but with roughly 10x the capacity. And it’s made of NAND, not DRAM like HBM. (NAND is the cheap stuff.)

The first samples of the memory itself are expected very soon from Sandisk, sometime in the second half of 2026. Samples of the first AI inference devices built with HBF follow in early 2027…

…NAND flash has traditionally been used for information storage, i.e. where data lives when it’s not being used. Cheap, dense, and non-volatile. But slooooooowwwww….

But memory is where the accelerator keeps information handy during computation, and it has to be fast enough that compute never waits…

…So how could NAND flash possibly be used as memory?

Well, could there be a way to get the right data to the accelerator in time for computation, even with NAND’s slow reads? If you start the read it WAY before the accelerator needs the data, it could work?

In addition to intelligently scheduling when data is requested, one can also try to achieve high bandwidth from NAND. Bandwidth is the amount of data that arrives per unit of time. So if you know the read is going to take a long time, well, might as well have a bunch of reads running in parallel, if possible, right?

HBF unlocks high bandwidth by stacking NAND dies to increase the “width” or the parallelism. Each die is divided into many sub-arrays, which are small blocks of NAND that can each be read at the same time, independently of one another. An HBF stack places 16 of these dies behind a single interface, thousands of bits wide, so a huge number of sub-arrays are available to read at once. Any single read still takes thousands of nanoseconds, but thousands of reads can run in parallel, so a lot of data can be moved simultaneously…

…Conventional flash ships about 14 GB/s behind a PCIe 5.0 NVMe controller. Packaged as HBF, the same material delivers 1.6 TB/s. Roughly 100x the bandwidth, from packaging alone…

…HBF’s latency is still 10-100x slower than HBM. But if the accelerator knows what data it needs in advance, it can prefetch it and avoid waiting on any single slow read. The argument for HBF is that inference decode is the perfect workload…

…Inference cost scales with GPU count, and for today’s massive frontier models, GPU count is often driven by memory capacity per GPU.

Why?

Well, you need enough HBM to hold all the weights, but HBM is co-packaged with the accelerator; a fixed amount of HBM is bonded into each GPU package. So you can’t add memory without adding GPUs. Hence, bigger models mean more GPUs.

Of course, weights aren’t the only thing the accelerator needs to store in memory and acccess quickly and often. The KV cache and activations sit in memory too, and both stay in HBM or DRAM.

But the KV cache takes new writes every token; NAND’s endurance can’t handle that. NAND has much lower write endurace.

And activations need low-latency random access that NAND is too slow to give.

So HBF is for storing model weights.

And frontier weights are huge and always wanting to be even bigger. Wouldn’t it be nice to hold the model in significantly cheaper memory than HBM?

Recall that a 70B parameter model at fp16 (2 bytes per parameter) requires 70 × 10⁹ × 2 = 140 GB just for weights. A 1T parameter model thus needs 1,000 × 10⁹ × 2 = 2 TB. But that’s a lot of HBM; today’s shipping HBM4 stacks hold 36 GB (12-Hi); 16-Hi parts push 48 GB, and the JEDEC spec tops out at 64 GB. So a large model needs many stacks, which means many GPUs. Expensive!

That also means more interconnect (to move data around all those GPUs), which requires more power and, ultimately, a higher cost per output token (watts and $).

But HBF can provide 512 GB of capacity per stack!

So 512 GB per stack versus 48 GB for HBM4 is roughly 10x the capacity at the same bandwidth…

…For the same bandwidth, HBF has up to 8-16x the capacity at roughly 2x the power.Yes, HBF has worse bandwidth per watt, but for model weight storage one can argue the metrics that matter most are capacity and capacity per watt, where HBF wins.

SanDisk also claims a similar cost to an HBM stack despite 8-16x the capacity, which works out to roughly 10x lower cost per GB.

I thought NAND was way cheaper?!The raw NAND is cheap per bit, but an HBF stack isn’t cheap.

3. A Penny on the Dollar: Li Lu’s Bet on Russian Privatization, Part 1 – Tim Isgro

Starting in October 1992, each Russian citizen would be given a paper voucher that could be used, in government-organized auctions, to purchase a portion of a Russian business. 147 million vouchers would be printed, one for each Russian citizen, and distributed until January 1993 via local branches of the State Savings Bank. To obtain the vouchers, a citizen needed only to pay a nominal fee of 25 rubles (equivalent to 10 cents US)…

…By January, 144 million of the vouchers had been picked up and were circulating in public.

Importantly, the vouchers were freely tradable, which could help poorer Russian citizens monetize their vouchers if they chose to (and many did), and free tradability helped attract larger pools of capital that were needed in privatizing the economy.

Once Russian citizens had their vouchers, they could either hold them and use them at the auctions or they could trade them immediately for cash, typically on the streets of Russia. As soon as the vouchers began trading, it was apparent that they were trading for very low prices…

…As the chart shows, from 1992 to 1994, these vouchers were trading on the streets of Russia for anywhere from 5 to 25 USD each…

…144 million privatization vouchers were distributed to Russian citizens. These vouchers were mandated by the government to be converted into the ownership of 29% of all businesses in the country…

…Using a generous value of $20 per voucher, that implies the entirety of Russian businesses was worth $14.4bn at the time (equal to $20 x 144,000,000 vouchers / 20%). As if such an extreme undervaluation needs a comparison, the market cap of Exxon alone, a single US company, at its stock price low in 1993 was $71.7bn…

…Why were the vouchers trading at such incredibly low values?…

…First, inflation was rampant in Russia at that time. Prior to the fall of communism, the prices of many goods and services were set by the state. Starting in 1989 and in a few stages in the years following, Russian leaders slowly freed prices from government control. The effect was a massive inflation that gripped the country in the 1990’s…

…Second, and perhaps most obviously, the primary owners of the vouchers were ordinary Russian citizens. Now, the voucher auctions were designed to be as simple as possible, to the point where a citizen did not even need to know anything about a company to participate…

…Finally, and for interesting reasons, the vouchers were given a “face value” of 10,000 rubles (approximately 25 USD as of the end of 1992, albeit highly volatile)…

…The apparent effect of that 10,000 ruble face value (again, about 25 USD as of late 1992) on the trading price cannot be overstated.

It inadvertently created an anchor around which the vouchers traded…

…From December 1992 to June 1994, 15,052 Russian businesses were taken private, in whole or in part, almost all of them at prices that were incredibly low compared to their counterparts in other countries.

Boycko, Shleifer, and Vishny describe two of the bargains:

1. VAZ, the auto maker of the popular Lada cars, came out of its auction with a total market value of $45 million. As a point of comparison, in 1991, Fiat reportedly offered the Russian government $2 billion for the company.
2. Gazprom, the gigantic Russian natural gas monopoly, emerged from its auction with a market value of $228 million. This was roughly 1/1000th the value of put on the company by foreign investment banks, presumably by comparing it to other natural gas companies around the world…

…1994, for example, was an incredibly volatile year. The ROS index of 30 Russian stocks (created by CSFB) surged from 116 at the start of the year to a peak of 1,669 in September of that year. Yes, the index was up 1,338% in nine months. That is not a typo. It later dropped to a low of 443 in April 1995…

…So, 1994 was an important year for Russian privatization as an investment. It marks the end of the “voucher period” when Russian vouchers were trading at absurdly low prices and marks the deeper involvement of foreign investors (and the re-pricing and volatility that came with them). It is the time when the price of Russian businesses went from absurdly low to merely very low.

4. A Penny on the Dollar: Li Lu’s Bet on Russian Privatization, Part 2 – Tim Isgro

In Part One of our case study of Li Lu’s investment in Russia, we discussed the the fall of communism, the government plan to privatize Russian businesses, and the voucher system that was used to convey interests in those businesses to private citizens.

Now, I would like to focus on one of the two companies mentioned by Li to try to get a more specific sense of what he saw at the time. That company is Lukoil…

…Lukoil was formed in 1991 with the merger of three companies, the Langepas Oil Company, Urai, and Kogalym (the “Luk” in Lukoil), and in November 1992, Boris Yeltsin officially designated Lukoil one of three integrated holding companies. In those early days of the company’s formation and Russian market privatization, the trading of stocks was in its infancy and systems were rudimentary…

…Lukoil was cheap in those days, but how cheap was it? Let’s take a look at the valuation of Lukoil versus that of another leading oil company at the time, Exxon.

While Lukoil’s trading prices implied a market cap of 20 cents to 50 cents per barrel of proven oil and gas reserves around the time of Li’s purchase, Exxon’s valuation implied a market cap around $6 to $8 per barrel. In other words, Lukoil was trading about 5% the value of Exxon on a reserve basis ($0.35 divided by $7). At the time, the price per barrel of crude oil on world markets was around $20.

5. A Penny on the Dollar: Li Lu’s Bet on Russian Privatization, Part 3 – Tim Isgro

Li speaks briefly and vaguely about the market prices of Lukoil and Gazprom around this time and is not very specific about the exact time when he bought and sold the stock, but he does provide some clues:

Forget about the earnings. Just… the assets on the balance sheet. At the time oil prices [in the world market], I think the four or five year average was around twenty dollars [per barrel], and the [value per barrel of proven oil reserves for Lukoil was] at really low prices… about 10 cents to 20 cents per proven barrel of oil on the balance sheet and that’s not even counting the earnings… This is how low it went. It was ridiculous.

From this quote and others throughout the talk, it seems that Li’s main focus was just how cheap Lukoil was trading relative to how much proven oil reserves were on its balance sheet. He seems to have virtually ignored any loss (or any income) that the company was making at the time, reasoning that such an extreme undervaluation relative to the company’s assets dwarfed the numbers on the income statement…

…All in all, with the limited information we have, I think the most reasonable conclusion is that Li made his initial investment in Lukoil somewhere in early 1995 to mid-1996 when the stock was trading around $3 billion to $5 billion in market cap and $0.30 to $0.50 per proven barrel of oil and gas reserves, and he sold it somewhere around mid-1997 to mid-1998, in the region of $12 billion to $20 billion of market cap and $1.25 to $2.00 per proven barrel of reserves. (Recall his comments from earlier, that he sold “two years after” he first bought it and that the $2.00 price per proven barrel of oil no longer looked protected.)

With those crude, round numbers, his investment would have made him anywhere from 2.5 to 6.6 times his money in just over two years of time…

…Li talks about the dramatic cheapness of Lukoil and other Russian stocks at the time, and he was right to some extent. But that cheapness had its limits. Below, I show the price of Lukoil’s stock from 1993 to 2021. I also show a variation of the chart with Lukoil’s market cap versus the value of its proven oil reserves (a measure of Lukoil’s “cheapness”).

The charts paint a picture that is difficult to rectify: For virtually the entire period from 1993 to 2021, Lukoil appeared to be cheap, trading at a market cap that almost never valued the entire company greater than 8% of the value of its proven oil and gas in the ground. Exxon (and later Exxon-Mobil), by comparison, averaged a market cap of 34% of the value of its reserves from 1993 to 2021.

So, at all points, an investor might have thought that Lukoil was cheap. And yes, buying the company’s stock in 1995 and holding it for two years, like Li, would have produced a great return. Even holding it for 10 years, from, say, June 1995 to June 2005 would have produced an annualized return (excluding dividends) of 21%, as the stock went from $5.21 to $34.75.

However, the next 10 years, through June 2015, would have only produced an annualized return of only 3%, despite the company appearing cheap in June 2005, when its market value was only 4.7% of the value of its reserves…

…This case study was particularly enjoyable for me because the lessons are so difficult to tease out. Simply buying Lukoil stock at any point in its history because it was cheap relative to other companies around the world would have been a mixed bag. Buying in the 1990’s or early 2000’s would likely have worked out great. Buying in the late 2000’s or the 2010’s would likely have been poor. At all times, Lukoil looked cheap versus Exxon and other western oil companies.

It is very difficult to know how to think about this issue, but one thing to keep in mind is the timing of Li’s investment. In the early to mid 90’s, Russia was emerging from communism and still getting accustomed to the cultural shift toward capitalism and democracy. One could argue that, although corruption was still rampant, the prevailing winds were blowing in the direction of a country getting more used to democracy and slowly reaping the benefits of capitalist markets. These trends could serve as a gradual but important kind of catalyst to close the gap between price and value. In Russia, for example, these changes would slowly lead to more Western investors participating in Russian markets through the 1990s and 2000s.

But it’s important to realize too that the lack of such change (or timing) could make for a difficult investing situation, whereby an investor thinks a stock is cheap by some measure but that situation sticks around for many years.

So one takeaway from Li’s investment is that extreme cheapness is a great thing to hunt for, but seek to have it come along with a changing situation or an outright catalyst.

---

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 no vested interest in any companies mentioned. Holdings are subject to change at any time. 

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles 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, Meta Platforms, and Microsoft. Holdings are subject to change at any time. 

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 28 June 2026:

1. The State of the AI Economy – Azeem Azhar, William Gildea, Hannah Petrovic, Nathan Warren and Marija Gavrilov

$110bn trailing 12-month revenues – now at a $175bn pace…

…AI is scaling three times faster than any IT wave…

…AI demand is reigniting a moribund US power sector

1950-2008: +6 TWh/month annual growth

2008-2024: ±0 growth

2024-today: +9 TWh/month annual growth…

…Against GDP, AI revenue is still a rounding error

Still tiny: AI revenue is equivalent to 0.42% of US GDP (vs IT sector’s 9.4%)…

…Seven in ten GenAI claims focus on cost savings or efficiency

Claimed AI outcomes 

S&P 500, Q4 2022 – Q1 2026

Revenue gain: 6%

Conversion improvement: 7%

Quality improvement: 18%

Throughput increase: 22%

Time savings: 23%

Cost reduction: 25%…

…Revenues cover the ongoing expense, not yet the cumulative bill

Q4 2025: Quarterly revenues first exceed CapEx depreciation…

…Still ~half-covered: cumulative revenue has nearly covered cumulative depreciation, but still has to cover the expected headroom…

…AI infra revenue now just clears today’s depreciation hurdle

GenAI revenues now cover the quarterly depreciation of AI infrastructure. Q1 26 headroom reached 19% for hyperscaler/neocloud revenues and 32% across all GenAI revenues.

Coverage remains thin. Depreciation absorbs roughly 81% of hyperscaler/neocloud GenAI revenue and 68% of total GenAI revenue before additional costs.

The next test is incremental coverage. As committed AI capex enters service, the depreciation base will rise. Revenue growth, utilization and pricing must continue to compound or headroom will compress again…

…Gross rental yields suggest useful lives extend past six years

Older GPUs earn yields long beyond their six-year depreciation life…

…This efficiency is increasing monetization per GW of capacity while revenues per token fall

Revenue per trillion tokens has fallen since its 2023 peak, mirroring price declines.

Efficiency gains drive lower token prices, which are more than offset by higher demand…

…AI demand is more revenue-validated than any prior platform shift. The investment case comes down to whether falling prices can move enough token volume to earn a return on CapEx.

2. Morgan Stanley Pitches Clients on a New Market for Data Center Loans – Dakin Campbell

Over the last few months, Morgan Stanley has suggested to clients that the next time they need to raise money for data center projects, they consider the leveraged loan market rather than the bond market, according to a person familiar with the matter who asked for anonymity to discuss private conversations.

Leveraged loans are those made to companies that don’t have investment-grade credit ratings, typically because they don’t have businesses that throw off lots of cash or they already have lots of debt. Such borrowers could include AI firms like OpenAI or new cloud providers such as CoreWeave…

…Leveraged loans are typically underwritten by an investment bank like Morgan Stanley. Most are then sold to financiers that bundle the loans into a single pool. That pool is then sliced up and resold to other investors based on their risk tolerance. These pools are known as collateralized loan obligations…

…Last month, Morgan Stanley brought the first AI-linked offering to the leveraged loan market when it sold $3.1 billion of notes on behalf of CoreWeave, which said it would use the proceeds to buy chips for OpenAI and Cohere. Investors placed more than $19 billion of orders, Bloomberg reported…

…Until now, most data center financing has been done via the bond market, either as junk bonds or—in the case of cash-rich tech firms like Google—less expensive investment-grade bonds…

…Other questions include the identity of the company actually leasing the space in the data center, and whether loans to finance chips get paid down on a schedule parallel to the chips’ expected useful life.

CLOs are a type of structured credit product, similar to the collateralized debt obligations that bundled mortgage loans and derivatives in the run-up to the financial crisis—debts that then suffered tens of billions of dollars in losses. CLOs haven’t experienced a similar blow-up, but many industry watchers worry that they contribute to financial instability by spreading the risk into corners of the financial system that can be hard to track.

3. China’s tribute system and the new world order – Ray Dalio

China is earning huge amounts of money from its exports, so Chinese companies and banks are building up large capital surpluses and accumulating buying power. This is exerting upward pressure on the Chinese renminbi relative to the US dollar and leading to its increased use for trade and capital transactions. Chinese investors and capital markets are emerging as competitors to their American counterparts…

…The tribute system was informed by Confucian values — in particular the idea that order comes from having clearly defined hierarchical roles. Relations within it are not between equals, but between superiors and subordinates that recognise their relative positions. The more powerful ones in the hierarchy should treat the less powerful well, and the less powerful should treat the more powerful well, so that there is harmony. If a lesser power treats the greater power inappropriately, the more powerful one punishes it, typically not violently but through pressure and deception. As Sun Tzu wrote in The Art of War, “to subdue the enemy without fighting is the acme of skill”…

…A military blockade that stops chip exports is just one of many potential pressure-points that China can exploit, but it is notable because the Chinese have a plan to be self-sufficient in chip production by late 2028, while the rest of the world will remain dependent on Taiwan.

Given these circumstances, China could put the US into the awkward position of needing to choose between fighting or not fighting, with each choice not to engage leading to the perception of diminished American power, so that China can gain ground by simply making threats. 

4. Is Ray Dalio correct that China is reviving the tribute system? – Arnaud Bertrand

China’s ancient tribute system – called 朝贡 (cháogòng) in Chinese – is typically very misunderstood in the West: we typically think it involved tributary states paying some form of “tribute” to China in exchange for protection – the way medieval vassals would pay fealty to a lord in Europe…

…The system was basically a quid-pro-quo where China would get “得名” (dé míng, literally “getting name/prestige”) while tributary states would get “得实” (dé shí, literally “getting substance/material benefit”) in exchange. It was about China paying huge amounts of money and other material benefits for the recognition of its centrality…

…Very concretely the way it worked is that tributary states would pay largely symbolic tribute to China (like local specialties and curiosities, the system codified that tribute should be “easy to obtain and not costly”, 必易得而不贵) and they would in exchange receive 3 layers of economic benefits:

Immediate payback in the form of money and expensive goods (silk, brocade, porcelain, tea, silver, etc.), which value was typically dozens of times the value of the tribute received by the emperor The right to trade during their tribute visit: the envoys’ entourage could trade with specially licensed Chinese merchants at the Huìtóngguǎn (会同馆, the official guesthouse in the capital) Most importantly, and that’s where the real money was, they would be granted the right to trade at Chinese ports. Under the Ming maritime prohibition, tributary status was the only legal entry point into the Chinese economy…

…He is however wrong to describe the tribute system as one fundamentally based on pressure and intimidation. As we’ve just seen, it was pretty much the opposite: the basic idea was to be so generous that everyone wants in (to the extent that countries would literally fight to be tributaries), not so threatening that nobody dares leave…

…That being said, he is ironically correct – I think – that there is some form of revival of a tribute-like system but not in the way he understands it: China will (and does) use trade – its “generosity” – as a gravitational force to pull countries into its orbit. Not by threatening to cut them off, but by making the relationship too valuable to walk away from. THAT is much closer to how the actual Chaogong system worked…

…Which, incidentally, is why you can be extremely confident that China will go to enormous lengths to develop its internal market, and why the current situation where China runs huge trade surpluses is facing mounting pressure to change from within China itself. If countries don’t feel they’re benefiting enough from trade with China, the entire logic collapses. That’s why developing domestic demand isn’t some target China sets itself to assuage Western demands, as some claim: it’s genuinely a strategic imperative.

It’s also why it’s ironic that the West is so keen on pushing China to boost domestic consumption: in effect, it means we’re already in a de-facto Chaogong-like system and they’re asking that the carrot be bigger.

5. Oil Prices Make a Stunning Retreat to Prewar Levels. Where Do We Go From Here? – Collin Eaton and Benoît Morenne

The U.S. war with Iran—and the economic war the latter waged in return—was supposed to be an apocalyptic moment for the oil market. Instead, oil prices are on the cusp of falling back to their prewar levels.

Their stunning round trip, just 11 days after President Trump reached a 60-day deal to reopen the Strait of Hormuz, has disrupted widespread expectations that the global oil market’s recovery would take months, at minimum…

…Tankers loaded with crude are leaving the waterway in droves; gulf countries are racing to resume crude exports; and some of the largest buyers of crude on the planet are proceeding without using as much oil. Analysts at JPMorgan Chase said this week that global energy flows had shifted in ways they hadn’t expected.

“The market has rebalanced through a meaningfully different mix of demand losses and inventory withdrawals than we initially assumed,” they said.

The reprieve could be short-lived. Some oil analysts are warning that the sinking prices don’t fully reflect how tight the market remains after months of draws on global oil inventories, which are now flirting with operational limits…

…Tanker traffic through the strait has climbed swiftly since the U.S. and Iran struck an accord on June 14. A postwar record of 78 tankers sailed through the waterway on Wednesday, up from a previous high of 49, according to S&P Global. That represents 57% of prewar traffic levels…

…Oil demand in China, the world’s largest importer of crude, appears to have fallen faster than JPMorgan analysts anticipated, implying that its economy might be adapting to higher energy prices more efficiently than experience would indicate, they said…

…Whether China picks up new purchases in the coming weeks will have a huge influence on the markets. Analysts said the country might not want to reduce its strategic reserves further…

…Over the past three weeks, roughly 2 million barrels of oil a day has come back on to the market, with Iran pumping out barrels faster than Saudi Arabia and the U.A.E., according to the research firm Rystad Energy. But it will likely take until October for Iraq, Kuwait and other gulf countries that had to slash production to pump oil at full speed, analysts said.

These barrels of oil aren’t immediately available to stocks around the world, which are still being depleted.  

---

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 no vested interest in any company mentioned. Holdings are subject to change at any time. 

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 21 June 2026:

1. Tenneco Automotive: Charlie Munger’s $80 Million Bargain, Part 1 – Tim Isgro

The story of Charlie Munger’s investment in Tenneco Automotive is a fascinating one. And as far as I can tell, it’s only been told in a cursory way before now.

Munger made the investment in 2001 and it likely returned to him somewhere between 4 1/2 to 7 times his money and an annualized return over three years of 65% to 93%…

…What remained at the end was Tenneco’s automotive business, which sold emissions products (exhaust systems) and ride control products (like shocks and struts). It was this business, Tenneco Automotive Inc., that Munger was considering in 2001.

To say the above series of transactions dramatically changed the nature of Tenneco’s business is an understatement. The company went from being a large, diversified conglomerate with $13.2 billion in revenue in 1993 to a smaller single-line automotive businesses with just $3.5 billion in revenue in 2000…

…After all these spinoff transactions were finished, Tenneco was left with approximately $1.5 billion of long-term debt. After reading the history of Tenneco above, I suspect the automotive business was a victim of circumstance with respect to its debt load, being the last business standing after management spun off or sold five others…

…Focus for a moment on the company’s Operating Income (EBIT) and Interest Expense. Prior to the spinoff of Pactiv, the company was doing well, earning $633mn in EBIT in 1998 and spending $240mn in interest payments, but after the spinoff, the company was earning only $115mn in EBIT and spending $186mn in interest payments…

…To add insult to injury, Tenneco’s revenues were also suffering from the 2001 recession.

Tenneco served two broad sets of customers, original equipment manufacturers (auto makers) and the aftermarket (auto repair shops). Both were suffering lower sales…

…Not only were the interest payments on Tenneco’s debt too much for the company to handle in the years after completing its spinoff, but, on top of that, principal payments were starting to come due in 2001. The annual report from 2000 lists those upcoming maturities as $54 million, $109 million, and $99 million for 2001, 2002, and 2003, respectively. From the perspective of a casual analyst or observer, it was not clear at all how Tenneco could make those payments or how likely they were to work with their lenders on renegotiating terms.

2. Tenneco Automotive: Charlie Munger’s $80 Million Bargain, Part 2 – Tim Isgro

Tenneco produced auto products in two business segments: Emission Control and Ride Control. In both of those business segments, it had brand names with an excellent reputation and market share…

…Moreover, Tenneco’s list of original equipment manufacturers was large, including just about every major auto maker in the world. And the largest automaker (GM) accounted for only 16.6% of the company’s sales, indicating the sales were nicely diversified…

…Munger understood the great reputation of Tenneco’s products, as indicated by his brief comments at the Daily Journal meeting, when he stated:

I kind of knew based on experience how sticky some of that auto secondary market was, and how many old cars needed Monroe shock absorbers.

I think this point is critical to understanding Munger’s willingness to purchase these securities. Since customers loved and needed Tenneco’s products, the company still had fundamental value as an ongoing concern…

… Importantly, and perhaps underappreciated by the market, Tenneco was still in the midst of a major transition in its business. It had gone through five major spinoffs or sales of business units since 1993, it was facing its first recession since that time, and it was coping with all the debt it was saddled with after those spinoffs.

But digging in a bit to the company’s annual and quarterly reports makes it clear that management was keenly focused on right-sizing company expenses and running a more efficient organization…

…So, as of the end of 2000, management expected to generate a total of $92 million of savings by right-sizing its workforce and by adopting more efficient processes and practices.

In fact, it was already becoming evident by Q3 of 2001 that those efforts were working better than expected. Figure 10 below (which is Figure 8 reproduced) shows annualized operating costs (plus DD&A) that were $104mn lower than those for the year 2000. And those lower operating costs boosted EBIT by 32% to $152mn…

…I think these positive points are ultimately what caused Munger to believe that the company’s bonds, around a price of 35, and the company’s stock, around a price of $1.55 per share and a market cap of only $59mn, were way too cheap. For reference, the entire enterprise value of the company was $1.1bn, a figure I arrive at by conservatively assuming that the company’s debt is valued at par (apart from the 11 5/8% bonds, which I value at 35 cents).

I think Munger saw a very difficult financial situation for the company, and he probably acknowledged that a further, prolonged downturn in the economy and/or a group of unfriendly bank lenders could have pushed the company into bankruptcy. And in bankruptcy, in the wrong economic environment, it was quite possible that his debt and equity got wiped out.

The fact that Munger did not invest fully in Tenneco’s equity, which was more likely to be wiped out in bankruptcy, and chose to split his investment between bonds and equity, shows that he realized this was a possibility…

…I think Munger likely reasoned about how a potential bankruptcy might play out, and I think this was the most important point of all, prompting him to make his investment.

If Tenneco was forced into bankruptcy, its lenders would then have to decide on the best course of action that might get them a full recovery on their lending amounts. The total amount of long-term debt outstanding was $989mn plus the $500mn of subordinate bonds which Munger would invest in.

Tenneco’s lenders would rightly ask themselves: How are we best off to recover our $989mn?

1. We could force a liquidation of the business and attempt to be paid in full. That would involve a few years of wind-down work, staggered employee layoffs and plant and equipment sales, along with the severance and interim operating costs that come along with it. Plus, we would also need to engage in a process to sell the valuable Walker and Monroe brands, two of the most valuable assets the company had.

Or…

2. We could effectively realize the value of those brands by recapitalizing the company and operating as usual. One way to do that might be to forgive Tenneco’s debt completely, take an equity stake in the new company without debt, and then sell the equity in the new company to make ourselves whole on the lending amounts…

…We also know that Munger made “$80mn” on the investment. But we don’t know specifically how much he invested in either of the securities…

…Assuming Munger invested somewhere between 25% and 75% in Tenneco’s bonds (and stock), he likely made anywhere from 4.5 times to 7.2 times on his investment in three years, from December 2001 to the time the bonds were called in December 2004 (and when he likely sold his stock as well). Those returns imply an annualized return of 65% to 93%…

…I think there is one clear takeaway from Munger’s Tenneco investment.

When you encounter a company with a quality, in-demand product and/or a great brand, and that company is suffering, look twice. 

3. Systems of Record Won the SaaS Era – Clearinghouses Will Win the Agents Era – Jamin Ball

In financial markets, the clearinghouse sits between different parties that aren’t able to fully trust each other. The clearinghouse verifies / authorizes / settles trades, and ultimately keeps the receipt. Nobody really loves the clearinghouse, but it’s clear it has to exist for the ecosystem to transact.

Now think about where enterprise software is heading. Agents from tons of different vendors, acting autonomously, touching your most critical data, and even in the future spending real money. Some company has to sit in the middle of all that and decide: which agent is cleared to act? On what data? With what limits? And can you prove what happened after the fact? Whoever holds that seat holds incredibly “strategic real estate.” (and every founder I’ve worked with has probably heard me discuss strategic real estate over and over). That’s the clearinghouse.

This may sound counterintuitive, but owning the clearinghouse for agents (given agent companies themselves will want to be the clearinghouse) may create a deeper moat than the one systems of record had. A system of record controlled your data. It kind of controlled your workflows (but not always, oftentimes someone else controlled the workflows, but the data in the system of record was a critical part of the path). The Clearinghouse controls four things: memory (what your agents know), context (what they see and how it’s served), execution (what they’re allowed to do), and governance (who’s allowed to do what, plus the audit trail behind all of it). If migrating off a system of record was painful, migrating off the thing that holds your policies, your permissions, and your entire audit history is probably harder (especially when the agents start to handle more and more of the work). AND – I think these agent companies that become The Clearinghouse will start to look more and more like systems of record in their own right. Data in systems of record were oftentimes transactional data. Data in agent systems of records (ie Clearinghouses) will be agent traces, agent evals, agent telemetry data, agent A/B data, etc

4. Automation’s Asymptote: Part 2 – Abdullah Al-Rezwan

Tom Reed wrote a very good piece last month arguing that we may be pursuing what he calls “Goodhart Singularity”. Reed’s counter to automation doom is disarmingly simple: you cannot get good at solving problems without access to a source of problems, and the only source of most problems is slow, expensive interaction with the real world. Without that contact, the recursive loop produces something far less impressive than advertised. From Reed’s piece:

“The output of the R&D produced by an isolated datacenter of geniuses would be a mere Goodhart Singularity.4 An isolated AI improving itself against benchmarks would only appear to be approaching superintelligence, while actually optimising for eval performance that fails to generalise beyond the lab.”

Why would self-improvement stall outside the lab? Because models get good at what they practice, and for most economically valuable work, there is nothing to practice on. Reed’s most clarifying observation is about what kind of data exists at all:

“For most tasks in the economy, the pretraining corpus contains writing about the task, but not a record of the task itself. This is of course one of many reasons coding has progressed faster than other domains – code is one of the neat cases for which the task itself is almost entirely reducible to its token trace.”

The internet contains commentary and advice in abundance, but the actual steps of closing an M&A deal or deciding which drone prototype to ship were never serialized into tokens. The natural rebuttal is that a sufficiently smart system can simulate whatever data it lacks. Reed is skeptical that simulation is a viable path:

“Consider that almost half of SWE-Bench submissions accepted by AI auto-graders would be rejected by the actual human maintainers of the relevant repositories. The fact that you can pump SWE-bench scores without increasing actual merge rates is, to me, suggestive of the situation the datacenter-genius will find itself in.

The great Zhengdong makes this point about the progress of AI research itself. Not only are “evals” the only things that models are capable of getting good at, but “the researchers [themselves], they just wanna optimise… they just want an important problem to solve, a clear evaluation that measures progress towards it, and then they just wanna optimise it.” I suggest that AI companies need real-world deployment as a source of problems, or else they will have no good targets for optimisation.”

5. Mao’s economic record wasn’t bad, actually – Arnaud Bertrand

One number for you: under Mao, China’s GDP PPP per capita (meaning per person) was multiplied by about 2.5x from just above $400 in the early 1950s to nearly $1,000 in 1978. These figures aren’t from a “communist source”, they’re taken straight from a report by the Congressional Research Service, the research arm of the U.S. Congress…

…This is confirmed in another report by the extremely serious National Bureau of Economic Research (NBER), one of the most prestigious economic research institutions in the U.S., who found in a report entitled “The Economy of People’s Republic of China from 1953” that “the Chinese economy in 1952-1978 grew rather rapidly” with an average annual growth rate of real GDP of 6%. This equates to the overall Chinese economy being multiplied by 5 over the Mao era, which is consistent with China’s GDP per capita nearly tripling since the Chinese population simultaneously increased by 75% during the period (5 divided by 1.75 equals 2.85)…

…The data is overall clear: during the Mao era, China outperformed both its most comparable peers. It grew roughly 25% faster annually than India (5-6.7% vs ~4%) and modestly faster than Indonesia (5-6.7% vs 4.8-4.9%). Which means that whatever criticisms one might make of Mao’s policies, the prevalent Western narrative that he presided over an “economic catastrophe” is demonstrably false. The reality, confirmed by American research institutions, international databases, and comparative studies alike, is that Mao presided over significant economic expansion that exceeded comparable peer nations.

Sure, it wasn’t all plain-sailing, to say the least. For instance during the Great Leap Forward, according to the Penn World Table data, China’s GDP contracted by 20.8% from its 1959 peak to the 1962 trough – a severe three-year recession that took until 1965 to fully recover from. Similarly, at the beginning of the Cultural Revolution, GDP contracted by 5.9% from 1966 to 1968, with back-to-back annual declines of 3.3% and 2.7% before rebounding strongly with 9.9% growth in 1969…

…We shouldn’t dismiss the human toll that the Great Leap Forward inflicted. It remains the most severe policy failures in modern Chinese history, causing genuine excess mortality and widespread suffering. But we shouldn’t exaggerate the catastrophe either: probably the best way to assess mortality rates during the Great Leap Forward is to look at population numbers and reconcile them with birth rate data (which dropped from 37 per thousand in 1959 to just 21 per thousand in 1960…

…But let’s be clear though: the Great Leap Forward was a largely man-made economic catastrophe stemming from disastrous policies that backfired spectacularly. Mao didn’t intend to cause a famine, but his policies – including unrealistic production quotas and the diversion of agricultural labor to backyard steel furnaces – undoubtedly did. He himself acknowledged some responsibility for the disaster, as did the Party officially, with Liu Shaoqi (then Chairman of the PRC) stating at the Seven Thousand Cadres Conference in 1962 that the famine was attributed to “thirty percent natural disasters, seventy percent man-made problems.”…

…Overall, China’s GDP nearly doubled over the entire 10-year period of the Cultural Revolution and the 1969-1975 period at 6.86% annual growth was the fastest sustained growth period during the Mao years, even exceeding the celebrated First Five-Year Plan period (6.53% average annual growth). This really goes against the widespread perception that the Cultural Revolution was an economic disaster comparable to the Great Leap Forward: not only it wasn’t, but China’s economy was actually booming during the period!…

…This is what resolves an oft-discussed paradox (discussed, for instance, by Branko Milanovic here): How could a “thoroughly inefficient system” create the basis for explosive subsequent growth? The answer is that the Mao era, despite its inefficiencies and disasters, created specific tangible foundations – human capital, physical infrastructure, industrial capacity, organizational systems, and transformed property relations – that made the reform era’s success possible.

You couldn’t have had the TVE explosion without the organizational legacy of communes. You couldn’t have absorbed foreign technology without an educated workforce. You couldn’t have rapidly expanded manufacturing without existing industrial infrastructure and millions of workers with basic industrial skills. You couldn’t have sustained 10% growth rates for three decades without the healthcare improvements that gave China a healthy, productive workforce. 

---

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 no vested interest in any company mentioned. Holdings are subject to change at any time. 

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 14 June 2026:

1. Gas Prices, Stock Bubbles, Grad Advice — And Teaching Personal Finance In School (Transcript here) – Morgan Housel 

I want to start with what is the biggest economic news story of this year: the war in Iran. For most of you listening or watching, the biggest impact that’s had on your life is the rise in gas prices and oil prices. I want to make a very nuanced point here about making predictions about the future, which is so common in economics, and so difficult and humbling.

When the war in Iran first started about three months ago, it was very common among the smartest, most astute, most educated economists, oil analysts, and talking heads to make predictions along these lines: if the Strait of Hormuz is closed for another week or two, you’re not going to see a rise in oil prices — you’re going to see an explosion of oil prices. Not $100 a barrel, but $150, $200, $250. Not $4 gas, but $7, $8, $9 gas, with flights being cancelled. Those predictions have been made for months, and it was always along the lines of “if it stays closed for another week or two, this is going to happen.”

I want to make this point without minimizing what’s happened to gas prices all over the world and what could happen in the future. I don’t want to say, “Look at all these people — they were wrong,” because the price of oil today is about where it was three months ago when the war first started. It surged and then plateaued at this level. That is something almost no one watching this three months ago would have predicted. Virtually everybody, if you had told them we would be three months into this war with the Strait of Hormuz closed, would have said this is going to be a Mad Max scenario in oil. And so far, as I record this, it has not been.

I want to make an important point here without making any predictions about what might happen next — almost the opposite point, about why these kinds of things happen. There is such a long history in economics, in politics, and in any kind of social world that makes predicting what’s going to happen next so hard, even when it seems like the most rational conclusion. It’s so appealing and so easy to make simple predictions: if X happens, Y will be the result. Very appealing, and I think very comforting, because when you make a prediction like that, it gives you — or the person listening to that forecast — a sense of control in a world that is uncertain, if not unpredictable.

I say this with the glory of hindsight and nothing else; I would not have known any of this three months ago. But from my understanding, a lot of why oil has not yet reached those Mad Max levels, despite being three months into the Strait of Hormuz closure, comes down to a few reasons. Number one, the United States is exporting oil and gas like never before, which has taken some of the supply-crunch pressure off. Number two, Saudi Arabia has a series of oil pipelines that have been massively extended and expanded over the last three months — one big pipeline going to the Red Sea has gone from 2 million barrels a day to 7 million barrels a day, taking a lot of pressure off oil that used to go through the Strait of Hormuz. Number three, China has massively decreased the level of its oil imports. And number four, all over the world, we’ve been draining down oil stocks and reserves. Can that last forever? Of course not. I’m not making any predictions about what’s going to happen next.

The point I want to make — in a much broader way that applies to so many more things in the world of money and economics than just the Strait of Hormuz and oil prices — is that it is extremely difficult to know how people are going to adapt and evolve to a change in the economy. It’s very easy to say “if X, then Y,” and it makes a lot of sense and it’s very comforting. It’s much more difficult to say, “If the Strait of Hormuz closes, then people are going to adapt in this way, and this way, and this other way, and therefore we don’t really know what the end result is going to be.” There are so many cases like this, whether it’s housing prices, stock prices, whatever it might be.

I’ll give you one example that was crazy at the time. 2009 was one of the worst years in economic history since the Great Depression — absolutely dreadful, during the financial crisis. Stocks finished up that year. They increased. It’s so easy to say that the second coming of the Great Depression would be bad for the stock market; that’s a very easy prediction to make. It was much more difficult to see how people, prices, and valuations would adapt and respond in that era.

I was thinking about this recently because gas prices in my town went up tremendously in the last three months, but they’ve been about the same for the last two and a half. They exploded at the beginning of the war and then plateaued. Looking into how the global oil market has adapted and evolved — and again, maybe that doesn’t last forever, I’m not making any prediction — it’s so important to have a sense of humility about how complex the global economy is, and about people’s ability to adapt in ways you never saw coming. That’s what makes predictions about what’s going to happen this year, next year, or over the next month so difficult.

The last thing I’ll say about the psychology of making predictions: it is very common that the higher the stakes, the more people are willing to believe forecasts. When the stakes are really high — gas prices could explode so much that you can’t afford your commute, or your flights are cancelled — people are willing to believe anybody who says, “I can tell you what’s going to happen next.” That becomes very appealing. The irony is that when the stakes are that high and things are moving that quickly, that’s when forecasts become the least reliable. The demand for forecasts increases exactly when the forecasts themselves become least reliable, because people are adapting and changing so quickly. That is why there is such a long history of economic forecasts for things that never happened.

2. Avoiding Death on the Yellow Brick Road – Joe Schmidt IV

The Yellow Brick Road is our shorthand for the path the labs are walking, where they’re committing extraordinary resources. The reason the labs are best-suited for problems like code generation, writing, or image-creation is because these problems improve with raw model capability: every dollar spent on pre-training and post-training improves product quality. Meanwhile, the rest of Oz is inhabited by more complex, often vertical problems, that aren’t as simple as giving a business user a horizontal tool with access to standard tools and computer use. The value comes less from the underlying model’s raw capability (though that’s still important!) than from the scaffolding around it that makes the output trustworthy, compliant, and operational inside a specific industry…

…The labs will certainly improve, but I’d argue there are a few ways the rest of Oz can defend themselves over time:

Data and learning flywheels: A lot of what you internalize isn’t in any training set — unwritten industry norms, undocumented standards, the tribal knowledge that lives in practitioners’ heads. None of it is on the public web. No amount of training compute substitutes for being inside the workflows where this knowledge actually lives. There are two flywheels stacked on top of each other here: an across-customer one — patterns that compound as you see more variants of the same problem — and a within-customer one — the why behind specific decisions, the unsaid exceptions, the firm’s own rules of thumb that only surface through real interaction with the system…

…A horizontal agent could in principle build the same learning infrastructure. The reason it doesn’t, beyond pure focus, is UX: capturing this kind of knowledge depends entirely on the workflow surfaces you give the user, and vertical players can shape those surfaces around exactly what their workflow needs to surface. Horizontal tools can’t. Eval sets, labeled outputs, and edge-case taxonomies can compound into a vertical-specific data flywheel which can fuel fine-tuning the next entrant can’t generate without comparable production exposure. Whether this is possible depends on data rights, the volume of production exposure accumulated, and the structure of customer contracts, but pattern recognition accrues regardless.

Managing model variability and complexity: The labs are already routing internally — different model classes for different requests, ensembles under the hood. What they can’t do is route across vendors, or evaluate a competitor’s model for a specific sub-task, or use an open-source fine-tune for the narrow piece where it’s actually best. The Rest of Oz company picks the right model for each sub-task across the entire model market, not just what its parent lab ships. It also does the work nobody wants to do — re-running evals on upgrades, recalibrating prompts for the customer’s edge cases, rolling out without breaking production — every time a new model lands. The labs aren’t doing this on the customer’s behalf; they sell you their next model and tell you to migrate…

…Cost optimization: Running every query through Opus 4.7 is the fastest path to negative gross margins. The best Rest of Oz companies route across tiers of models — frontier models for the hardest tasks, mid-tier for the bulk, smaller custom or fine-tuned models where they’ve earned the right to use them. Some are now post-training their own models on top of that, optimizing them for the narrow slice of work their customer cares about and serving them at a fraction of the cost of a frontier API call…

…Governance: There is considerable value in becoming the control plane for how their customers run AI in that vertical – the place where permissions, auditing, what-the-agent-is-allowed-to-do, and what-the-agent-actually-did all converge. That control plane is built out of use case specific guardrails that look completely different across industries and job types. Because they own the tools, the workflows, and the data the agent touches end-to-end, they can provide deterministic outcomes in ways horizontal tools will struggle to. They are also the entity that absorbs the regulatory complexity for the end buyer — FRCP and bar rules in legal, HIPAA in healthcare, SEC and FINRA in finance, state insurance regulations, and so on. A horizontal player can’t credibly do that without becoming a hundred different verticals at once. CIOs want to have a partner that contractually states they are handling compliance for the agents they are providing.

All of these come back to the same thing: focus. That could be a vertical (insurance, legal, accounting) or a function done deeply (sales, customer support, finance). Either way, the work needs a team that’s heads-down on one customer set — its workflows, its edge cases, its regulations. The labs aren’t built for that. They have to be everywhere, for everyone, which is how they built the Yellow Brick Road in the first place. 

3. Sergey Brin: Where Frontier AI Is Headed | Unscripted Q&A @ AGI House × Google DeepMind (Transcript here) – Rocky Yu and Sergey Brin

Sergey Brin: That’s a great question—what’s next after we hit AGI? Everybody is pretty focused on accelerating the growth in AI right now. You’re right: we started with the web and internet search, went through the mobile generation, which was another big explosion, and now AI is a huge new industry trend. What comes after that? I think if you can answer that, you’ll have a fantastic company on your hands…

…Audience Member: I have two questions. First, now that we talk about superintelligence, and AI can help us drive cars and do office work—what kind of thing do you think only humans can do after superintelligence? Second, 20 years ago Google was famous for connecting people, and now it’s a company focused on AI. So my question is about strategy: what do you think Google’s role will be over the next 20 years?

Sergey Brin: Small questions, I guess—what is humanity’s role in this world, and what is Google going to do for the next 20 years? The definition of intelligence has always shifted with what machines can do versus what people can do. For a long time, chess was the measure of intelligence, and then Deep Blue beat Kasparov in the 1990s. The interesting thing is that people kept playing chess. How many people here know who the top-ranked human chess player is? Anyone can yell the name—I’m assuming it’s Magnus Carlsen, and people bounce up and down. But how many know the top-ranked AI program?

Audience Member: Stockfish?

Sergey Brin: That’s the most popular—is it number one? You don’t think AlphaZero can beat Stockfish? Okay, well, you’re the only one who named the top chess program; let’s point that out. My point is that computers doing things well hasn’t stopped humans from getting better and better at them, getting more recognition, and enjoying them. We’ve adjusted our view over time—it used to be that chess was the intelligent thing, then Go was the intelligent thing, then poetry or painting. I think we’re going to find that AIs can do a whole lot of surprising things, but they also help advance people in doing those things. Since AlphaGo, the game of Go has advanced a lot—the players who played against Lee Sedol became vastly better afterward, and Ke Jie did too after he played AlphaGo. It pushed the state of the art. So people will be able to enjoy and do a lot of things even with AI assistance. As for the 20-year question—I don’t know. I think we should let somebody else ask. That’s a big one.

Audience Member: Do you believe transformers are sufficient for AGI?

Sergey Brin: Great question. I’ve asked myself that a bunch of times. Transformers have been weirdly flexible—we use them for image and video in addition to text, and they’ve exceeded their original capability. To be fair, they’ve also changed along the way: we have sparse transformers and a lot of little details that have shifted, so it’s not exactly the same thing as the transformer paper. If I had to guess whether something close to that could be AGI, I’d say yes—just because they’ve been able to evolve so much. But they are changing; it’s not the exact same thing as the original transformer paper…

…Audience Member (Boris): What’s your perspective on how world models can help reach AGI?

Sergey Brin: World models are basically video models. People talk about AGI pretty broadly. I think of AGI as the idea that the AI can actually improve itself. Other people—and they’re probably more correct—think AGI means the AI can do anything a person can do. Those are two different things. To do anything a person can do, you absolutely need to understand and interact with the physical world. So being able to dream or imagine what’s going to happen in the world if you do something, and to comprehend it, is obviously important. If you’re going to do everything—and that extends to robotics—world models are key. You all have probably had more time to play with our Gemini Omni model than I have, honestly, because I’m deep into the self-improvement game. But we’ve been working on that for a long time, and Omni is the latest version. Omni is also pretty cool because it’s the same Gemini—we train it with all the text and all the other things, exactly the same way. The fact that these converge is amazing. But yes, you need that capability for the ability to interact physically.

4. Blackstone Investors Ask to Pull $4.4 Billion From Private-Credit Fund – Matt Wirz

Investors in Blackstone’s flagship private-credit fund, known as Bcred, asked to redeem 10% of their shares in the second quarter, up from about 8% in the first quarter. That amounted to investors asking for $4.4 billion.

Blackstone will limit redemptions from the $79 billion fund to 5%, a reversal from its strategy in March when it opted to pay the full amount requested. The about-face highlights rising financial strain on managers of large private-credit funds marketed to individual investors who continue to ask for their money back…

…“BCRED remains well capitalized, and repayments [from loans] and inflows have outpaced shares repurchased,” the firm said Thursday. It said the fund’s structure, allowing it to limit redemptions, is a core feature that is meant to trade some liquidity for long-term performance…

…Wealthy individuals piled into private-credit funds—known as business-development companies, or BDCs—which invest in high-interest loans to midsize companies and distribute most of the income they collect to shareholders via dividends. The boom ended this year when investors turned bearish over increasing loan defaults and the potential for future losses from lending to software companies.

The Blackstone fund is the largest of the bunch, surging to a high of $82 billion at the end of 2025, but it is now shrinking, cutting into the fees the firm can collect. 

5. The AI Price War Is Here, Piling Pressure on OpenAI and Anthropic – Bradley Olson and Tina Li

Big companies and startups, chafing at rapidly escalating artificial intelligence costs, are increasingly turning to tools that tap in to cheaper AI models, including some from China. That’s raising pressure on industry leaders OpenAI and Anthropic to lower their prices, a prospect that could hurt their ability to grow into profitable enterprises…

…The ecosystem allows autonomous AI systems, or agents, to use cheap models—including those made by Chinese companies like Alibaba and DeepSeek—for many functions. The agents only tap the most capable versions of OpenAI’s ChatGPT and Anthropic’s Claude for more complex tasks. That can reduce costs for some AI-assisted work by as much as 95%, according to executives using the tools.

“Once we find something that is working well and engineers love, we find ways to make it cost effective,” said Dan Robinson, founder of Detail, a startup that identifies bugs. “There’s really an embarrassment of riches right now coming out of the open source labs.”

Robinson shifted 90% of Detail’s workload from Claude and Google’s Gemini to custom models and GLM, a family of models developed in China…

…OpenAI is considering drastic cuts to the prices it charges AI users, ahead of similar cuts the company expects at Anthropic, The Wall Street Journal reported. The company sees itself as having an advantage in such a scenario because it spent massive sums in the past year to secure access to computing resources at far lower prices than what’s available now…

…Open-source Chinese models have been rising in popularity across American businesses. DeepSeek’s share of AI usage rose from 1% in April to 17% in May on the startup Vercel’s platform, the company said.

On OpenRouter, another startup that processes AI queries, DeepSeek has been the most-used AI company since mid-May. Among their highest-spending customers, open-source token usage grew four times faster than closed-source between fall 2025 and spring 2026, OpenRouter said. The company has also seen more than 500 organizations swap from proprietary to open-source models…

…Anthropic’s recently-released Fable 5 model is more than 50 times more expensive per token than DeepSeek’s V4 Pro, for example.

But the top proprietary models from companies like OpenAI, Anthropic or Google remain four to six months ahead of open-source competitors, researchers say. In some cases that means they can complete a complex task using fewer tokens, equating to a lower total cost…

…Many companies have begun to design their own AI models using open-source alternatives and say they are managing to reduce AI costs. When companies build in-house models and train them with company data, their performance can improve or even exceed the capabilities of frontier AI models, executives say.

---

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). Holdings are subject to change at any time. 

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 07 June 2026:

1. Most of the Economy Won’t Run on the Best Model – Rihard Jarc

When a company hires an accountant, it does not go out and hire a PhD in pure mathematics to reconcile the ledgers. Not because the PhD couldn’t do it — they obviously could, and probably faster — but because it makes no economic sense. The PhD is overqualified, which is just another way of saying they are too expensive for the value the task produces. The economic output of bookkeeping is capped. There is only so much upside in getting the books done. So you hire the cheapest person who clears the quality bar, and you pocket the difference…

…If you are running a drug-discovery program, you absolutely want the PhD — in fact you want five of them, plus a Nobel laureate consulting on the side. Why? Because the economic output of a single discovery is enormous, almost unbounded…

…This is, I think, exactly how the AI model market is going to bifurcate…

…Today, essentially everyone uses the state-of-the-art (SOTA) model for everything. You want to summarize an email? SOTA model. Classify a support ticket? SOTA model. Extract three fields from an invoice? SOTA model. We do this for one simple reason: the frontier models have only just crossed the threshold of being broadly truly impactful for knowledge work, and when something has only just started working, you reach for the best version of it you can find. You don’t optimize cost on a capability you weren’t sure you had last quarter.

But I believe this is a transitional behavior, not a stable equilibrium…

…We have a rapidly falling price for any given level of capability and frontier that is already shrinking in size in terms of what is actually being deployed, and we have companies burning through their annual token budgets in a matter of months.

As such, I believe that for the overwhelming majority of economically valuable knowledge work, the correct model is not the SOTA model. It’s the cheapest model that clears the task’s quality bar. And as pilots move into full production (which is the stage we are in today) — where you’re suddenly paying for millions or billions of tokens a day instead of running a demo — intelligence-per-dollar becomes the only metric that survives contact with a CFO…

…The sellers of new compute (semis) are only winners in a world of continued high-cadence spending on new compute. And my thesis specifically questions whether that cadence is necessary. So let me lay out the two states the world can be in, because the asymmetry between them is the whole argument.

Scenario 1: Capex falls or stabilizes. If you can squeeze an order of magnitude more useful tokens out of the hardware you already own — because models got smaller, cheaper, more efficient and verticalized — then you no longer need to spend $100bn+ every single year just to stay relevant. In this world, the owners of the installed base win and the sellers of new compute lose. Hyperscaler free cash flow inflects sharply upward, because capex was the one thing suppressing it. Multiples re-rate higher as the cloud business converts from a capex incinerator into a cash machine running largely paid-for, partly-depreciated hardware. And the semis de-rate, because the market finally realizes the upgrade treadmill has slowed.

Scenario 2: Capex stays high — and revenue explodes. This is the Jevons-paradox-on-steroids case. Demand is so strong that hyperscalers do both: they extract enormous output from cheap, long-lived existing hardware and keep buying new gear. Here everyone wins at once — but the hyperscalers win more, because their incremental revenue now lands on a cost base that is partly depreciated and dramatically more efficient per token. Operating leverage goes vertical.

2. Calling the Top – Dirtcheapstocks

Spacex is set to go public next month.

I read the S1 and felt like I was watching “Whose Line is it Anyway?”. You know, the show where everything’s made up and the points don’t matter…

…Spacex is eyeing a ~$1.8 trillion valuation, from the latest reports I’ve seen. SPCX did $18.7B of revenue and generated a net loss of $4.9B in 2025. Free cash flow was severely negative: -$15.8B (adjusting for stock-based comp).

So the business is valued at ~100x revenue, and revenue has been growing at a ~34% CAGR over the last two years. Q1 2026 revenue grew 15% yoy.

The business has never sustained profitability, as evidenced by a $41B accumulated deficit…

…If you pay $1.8T for a business, and want a 10% return, you need it to send you $180B in year one. If it sends you $0 in year 1, you need it to produce $198B in year 2, and every year after that until the end of days. If year 2 also produces $0, you need year 3, and every year beyond that, to produce $218B.

I don’t think it’s likely that SPCX will reach profitability in the next couple years…

…According to ChatGPT, General Motors (in the 1950’s) was the largest company in American history when measured on GAAP revenue as a percent of GDP.

This isn’t a perfect metric, but I think it helps us get a rough feel for how large a company can become as compared to the ecosystem in which it exists.

GM’s revenue was equal to ~2.3% of American GDP. This shouldn’t be surprising as GM had ~50% market share in the second most expensive asset Americans owned…

…Now let’s take this metric and apply it to SPCX.

U.S. GDP is ~$32T today. Historically speaking, it would be difficult for a single business to earn more than $750B in annual revenue.

But SPCX will conquer the world (and Mars), so let’s assume it shatters the record. Maybe SPCX revenue can be 3% of GDP, beating out every business in history by 30%!

That would imply SPCX revenue of $960B. So what kind of profit margin can we expect for this business…

…But let’s say SPCX is a killer business at scale and it can achieve 20% operating margins, and 15% net margins. And let’s say it takes us 10 years to work our way there.

So, at a $1.8T valuation, we need $180B of cash in our pocket this year to generate a 10% return.

If we are unable to earn an cumulative profit above $0 for the next 10 years, then year 11 (and every year after that) needs to pay us $466B!

Alright, so we need $466B of profit in year 11. At 15% net margins, that means we need $3.1T of revenue.

If nominal GDP compounds at 7% for a decade, then GDP will have grown to ~$64T. So, SPCX in year 11, will need to have grown its revenue to ~4.8% of GDP ($3.1T / $64T) – a percentage more than double any company in history.

To get to $3.1T of revenue in 10 years, SPCX will need to grow its top line at 67% annually. The past couple years have shown revenue growth in the 30’s…

Hmm, this is getting difficult.

3. X thread on the difference between HBF (High Bandwidth Flash) and HBM (High Bandwidth Memory) – Eugene Ng

HBF is essentially HBM but with NAND flash dies instead of DRAM. It uses similar 3D stacking and TSV technology, delivering 8-16x higher capacity than HBM in a comparable footprint, while offering similar bandwidth, much lower cost per GB, lower power, and acceptable latency for read-heavy AI inference workloads (e.g., massive model weights, long context windows, and large KV caches)…

…HBF Shines in Inference: AI inference (LLM serving) is dominated by read-heavy, capacity-bound tasks, loading huge models, managing long contexts, and high-throughput batching. HBF excels better than HBM…

…Limitations: HBF has significantly higher latency (~10 µs, roughly 100x slower than HBM), slower write performance, and limited endurance (~100k write/erase cycles), making it unsuitable for frequent updates during training…

…Training vs. Inference Shift: As inference grows faster than training in overall AI compute, hybrid HBM + HBF setups are superior to HBM alone. HBM dominates training, while HBF’s capacity and cost advantages break the “memory wall” for cheaper, higher-throughput inference at scale…

…Bottom line: HBF expands the total AI memory TAM without cannibalising HBM. It creates a new high-value inference tier, making the overall market more competitive, multi-layered, and resilient, which is great for innovation and supply diversity.

4. Project Glasswing: what Mythos showed us – Grant Bourzikas

Mythos Preview is a real step forward, and it’s worth saying that plainly before getting into anything else. We’ve been running models against our code for a while now, and the jump from what was possible with previous general-purpose frontier models to what Mythos Preview does today is not just a refinement of what came before.

It’s a different kind of tool doing a different kind of work, and that makes a clean apples-to-apples comparison to earlier models difficult. So rather than trying to benchmark Mythos Preview against general-purpose frontier models, it’s more useful to describe what it can actually do, and two features that stood out across the work we did with Mythos Preview:

• Exploit chain construction – A real attack rarely uses one bug. It chains several small attack primitives together into a working exploit. For instance, it might turn a use-after-free bug into an arbitrary read and write primitive, hijack the control flow, and use return-oriented programming (ROP) chains to take full control over a system. Mythos Preview can take several of these primitives and reason about how to combine them into a working proof. The reasoning it shows along the way looks like the work of a senior researcher rather than the output of an automated scanner.
• Proof generation – Finding a bug and proving it’s exploitable are two different things, and Mythos Preview can do both. It writes code that would trigger the suspected bug, compiles that code in a scratch environment, and runs it. If the program does what the model expected, that’s the proof. If it doesn’t, the model reads the failure, adjusts its hypothesis, and tries again. The loop matters as much as the bugs it finds, because a suspected flaw without a working proof is speculation, and Mythos Preview closes that gap on its own.

Some of what we describe above is not entirely unique to Mythos Preview. When we ran other frontier models through the same harness, they found a fair number of the same underlying bugs, and in some cases they got further than we expected on the reasoning side too. Where they fell short was at the point of stitching the pieces together. A model would identify an interesting bug, write a thoughtful description of why it mattered, and then stop, leaving the actual chain unfinished and the question of exploitability open.

The Mythos Preview model provided by Anthropic, as part of Project Glasswing, did not have the additional safeguards that are present in generally available models (like Opus 4.7 or GPT-5.5).

Despite this, the model organically pushes back on certain requests – much like the cyber capabilities that made it useful for vulnerability hunting, the model has its own emergent guardrails that sometimes cause it to push back on legitimate security research requests. But as we found, these organic refusals aren’t consistent – the same task, framed differently or presented in a different context, could produce completely different outcomes…

…When we first started AI-assisted vulnerability research last year, our instinct was the obvious one: point a generic coding agent at an arbitrary repository and ask it to discover vulnerabilities. This approach works, in the sense that the model will produce findings, but it doesn’t work in producing meaningful coverage of a real codebase and identifying findings of value…

…Four lessons came out of running the work at scale, and each one pointed to the need for a harness that manages the overall execution:

• Narrow scope produces better findings – Telling the model “Find vulnerabilities in this repository” makes it wander. Telling it “Look for command injection in this specific function, with this trust boundary above it, here’s the architecture document and here’s prior coverage of this area” makes it do something much closer to what a researcher would actually do.
• Adversarial review reduces noise – Adding a second agent between the initial finding and the queue – one with a different prompt, a different model, and no ability to generate its own findings – catches a lot of the noise that the first agent would miss if it just checked its own work. It turns out that putting two agents in deliberate disagreement is way more effective than just telling one agent to be careful.
• Splitting the chain across agents produces better reasoning – Asking “Is this code buggy?” and “Can an attacker actually reach this bug from outside the system?” are two different questions, and the model is better at each one when you ask them separately, because each question is narrower than the combined version.
• Parallel narrow tasks beat one exhaustive agent – Coverage improves when many agents work on tightly scoped questions and we deduplicate the results afterward, rather than asking one agent to be exhaustive.

Each of those observations is about model behavior, and put together they describe something that isn’t a chat interface anymore. It’s a harness that helps you achieve the final outcomes.

5. Open-source agents with frontier advisors: matching frontier performance through training and harness engineering – Fireworks AI

On LAB’s continuous mean-score metric, GLM 5.1 ranks highest among the open-source models we evaluated, at 0.8921 mean score putting it directly alongside frontier: Claude Opus 4.7 at 0.911, GPT-5.5 at 0.892. Kimi K2.6 (0.863) and DeepSeek V4 Pro (0.871) come in just below, both still clearly viable for production legal workloads.

On the LAB all-pass metric, the production-readiness measure, the closed frontier holds a small lead: Opus 4.7 at 14 / 100, GPT-5.5 at 11 / 100, GLM 5.1 at 12 / 100. That gap is where the rest of this post lives; the two interventions we describe below close most of it.

Cost is the headline. GLM 5.1 reaches its 0.8921 mean for $121 across the 100-task run. GPT-5.5’s nearly identical 0.892 costs $560. Claude Opus 4.7’s 0.911 mean and 14 / 100 all-pass runs $954, roughly 8× any open-source candidate.

“The customer ask is no longer ‘how do we get the smartest model on every query.’ It is ‘how do we get frontier-quality outputs on the queries that need them, and a model we control on the queries that don’t.’”…

…A single LLM call is the wrong unit of work for a legal task: reasoning chains run long, citation discipline is unforgiving, and under all-pass grading any missed criterion costs the entire task. To solve the problem, the team built a small, opinionated multi-agent harness with the open-source worker at its core. The configuration is straightforward: open weights at the core, orchestration the team can inspect and tune, and the frontier model invoked as a callable tool rather than a load-bearing dependency.

A frontier advisor as a callable tool. Treating Opus 4.7 as an advisor the worker can call on hard sub-tasks unlocked the cost savings on the harness. The GLM 5.1 worker does the bulk of the reasoning, drafting, and tool calls. There is no external router or orchestrator. The worker pulls the advisor in itself, wherever it needs a second opinion: retrieval, drafting, validation. Across the run, the advisor is invoked just 0.83 times per task on average — sparse-but-targeted use. That captures most of the quality lift of running the frontier end-to-end, at a small fraction of per-query cost, and it gives us a tunable cost/performance knob: dial advisor calls up on complex matters, down on routine ones.

The harness traces show a recognizable pattern. The worker’s turn count rises meaningfully versus a GLM 5.1-only run: the model reaches an uncertain step (typically during validation, occasionally mid-draft), calls the advisor for guidance or review, then resumes the trajectory with additional turns informed by the response. The advisor is doing less of the writing and more of the steering; the worker is doing the rest of the work it would not have known to do on its own. Sparse advisor calls, denser worker activity downstream of them.

The harness moves GLM 5.1 from 12 / 100 all-pass to 18 / 100 — higher than Claude Opus 4.7’s 14 / 100 — at $368 across the 100 tasks, roughly 39% of Opus’s $954 standalone cost (Figure 1). Against Opus the comparison is clean on both axes: −$586, +4 tasks all-pass. Against the GLM-only baseline, the advisor adds +6 tasks all-pass for +$246 — the cost increase is real, but it is the cost of beating Opus while still running the open-source worker at the core.

---

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 no vested interest in any company mentioned. Holdings are subject to change at any time.