What We’re Reading (Week Ending 22 November 2026)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

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 22 November 2026:

1. Obituaries – Ben Thompson

The first part of the AI buildout was funded by free cash flow, particularly from the hyperscalers. Then, during the second half of last year, nearly everyone started turning to debt. Now — starting with Google this past summer — the primary funding mechanism is equity. In the long run, this should come full circle, and future build-outs will once again be funded by free cash flow, generated from the infrastructure that is being built now.

The problem for Nvidia is that the companies most able to navigate this progression are the hyperscalers, which have the resources to challenge Nvidia’s hold on the chip market. To that end, Nvidia has been very eager to ensure that it is selling not only to the hyperscalers, but to other would-be competitors, like the neoclouds; there will likely, at some point, be serious questions raised about exactly how Nvidia went about allocating compute in pursuit of this goal.

The neoclouds, however, are at a meaningful disadvantage when it comes to funding. First, their debt is significantly more expensive, given that their only business is renting GPUs (as opposed to, say, running the most valuable advertising businesses in the world, or the world’s largest e-commerce platform, or dominating enterprise software). Second, their equity is increasingly worthless, given the significant risks attached to that credit. If a neocloud can’t pay, it’s the equity holders that get wiped out first.

This led to what I termed Nvidia’s Risky Business: Nvidia increasingly used its cash flow and credit rating to backstop and fund the companies who were buying its chips. Bears called this circular financing; bulls called it a necessary bridge to a future where AI revenue paid down these companies’ debt and drove massive upside. What is undoubtedly true is that this was the sort of risky business that is uniquely possible for a multi-trillion dollar company that is completely controlled by one man, who has made an entire career of pushing the limits and making risky bets — and seeing them pay off.

It’s also a series of bets that are unlikely to continue. The neoclouds should be very nervous about losing their patron, along with their creditors and equity holders. The other problem is that the revenue necessary to make all of this build-out work is something on the order of a trillion dollars industry-wide, and it needs to materialize sooner rather than later; if it falls short, it’s the hyperscalers who will prioritize their own compute first, and they will likely be seen as a safe haven for everyone committing to spending. In that world, Nvidia’s neocloud hedge — absent Huang’s willingness to shore the ecosystem up — is liable to turn into a millstone, with their assets flooding the market even as the supply chain ramp-up brings ever more supply to market…

…The chief objection that China hawks in particular make to the idea that it is a good idea to sell China chips in order to prevent them from surpassing the U.S. is that China is already heavily invested in building their own ecosystem. This is certainly true, but I think it underplays the complexity of chips and what is necessary to actually compete with the West.

Specifically, the Chinese government can only control the supply side, by spending money and making top-down imperatives; building an ecosystem, however, whether that be for AI specifically, or semiconductors broadly, requires demand as well. Progress in models and systems and chips depends on a feedback loop between the user and the developer, or the customer and the manufacturer. The demand side, however, is much more short-term oriented than the Chinese government might prefer. The model companies’ concern is the other model companies: that means if they are distilling, we must distill too. If they are using Nvidia chips, then we must too. This matters up and down the stack, just as it has always mattered in technology, which really is different in this regard from almost every other industry.

A useful analogy is to consider the Internet. Why is it that country after country can profess a desire to have an indigenous alternative to the American tech giants, yet miserably fail in every attempt to generate it? It’s because the tech giants are only a click away! In this regard, China’s Great Firewall was one of the most brilliant geostrategic moves ever: by virtue of cutting off access to the American tech companies, China created the demand that pulled China’s own tech giants into existence. You need demand, not just desire.

This is why China itself is currently banning the import of Nvidia’s two-generations-behind H20 chips, after the Trump administration foolishly banned their import, only to reverse themselves while bragging about the reason to do so — i.e. keep China addicted to Nvidia. China is replaying the Great Firewall playbook: force its internal demand to use its much worse alternatives, because that is the only way those alternatives become better. And, in the long run, those alternatives, and the entire ecosystem that supports them, will flood the world.

Huang was, to be perfectly clear, talking his book when it came to his insistence on selling chips to China, or, in the case of the current controversy, allegedly turning a blind eye towards nominally legal buyers who were intent on violating the chip ban; that doesn’t mean, however, that he was wrong about the long-term effects, at least if AI is a normal technology (and I haven’t gotten to the risks this imposes on the Taiwan Strait). What is clear now, however, is that the doomers and the hawks have won: AI, at least in terms of policy, will be treated as anything but normal, and in the fullness of time, it is the U.S. chip ecosystem that will pay the ultimate price.

2. How Often Do Bonds Beat Stocks? – Ben Carlson

I looked at the rolling 10 year returns for the S&P 500 and 5 year Treasuries going back to 1926:

Stocks beat bonds in most 10 year periods but it’s not guaranteed. There have been times when bonds beat stocks over 10 years…

…Another way to look at this is to see how often stocks have beaten a 5% bogey, which is around the current level of government bond yields.

This number was higher than I anticipated. Stocks earned more than 5% per year in 85% of all 10 year time frames…

…But the 15% of the time when it has been below 5% wasn’t much fun to live through as an investor…

…The 10 year annual return for the S&P 500 through the end of October was more than 15% per year.

The 15 year annual return is also close to 15% annually. From the lows of the Great Financial Crisis in March 2009, the index is up nearly 17% per year. That’s almost 18 years of extraordinary gains.

This is a generational bull market. The hard part about handicapping generational bull markets is that they can last a lot longer than you think but also tend to be painful when they end.

3. An Interview With Eric Seufert About Agentic Commerce – Ben Thompson and Eric Seufert

The thing about Amazon that you have to recognize, especially in the context of agentic commerce, but even more broadly, is that Amazon is the identity layer for the Internet. It is the identity backbone for the entirety of the open Internet. That is the open Internet, it’s Amazon, it’s Amazon’s bid, it’s Amazon identity.

Really? Explain this. Why is it Amazon and not Google or Facebook, which people would assume?

ES: Because they are walled gardens. Amazon has deterministic identity on 90% of US households. And so when you see a deal that Amazon does with Roku, with Disney+, with Netflix, with Tubi, with OpenAI, with Pinterest, the reason that those other parties are willing to allow somebody to inject demand into their bid stream, into their ecosystem, is because Amazon has the identity, they know who they are. And so when Amazon can connect that person to their data set from Amazon Prime, they’re going to target ads better than almost anyone else could.

Wasn’t this formerly Google’s domain? Are you saying they’re taking over what Google Networks used to be?

ES: Google Networks still exist, but they weren’t doing the targeting — they were effectively allowing the ads to be served, the ads would be served against cookies. But what Amazon is doing is different, because Amazon says we don’t need a cookie.

Because you’re signing into the Amazon account and linking it to that other service?

ES: You might be. In some cases, you are. In some cases, you aren’t. Amazon says we don’t need cookies, we don’t need device IDs in order to do that identity, we just have all these other signals that we can place the user against…

…So continue on this point. You wrote, “If third-party agents are effectively fetching and restoring the output from retail platform agents, meaning the purpose they serve is principally discovery, and retailers can play agents against one another on commissions, what monetization model are third-party agents left with for the commodified purpose they serve in e-commerce?” I won’t answer the question, tell me what you mean, and then give me the answer to the question.

ES: Let me take one step back and just explain the context there, because it’s really important.

So what happened was, Amazon announced that they’re going to make Alexa for shopping available in text messages, so it’s basically just opening up any interface to connect to Alexa for shopping. Well, that actually was after Walmart came out with these agentic principles for agentic use on its own website.

Yes, walk us through that. I was going to get to that next. Walk us through the Walmart thing. Everyone’s like, oh, Walmart’s going to step in for Amazon, what is actually happening there?

ES: So that was how it got interpreted, but you spend three minutes reading what they wrote and you realize it’s the exact opposite. What they did was they set these very strict rules for agentic checkout so if you want to check out on Walmart.com, you have to abide by these rules. You can still do agentic use outside of this API, just browser-based, but then you can’t do checkout. If you do that, you’re blocked from checking out, a human has to intervene. But if you want to check out, you’ve got to abide by these agentic principles and what they say is, you essentially collect the request from the user, you collect the prompt from the user, and you give it to Sparky.

Sparky is Walmart’s agent, so you basically hand the request over to Sparky, and Sparky, as an agent, then manages the request, fulfills the request, and then hands you the output of Sparky.

Right. It’s like the most power-using form of telephone tag ever, let’s burn up a few GPUs to my telephone.

ES: Exactly. It’s an API into Sparky, what they’re saying is, no matter what agent you’re using, you’re using Sparky, Sparky’s behind the curtain. What does Sparky include in its output? Ads. So now you’ve got essentially this throughput that just forwards to Sparky, Sparky returns its Sparky response, and then it potentially has ads in it.

Now the other agent can do, I guess, what it wants with that. It could distort it, or it could change it, re-rank it. But then you might mistrust what your agent’s doing. And so the reality is, when these platforms are making their catalogs, or the platforms themselves, open to external agents, really what they’re doing is they’re creating an API connection to their own agent, their own agent’s ultimately going to mediate the request.

Got it. And do you think that Amazon will do something similar in the long run?

ES: Well, they kind of have, by making Alexa for shopping available in text messages. But I don’t know, we’ll see, because Amazon’s a little bit more particular. I think it actually makes more sense for Walmart to do that, because Walmart would value the incremental e-commerce transactions more than Amazon would.

Also, the sort of person using an agent is not the typical Walmart customer, so it’s like all greenfield to them in many respects.

ES: Yeah, could be true, could be all new customers too. But my point there was, now you’ve got all these agents that are essentially competing to forward requests to Sparky, and to Alexa for Shopping, and to whatever else platforms are going to integrate this way.

Now, what could they charge for? How could they charge money for that? How could they charge a subscription for that? You’re just basically an interface on top of Sparky or Alexa for Shopping and I think you’re even starting to see Shopify kind of force people to come to terms with the fact that you’re not going to be able to charge some meaningful transaction fee. [Shopify CEO] Tobi [Lütke] said recently in an interview, you’re not going to get 2%, he said 2% is unrealistic as a transaction fee, as a take rate.

Now, keep in mind, OpenAI was trying to charge 4% with Instant Checkout, so now we’re down to 2% being unrealistically high. So what do you get? He said, welcome to the world of basis points, so you’re talking, what, 50 basis points? 25? What value is there?

What does Apple pay? Like 15 basis points, something like that?

ES: So now what do you do? You’re essentially just a tool for discovering options. How do you monetize discovery? Ads.

Isn’t all this a very supply-centric analysis? At the end of the day, to go back to my example, I don’t want to use any other agent, I want to use my agent. Isn’t this sort of the Meta bet — that Muse is going to be their agent, they don’t want to use another agent, and if they get sufficient people, then people will have to accommodate them?

ES: I would go back to my response there, which is, well, you’ve got an agent that could fulfill this workflow, it’s just not your agent. There’s an agent that could fulfill it, it’s called Alexa for Shopping, and so would that be more of a draw? If you knew, before you’d ever started using Claude Code, that — I’ve got this workflow, I can’t fulfill it all in Claude Code and reach the biggest retailers, I could fulfill it all in Alexa for Shopping, and reach and get to use Amazon for purchasing. Would you have chosen Claude Code? Assuming a totally equivalent underlying sophistication and power of the LLM.

Um, no. Because Claude Code is like, it’s all my projects tied together. But I see what you’re saying. Like, Claude Code is designing the 3D files that are going to the 3D printer, and then knowing the bolts that I need to put these pieces together. It’s a fully integrated thing that’s centered around me and what I want to do, that’s where my preferences start.

ES: Fair enough. I take your point, and I agree. People have said I’m being too dismissive of what consumers want here. And consumers will define how this evolves, I’m not disputing that.

Right. And to be clear, I’m not saying I’m the typical consumer. But is there an extent to which I’m the typical consumer in ten years, or five years, or whatever timeframe you want to put on it? I guess there’s a few — and I guess in Meta’s defense, the argument would be, if they actually get critical mass, if Muse actually becomes integral to their life, then Muse does start choosing retailers based on whether they work for Muse or not and you have some meaningful share going to a Walmart, even if it’s through this media interface, and not to Amazon. Amazon’s carrying all these huge fixed costs, they need volume. At some point, they just have to give in, I think that would be the Muse bull case, and I assume you think that’s pretty unrealistic.

ES: Here’s my response. I don’t know if it’s realistic, we’ll see, but here’s my response. The degree to which these retail platforms will open up with agents is a function of how much leverage they have. Right now, I would say that Amazon has way more leverage than Muse has. I think the mistake here is collapsing agents as a category down to Muse, because that’s not true. There are other agents and there will be more and you have to assume that every big consumer tech company is going to have an agent, so let’s put that aside for a second. Who’s got more leverage? I would argue that Amazon’s got more leverage. Muse probably needs Amazon more than Amazon needs Muse.

Now, let’s say that’s not the case. Muse gets so much market share with the agent market, and the agent market becomes really big, then Amazon has to capitulate. But then what happens? Because Amazon has market power now. It doesn’t have market power on the basis of a permanent, defensible basis of its user base. It has market power because it provides better fulfillment speeds at lower costs than any other retailer can.

Now let’s say Amazon capitulates. Can it do all three of those things? Open up its catalog to third-party agents, deliver faster fulfillment at cheaper prices than any other retailer — I don’t think it can. Something’s got to give, especially then if Muse says we’re going to take a transaction fee, because if they take a transaction fee, that’s going to increase the price for the consumer in the first place. So the consumer then has to say, what is the convenience worth?

Right. Then Amazon’s going to have to charge more as well, because they’re not showing an ad, so it’s like a double whammy.

ES: Right, exactly. So you get this whole economy collapsing.

But the other piece of this is, will Muse become that one dominant agent? Would it dominate the entire market? They capture 60, 70% market share? I don’t know, because it’s already a crowded space, a lot of other companies are interested in competing here, and on what basis are they going to compete? Well, probably on cost. So they’ll bring the subscription costs down. And what happens if subscription costs are zero? They take the take rates down, what happens if take rates get to zero? Ads.

4. This Is The Problem With AI’s Most Bullish Chart (Transcript here) – Joe Weisenthal, Tracy Alloway, Stanley Lee, and Paul Kedrosky

Stanley Lee: Thanks for having me on. It’s definitely a topical time to be talking about all of this. Liquid Compute is a startup focused on building financial infrastructure in the compute space. We run a marketplace for people to buy and sell capacity, which can obviously be used to relight capacity if there eventually are defaults. We maintain a rental index that tracks GPU rental prices, which we use with a lot of the lenders we work with on financing, potential hedges, and ways to lay off residual value risk. Nobody really felt the need to do that a few months ago, and everybody is starting to feel the need now. And we work on financing for a lot of these buyers and sellers.

One of the biggest things that hit me was that I didn’t see the GPU marketplace as just a marketplace of buyers and sellers. It truly is a credit and financing ecosystem. If you look at buyers and sellers of compute at the below-investment-grade level, the buyers are potential creditworthy off-takers for high-yield and leveraged finance lenders to underwrite, and the operators are potential borrowers. They’re either borrowers now for this cluster or, if they do a really good job, they’ll be borrowers for their next cluster…

…Tracy Alloway: Do people in the AI industry understand the financing aspect of all of this? It feels like 2026 is the year they’re having to wake up to it. We talk a lot about the hyperscalers, big cash-rich companies that never had to borrow much in the market before, and now we’re seeing them take on loans and bonds and issue new equity. Are people waking up to the reality of the financing aspect?

Stanley Lee: I think absolutely. At the top of the market, we as a country and a society have created this dual-track economy where AI capex sustained so much of GDP growth, which you’ve talked about in the past. And it’s trickling down to the middle market. A lot of it has been funded by venture equity, and private credit has been financing a lot of these mid-market GPU loans. A lot of people in tech and in Silicon Valley did not care that much about credit. Credit is very boring. It’s negatively convex. It was not something people building innovative solutions had to care about. They’re definitely focused on it right now…

…Joe Weisenthal: Going back to the B100 chart or the H100 chart, or any of these charts: how much are they affected by the phenomenon you’re describing, where the people buying compute regularly, maybe on the spot market, can’t lock in long-term agreements? What’s the significance of that for what looks like sustained demand?

Stanley Lee: I think they’re drastically affected. Most of these charts, if they’re worth anything, will exclude hyperscaler pricing. You’re looking at a lot of neocloud pricing and where transactions are actually happening. And if the chart is going up and to the right because credit is getting harder, that is not a healthy or sustainable environment. If there’s a breaking point and people start to run out of money, that’s where the floor can drop out. The spot market is pretty thin. A good question I got from one of the large private credit firms was: how many bids do you think really need to get hit for the floor to fall out in these graphs? The answer is probably a lot fewer than people think.

Tracy Alloway: This is what I wanted to ask. Talk to us about the standard covenants in some of these deals. Are there trip points where things become problematic and the borrower has to start renegotiating the loan or returning collateral to the lenders?

Stanley Lee: In the middle market, it’s a lot less about covenants. The covenants that matter most are your lien on the hardware, the ability to access it, and a tripartite agreement with the data center so you’re able to relight the contract. But if there’s a default, there’s a default.

Tracy Alloway: And the default might come from something like an LTV ratio?

Stanley Lee: We’re already seeing LTVs come down. We were seeing 80% LTV loans get priced over the summer. Post-Labor Day, it was more like 70%. Now people are asking for 50% to 60%. The other thing I’d say, and this is where a lot of the more punitive covenants are coming in, is that a lot of these mid-market neoclouds are coming to credit for emergency rescue and bridge financing. If you’re a lender, you want to be protected. First of all, I think these are extremely interesting investments. But they have extremely tight covenants, and if there is no takeout in three to six months, and that takeout relies on future capital from private credit anyway, lenders will foreclose and seize the equipment.

5. Why Everything Is Suddenly Breaking All at Once – Derek Thompson, Nicole Perlroth, and Jennifer Pahlka

Nicole Perlroth: We have a great digital deficit in this country. The digital deficit is that we have been rolling out flawed code, leaving that flawed code in systems, and we haven’t been updating it for years. We’ve also been baking code into every nook and cranny of our economy, critical infrastructure, our homes, our cars, our bodies in some cases. No one ever stopped to think that we were just adding and adding to this digital deficit that would one day come due.

With AI, it is now coming due in a big way. AI models have tremendous capabilities to find vulnerabilities in code and exploit it to break into systems on the other side. So all of these errors that we rolled out in this code and never addressed are now being found by these powerful models and stitched together in such a way as to attack systems.

We are finding three times the number of vulnerabilities using AI this year than we were in the same period last year. Attackers are exploiting those vulnerabilities using AI at a much quicker rate than we’ve ever seen. By the time a company learns it has a flaw in its code and rolls out a patch, it turns out that, on average, attackers have already been exploiting that flaw for seven days.

What this means is that while we used to be able to live on this Patch Tuesday cycle, where companies would roll out fixes for vulnerabilities in their code and customers would have time to update their systems, now we’re on the attacker’s schedule. So companies and government agencies are finding it very urgent to install these patches and software updates. It’s really easy to find a vulnerability now and exploit it. It’s still really hard and time consuming to properly install a patch. And when you install a patch, you do it in these bespoke environments. No environment is the same. You can’t just push a big red button and everything’s fixed. Usually something will break on the other end.

What we have seen over this past week of outage after outage is companies and government agencies installing these patches and software updates to fix vulnerabilities that were discovered with AI, and they are breaking things on the other end. So these are not cyberattacks, but the impact is the same.

Thompson: You made two important points that I want to make sure I understand. Advanced AI has offensive capabilities to discover cyber vulnerabilities and attack. But what we’ve seen over the last few days is not hackers breaking into systems. Rather, sophisticated AI tools are finding vulnerabilities, and the patches themselves are coming in so fast that they’re causing their own breakages. Is that right?

Perlroth: Two things are true at the same time. AI is absolutely increasing the number of incidents and the velocity of these incidents. And it’s much harder to triage these incidents because when you’re hit with an agent, it doesn’t follow the same workflow as a human attacker.

What happened with these outages that are almost simultaneous, the FAA, the bank disruptions, United Airlines, the NSA net outage, is that those are companies or organizations that are trying to get ahead of this huge tsunami of AI-powered attacks by installing software updates as quickly as they come. And those software updates have not gone through the proper quality control and testing that they might have in an earlier pre-AI era. And they are breaking things on the other end.

Thompson: Are there actually more outages now than in previous weeks or months, or are we paying closer attention because we’re sensitive to the imminence of a cyber apocalypse? Does anyone keep a baseline for how often critical systems fail?

Perlroth: I saw some people use AI to conduct a timeline analysis of whether we’ve had these outage clusters before. And we absolutely have. It’s not the first time we’ve had simultaneous outages or near-simultaneous outages. So I do think you’re right that a lot of this is baked into the attention that is on AI right now. I don’t know that we’ve had outages in these very public systems, critical infrastructure, that would capture our attention in the same way. For years I have been warning about two critical infrastructure systems in particular that kept me up at night.

People ask me: What keeps you up at night? And I’ve always highlighted two potential scenarios. One, an attack on our water treatment facilities that would contaminate the water or lead to a water shutoff. And two, an attack on air traffic control.

Thompson: What makes them so vulnerable?

Perlroth: A few years ago, there was a hack, and we still don’t even know whether it was a hack, on a water treatment facility in a town in Florida just outside the city hosting the Super Bowl that year. The way the story was initially relayed to us, an engineer was watching his screen and this ghost mouse moved across his screen, moved over to the chemical controls, and increased the amount of lye in the water and could have potentially poisoned everyone downstream.

When we started digging into this, it was pretty terrifying to get a sense of the state of cybersecurity around our water treatment facilities. There is unfortunately a very big story of the haves and have-nots in cybersecurity. Companies like JPMorgan, Goldman Sachs, the banks, and certain critical infrastructure providers have the budgets to be able to buy every bell and whistle in cybersecurity. But the little guys don’t.

Water treatment facilities are notoriously underfunded, understaffed, and often have to rely on outside contractors who remote in to manage things like the chemical levels in the water. And that naturally creates an opening for hackers. Often the people who set these systems up were not thinking about the fact that one day an Iranian hacker might want to contaminate the water supply in the United States. So often they do it with the poorest cyber hygiene—like using “password” for a password, or being without two-factor authentication.

We actually can’t have a system of haves and have-nots in cybersecurity. We’re an ecosystem. And if the most critical part of that ecosystem breaks, it’s going to break the big guys too. Water is the one that is probably the most susceptible to an attack, even though it’s the most critical.


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 Amazon, Apple, Meta Platforms, and Shopify. Holdings are subject to change at any time. 

Leave a Reply

Your email address will not be published. Required fields are marked *