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 25 October 2026:
1. The Numbers Behind the Bond Bear Market (with a Look at Equities) – Ray Dalio
In one sentence, countries “go broke” when:
1. The supply of debt to be sold is much larger than the demand for it and/or2. Debt service payments (i.e., interest and principal payments) rise relative to incomes to levels that painfully squeeze out other spending.
Both of these things are now happening and are projected to happen increasingly to the extent shown in this report.
Classically, these circumstances a) raise interest rates, tighten credit, hurt equity and other markets, and reduce borrowing and economic activity and/or b) lead governments to accommodate more money and credit creation to temporarily alleviate the debt, which devalues money and pushes the problem down the road. Most typically, debt and debt service payments are reduced relative to incomes through a combination of debt restructuring and debt monetization, and as a result, money and debt decrease in value. In any case, this dynamic is bad for the holders of debt assets…
…The punchline is that it appears very likely that there is and will be much too much supply of financial assets (both debt and equity) relative to the demand for it…
…At the end of September 2026, there were about $32 trillion of marketable Treasury securities outstanding, 8% more than a year earlier. Through September, Treasury already issued approximately $1.5 trillion of net new marketable debt and is expected to issue $2.1 trillion of net new marketable debt over the full year of 2026. That is a huge amount of debt to sell at a time when the government’s debt and debt service burdens are extremely high relative to its income…
…The amount of short-term debt that has to be sold has been rising in an unusual way for two reasons: the size of the Treasury debt is rising at a fast pace and the share of debt that is short-term is rising at a fast pace because the demand for Treasury’s long-term debt has been falling relative to the supply. In an attempt to deal with this worsening supply/demand balance for bonds, Treasury has been shortening the maturities of the debt it has been selling…
…Foreigners as a whole are the largest holders of the stock of debt, with 32% of the debt, and the Fed is the next largest holder at 14%. Together, they own nearly half while Americans of all sorts own a bit more than half. You can see by comparing the percentage of holdings with the recent percentages of their purchases in 2025-26 that the support for the Treasury market has been shifting away from foreigners to domestic buyers…
… Notably, the composition of demand is changing because the demand for longer-duration debt is clearly weakening. As explained above, foreigners and the Fed together hold nearly half the debt. That’s not a healthy mix for two reasons: 1) such a large amount of debt bought and held by foreigners raises risks, and 2) the Fed’s large role in the Treasury market to bring about balances indicates that the bond market is not in a healthy free market/balanced state…
…It is officially estimated that foreign investors currently own about $9.3 trillion of Treasury securities, making them the single largest broad category of Treasury debt holders. It is also notable that their rate of buying Treasuries is slowing while the supply of Treasuries is increasing…
…In other words, China and Japan (which are two of the largest beneficial holders of US Treasuries) are the most important holders to watch because their incentives have shifted from accumulating to reducing their holdings of that debt…
…While most discussion about the US Treasury supply focuses on the deficit and therefore on new debt issuance, which of course is a huge amount, the quantity of maturing debt that has to be rolled over is much larger. Because the numbers are huge, this is a big potential negative that is not well understood, so I believe it is worthy of a quick illustration.
While debt rollovers typically don’t cause a problem because owners normally reinvest most of their maturity proceeds, that can’t be assumed, especially nowadays. Rollovers can have a big impact on the supply/demand picture. The table below shows the connection between the reinvestment rate and the amount of demand shortfall that could result from reduced reinvestment in billions of dollars. From a supply/demand perspective, these reduced rollover rates are the equivalent of increases in the budget deficit…
…Debt service payments = principal payments + interest payments.
As we looked at principal payments before, let’s look at the interest payments, which, very simply, can be estimated by multiplying the change in the interest rate by the amount of debt. With approximately $32 trillion of marketable Treasury debt, if the average interest rate paid on the entire stock eventually rose by one percentage point (as bond yields have done over the last seven months), the ultimate annual interest expense would rise by roughly $310 billion per year, which over 10 years would add about $3 trillion of supply and debt to the US government balance sheet.[6] (This doesn’t happen immediately because existing fixed-rate notes and bonds continue paying their old coupons until they mature.)
The typical debt squeeze dynamic that leads to overindebted countries “going broke” occurs when more debt leads to higher yields, which leads to higher interest expense, which leads to larger deficits, which leads to more debt, etc. When debt levels and interest rates go up together, it is a double debt service whammy. The increased debt service costs lead to further decreases of demand for the debt; credit market players call that a “death spiral.”…
…In summary, it appears to me that we are in the part of the cycle that follows relatively big interest rate increases but before the previously mentioned adverse impacts on markets and interest-rate-sensitive parts of the economy occur. At this stage, interest rates tend to continue to increase until the supply/demand imbalance starts to shrink toward sustainable levels, which shows up in the weakness in interest-rate-sensitive markets and economic sectors. As the supply/demand picture isn’t improving and neither the markets nor the economy appear to be weakening to improve the imbalance, the bond bear market probably has a long way to go, and we should be alerted to the other markets’ possible weakness. In this part of the cycle, policy makers typically tighten monetary policy in order to try to tap the brakes without popping the bubble. That is now happening in a pretty classic way. If the past is prologue, higher interest rates will 1) produce increasingly severe capital losses for bondholders that will continue until they produce bad equity markets and a bad economy, 2) raise financing costs that will ripple through the economy, and 3) lead other borrowers and sellers of competing financial assets (like equities) to compete harder for money, so money will become tight, reducing the demand for goods, services, and financial assets…
…There is a lot more to look at in order to get a complete picture of the supply and demand for equities, which I will not do here because this note has already gone on too long. But suffice it to say that the supply of financial assets (both debt and equity) being offered appears much larger than can be absorbed by the demand for it.
2. The Market Is Repeating the 1999 Pattern. Is the Lost Decade Beginning? – Outlier Capital
The pattern is repeating. Anyone who says it isn’t has not looked at the chart.
A general-purpose technology. A capex boom large enough to move GDP. A market priced for a decade of profits that have not all arrived. A chorus, getting louder this week, that the next stop is a crash and then a lost decade…
…The Shiller CAPE has been above 40 since May, the second time in a century and a half. The first time was the dot-com top.
Oxford Economics put AI-related capex at 27% of US fixed investment in the second quarter, with the three-year build already above 2% of GDP. Among technology booms of the last two hundred years, only the British railway mania of the 1840s was larger…
…The bond market is not cooperating with the story. The 10-year Treasury is at 5.3% today, up from roughly 4.1% a year ago, the highest in this cycle. The 30-year is near 5.7%. The Fed hiked in September, to a 3.75–4% funds range, and the median dot implies another move before year-end. The 2-year is back above 4.7% for the first time since the summer of 2024…
…This is the setup people mean when they forward the three-column chart. Electrification. The internet. Artificial intelligence. A vertical narrative, a multiple that cannot be defended, and a decade in which the buyers of the top go nowhere…
…In the second quarter the S&P 500 grew earnings 52% year over year, the fastest in five years. The Magnificent Seven grew 118%. The other 493 grew 32%, the best print for that group since late 2021. Strip out Alphabet and Amazon, the two names that distort the picture most, and the index still grew earnings above 30%.
That is not a one-story market hoping the story arrives. That is a market in which the story is already in the income statement, and not only in seven stocks…
…The guidance is the tell. Of 116 companies that have issued a number for the third quarter, 72 have guided up and 44 have guided down. That is the highest count of positive preannouncements since FactSet started tracking the series in 2006. In 1999 the multiple was covering for earnings that were missing. Here the earnings are being revised up into the print, and the multiple is the thing giving ground.
The forward price-to-earnings on the S&P has fallen from about 22x to about 19x , in line with the 10-year average and under the 5-year average, while the index sits roughly 1% from a record. The S&P closed October 2 at 7,723. Price barely moved over the summer. The earnings estimate did. That is the opposite of how a bubble finishes. A bubble finishes with the price running and the earnings apologizing.
The CAPE above 40 is real, and anyone waving it at you is not inventing it. They are using a 10-year trailing, inflation-adjusted multiple to answer a question about next year’s operating earnings. A high multiple on earnings that never arrive is a lost decade. A high trailing multiple on a market whose forward multiple is already back at average, because the earnings showed up, is a market that can earn the valuation from behind. Expensive is not the same as hollow…
…What Would Make This 1999
Few things. None of them is a 5.3% yield on its own.
Capex stops converting into revenue, and the $300 billion gap becomes a write-down rather than a lag. The other 493 stop growing and the entire earnings number collapses back into 7 names. The next leg of the index is the multiple, not the profit. The delivery queue turns into cancellations. The labs’ negative cash flow migrates from the private round into the listed buyers’ income statement and stays there.
3. Brain-Computer Interfaces: The Next Compute-Medical Frontier Enters the Production Era – Abhishek Jain
Neuralink’s N1 implant is a coin-sized device that replaces a small section of skull and interfaces with the brain via 64 flexible threads carrying 1,024 electrodes. The threads are inserted by a surgical robot called R1, which places them into the motor cortex at depths of a few millimeters. The device records the electrical signals from individual neurons, transmits them wirelessly to an external computer, and returns machine-learning-decoded output as digital commands.
Two things distinguish this approach from every other BCI in clinical use. First, the channel count. At 1,024 electrodes, Neuralink’s N1 records from roughly sixty times as many sites as Synchron’s Stentrode and slightly outnumbers Precision Neuroscience’s current Layer 7 array. Higher channel counts allow richer decoding of intended movement, which is decisive at the frontier of what the technology can currently do.
Second, the surgical robot. Human hands cannot reliably place hair-thin threads to sub-millimeter accuracy without damaging vasculature. Neuralink built R1 specifically to remove that limitation. In December 2025, Musk announced the company will move to an almost entirely automated surgical procedure in 2026, including a dura-piercing thread design that eliminates the need to open the outer brain membrane. If the automation lands as described, procedure time falls dramatically and the surgeon-supply bottleneck weakens…
As of June 2026, Neuralink has implanted 26 patients across four countries, spanning five clinical programs: PRIME (motor control for quadriplegia), CONVOY (robotic arm control), VOICE (speech restoration), Blindsight (vision restoration), and a newly announced hearing initiative. Trial sites are active in the United States, United Kingdom, Canada, and the UAE.
Zero serious adverse device events have been reported across the entire clinical footprint. That is a meaningful figure…
…Neuralink’s public roadmap is a straightforward multiplication of the current architecture. The N1 delivers 1,024 electrodes today. Musk stated on January 28, 2026 that a next-generation implant with roughly three times current capability is expected by late 2026, targeting the range of 3,000 channels. This claim is not independently verified and is not tied to a specific regulatory filing. The longer-horizon targets are 10,000 and eventually 25,000 channels per implant…
…The specific bottlenecks are known. Electrode-thread yield at scale is a serious manufacturing problem, as each thread must be defect-free across its entire length. R1 surgical robot fleet size limits the number of procedures the company can perform per week. Trained surgeon supply constrains the geographic footprint. ISO 13485 quality-system compliance, service logistics for explant and replacement, and long-term device support are all operationally material and rarely emphasized in company communications.
For context, the closest medical-device precedent is deep brain stimulation for Parkinson’s disease, where Medtronic and Boston Scientific have spent decades building the manufacturing, surgical training, and reimbursement infrastructure that permits routine commercial deployment. Neuralink is attempting to compress a comparable curve into two or three years, which is either an operational achievement of the first order or an aggressive timeline that will slip.
4. Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Supercluster – Kevin Lee, Yashar Bayani, Ahmed Sharif, Yee Jiun Song
Right now, in New Albany, Ohio, Meta is undertaking its most ambitious AI infrastructure project to date. Prometheus, our gigawatt-scale supercluster, is an advancement in our infrastructure that is training the next generation of foundational models that will bring new AI experiences to billions of people…
…Large training jobs require everything to work optimally to sustain their performance; an issue on a single host can slow down or stop the entire job, lowering its ‘goodput’. By focusing on improving our systems’ stability and reliability we were able to achieve more than 85% goodput for pre-training runs and more than 90% goodput for reinforcement learning (RL) runs of Muse Wonder. During these training runs, we were able to achieve an overall wall-clock mean time between interruptions (MTBI) of ~47k GPU hours. Since then, we have landed further changes across our fleet including firmware fixes to increase MTBI to more than 54K GPU hours, and we are constantly improving it…
…Today, Prometheus comprises an array of hundreds of thousands of GPUs across multiple interconnected buildings, including traditional and AI-optimized data center buildings and rapid deployment structures, all tightly interconnected by miles of fiber and functioning as a single machine.
This capacity allows us to pursue larger models and longer context windows than were previously practical. At full scale, Prometheus will comprise roughly 1 million distributed GPUs, all functioning in lockstep…
…Prometheus currently deploys two NVIDIA GPU platforms, with a third (based on NVIDIA’s Vera Rubin NVL72) arriving soon. Each represents a generational leap in performance and efficiency, leveraging our NVL72 rack architecture where 72 GPUs act as a single accelerator within the same network fabric…
…In traditional web services, the data center network transports requests and responses between services whose instances typically don’t require high bandwidth, high performance, and tight, lock-step coordination across many servers. Large-scale AI training requires all these, meaning the network that connects them is now an important part of AI workloads. To best address this, in addition to the standard data center network connecting servers, several years ago we added a second dedicated data network based on Ethernet, called the Back-End (BE) network, to enable direct GPU-to-GPU communication both within and across data center buildings.
Large pre-training jobs partition the training into data-parallel model replicas, creating a hierarchy in network usage: Latency-sensitive and bandwidth-sensitive operations concentrate within a replica, while traffic that is more tolerant of higher latency and less bandwidth is scheduled between replicas. Our network is designed to match this hierarchy, with low-latency, high-bandwidth fabrics within buildings and an aggregation network connecting the building fabrics…
…At the building level, each Ethernet-based fabric interconnects up to 20,736 GPUs through a high-bandwidth, low-latency multi-layer topology. Our initial design used a Disaggregated Scheduled Fabric (DSF) to mitigate the network congestion caused by training collectives that typically have a small number of large, long-lived network flows that collide under traditional load-balancing techniques; it does this by utilizing sub-packet-level multipath routing and packet re-ordering in the fabric along with virtual output queues and credit-based end-to-end flow control…
…The Backend Aggregation (BAG) network is a multi-layer hub-and-spoke network that connects building-level fabrics across the Prometheus cluster, spanning 2 km to over 100 km. It is built with Arista 7800-series chassis running deep-buffered Broadcom Jericho3 silicon, delivering multiple Pbps of bandwidth per region pair.
Running pretraining over this network required a dedicated optimization effort with Meta Superintelligence Labs to ensure training was not slowed down, especially given speed-of-light constraints at these distances.
Pulling this all together, the complexity of the Prometheus network involves millions of miles of fiber strand, tens of thousands of switches, and millions of optics — from 2×400G FR4 LITE short-reach, low-cost pluggables inside the buildings to 800G-ZR coherent optics for links beyond ~10 km…
…The transition from the H100 generation’s NVL8 architecture to NVL72 has fundamentally redefined our approach to infrastructure resilience. In the NVL8 era, the scale-up domain — the group of GPUs tightly coupled via NVLink for tensor parallelism — was limited to eight units. With NVL72, that domain expands to 72 GPUs, increasing the fault blast radius by 9X.
The larger scale-up domain is what gives NVL72 its bandwidth advantage, but it also makes the 72-GPU rack the unit of allocation we plan capacity around. Training parallelism maps most efficiently onto power-of-two group sizes, so each NVL72 domain runs a 64-GPU slice and keeps its remaining eight GPUs as in-rack spares. When a GPU fails, a spare takes its place immediately, so the job restarts from its last checkpoint instead of waiting on hardware repair, and the rack stays in the job. The in-rack redundancy pushes the effective rack fault rate down to the floor set by failures that can remove an entire pod; for rarer whole-rack faults we hold roughly 2% additional racks in reserve. For a 32,768 active-GPU training run, this means provisioning 523 NVL72 racks, or 37,656 GPUs, about 13% above the active footprint. Most of that margin (11%) falls out of the 64-of-72 slice geometry rather than being an absolute reliability tax.
Physical complexity further intensifies these reliability requirements. A single NVL72 domain incorporates 5,184 tuned differential-pair signaling links operating at 200 Gbps, each traversing multiple potential points of failure, including GPU SerDes, compute tray PCBs, cable cartridges, and switch tray interfaces. Maintaining consistent signal integrity across this dense fabric — at 200 Gbps per lane — represents an unprecedented engineering challenge at this scale…
…Deploying hundreds of thousands of first-generation accelerators reveals quality issues that are hidden at lower volumes. The GB200 deployment has transformed hardware maturation into a primary engineering discipline for our teams and partners.
5. Markets Are Always Different This Time – Ben Carlson
A friend was cleaning out one of his rental properties and came across the business section in an old newspaper from 1960:
The chart shows the make-up of the S&P 500 by sector at the time:
- 425 industrials
- 25 rail companies
- 50 utilities
Things look just a tad different today.
This is what the S&P 500 company breakdown looks like now:
- Industrials 83
- Financials 76
- Information Technology 75
- Health Care 60
- Consumer Discretionary 47
- Consumer Staples 34
- Utilities 31
- Real Estate 30
- Materials 24
- Energy 21
- Communication Services 22
Yes the stock market is more concentrated in the top names now, but there is far more diversification in the types of companies that make-up the index.
This is also one of the reasons it’s so difficult to compare historical valuations to today’s markets. The companies and sector make-up of the index are completely different…
…Maybe stock market returns will be lower going forward on a gross basis than they were in the past. But on a net, after cost basis, investors have never had it better than today.
Reg FD is probably the most important change to the market structure this century that no one ever talks about. Once regulators required any significant corporate information be made available to all investors at once, it made it much harder for hedge fund managers and other large active investors to win using proprietary information.
There’s a reason there haven’t been nearly as many legendary investors beating the markets lately. The playing field has been leveled.
It’s also true that the low fruit has been picked in many ways because market knowledge and data are so much more ubiquitous today.
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 Meta Platforms. Holdings are subject to change at any time.