We are back with yet another post about CPRT. Below is a selected list of earlier articles we wrote:

1. History & business economics

2. More business economics

3. Details on auctioned vehicles

4. Q3 FY2026 update

Link #1/2 are probably the most important as it discusses the fundamental drivers of the business. But we realized that most of the content is focused on the supply side, not very much has been written about the buyer demand side.

The choice to focus on supply is not misguided since CPRT is indeed very dependent on auto insurers sending totaled vehicles.

Although CPRT also acts as a principal by buying totaled cars outright and selling them, this segment is not high ROIC since they have to carry the inventory on their balance sheet.

In contrast, the consignment model has far better returns, where CPRT acts as an agent for insurers who sends those damaged cars.

About 80% of units that come to CPRT are from insurers. The remaining 20% comes from non-insurance:

1. Blue Car. This segment supplies damaged cars from banks, rental car companies, and fleets. Rental car suppliers are particularly interesting, because every day that a rental car is off the road translates to revenue lost. So a rental company is time sensitive when thinking about whether to repair or auction a damaged car. The more time passes, the more likely they will auction it. Furthermore, rental cars are unlikely to fetch a good rate if they are even slightly damaged.
   
   Rental companies also have shorter turnover time than an insurer. As this segment continues to become larger, it can show up in metrics like “cycle time” which fell -9% in Q1 2026.

2. Dealers. Copart Dealer Services (CDS) and NAPA (National Powersports Auctions) for dealers that repossessed vehicles for whatever reason.

3. Copart Direct. This is the principal model mentioned above.

Therefore, it’s obvious why we spend much time describing the economics of CPRT’s supply side.

For a more complete understanding, let’s explore the buyer demand side as there are also very interesting details that we should appreciate.

Types of Buyers

The demography of buyers are fragmented since the top 10 biggest buyers make up only a low single-digit percentage of all vehicles CPRT sell in the US. Buyers can be licensed vehicles dismantlers, rebuilders, repair licensees, used vehicle dealers, and exporters.

The general public is actually only a small fraction of participation. To buy salvaged vehicles in the US, the individual has to go through a registered broker. These brokers hold licenses that allow individuals to bid under their credentials. The process is costly because these brokers charge their own fees too.

Therefore, the auction game is played by companies rather than individuals.

CPRT sells membership tiers for access to their auction platform. There’s 3 tiers: Guest, Basic, Premier. There are some countries like the UK where they only offer Guest and Basic tiers, suggesting that the market is still not matured.

The difference between tiers mainly is their fee structures and quality of access.

1. Guest is free of charge. You can browse and receive updates, but cannot bid.

2. Basic cost $99/year. You can bid on one vehicle at a time with limited buying power that requires a deposit. You also get free yard previews during designated business hours.

3. Premier cost $249/year. You have bid limit of up to $100k per day, can bid on multiple vehicles and have priority customer support.

One of the more important buyer groups are exporters. These are companies that buy cheap salvaged cars from the US and sell them for more in other countries. They are important because they indirectly bring buyers from all over the world, contributing to the liquidity that CPRT’s two-sided auction platform requires to function well.

Exporters exist because the price of a new car is expensive relative to purchasing power in countries like Latin America, Africa, or Eastern Europe. Driving a damaged vehicle makes economic sense for some people living there. Not only is it cheaper, but they also have slacker legal requirements.

For example, cars with faulty airbags cannot be legally driven in the US, but they are fine in Mexico. Or the car has a damaged electronic dashboard, that’s no problem for someone in Bulgaria.

In volume terms, 39% of vehicles sold in US auctions go to an international buyer. In price terms, that number is ~50%!

This is a symbiotic relationship; CPRT and exporters benefit from each other’s existence with each party extracting surplus from the value chain.

So CPRT buyer base is really very diversified. Even though the income statement says that 85% of revenues are from US region, but the underlying demand is actually coming from overseas.

Because the pool of international buyers is so large, they collectively bid up prices, driving up the average selling price (ASP). This implies that US insurers who supply the cars get a higher return — they could not have achieved this without the existence of CPRT, and so the relationship is very dear.

Buyer Fees

You can find the details on fees structure here.

There are 2 main variables that affect the final fee:

1. Non-clean or Clean title.
   
   (a) Non-clean title is another term for “salvage vehicle”, meaning that it was declared a total-loss by an insurer at some point (even if it’s repaired and roadworthy). CPRT obtains the title on the insurer’s behalf — this is a complicated process as the rules differ by state, but is also a value-added service that CPRT offers.
   
   (b) Clean title means the vehicle has never been declared a total-loss.
   

2. Standard or Preferred pricing. To qualify for Preferred, a buyer must have bought 25 vehicles of at least $75,000 over 12 months, across at most 5 accounts (this criteria covers roughly all of its professional clients). Those who don’t meet these criteria fall under Standard.

The rest are fixed: gate fees $95, and environmental fees $15.

Put together, for pro clients their fee schedule vs. bid price looks like this (assume secured payment, exclude title shipping and auxiliary charges):

The fee schedule is regressive in nature, thus we get convexity in the fee revenue curve:

As a result, the gradient at each bid price is different. This has some nuanced implications on unit economics. We cannot assume a linear relationship between ASP movement and fee revenues.

An extreme example would be if CPRT sells a hundred $250 wrecks, it collects $21k in fees. But if CPRT sell a price equivalent $25,000 wreck, it only collects $2,125. We should keep in mind that higher number of low-value wrecks require more towing and yard space which doesn’t immediately mean high operating profits.

In other words, CPRT unit economics is driven more by mix than average prices.

Since most of the car bids are above $15k where fees % are capped, if used car prices rise, all else equal, CPRT will earn less incremental revenues per unit relative to ASP (absolute dollars will of course increase).

This was what happened in the recent quarters where revenues per unit increased less than ASP. The instinctive reaction would be to say CPRT has lost pricing power to competitor IAA.

Pause for a moment… what happens if ASP fell instead?

Yes, their revenues per unit will decrease less than ASP:

On the surface it can look like CPRT is structurally designed to be less sensitive to changes in ASP — this is only true if number of units sold increased.

Falling ASP will trigger insurers to deem more vehicles as total-loss. A lower sensitivity to falling ASP is an illusion of resilience if CPRT cannot sell more units in such a scenario.

We know this is complicated. Analysts have been asking questions about mix and pricing for the longest time ever, and management always point to the number of units sold as the primary driver of growth.

To hit on this topic further, we know that PGR stopped sending units to CPRT, but that might not be a bad thing.

Despite PGR being the largest auto insurer, we suspect CPRT is less interested in PGR's salvage because of the lower ASP it insures compared to its competitors. In fact, CPRT intentionally rejected price negotiations with PGR.

This low-ASP high-volume strategy doesn’t fit with CPRT, whose strategy for the past few years has been focused on high value salvage to lower earnings volatility, since the high end of the fee curve is flatter. Note: This is guesswork from the loss of PGR coinciding with ASP at all-time high.

At this point you should start to notice the information asymmetry that exist between buyers and suppliers.

• Insurers (suppliers) have to negotiate long-term contracts with CPRT and these terms are trade secrets. Insurers cannot go directly to buyers because it is not their core business model and they don’t have the infrastructure to do so.

• Buyers have no negotiation power with CPRT, they are price-takers.

CPRT sits between them with information advantages about both parties, hence exploiting this asymmetry through pricing power. However, CPRT is also constrained because insurers are the lifeblood of the business, there are only a few large auto insurers so keeping them happy is a priority.

Competitor IAA

We know CPRT and IAA operate in a duopoly, but are they competing on price? What is IAA fee schedule?

No surprises, both exercise rational pricing otherwise the duopoly wouldn’t have lasted. IAA fees are almost the same, they actually have slightly higher fixed service fees.

If they are not fighting over prices, then where is the battleground?

Service quality — especially on the supply side.

Insurers want to get rid of salvage cars fast (at a good price) to recoup some of the claims losses. So, there are 2 important battlegrounds:

1. Fast turnover + high bid prices. This means efficient towing, title handling, deep auction liquidity.

2. Response to catastrophic disasters (hurricanes). Obviously, these events contribute a lot of salvage units. Whoever can response effectively wins the battle.

Theoretically, insurers are not stuck with IAA or CPRT, but in practice switching carries logistical cost. The value proposition for winning is to provide insurers the best salvage price and service.

Cyclical Problems

For the ending notes, we think it's worthwhile to remind ourselves of the recent problems that CPRT is facing. Afterall, the stock wasn't cut in half without reasons.

We start from 2023 when auto insurers realized that inflation was causing inadequate reserves. So they adjusted prices upwards for a huge +20%.

Underwriting margins were restored in 2024, but consumers felt the pinch and started to drop renewals and deductibles. The entire salvage pool shrunk.

In 2025, PGR gained market share from State Farm and became the largest insurer. At the same time, they used this as leverage to negotiate for better terms which CPRT refused. So PGR shifted most of their salvage to IAA.

Unit volumes fell for a few straight quarters, and the market feared that IAA is going to beat CPRT.

These are cyclical issues.

Clearly, insurers hiked prices too much (typical over/undershoot estimation) causing the salvage vehicle pool to shrink across the board, and CPRT was disproportionally affected because it disagreed with lower supplier fees with PGR.

Currently, PGR is no longer growing inforce policies as quickly. In fact, it has slowed down significantly:

PGR must slow down because it's impossible to grow fast and yet be profitable in a commodity-like auto insurance industry.

We think that the market has once again exaggerated the issues. We should watch if the cycle to turns soon and whether CPRT unit volumes increase.

The first ATM was installed in 1969 by the Chemical Bank, predecessor of JPM.

Naturally, there were very stark predictions that hundreds of thousands of bank tellers were going to be redundant. But that didn't actually happen initially for the next 40 years, look at these two disjointed charts below.

1. The number of tellers employed increased after ATMs were adopted:

2. The number of bank tellers only fell rapidly after 2010:

This seems counter intuitive, why did tellers jobs suddenly decline in 2010?

Just so you know, it wasn't due to the Great Financial Crisis (GFC) as number of tellers were about same from 2008 to 2010.

In fact, this story about ATMs is often misused by politicians.

In 2011, Barack Obama claimed that ATMs had decreased the number of tellers, in order to suggest that the elevated unemployment rate during his presidency was due to productivity gains from technology.

To explain why his claims were incorrect, we should explore the history and find out what caused the disappearance of bank tellers.

Background of ATMs

Before the 1970s, the operating model of banks required physical branches. Labour was a big cost item simply because they needed bank managers, loan officers, and bank tellers.

The relationship of the customer with the bank literally ran through a bank teller. Whether it was withdrawing and depositing money, checking balances, or opening accounts… everyone had to talk to a teller.

Since this was a high volume activity, banks had to hire a large number of tellers.

These tellers required only a high school education, some on-the-job training, and they were good to go. Post WW2, America was enjoying prosperity and rising wages, although individual tellers were not costly, but these staffing requirements don't benefit from scale. As a bank expands with more branches, they must employ more tellers too.

It was during this period that the word "automation" was coined in 1946 by Delmar S. Harder, an engineering manager at the Ford Motor Company.

As wages rose, businesses sought ways to optimize cost, thus an economic movement sparked off the concept of “automation”.

One of those ideas was to get the consumer to self-serve, reducing the need for service staff, giving rise to supermarkets, laundromats, vending machines, self-service gas stations, and fast food restaurants like McDonald’s.

Banks also latched on this trend even though they were a relatively labour-light business. It was actually particularly apparent in Europe that some innovation was needed because labour unrest among bank employees was an ongoing headache. Financial sector employees were actually some of the most militant white-collar workers, just look at the strikes (1966—76) which closed down all the clearing banks in Ireland.

The much hoped for technology was around the corner.

The first was IBM’s invention of the magnetic stripe card in 1960. This was a thin strip of magnetized tape bonded to a plastic card that stored data. A machine could read the data when swiped through a card reader.

The second was the mini-computer in 1965, which dramatically reduced the price from over $100,000 to $20,000.

Tech companies saw the possibility of automating the teller's job as these two ingredients would form the recipe for the first ATM.

IBM being the largest tech firm of its time grabbed this opportunity.

All Day Banking

In 1972, Lloyds Bank collaborated with IBM to introduce a self-service kiosk called IBM 2984 that dispensed cash. It was advertised as a bank that never closed.

The IBM 2984 accepted bank credit cards and required customers to enter a personal identification number, a relatively new feature at the time. That information, along with account data encoded in the magnetic stripe, was transmitted to a bank computer that checked the validity of the transaction.

Lloyds Bank named its cash-dispensing terminal Cashpoint, and the machines became so popular that the word “cashpoint” became a generic term in England to refer to any ATM.

In 1973, IBM introduced its second-generation IBM 3614. It enabled truly automated banking; customers could check their balances and make deposits and payments in addition to cash withdrawals. Linked to a bank’s central mainframe computer by telephone lines, IBM 3614 found itself in many supermarkets and department stores.

This enabled banks to extend their reach beyond branches without having to invest in expensive infrastructure. Department stores also benefited from having higher traffic and customers with ready cash to spend.

However, IBM was more interested in mainframe computers where they dominated the industry. By the 1970s, IBM reported 20 consecutive years of profits!

The company decided that the ATM market was too small to pursue and ceded the nascent ATM industry to a company called Diebold, which in turn became the world’s leading manufacturer of ATMs, but racked up lots of debt in the late 1990s and went bust in 2023.

Citi Never Sleeps

Despite delivering convenience to customers, ATMs were not widely used. In 1977, Citibank boldly invested $50m to install ATMs all over New York City.

As expected, people avoided the new technology, preferring to speak to human tellers.

Then came an unlikely break during the historic January 1978 New York City blizzard, which blanketed the city in 20 inches of snow.

Physical bank branches were forced to close for days and ATM usage surged. Citibank quickly capitalized on the accidental success and came up with the marketing slogan: Citi Never Sleeps

Remarkably, after nearly 50 years, Citibank is still using this branding message today.

Fun fact: “The city never sleeps” is also a famous nickname for NYC first used in a 1912 newspaper article.

Fate of Bank Tellers

If you were a bank teller at that time, it would’ve been natural to worry about losing your job. The cost advantages of the ATM were obvious. They also worked without mistake, around the clock without breaks, and were much faster than humans.

Banks loved it. They were low maintenance coupled with opportunities to charge fees for out-of-network transactions. ATMs were also not legally considered as branches, so banks could scatter their ATMs all over the country without breaching interstate regulations.

In 1975, there were only 31 ATMs per 1 million Americans. By 2000, that number had grown to 1,135. However, tellers per capita did not experience a drop off at least for another 20 years:

Bank tellers did become a smaller share of total employment, since the increase in tellers was smaller than in other roles. But at no point in the period 1970 — 2010 did the number of bank tellers actually enter a prolonged decline.

So what happened?

As the routine cash-handling tasks of bank tellers receded, technology also enabled a broader range of bank personnel to become involved in “relationship banking”. Banks recognized the value of tellers enabled by technology and converted them to sales, forging relationships with customers and introducing them to additional bank services like credit cards, loans, and investment products.

This is called Jevons paradox where technological improvements that increase the efficiency of a resource’s use lead to a rise, rather than a fall, in total consumption.

In other words, the ATMs was a complementary technology, but not a competing one.

What Made Tellers Obsolete?

The bank tellers parable went quite far. It was commonly cited as the “myth of technological unemployment”. Even Eric Schmidt retold the ATM story in 2017 as one reason why he doesn’t believe in technological job loss.

But everything would change in 2010.

Enter the iPhone.

On a superficial level, this piece of tech has nothing to do with banking. But think deeper and you would realize that the iPhone was a paradigm shift. It changed how people conducted their daily affairs, now through a smartphone.

Banks knew by the early 2010s that the future was in mobile banking.

In this new world, people will do banking through the internet and their phones. Taken to an extreme, it could be argued that physical bank locations were unnecessary. In fact, digital-first financial platforms like Revolut and Klarna emerged without a single walk-in retail branch; everything was done in apps.

Visits to bank branches rapidly fell and banks redesigned the banking experience around the digital interface.

As branches disappeared, so did the teller.

ATM had been an innovation within the existing world of physical banking, and thus its replacement of the bank teller could inevitably only be partial. But the iPhone completely changed the way of banking, and there was no real need for the bank teller.

As bank tellers were eliminated, entire new sectors of jobs were created:

• Software developers

• Customer service help-desk

• UX designers

Lessons

When technology substitutes away tasks, it doesn’t really make an employee obsolete. The real force of technology is the ability to create new paradigms.

If you drop a new technology into a human-shaped workflow, it can indeed increase productivity, but it cannot completely replace humans because the workflow was created with frictions that require human involvement.

The ATM substituted tasks, while the iPhone made them irrelevant!

This train of thought can be extended in AI where people are worried that it will replace their jobs. Yes, AI systems can be dropped into human workflows and learn how to do them autonomously. The expectation is that white-collar workers will be displaced.

We are skeptical that simply implementing new technology into human-defined workflows will give the results people expect. Usually real disruption happens when people figure out how to organize work around technology, rather than slotting into what already exists.

The world is still in the phase of “slotting” AI into human-defined workflows. We don’t expect much real job displacement as long as we are stuck in this regime.

A new paradigm can happen eventually, but we don’t know how these new structures will look like.

Some general comments related to investing… In times of change and uncertainty, investors tend lean on the history of past technologies for comfort.

This is a mistake.

The ATM story tells a comforting narrative — it is indeed quite difficult to be disrupted. We want to believe that new paradigms won’t hurt our businesses, until suddenly they do.

We have written about Tencent here, where we touched briefly on the history of WeChat and its founder Allen Zhang Xiao Long (张小龙). It is not a stretch to put Allen on the same stage as the legendary Steve Jobs. At least within China’s tech world, he is renowned as the creator of WeChat in 2011, the super-app that boasts 810 million users in China today.

There is much to learn from Allen about product design, particularly his approach to a product-driven culture rather than a design-driven one. For the last decade, WeChat has been at the forefront of social innovation. For example:

• Red Packets — sending money as a means of communication.

• Mini Programs — services within WeChat that don’t require separate app download.

• City Services — booking doctor appointments, reporting incidents to police, booking flights, taxi etc.

As a result, WeChat is more than just a messaging app. It’s more like a portal, or operating system.

This post explores the design principles that Allen Zhang adopted for WeChat.

WeChat is Your Friend

Only when we treat users with genuine empathy will our products be used for a longer time.

Product design should not be reduced to processes that can be continuously optimized by data-driven teams. Allen believes there is an amount of inspiration that optimization cannot solve for.

The fierce focus on user experience leads to a tendency to avoid monetization strategies that might negatively impact daily use. His philosophy on monetization is clear:

If WeChat was a person, it would be your best friend based on the amount of time you spend on it. So, how could we put an advertisement on the face of your best friend? Every time you see them, you would have to watch an advertisement before you could talk to them.

Unlike many other Chinese apps, WeChat does not charge subscription fees for an enhanced user experience, it also does not show ads upon launching the app. Despite all the advertising revenue potential, WeChat limits ads in its social feed to just 3 per day.

Another innovation WeChat chose not to monetize is Mini Programs.

They allowed Allen complete autonomy in how he wanted to build “WeChat as a friend”.

In 2017, Mini Programs was launched, it required no downloads, no accounts, no approval process, and no fees. Anyone could build their mini app for almost no additional cost.

All these millions of small businesses in China can now print a QR code and have an online ordering, loyalty points, delivery tracking, payment system working.

All of these functions within the same messaging app they use for their private lives.

Contrast this with Apple who owns their app store, charging commissions without incremental cost for every transaction. The profitable playbook was out there, yet Tencent didn't force WeChat to copy it.

Social Efficiency

People only have 24 hours in each day. The internet’s goal should not be to reduce our lives to spending all our time on mobile phones.

His idea that the purpose of technology should be to improve user’s efficiency is almost the opposite of what the Western social media industry focus on.

Allen dispels the common understanding that the more time you spent in app means it’s a good thing. For him, WeChat is a tool that must help users get the most useful information in the shortest possible time. WeChat surely has many ways to increase app usage time, but that doesn't mean it will benefit users since it lowers their “social media efficiency”.

They care about a core question: Is WeChat the most efficient social network?

Because when they solve for this, then WeChat is the best tool for users.

For example, WeChat doesn’t give sent/read indicators when messaging. The goal is to send your message then leave the app.

The value proposition simplifies how users maintain relationships, which in turn should give them more time away from their phones.

Put another way, Allen sees his product as a productivity tool rather than a social destination.

Don't Chase Traffic

What extends from his vision of WeChat’s purpose is how KPIs are counterproductive. For Allen, his team is not adopting a data driven approach.

His team does have key metrics, but they are mostly evidence based rather than a single number. The good metrics reflect an outcome from understanding users; it’s never the primary goal.

Our teams have developed the habit of thinking about the deeper meaning behind each function and service. If a function is done purely for traffic, and we can’t think of what value it brings to the user, this feature is problematic or not long-term.

The industry at large focuses on traffic, product managers usually have KPIs based around producing traffic. If you incentivize people this way, then you cannot expect to create the best product, because they will find all possible ways to increase traffic.

Although WeChat has an enormous user base, they don’t leverage this to promote new features. Chasing traffic for the sake of it isn't WeChat’s game.

During the first five months of WeChat, we didn’t promote any new features. If a user is not willing to share features naturally, it is meaningless how much we promote products.

Decentralize

WeChat’s design is deceptively simple, the navigation bar only has 4 tabs:

1. Chats

2. Contacts

3. Discovery

4. My Profile

But WeChat can be used for many things like shopping, news, travel, games… So how is it that their user interface is so clean?

Consider that most platforms have a three-sided relationship with the user, publisher, and platform operator. For publishers, they are incentivized to pay the platform which in turn encourages users to click on a curated list of top trending products. This often results in big publishers getting bigger.

However, this centralized way of operating doesn’t guarantee that those promoted products are the best.

WeChat strives for decentralization, taking the role of an administrator, providing a space where publishers can earn their own growth by providing the best value to users.

So, users have to actively discover all these third-party services themselves. It might seem to be a poor discovery experience, but if they didn’t decentralize, then Tencent would have monopolized Mini Programs. This benefits the short run but hurts the ecosystem in the long run.

Forcing self-discovery earns trust in a unique way. Trust from customers is the most difficult and greatest achievement of any business.

Despite not promoting publishers, in less than 2 years after launching Mini Programs, there were over 1 million mini-apps and over 200 million daily active users!

Brands have discovered that user trust is a key in turning a network into a transaction platform. It’s the reason Apple and Amazon have been able to expand into new revenue streams so quickly.

In the West, companies use strategies like encryption and real-identity policies to reinforce a sense of security. On the other hand, WeChat uses a different set of tactics.

For example, group chats are discovered entirely by word-of-mouth or QR codes. You can't search for public groups. The QR is automatically disabled once a group reaches 100 members. Once the group grows to that size, users can only join if invited by a friend, capping the group at 500 members.

This means that even large public group chats have a built-in social filter, since anyone who joins the group is likely the friend of at least someone in the group. As a result, every group feels like a secret party.

Cultivating a sense of safety is key. Unlike WhatsApp and Signal, which reveal users’ phone numbers to the group, WeChat allows its users to use aliases. Anonymity acts as a privacy shield, giving users control over how their identities are displayed (phone number and real-name verification required to sign up).

Trust is the key reason why people are willing to put in their payment credentials in WeChat for everyday mobile payments.

Ending Thoughts

Allen’s idealistic product-driven philosophy is elegant, but has its own challenges too.

For one, his ideas are critical for long-term survival, but in the short-term it burns cash. If WeChat was a standalone company, would it be able to postpone monetization for so long?

It is also fascinating to wonder how such an opinionated and stubborn leader minimize friction at work?

Good products require a certain level of dictatorship, otherwise it will contain a lot of different opinions and fragment product character.

When Allen Zhang made clear he would not relocate to Shenzhen, would not attend management meetings, and would not conform to corporate expectations… Pony Ma didn’t try to consolidate Allen's team.

Instead, he established the Tencent Guangzhou Research Institute so that Allen could stay in his provincial office — this talent retention tactic is rare in corporations.

Intro

Did you know that in 1850, sewing a standard men’s shirt took about 10 to 15 hours? It was only right after 1850, by the introduction of foot-powered sewing machines, it reduced the time taken to just 1 hour.

This is an old Singer model sewing machine at home, some of you might have one of these ancient artefacts laying around.

Named after the man who invented the machine, Isaac Merritt Singer. His story is one of rags to riches, never successful in anything he did up until he tinkered with the idea of a sewing machine.

He was a flamboyant man and had relationships with 5 different women, and had 24 children. It’s not a stretch to say that he was a womanizer who also actually liberated women. Just look at his fashion sense for someone who lived in the 1800s:

Early Life

Isaac Singer was born in Pittstown, New York, 1811. His parents didn’t have a good marriage and divorced when he was just 10 years old. This trauma may have caused his polygamy during his adult life.

He was then put under the care of his new stepmother, but he didn’t like her and so he ran away from home at 12 years old. Isaac went to Rochester and stayed there for 7 years, becoming an apprentice at a mechanic’s shop when he was 18.

Isaac wrote this down in his diaries:

November 29, 1821

Papa and Mammy left each other today. They said something about a “d’vorce,” but I didn’t know what that meant. I’m writing now because before Mammy left, she gave me this book, so I’m writing in it now. ‘Tis a scary place without Papa. I will write again.

April 9, 1822

I couldn’t stand all the noise and sadness at home so as soon as I could get my hands free, I left. I’m old enough, anyways. I’m twelve. I joined a traveling stage act and they was very kind to me. Also, I do odd jobs to make some money to send to Papa — that’s how good of a son I am.

By some accounts, he was a talented mechanic, but such menial work could not contain his ambitions. He wanted to do something else… art and theater attracted the young Isaac and he dreamed of becoming a Shakespearean actor.

So, in 1830, young Isaac joined one of the many travelling theatrical troupes. This marked the start of his nomadic life, moving from town to town performing. His favourite fictional character was King Richard III from William Shakespeare’s (1592) play.

The same year, he married his first wife, Catherine Haley, who was 15 years old. Isaac was poorly paid and in 1836 they moved back to New York City. They had 2 children together, son William Singer, and daughter Lillian.

He would record these events in his diary:

March 28, 1827

I met a pretty girl today! She said her name is Catharine. She’s only thirteen now, but we can wait a little while longer to get married. She sure can cook. I visited her house and we had soup, fish, roast mutton, currant jelly, and vegetables. For dessert, she made pudding. She’s such a nice lady.

July 8, 1830

I’m a man today! This is because I married Catharine Maria Haley (she’s fifteen years of age and quite pretty!). We didn’t have a grand wedding like most people do because I didn’t have much money. So we just had a little ceremony and now we’re husband and wife! ‘Twas a grand day.

Inventing The Sewing Machine

Isaac followed his troupe down south to Baltimore where he met Mary Ann Sponsler. He found a pretty girl and couldn’t resist cheating on Catherine back home, and so he journaled:

February 11, 1836

I left with the company of theatrical players and we toured through Baltimore, Maryland, where I met the charming eighteen year old Mary Ann Sponsler. I can’t believe it, but I actually proposed!! She agreed to wait for me to return since I had some business in New York City. Yes, I’m still married to old Catharine, but I suspect she is having an affair with another man.

It wouldn’t be long after Isaac separated with Catherine, while Mary Ann moved in with him to New York. Although the two of them never married, the couple would raise 10 children, 2 of whom died at birth.

Realizing that theatre was not going to earn him money, Isaac went back to his mechanic roots and started inventing machines. In 1839, he invented a rock drill patent which sold to the I&M Canal Building Company for $2,000.

Strangely enough, he decided to use this money to start his own acting group! It was called “The Merritt Players”.

By 1844, he had 4 children, adding one child per year since the couple moved back. Then in 1846, another child was born, the family was getting large and finances were thin.

Isaac’s next 4 lines in his journal lamented about having so many children. Well, rubber condoms would only come in 1855…

April 13, 1844

Fanny Elizabeth Singer born today. I confess: I am getting tired of all these children who have to be fed and clothed.

December 14, 1846

Jasper Hamlet Singer born on this winter day. I hope I can soon get a patent to support this large family.

June 1, 1847

Mary Olivia Singer born today. I’ll leave it at that.

June 13, 1848

Julia Ann Singer born today. I hope this is the last child.

He continued to work on his inventions and made a carving machine that could carve the wooden type-board used by printers.

The endeavour attracted bookseller and publisher George Zieber, who agreed to loan Isaac $1,700 to work on his carving machine. They found a Boston machine shop owned by Orson Phelps and set up shop there. Unfortunately, metal type-boards were becoming popular and Isaac failed in this venture.

Zieber continued to finance poor Isaac who had to feed two “wives” and 8 children.

Isaac was already 37 years old, but a lucky break would come from Phelps’ machine shop. The first mechanical sewing concepts and functional designs were created in Europe, but the tailoring guilds wanted to protect jobs and thwarted the adoption of sewing machines. These guilds were outlawed in 1791 after the French Revolution.

And so the technology travelled across the Atlantic Ocean to American soil. Phelps was making one of these machines but they were unreliable, so they turned to Isaac and asked him to improve on it.

Isaac was uninterested and said:

What a devilish machine! You want to do away with the only thing that keeps women quiet, their sewing!

Despite the objection, he was desperate for money, so in 1850 the three men formed a partnership of equal rights. Isaac and Zieber worked hard on the sewing machine and finally succeeded in a good design.

In 1851, they incorporated the foot-powered pedal and side turning wheel. Isaac’s sewing machine could do 600 stitches per minute.

They drew up the patents and went out selling their machines on horse carriages. Isaac’s theatrical personality made him a good salesman, and the Singer Company was born.

Corporate Drama

Rapid success followed within 2 years, they were exhibiting in Europe boasting incredible productivity gains reducing time taken to make a men’s shirt from 14 hours to 1.

Competition was fierce. Rivals Grover & Baker and Wheeler & Wilson came with patent infringement lawsuits.

Singer sewing machines sold at $125 each when the average family made $500 per year. Another competitor, Elias Howe, priced their models at $300. These machines were expensive and slow to be adopted.

Zieber and Phelps became discouraged and wanted to sell their stake to a new investor, but nobody was interested. Isaac himself offered to sell his ownership for $1,500, but Zieber talked him out of it. Phelps did manage to sell his shares to a new person, but he didn’t get along well, so Isaac and Zieber bought his shares.

Meanwhile, Isaac hired lawyer Edward Clark to work on their patent lawsuits. At this time, Zieber fell sick and Isaac concocted a lie, saying that the doctor thought Zieber was probably going to die.

It was under this ruse that Edward Clark and Isaac Singer bought out Zieber’s shares for $6,000.

The “sewing machine wars” would come to an end when another lawyer suggested that the 3 companies pooled and share their patent. This created the first “patent pool”, a technique that was later used in automobiles, movies, and radio. Now they could use the patents by paying the pool $15 for each machine made, and let others use the patents for similar fees, splitting the profits equally.

Many Wives

It turned out that Edward was a better businessman than lawyer. With the patent problems behind them Singer made 2,564 machines in 1856, up from 883 the year before.

By 1858, Singer was the third largest manufacturer producing 3,591 machines, behind Wheeler & Wilson’s 7,978 and Grover & Baker’s 5,070.

Just 2 years later, in 1860, Singer sold 13,000 machines.

Growth was booming and made Isaac a wealthy man. His lover Mary Ann was popular among the fanciest stores in town, the family had the largest carriage fitted with seats for his large family, beds for his kids, and a toilet.

Isaac finally divorced his first wife, Catherine, on January 1860 after 30 years of estrangement. It was Isaac who divorced her for adultery. Mary Ann was pleased with the arrangement and kept the divorce papers, then she asked for marriage — which of course Isaac had no intentions.

Then 7 months after the divorce came a bombshell. Mary Ann’s lawyer filed a document alleging that Isaac Singer:

Repeatedly beaten and choked her to insensibility, frequently forcing the blood to flow in streams from her nose, mouth, face, head and neck.

Oh! Isaac was a wife beater (are you surprised)!?

It had happened before, in the same new house. The previous year he had beaten her over an uncovered box of matches left in the bedroom; their daughter Voulettie threw herself between them, and he beat both women until they lost consciousness.

Oh boy, you'd believe that was bad enough… one day, Mary Ann was riding in her carriage and caught Isaac cheating with one of his mistresses in another carriage.

She screamed.

Edward Clark horrified that this scandal would hurt the business told Isaac to flee the country. It turned out that Isaac actually had 4 families, all living in Manhattan, with a total of 16 children!

Before getting himself arrested for bigamy, Isaac would marry yet another woman named Isabella Eugénie Boyer from France in 1863. Both of them fled to Paris, then to the Devon coast in England. Isaac was 52 and Isabella was only 22, they managed to have 6 children.

If you’re still counting, Isaac has 24 children so far!

Back to Business

On the other hand, Edward ran the company very well. It won’t be long that Singer overtook the two rivals in 1867. By 1870, they were selling 127,833 machines.

He developed a program where people could rent the sewing machines with a small monthly payment. This was the first use of “installment plans” by a major American company.

Singer was one of the earliest retail chains opening hundreds of offices across the country.

Then they went overseas, first setting up a big factory in England, then Russia.

At this point in time, Edward and Isaac disliked each other, but Isaac’s flamboyant style was still useful. He held a huge event in New York, inviting their customers and employees to free food, while seamstresses demonstrated sewing machines on stage.

Isaac Singer died in 1875 at the age of 63 at his British estate. He left behind $13m and his wealthy children lived comfortably. But over the next 150 years their inherited fortune disappeared by the fifth generation.

Later in 1882, Edward Clark died, leaving behind a fortune of $25m and lots of real estate in NYC.

After both founders were gone, the business continued to prosper.

Efficiency in production caused prices to drop rapidly from $100 to $30. Demand skyrocketed, selling over 500,000 machines in 1880.

By the early 20th century, the company was making more than two million sewing machines a year, over half of them in England. Estimates place Singer’s market share at 80% of global production (1880 — 1920). Over 1000 different models for industrial and home use were made.

In 1900, to flex their financial prowess, they built the tallest skyscraper in the world: The Singer Building. Construction was completed in 1908 at over 600 feet. It only held the tallest honor for a year, until the Metropolitan Life Insurance Company built its Manhattan Tower. The Singer Building was demolished in 1969 due to inefficient infrastructure.

At the same time, Singer built a headquarters for their Russian operation in St. Petersburg. Since the city had limits on how high a building could be built, the architects crowned it with a glass globe. Before the Communist Revolution, the building briefly served as the American embassy. Today, it is home to St. Petersburg’s largest and most famous bookshop Dom Knigi.

How It Ended

By the 1920s, Singer had 1,700 stores in the United States and 4,300 overseas, supported by 60,000 salesmen.

Even into the 1950s, Singer continued to hold 40% market share in the US.

Then the 1960s conglomerate trend came where large corporations bought unrelated businesses using cheap credit. Singer’s management was caught up in this, they bought into calculating equipment, automated point-of-sale cash register systems, aerospace and defense industries.

In 1966, sales crossed the $1b mark.

In 1969, sales doubled to over $2b.

However, Singer made all of its profits from sewing machines. The high revenue numbers came from consolidating unrelated businesses that produced no earnings.

At the same time, European competitors Pfaff, Bernina, and Necchi were on the rise, along with Japan’s Brother.

To fix this mess, in 1986, the sewing business was spun-out. After 135 years, Singer was one of the most recognized American names in the world.

The new entity was publicly traded subsidiary called SSMC, which was later acquired in 1989 by Semi-Tech Microelectronics.

After buying the Singer sewing machine brand, founder James Ting used a complex web of interconnected publicly traded companies spanning Canada, Hong Kong, and Bermuda to buy up struggling consumer brands. This included German sewing machine maker Pfaff and the electronics brand Akai. To finance this aggressive growth, Semi-Tech piled on hundreds of millions of dollars in high-interest junk bonds.

Then came the 1997 Asian Financial Crisis causing both Singer and Semi-Tech to file for bankruptcy in September 1999. At liquidation, Semi-Tech had just $600k in cash left against $650m debt.

Turned out that James Ting was involved with accounting fraud. He fled to Hong Kong where he was arrested in 2003, but ultimately managed to walk free after his conviction was overturned. His legal battles continued for years, leaving him bankrupt in 2016.

Singer was restructured and merged with Husqvarna Viking and Pfaff, all of them operate under SVP Worldwide (owned by PE firm Platinum Equity) today.

CSU just released Q2 2026 results and we are seeing no signs of AI fears materializing in their financials, nor did management express concerns over business prospects.

The very first red flag that people check is the FX-neutral organic growth, which came in at a weak +2% YOY, much lower than +4% in Q1. However, there are 2 things we must understand:

1. CSU is a serial acquirer that operates in VMS space. It’s known that growth has been mostly from acquisitions, organic growth is not as important as capital deployed. As of Q2 2026, CSU reported $1.6b for acquisitions. Given that their hurdle rates are unchanged, this is very likely to produce above 20% incremental returns. Furthermore, they have $818m in committed considerations for acquisitions. In total, we are looking at $2.4b deployed so far and even more for full year 2026. This quantum of capital deployment is a record high.

2. For serial acquirers, there will be a lot of accounting noise. Q2 2025 had one-off items because IFRS 15 requires software revenues to be recognized based on when control of the product is handed over to the customer, rather than when cash is actually paid. This applies to Altera which signed a few big new contracts causing large revenues recognition, so YOY comparison is negatively affected.
   
   Lumine’s organic growth was also just +1% because of they are working through cost restructuring of newly acquired Synchronoss.
   
   A South American business lost one large customer which attributed -0.3%.
   
   If we adjust out these one-off items, the organic growth normalizes back to +5%, in line with averages.

You might question why are we stripping out Altera’s impact when its revenues fell by -22% YOY with declining trend:

Remember that CSU is about achieving high IRR. Again, they don’t prioritize organic growth, IRR depends on speed of cash inflow — basically time value of money.

Altera was acquired in May 2022 for $670m, financed by $335m of non-recourse debt and cash $335m. LTM Altera produced $104m of free cashflow (FCF), and over its lifetime under CSU it has generated $400m. It already paid off the cash outlay!

To exceed 15% IRR in 10 years, Altera can afford to perform badly because significant FCF came in early:

Despite being a declining asset, the IRR for Altera is on track to pass the hurdle rate.

Moving to the AI hot-topic, management’s message was consistent, practical, and candid. Rather than positioning AI as a risk/opportunity, Mark Miller simply stayed true to the decentralized nature of CSU:

You can build products fast, but selling them is a whole other thing. In most cases, we are using a blowtorch to light a cigarette. […]

What we are not doing is running an AI program out of head office. Our business unit managers understand their verticals far better than we do, and they are making those calls themselves at their own pace, funded out of their own P&Ls. We are not going to give you an AI target, an AI revenue line, or an AI timeline. If we start reporting numbers like that, we will start managing to it.

So, AI is indeed very powerful today, but it actually means little for CSU’s customers. These are small businesses who care more about consistency than new features. Creating incredible software was never the business model, but solving customers’ problems is the reason why maintenance recurring segment is 77% of total revenues and grew +20% YOY:

The finance team has created a new account code to track AI expenses. It’s a small operational detail that shows management wants to get visibility on their expenses.

Because of how CSU incentives focus on IRR they won't aggressively innovate, unless they have confidence that a new feature can sell. CSU typically waits for their customers to request something before building it.

In this context it makes sense why AI doesn't have much of an impact yet, as most small businesses with limited budgets wouldn’t know what to request from AI.

Capital Allocation

An important point from the earnings call was that hurdle rates are unchanged. On the back of that assurance, we are seeing higher pace of capital deployment.

We are also seeing higher prices paid, particularly 4x revenues for DerbySoft (travel tech & digital marketing company). This is much higher than the typical 1—1.5x revenues multiple.

We should check the ROIC. It is very encouraging to know that the invested capital base grew from $2.8b to $9.1b in the last 5 years, yet ROIC is still comfortably above 20%.

Adjusted FCF of $2.5b is a record high growing +20% YOY with 20% margins:

On private market software valuations, CIO Bernie Anzarouth reminded us that prices are still not matching public software companies:

I think the competition is still very robust. No one is giving up on vertical software, whether it is large or small. We are seeing some weaknesses at the high end in pricing, but it is still very competitive. So it is not like we are increasing our win rates or anything like that. It is same old, try to get as much as we can.

He also noted that the wave of private equity roll-ups funds that emerged 5 to 10 years ago are starting to reach end-of-lifecycle, and CSU expects to eventually absorb some of those portfolios as sponsors exit.

Overall, we think it was a good earnings report. At current market cap of US$47b, the starting FCF yield looks good at 5.3%, so we are maintaining our ~10% allocation at $2788/share (click here for more valuation details).

Details on Lumine

We want to shine some light on the +1% organic growth of Lumine, let’s look at the revenues breakdown:

Maintenance and recurring segment is the bulk of total revenues. Great news, growth was +34% YOY. This is high margins, high scale type of sales.

Professional services fell -7% which is actually also good news. Because this segment involves low margins, low scale consulting work.

We should think quality over quantity!

Lumine’s playbook is to buy carve-outs, make operating improvements, and reap the margins expansion. Then, use the resulting cashflows to reinvest into more acquisitions.

A good example is the recent Synchronoss acquisition. In Q1, margins compressed as Lumine took on a bloated cost structure from Synchronoss. This caused general and admin (G&A) costs to increase +95% YOY.

What happened in Q2?

G&A costs only increased +34%, this is a good sign that Lumine is doing well in cost restructuring.

Capital deployment is also very high:

1. Synchronoss $309m

2. Imagine Communications $136.8m

3. Quortex $96.8m

That’s a total $543m deployed so far. Compare it to Lumine’s market cap of $4.9b and you start to see the potential future returns.

To fund #2/3 deals, Lumine used $110m on its corporate credit facility, leaving $220m drawn against $360m total capacity. That still leaves $140m for further acquisitions for the rest of this year.

Due to the additional debt, FCF decreased -17% YOY to $60.4m. Interest paid increased to $4.2m from $3.9m, and transaction costs on debt (essentially zero in the prior year) hit $1.9m this quarter.

Working capital absorbed $20.3m, driven primarily by $8.4m decrease in payables.

We should check that these costs eventually normalize as acquired businesses start to return cash.

Historically, FX-neutral organic growth decreases after acquisitions, then recovers with 2 to 3 quarters.

• WideOrbit (Q1 2023)

• Motive and Axyom.Core (2024)

• Synchronoss (Q1 2026)

Of course, being a spin-off from CSU, the strategy is similar — prioritize IRR on cash deployed. If Lumine continues to do carve-outs at this pace, organic growth will remain perpetually depressed.

This is not a concern if ROIC is high. So far, the results are very encouraging:

Jeff Bezos, founder of Amazon, famously said that customers will always want cheaper things delivered to them at the fastest possible time. This idea has pushed Amazon's e-commerce logistics to dizzying sophistication, complexity, and scale.

Now with the idea of agentic commerce coming, we are testing the limits of logistics infrastructure to deliver a near infinite assortment of goods at even a faster pace.

Amazon aims to make this dream reality.

On the surface, it seems like a reasonable desire for a consumer, but there's a few wrong things about it.

Even if you think it is preferable at an individual level, there are good reasons to question the social value of the logistical complexity that it necessitates.

Home delivery of single-packaged items requires an entirely different cost structure versus freight trucks moving goods to consumer-facing warehouses where customers do the last-mile home delivery themselves.

This represents different ends of a spectrum, which has produced two companies with very efficient, successful business models: Amazon and Costco.

There's 2 main differences about them:

1. Costco was late to the e-commerce game, it doesn't invest so heavily in their distribution network.

2. Costco commits to a limited assortment, the opposite of Amazon with near infinite shelves.

To help us think about which company benefits society more, let's break the dream down into 2 components: variety and speed.

Variety

There are some cases which consumers might want access to high product variety. For instance, buying furniture of certain dimensions, or when making a major electronics purchase.

But in general, scrolling through options and reading through reviews online for every consumption choice is overwhelming.

We know that the brain loves to make easy decisions.

Costco does this by exercising constrain. Compared to the large number of assortments in Amazon or Walmart, it deliberately keeps variety low.

Wouldn't forced selection turn shoppers away and limited choices should reduce social surplus?

Not true, the popularity of Costco is evidence of the contrary. Membership renewals have always been above 90% retention!

Indeed, the pre-selection of items for sale in their warehouses is part of the value proposition. You are going to get a bulk purchase of a particular item for a good price, and you also won’t have to think about small differences in varieties.

The Costco experience makes customers’ brain comfortable, but only because they trust Costco gives the best value. Although it is not known to stock the cheapest goods, but it has the cheapest price on its goods. That's a critical value proposition.

Costco can achieve this because their employees have closer relationships with suppliers than any other big retailer.

This is a structural effect of Costco’s conscious choice to offer low assortment; fewer products means more time to negotiate each product.

As long as Costco is the middleman, suppliers are never going to rip off the end customer in terms of price and quality. This expectation is baked into shoppers’ minds when they step into Costco.

For as long as Costco keeps the promise, customer retention will remain high.

Another benefit to Costco is a negative cash conversion cycle (CCC). The CCC measures how long it takes to turn inventory into cash through sales. Both Amazon and Costco achieves negative CCC, meaning that they receive cash upfront faster than they pay suppliers.

Amazon achieves through the speed of its e-commerce delivery, quickly collecting payment and turning over inventory. Because it is reliable, they can negotiate longer payment terms with suppliers.

Costco achieves through fewer variety, when a Costco store receives a shipment of a particular item from a supplier, it is often going to sell every unit in that shipment in less than a month, since that it the only variety.

Speed

Is the fastest delivery really beneficial to society as a whole?

Going by first principles, logistical success in business means fundamentally the success with which goods are supplied to the customers who need them.

At the social level, logistical success can be measured in terms of cost efficiency. This cost can be understood in accounting terms as SG&A (selling, general & admin) expenses.

There are 2 things that distinguish an efficient logistics company:

1. It can spread SG&A over a large volume of sales, extracting maximum economies of scale.

2. Operations are reliable and not prone to disruption. The more complicated an operation, the more likely it is to fail. A simple operation also puts fewer demands on our already congested transportation infrastructure.

Yes, Amazon is able to deliver toothpaste quickly to your doorstep, but we think this is not something every business should strive towards.

Because for most consumption, it is more efficient for people to handle the “last-mile delivery” themselves by going to stores and buying in bulk.

Why so?

Think from the perspective of retailers… it lowers the need for price mark-ups by keeping overheads low and simplifying logistical operations.

We have evidence of this from Costco SG&A expenses which are consistently lower than any of its competitors at 10% of sales.

Amazon delivery costs about 40% to 50% of third party gross revenues.

One of the reasons for this is the simple nature of the Costco distribution network. At Costco’s depots, all in/outbound inventory is cross-docked in pallet quantities. Full pallets come in from suppliers on one side of the building, workers on electric pallet jacks move them from one side to the other, and full pallets are loaded onto trucks bound for stores. There is no pallet breakdown, no conveyor belts, no fancy robotics.

This setup also allows Costco to pay higher wages for its warehouse employees. For comparisons on per hour basis:

• Walmart pays retail workers $16.

• Amazon pays warehouse workers $19.

• Costco pays depot/warehouse staff $21.

This has social benefits:

1. Very low staff churn for a retailer at just 6% per year.

2. Exceptionally low inventory shrinkage from theft at 0.1%.

This is not just attributed to company culture, but the economic fact that when Costco minimizes overheads, they can pay workers more without hurting margins.

Conclusion

From the above arguments, we will make this bold claim:

Costco offers a far more logistically elegant and socially beneficial model of goods distribution than Amazon.

Amazon dominates when it comes to logistically-complex operations, but there is no inherent reason to prefer complicated operations to simple ones.

If anything, simplicity should be the rule.

Why do you need to figure out how to integrate robotic arms into a fulfillment operation dominated by autonomous mobile units when you can cross-dock full pallets?

Is it more logistically impressive to solve difficult problems or to eliminate the need to solve them in the first place?

Thinking deeper, Amazon's Prime membership might take away some society surplus because two-day delivery requires huge investments in their distribution network. This increases costs, and someone else has to pay for it — turns out it's an indirect tax to society.

However, it will be foolish to claim that Amazon's logistics have no positives. Same-day medicine delivery is obviously good for society.

If we’re looking for a generalizable model for the social distribution of goods, Costco offers an useful blueprint for everyday consumption while Amazon does not.

Intro

Every investor knows the airlines industry is notorious for destroying shareholders’ value. Perhaps you have heard Warren Buffett in the 2007 Berkshire annual meeting:

The worst sort of business is one that grows rapidly, requires significant capital to engender the growth, and then earns little or no money. Think airlines.

Don't take his word for it. The International Air Transport Association (IATA) has documented for years that airlines as a sector destroy investor value in the aggregate. The IATA’s 2026 outlook (before the Iran war) projected an average return on invested capital of 6.8%, against a weighted average cost of capital of 8.2%.

As the IATA’s report said:

The airline industry collectively does not generate earnings that cover its cost of capital.

This has been the case for a long time. From its deregulation in 1978 until today, the airline industry has cumulatively lost money.

Below is a McKinsey study of ROIC versus cost of capital, the airlines industry on average cannot produce adequate returns and thus growth actually destroys value:

The proof is in the numerous bankruptcies that have transpired:

• Pan American was the largest international air carrier for much of the 20th century. The first airline to fly worldwide, it pioneered innovations such as jumbo jets and computerized reservation systems, and introduced the first American jetliner Boeing 707 in 1958. Competition and deregulation caused its slow decline to eventual bankruptcy in 1991.

• Eastern Air Lines operated from 1929 to 1991 and went bust. It sold some routes to Donald Trump, who ran it under Trump Shuttle from 1989 to 1992, until it faced too much accumulated losses and closed down.

• TWA, the carrier of Howard Hughes, was absorbed into American Airlines after a third bankruptcy filing in 2001.

• Alitalia, Italy’s flag carrier, reported only a single year of profit since its founding in 1946. It received rescue packages from the Italian government totaling well over €7b before finally shutting down in 2021.

Between 1978 and 2005, more than 160 airlines filed for bankruptcy.

In 2005 (post 9-11 terrorist attacks), every one of the four largest American airlines (United, Delta, Northwest, US Airways) was operating simultaneously under Bankruptcy protection.

Why is the airlines business so bad despite selling a service that is used by millions of people everyday at a much cheaper cost, on safer, more comfortable planes?

It gets stranger when you look at other companies involved in aviation: Airplane manufacturers, aftermarket parts, airports… they all do well in aggregate.

We think that this is the economic nature of a competitive airline industry. The industry's equilibrium state is an unprofitable one. It's not because airlines are vulnerable to shocks or they’re poorly managed. The problem is the industry itself can either be profitable, or it can be competitive, but it can’t be both.

Theory of The Firm

We can breakdown the problem into both demand and supply side dynamics.

On the demand side, people don't really differentiate which airlines they take, preferring the fastest and cheapest.

On the supply side, airlines combine high fixed costs with the lowest marginal cost achievable only at very large scales.

If we agree that firms are profit maximizers, then they must seek to produce at an output which minimizes long run average cost (LRAC).

This is called “minimum efficient scale”.

On the left side of this point there's economies of scale to be earned, on the right side there's diseconomies of scale which is inefficient production:

Let's consider a theoretical example:

Suppose an industry such that the minimum efficient scale of a single firm is large enough, so that the market can support 2.5 efficient firms.

So, 2 firms can’t produce enough to satisfy demand, but 3 firms is too many for all to operate at full capacity. Yet you can't have half a firm, so what happens?

With just two firms, demand exceeds supply, prices are high and excess profits exist. This attracts a third firm to enter.

Now there are 3 firms, and supply exceeds demand. Someone has to operate below scale. Eventually, one of them will have to leave. We end up with 2 firms again, profits appear, competition enters… the cycle repeats.

You might already notice the issue. The dynamics in the example above happens because the minimum efficient scale is large. In other words, the number of efficient firms is low.

Imagine the counter-example: A bread industry which can support 200 bakeries, if 1 more bakery entered, no one would notice the difference.

The bread industry is fine because the number of efficient firms is high. Products are differentiable enough through brand and taste. Demand is steady. So the bakeries can settle into a stable equilibrium, each earning normal profits.

But when the efficient number of firms is low, adding or removing one firm causes significant equilibrium disruptions.

This is the crux of the problem.

Combine this with the volatile demand of air travel and airlines experience over supply the moment demand softens.

Fixed Cost

Why is the efficient output high for airlines?

Well, it's obvious that they have very high fixed cost.

Consider these fixed costs for an airline:

1. Aircrafts cost hundreds of millions of dollars.

2. Gate and parking fees in airports.

3. Labour is governed by law which protects the pilots and cabin crew.

4. Repair and maintenance that are mandatory by law.

5. Even fuel cost is fixed because airlines pay to heavily hedge against oil prices!

Since fixed costs are much higher than variable costs, flying a half-empty plane is as costly as a full plane. Airlines can't adjust capacity proportional to the decline in demand. As a result, flight ticket prices have the tendency to be competed down, thus profits erode to losses.

This happens due to the toxic combination of competition + high fixed costs. Players are forced to produce a high level of output relative to total demand to be efficient. In such a situation, competition will create boom-and-bust cycles.

That's why airlines go bankrupt so often.

Even so, bankruptcy doesn't solve the problem.

Under American law, Chapter 11 Bankruptcy protection allows a company to continue operating while it restructures its debts under court supervision. Airlines use this to renegotiate its rigid cost structure, from aircraft leases to collective bargaining agreements.

For example, when United Airlines filed for bankruptcy in 2002, it defaulted on its pension plan and left $3.2b for the next 5 years to be absorbed by the government’s Pension Benefit Guaranty Corporation.

Another example, US Airways which was under court protection for the second time since 2002, terminated the last of its pension plans in 2005. The federal government took over the responsibility of $3b worth of liabilities.

For airlines, bankruptcy rarely ends with liquidation. Those that emerge from bankruptcy proceedings, having transferred their liabilities elsewhere, can operate at a fundamentally lower cost basis than their competitors.

But that doesn't solve the competitive equilibrium problem. It just resets the floor at a lower level, from which a new round of ruinous competition can begin.

Airlines Cartel

If competition is bad, then low competition is one solution.

Yes, we can just look at how the railroads industry evolved. They have similar equilibrium problems and ended up in oligopolies.

Anti-competitive measures, like cartels, mergers and vertical integration, are necessary in these industries. This doesn't stem from corporate greed, rather the uncompetitive solution is the only stable equilibrium.

This realization is not new, people have already seen the numerous failures since 1930. So, way back in 1940, the US government created the Civil Aeronautics Board (CAB) to regulate the airline industry.

Essentially, this is a government-approved cartel.

The CAB told airlines which routes they could fly, when they could fly them, and what they could charge.

Barriers to entry were erected by law. There were zero major interstate airline admitted to the industry for 40 years (1938 — 1978)!

Fares were set to provide carriers with reasonable rates of return.

Under this regime, the airline industry was profitable, they competed on service and cabin quality, but never on price.

This situation would not last for long, eventually deregulation caught up…

In 1978, the Airline Deregulation Act stripped the CAB of its rate-setting and route-licensing authority. The regulated cartel turned into a competitive free market.

The result was waves of instability.

The first casualty came in 1982. Braniff International Airways was once the eighth-largest airline in the US, operating an extensive network of domestic and international routes through the South and Latin America in the early 1980s. In the deregulated environment, Braniff management pursued aggressive expansion, adding routes and capacity in hopes of capitalizing on the competitive environment. Instead, Braniff encountered overcapacity, weak pricing, and operational challenges. By 1982, Braniff was sustaining heavy and unsustainable monthly operating losses and had exhausted its credit lines.

The 1990s were relatively stable, and then the 2000s saw another wave of insolvencies. United, Delta, Northwest, and US Airways all went into bankruptcy protection after the 9-11 terrorist attacks.

The industry enjoyed a period of rare stability in the 2010s, perhaps prompting Warren Buffett to believe that airlines economics have finally stabilized. In 2016, he bought stakes in American Airlines, Delta, Southwest, and United. Then COVID pandemic shock came, although they didn't go bankrupt, but they accumulated massive amounts of debt and dilution to stay afloat.

Innovative Solutions

The market always has clever ways to pull itself out of a hole. Since there was no way to recreate a regulated cartel like before, airlines have sought to create a pseudo-cartel among themselves.

Alliances

We hear their announcements often in flights: Star Alliance, SkyTeam, oneworld.

They allow weaker airlines to coordinate scheduling, share revenues, and refrain from undercutting each other on high-value trans-oceanic routes.

Hub-and-spoke Model

By concentrating its operations at a few major airports, an airline can turn those airports into something close to local monopolies.

American Airlines carries 35% of all passengers starting or ending their trips at Charlotte Douglas, and 70% of passengers connecting through the airport. But they have only 6% at San Francisco.

In turn, United dominates SF with 44% share, but has minimal share at Atlanta International, which Delta has at least 70% share.

In effect, major domestic airlines have carved up their own territories, with each one operating a quasi-monopoly through which it produces the margins that cannot be earned in free competition.

Frequent Flyer Program

Now this one is really clever. Instead of trying to survive operating flights, why don't you treat flights as a loss-leader for something more lucrative?

The first frequent flyer programs were invented soon after deregulation as airlines scrambled for ways to lock in customers.

Frequent flyer programs are a different business model: Miles-as-currency operations whose value bears no relationship to the cost of the seats backing them.

Maybe that's why Buffett invested in Delta again recently, especially when Berkshire owns a big stake in AMEX?

Delta is the most profitable airline in the US. But it's not because of their flight operations… They have a partnership with AMEX in 1996 and launched a co-branded card with them in 2008.

In 2025, annual spending on Delta-branded AMEX cards generated $8b of revenues for Delta. This almost zero marginal cost business is much bigger than Delta's entire net income of $5b!

Delta would be in operating losses without the AMEX partnership, that's for sure.

In effect, Delta’s aviation business is a loss-leader for a much more profitable credit card partnership.

So Delta can be said to be successful because it morphed into less of an airline!

Conclusion

In today's modern world, air travel have become a product that most people can afford, not limited to the wealthy few.

This was gained off the back of equity holders and airline workers whose contracts are renegotiated when airlines fail.

In this light, the deregulation of 1978 should not be viewed as negative. Without it, prices of air tickets wouldn't be so low now.

This study teaches us something about economics:

Not all industries are able to attain a profitable competitive equilibrium when left to free markets.

We need aviation to exist as an industry, but we’re unable to have it survive. The natural tendency is toward anti-competitive consolidation, or some kind of innovative strategy like Delta's partnership with AMEX.

We have written about the business economics of GOOGL in the past here. This post will attempt to address the question of whether GOOGL is able to get a good return on invested capital (ROIC) on their enormous CAPEX investments.

As management continues to reiterate their bullish expectations, we have not seen them quantifying these statements.

It’s common sense to believe that earning high returns on hundreds of billions of dollars must be very difficult, especially against other hyperscalers like Microsoft, Amazon and META. These competitors have deep talent pool, generate lots of cash, and boast fortress-like balance sheets.

Then when we pull up GOOGL ROIC chart, it doesn’t look hopeful:

The business used to be asset-light and now evolved to asset-heavy, the last 5 years (2021 — 2025) saw CAPEX go from $25b to $91b. This year GOOGL is expected to spend more than double at ~$197b.

Assets In Service

However, we must be aware that the way ROIC is calculated above is useful only for looking at long-term trends. There’s nothing wrong with the academic way of taking net operating profit after tax (NOPAT) divided by average invested capital.

But we need to appreciate the nuance in the numbers, if we want to know what’s the return on the most recent CAPEX. Because in reality, not all CAPEX spending is productive at once — we must differentiate between assets “in service” and “not in service”.

The details are in Note 7. Property and Equipment (PPE):

Notice that as of Q2 2026, out of total PPE of $321b, there were $123b of assets not yet in service. In other words, these assets don’t contribute to earnings and should be removed from ROIC calculations.

For LTM, GOOGL had $56.9b of net incremental assets in service.

So how much incremental NOPAT did they generate over this period?

It was a very impressive result of $20.7b, representing +22% YOY growth and return on incremental invested capital (ROIIC) of 36% (=20.7/56.9).

This is a very good return considering the fact it is on $56.9b of incremental new assets, not existing assets! We can’t think of many businesses which are able to invest such large sums and still get 36% returns. Indeed, a wonderful business is one that can deploy huge amounts of capital at good incremental rates of return.

Now we can look at GOOGL reinvestment rate and estimate intrinsic value growth. Over LTM, reinvestment rate was 71% with ROIIC 36%, we estimate that GOOGL grew intrinsic value at +26%. It checks out with NOPAT growth of +22%. It still amazes us that such a large company can grow at over 20%.

Future Growth Rates

So far we have been only looking backwards in time. How about the future growth from all these CAPEX investments?

Let's model it with some assumptions:

1. $123b of assets not yet in service goes productive by Q2 2027.

2. Depreciation 10 years for these assets.

3. Additional new assets $240b (guidance $60b per quarter CAPEX).

Under these assumptions, the NTM incremental net assets in service will be $110.5b.

If ROIIC for next year remains constant at 36%, we can expect NOPAT to grow +27% YOY, which translates into intrinsic value.

Below is a sensitivity table with different ROIIC scenarios:

We think this is the baseline that investors can work from. There are a few upsides that can improve the outcome:

1. The model didn’t account for returns on new additional CAPEX.

2. TPU hardware sales that are expected to be material in 2027.

3. Earnings from renting compute from SpaceX (GOOGL wouldn't sign this $920m per month deal if foreseeable demand was weak).

Certainly, part of this rising CAPEX is due to very expensive memory chips which is a drag on incremental returns unless GOOGL can pass the cost to customers. Even if we assume $10b of assets is due to price inflation and cannot be passed on, the effect of NOPAT growth is not large (from +27% to +25%).

This supports the fact that management sees enough upside to continue investing despite high hardware prices.

Admitting Mistakes

In summary, we think the odds are in favour of GOOGL making a good return on investments. We are seeing very quick returns that materialized on assets that barely went productive for 1 year. It would be less optimistic if returns took years to appear.

We own up to our mistake of selling a substantial stake in the past (GOOGL used to make up 30% of our portfolio). Although we made a good return on the stock, we didn’t fully appreciate the nuances of assets that were in/out of service.

Such mistakes are valuable learning points — we call them “errors of omission” — they don’t show up anywhere on our financial statements, unless we discuss them openly.

For our fund holders, we can guarantee that there will be many more errors of omission in the future!

When FF7 Remake launched a few years ago, I bought a used Playstation 4 (PS4) just to play my favourite game. I was so excited that I bought the digital copy from the PS Store.

Satisfied with nearly 100 hours of play, I started to buy other games but it was always second-hand physical discs. The reason was simple, a disc allowed me to play the game then resell it, often at breakeven. I played many games for “free” this way… occupation hazard of a value investor.

Then, FF7 Rebirth released in 2025, the second installment in the trilogy. It is only available on PS5. So, I got a new PS5 using new credit card sign up promos… and bought the digital copy eagerly (again).

Sinking another 100 hours in, I completed the game and moved on to buy other games on discs (again). This time I stretched my dollar further — visiting Japan for second-hand discs.

Now, imagine my horror when I learnt that Sony is going to stop all physical discs from January 2028!

For one, I felt the days of value gaming on PS is over. On the other hand, the business analyst living inside me can’t help but to look at the digital gaming platform called Steam.

Nobody has grieved over how Steam controlled distribution of digital games.

So why did people react negatively to Sony?

Economics of Gaming

The announcement to stop producing physical discs went down terribly. Even Domino’s Pizza (UK) came up with this satirical social media post:

This move raises questions on ownership and limits community building by removing the ability to share games, as well as eliminating physical disc collectors.

Sony obviously knew it wouldn’t go well, yet it did so because it made economic sense.

The whole business model of Sony's gaming arm is to sell the console at a loss, then make it up by selling games.

In 2025, out of Playstation's total revenues of $28.8b, games software and network services made up $21b, while hardware was only $5.8b.

Its overall operating margin was modest at 10%.

Selling games is not an easy business, with just 10% margins, development and marketing expenses can quickly eat up profits.

For example, Spiderman 2 reportedly had a production and marketing budget of $315m. The standard edition of the game sold for $70. After accounting for retail cuts, digital store commissions, and licensing fees (paid to Marvel), Playstation’s net revenue per unit is estimated to be $45.

To just cover the $315m cost, it needs to sell at least 7 million copies.

That’s a high number, usually 2 to 3 million copies sold is considered a hit.

Since launching in October 2023, Spiderman 2 has lifetime sales of 16 million copies. Very impressive!

But there are games that go belly-up too. For example, Concord cost at least $200m to make and sold fewer than 25,000 units. It was permanently deleted from PS Store and physical discs were recalled.

So selling games is not a very predictable business, which brings us to the existence of PS Store. This is the digital platform where Playstation distributes games digitally. Consumers pay to download games and can keep their games on their online account.

The economics of owning the distribution platform are attractive:

• Playstation takes a 30% commission from third party games.

• To play online multiplayer, you need to subscribe to PS Plus for $80 to $160 per year.

• Almost all of the top 10 games have multiplayer mode.

Separately, Playstation also acts as a game publisher.

Below is the estimated unit economics of physical versus digital sales for a publisher:

Of course, the first-party publisher (Sony itself) earns the most. But more importantly, digital sales brings the most money for all types of publishers.

So are consumers angry because Playstation makes more money?

No, that’s not the real reason.

We have to examine the business of competitor Steam — the fully digital gaming platform on PC. Steam has an estimated 200 million monthly active users (MAU), while the PlayStation reports 125 million MAU, making Steam's active user base 60% larger.

Scale Economics Shared

Did you know that the economics of PC gaming is inferior to consoles?

Despite the flexibility to play any game as long as it’s available, PCs are much more expensive than a console. The PS5 retailed at about US$600 and will probably last 7 years before PS6 arrives, plus 3 more years of cross-over where games are available on both old/new consoles. Therefore, the total lifespan is about 10 years.

To build a PC that can last 10 years while maintaining game quality, it would likely cost US$2,000. Also, for a console you don’t have to think about upgrading components.

However, since Valve launched Steam in 2003, the PC gaming market has outpaced consoles:

Just like Playstation, Steam operates as a game developer/publisher, game distributor, and also sells hardware called Steam Deck.

On hardware, Steam Deck a hand-held PC gaming device running on their SteamOS (operating system). It looks like Nintendo Switch:

Steam’s corporate parent, Valve, is also a successful game developer. It owns titles like Counter-Strike, DOTA, Team Fortress. These games are decades old and still remain very popular today.

However, neither of these 2 segments is the golden goose. Actually, Steam platform lays the golden eggs. Look at the revenues generated by Steam over the years:

These are very high margin sales since the incremental cost of distributing a game on a digital platform is near zero. But did you know that Steam paid out 76% of revenues to developers in 2025? In other words, they only took commissions of $4.8b (24% * $20b).

For developers, the upfront cost of game development is high, but after that every sale is just a digital copy with no incremental cost. That’s why volume is the driver of profits, and discounts attract more volume.

This is the clever business strategy of Steam. As the PC market grows, they get economies of scale and pass the savings back to consumers. Does that sound familiar?

Yes, it’s the same strategy as Costco and Amazon!

Nick Sleep popularized this strategy as “Scale Economies Shared”.

By giving back a large portion of revenues to developers, Steam can offer gamers deep discounts:

They even have sales event schedule published ahead of time. For example, the first half of 2026 sales calendar was published on August 2025:

We can appreciate why the business model is very strong:

1. Steam attracts players through discounts.

2. Developers win by having access to one of the largest distribution in the world.

3. Steam pass back the savings from economies of scale to developers.

4. Developers can offer discounts for more sales.

5. More players join the network.

The flywheel effect continues to reinforce itself!

It’s Trust

Steam owns an ecosystem where the supplier, customer, and distributor wins. As a result, Steam extracts something more valuable than profits… that is trust.

Trust is the most powerful quality that a business can hope to get from their customers. That’s why Costco has a 92% renewal rate on their membership — shoppers trust that Costco will always give the best value.

Now we finally know why players are angry with Playstation. They simply don’t trust that Playstation will pass the savings to them.

To know why Playstation doesn’t earn trust, just look at my own behaviour when buying Playstation games. Remember in the beginning I mentioned the cheap discs? Well, the digital versions of those games were likely sold at full price (at launch) on PS Store, even though the game has been out for years already.

Peter Drucker, often described as the inventor of modern management, said this:

The purpose of business is to create and keep a customer.

In this context, the whole saga showed that Steam has an edge over Playstation.

We came across this table showing the average ambulance fees across a few developed countries:

Notice that ambulance fees in the US are the highest and by a large margin.

Why is this when America is a very rich and modern nation?

The popular opinion is that ambulance services are run by corporations, profit-chasing firms, that are out to squeeze the most money out of people at their most helpless moments.

That’s actually not true. Ambulance providers in the US operate on low margins and operators throw in the towel every year. Labour is the largest expense plagued with high staff churn.

The problem lies in old rules from Medicare that didn’t change, even though the cost structure of ambulance services is vastly different today. To fully appreciate the problem, we have to explore the economics and history of emergency medical services.

Ambulance Economics

We start with the ambulance. It’s the first thing patients encounter during an emergency. It’s helpful to think about the economics of ambulance services as quite similar to an option.

Consider the area which you’re living in, it probably has a fire department or hospital where ambulances are on standby. When someone needs to be rescued, they call a number and summon an ambulance.

Although there’s no explicit contract, the ambulance provider is obligated to come, since they are essentially selling an option to citizens — the power to deploy an ambulance.

Even if you never called an ambulance, simply having the option is valuable, because knowing help would come during distress allows you to take risks that you wouldn’t otherwise.

Therefore, everyone’s utilizing this option everyday… without paying any premiums!

Of course, the ambulance provider cannot sell this “option” for free. They have to keep the hotline open, pay salaries, maintain infrastructure and vehicles. It must be prepared to respond quickly during emergencies.

The product they sell is not the hospital ride. It’s the readiness, which has to be maintained every second on the clock, regardless whether anyone calls or not.

Similar to an insurance policy, people who are utilizing this option should pay a small premium, so that when someone needs to be rescued they don’t get hit by a massive bill.

But the emergency medical system doesn’t work this way in America. This is where we need to understand how its cost structure has changed.

Funeral Cars

Before the invention of the automobile, the injured have to travel by horse and cart to the nearest doctor. The mortality rate was quite high, many died on the way, or just gave up before the horses came.

Fortunately, when the automobile came during the early 20ᵗʰ century, ambulances became a possibility. People had access to a fast moving vehicle dedicated to bringing people to the hospital. Fun fact: The word “ambulance” gained its modern meaning during the 1854 Crimean War. Not too far away from the first gasoline cars in 1886.

But ambulances have a unique requirement, they must be long and low. The reason is because it is better for the injured to lay down. The only vehicle that met this function was the funeral car. Yes, the same one that carried coffins to the cemetery.

In the early 1900s, funeral homes across America commonly provided ambulance services. They would use the same car to transport both living and dead bodies. They were also open 24/7 since anyone can die at anytime. It was not unusual that one funeral car made the same trip to the hospital and cemetery in the same day, perhaps even for the same person...

It’s important to understand that funeral businesses stepped up not out of obligation, but out of service. At a time when many communities lacked the infrastructure for organized emergency response, funeral businesses filled a critical void. They didn’t claim to be medical professionals, but they understood urgency and care.

The real carrot was for the relationship; the family who called for the ambulance was likely to also use the same funeral service when someone died.

Under these circumstances, the cost of selling the ambulance option was cheap because the service was plain and convenient. It also served as a loss-leader strategy for funeral businesses.

Modern Medicine

As medicine advanced in the 1960s, it became possible to help patients on their way to the hospital. This led to laws and standards that required trained staff and fully equipped ambulances. Particularly, in 1965, the US government created two programs of social insurance:

1. Medicare (public health insurance for Americans above 65)

2. Medicaid (same program for low-income)

Both programs reimbursed medical bills with ambulance rides included. At that time, ambulance trips were cheap, so not much thought was put in. But medical improvements would eventually change the cost structure of emergency care.

Earlier in 1947, the first portable radio-based ECG telemetry was invented, as semiconductor technology improved, an ambulance could transmit a patient’s vital signs to the hospital.

In 1960, CPR (cardiopulmonary resuscitation) was developed at Johns Hopkins. The life saving technique was endorsed by the American Heart Association in 1963. Shortly later, the portable defibrillator was invented in 1965 by Dr. Frank Pantridge. The initial instrument weighed over 70kg and ran on car batteries, but this innovation improved the chances of surviving a heart attack outside the hospital.

Given this background, the funeral coach started to look inadequately equipped. As medical services became better, it gave rise to emergency medical services (EMS). The world’s first modern paramedics were the Freedom House Ambulance Service, established in 1967 Pittsburgh, Pennsylvania. Freedom House revolutionized EMS by training their teams to administer CPR, clear airways, manage trauma, and transmit heart rhythms to doctors.

Even the media was involved, in 1972, the action-adventure medical drama Emergency! debuted on NBC and ran for 122 episodes over the span of 5 years.

Public awareness was indeed building up.

Financing came from the Emergency Medical Services Systems Act of 1973 which created a new grant program to further the development of regional EMS systems. More than $300m was appropriated for EMS feasibility studies, planning, operations, and research.

Then in the 1980s, fire departments began to get involved since there were fewer fires to fight thanks to better building codes and smoke detectors. Did you know that, in 2020, 64% of American fire departments runs were EMS-related, only 4% were for fires.

All these developments transformed EMS into a business with high fixed costs. You might think the remedy is “economies of scale”, maximize utilization to spread fixed cost over a large revenue base.

But this strategy doesn’t work for ambulances. As described earlier, the economics of ambulance services is like selling an option that they are obligated to honour, even if capacity is idle. The utilization rate of a well-run EMS is only 30% to 50%. If the utilization is higher, then there is a risk of missing emergency calls — a disaster nobody wants to bear.

There’s nothing wrong with changes in how the industry worked. The error was not changing the way people paid for this service.

When Medicare was created in 1965, it reimbursed on a per-ride basis. Commercial insurers all followed suit. This rule didn’t change even though the cost of EMS had ballooned. In other words, ambulance providers have to charge individuals for round-the-clock readiness.

This is the exact opposite of insurance: A large pool of people paying for coverage that is claimed by few. So what happened is that the entire cost had to be recovered from the unlucky minority.

To make matters worse, Medicare covers the elderly who are the most frequent ambulance users. Since Medicare is the largest purchasers of medical services and backed by the government, it has the power to set prices.

Medicare Part B Ambulance Fee Schedule (AFS) establishes standardized payment rates:

• Basic Life Support: $330 — $350

• Advanced Life Support, Level 1: $330 — $350

• Advanced Life Support, Level 2: $480 — $510

• Ground Mileage: $9 — $10 per mile

Compare this to the average ambulance transport cost of $2673, we can see why EMS providers lose money on every run. To cover this shortfall, the provider charges the balance to the patient, this practice is called “balance billing”.

Well, Medicare/aid strictly prohibits balance billing… then who else pays?

There’s 2 groups who are left on the hook:

1. Uninsured

2. Privately insured

The uninsured are unlikely to pay since they are generally low-income in the first place.

Now, the more realistic scapegoat are those people who have private insurance.

Even this is problematic.

In ordinary medical care, when an insurer wants to bring a hospital into its network, it steers patients to the hospital in exchange for discounted rates. But ambulance rides is not a medical procedure, the insurer cannot make policyholders go to a specific ambulance. An ambulance provider responses to whoever that calls.

So the insurer has nothing to offer the ambulance provider. The rational response is to then charge insurance companies the full amount. And so 80% of ground ambulance rides are billed to insurance companies as out-of-network charges.

It’s up to the insurer to determine a reasonable amount, then remainder goes to the patient like a surprise bill.

Wrong Focus

We hope you understand now that the underlying problem is a poorly designed payment structure trapped by outdated 1965 Medicare rules.

The solution would be to operate like an insurance product, charging premiums in exchange for EMS.

The market has tried to use scale to solve this problem. Even the largest private operator of ground ambulances (AMR) earn thin margins. You can look at its parent company GMR financial statements — laden with debt and very low cash flows.

This system leaves everyone unhappy.

• Insurers don’t want to pay for things they never negotiated for.

• EMS charging high fees but earning small profits.

• Patients fear the ambulance surprise bill.

Most of the political attention has focused on the problem of surprise billing. But no one has attempted to solve the actual underlying problem.

In 2015, for example, the state of New York banned surprise ambulance bills by requiring insurers to pay out-of-network providers close to their billed charges. This means paying whatever the ambulance providers charged. But it help patient because insurers passed the costs to consumers in terms of higher premiums!

Don’t even think about capping ambulance charges either… that would just kill the whole ambulance industry.

There are good examples of how countries have solved this problem:

• UK and Japan simply fund ambulance services out of taxes.

• Victoria state in Australia sells memberships called “Ambulance Victoria” with unlimited exercise at the cost of AUD110/year per family.

• Some rural counties fund ambulance services entirely out of property taxes. In Tulsa and Oklahoma City, the government buys staffed unit-hours from an ambulance operator, while households can prepay a few dollars a month on their utility bill and pay nothing for ambulance services.

Fair Isaac Corporation (FICO) was founded in 1956 by Bill Fair and Earl Isaac. Bill has a background in engineering while Earl is a mathematician. Both met at the Stanford Research Institute while they were working as researchers for the US military.

Bill and Earl saw an opportunity to put their research methods in the private sector, so they invested $400 and started a consulting business. Their goal was to apply maths and early computers to improve business decisions. Eventually, they settled on consumer lending and built its first credit scoring system in 1958.

Both founders sent out letters to 50 major US lenders in hopes of selling their statistical methods to measure credit risk. Only one lender reached out, it was American Investment Company (AIC), the 4ᵗʰ largest consumer finance company at that time.

AIC allowed FICO to analyze their credit data. The data was simple, for example:

• How long borrowers lived in their current residence?

• Do they have a telephone?

• Age and marital status?

Despite this rudimentary dataset, after FICO applied their methods, it resulted in AIC reducing bad debt charge-offs by 25% with only 3% reduction in total loan volumes.

The remarkable achievement was that FICO managed to apply statistical methods to consumer credit before the digitization of credit data. Imagine that in the 1950s, people had to physically transfer materials back and forth between FICO and the lender, with storage rooms of credit operations across the country.

In effect, FICO was turning stacks of credit data printed on paper, into easy-to-use physical cardboard printed table that mapped out the odds of default. They produced a numeric score which correlated with the probability of default. Each lender could then raise or lower its own cut-off score to reflect their risk appetite. FICO created a standardized approach in making credit decisions that was both consistent and scalable, it eliminated the subjective biases of loan officers.

Over the course of the 20ᵗʰ century, there was significant consolidation of the US credit bureaus which controlled consumer credit data. This consolidation accelerated following the introduction of the Fair Credit Reporting Act (FCRA) in 1970, which introduced a regulatory framework that favored the larger credit bureaus that could afford to make the necessary investments in infrastructure to be compliant.

It was in this environment that today’s three credit bureaus emerged: Equifax, TransUnion, Experian.

These bureaus provided exactly what FICO needed — big centralized credit data. Combined with mainframe computers, the industry was able to digitize credit data. These factors resulted in FICO releasing PreScore in 1984.

Then in 1989, FICO introduced the FICO Score that we know today. It was used as a general-purpose credit score to standardize and automate the assessment of credit risk.

Rather than manually processing consumer credit applications, the FICO Score is an automated credit underwriting process that was fairer, more consistent, and expanded people’s access to credit.

The FICO Score also emerged at the right time. In the late 1980s, America had a recession that led to high number of credit card delinquencies. The credit card companies (Visa, Mastercard, Discover) started to use FICO which saw credit scoring jump to more than 50% of card products in the early 1990s.

Adoption spread among lenders, and seeing the effectiveness of FICO, the government sponsored enterprises (GSE) Fannie and Freddie directed lenders to use FICO Score in 1995 for all new residential mortgage applications.

This cemented FICO Score as the industry standard for US residential mortgages. More broadly, FICO Score also become dominant in other verticals such as auto loans and credit cards.

In fact, the idea that FICO credit scoring monopoly is granted by the government is false. The vast majority, approximately 99%, of FICO Scores used for decisioning across the consumer credit industry are non-mortgage. Total mortgage originations, including GSE-purchased mortgages, constitutes only 1% of all FICO Scores.

For all 70 years of FICO as a company, it only had 6 CEO changes.

From 1956 to 1983, both co-founders led their company. Earl Isaac died in 1983 at a young age of 62, before the FICO Score was launched. Bill Fair took FICO public on 22 July 1987 at a valuation of $14.6m, he worked as CEO until 1991 when he retired, unfortunately he would shortly die from cancer at age 72.

Larry Rosenberger, a FICO veteran who joined back in 1974 quickly rose to the rank of Executive VP in 1985, eventually taking over the CEO role in 1991. For his short leadership tenure until 1999, FICO grew revenues near 9x from $32m to $276m. After that, he wanted to get deeply involved in FICO’s new analytic activities and transitioned into specialized research roles. In 2017, he was named FICO’s very first Research Fellow. After 52 years, he’s still working at FICO today!

Thomas Grudnowski was CEO from 1999 to 2006. He came from consulting firm Accenture, and naturally expanded FICO into business analytics. They launched myFICO, a consumer service portal to let people directly access their credit scores. FICO shifted into selling software solutions, it was after their major $810m acquisition of HNC Software that FICO separately reported their Software segment in 2003. The housing boom leading up to the GFC contributed to the rapid revenue growth from $270m to $830m.

Stepping down before the economy crashed, Grudnowski passed the torch to Mark Greene. During the GFC, mortgage originations froze and revenues fell from $784m (2007) to $631m (2009), net profits also fell by -38% in those years. Greene would leave in 2012 after being unable to revive the business post-GFC, in 2011 revenues were still below 2009 levels.

William Lansing took over as CEO in 2012, that year FICO grew revenues by +9% and EPS by +42% YOY. He would lead FICO through shifting into SaaS, and the largest price hikes in FICO’s history, resulting in net margins increasing from 12% (2011) to 33% (2025). After keeping royalty rates unchanged in real terms for nearly 30 years, FICO Score went from 50—60¢ in 2022 to $4.95 per score in 2025.

This is where we land today with regulators and credit bureaus seeking to break FICO’s monopoly. The stock price reflects this uncertainty dropping from a peak of ~$2300 to $1250 currently.

Business Model

FICO reports in 2 segments:

1. Scores (59% revenues)

2. Software (41% revenues)

Although Software takes up a significant portion of revenues, the real profit driver is Scores. It is about 79% of operating income:

Furthermore, Scores is growing much faster, and more importantly carries much higher incremental margins. In 2025, Scores had op margins of 88%, that’s a very lucrative business profile!

Most of the Scores business is B2B (88%), with the remainder B2C.

Also, FICO revenues comes mainly from the US (95+%).

Since Scores is the main profit driver, we will focus on this segment as there is already too much to write about it.

Mechanics of FICO Score

The above are the factors that go into producing a FICO Score.

Payment History (35%)

This is the most significant factor affecting your credit score. It includes a record of whether you have paid past credit accounts on time. Late payments, bankruptcies, foreclosures, and other delinquencies negatively impact this factor. Regular on-time payments positively influence your score, while missed or late payments can cause significant damage, especially if they are recent.

Amounts Owed (30%)

This is often referred to as “credit utilization”, measuring how much of the available credit you are using. It’s calculated for individual credit cards and other credit lines. If you use a small % of available credit, then it will positively impact this factor.

Length of Credit History (15%)

This measures age of your oldest/newest accounts. Generally, a longer credit history will improve your score, assuming you have a history of managing credit responsibly.

Credit Mix (10%)

This is the variety of credit types. For example, credit cards, mortgages, installment loans etc.

New Credit (10%)

This is the number of recently opened accounts and number of recent inquiries into your credit report. Too many new accounts and inquiries within a short time will suggest lower scores.

Combining these factors through FICO algorithm, it produces the FICO Score that we are familiar with:

A higher score indicates low risk of default.

There are over 50 variations of FICO Score for various markets, but the general idea is that FICO uses predictive algorithms to uncover patterns in historical data, which standardizes credit outcomes by a single score.

Business Economics

Consumer Reporting Agencies

FICO algorithms can only work if there is data. So who has credit data?

It’s the three Consumer Reporting Agencies (CRA):

1. Equifax

2. Experian

3. TransUnion

The economics of FICO can be thought of as a licensing business. FICO licenses out their algorithm to the CRAs, they run it on their credit data and then sell the report to lenders (eg. JPM) through third-party processors (eg. Advantage Plus).

When a lender wants to make a loan, they will request a “tri-merge” report (referring to the three CRAs).

Therefore, the CRAs act as distributors in-between FICO and the ultimate end user (lender) of the score. FICO doesn’t store or handle any of the sensitive credit data.

Industry Standard

One of the big ways that data becomes useful is when it’s tied to other data. You get better outcomes when you have more useful data joins. The reason is simple: data is only as useful as the questions it can help answer.

Joining, linking, and graphing datasets together allows you to answer more questions. So, once FICO managed to join all these millions of datasets together that unlocks value in the form of predictive loan underwriting — it creates an industry standard.

Such industry standards are very hard to breach, and this FICO’s main competitive advantage.

There are 2 reasons why standards have durable advantages:

1. It’s very hard to make something an industry standard. It often requires a unique set of circumstances to happen.

2. Once a market builds around a standard, it’s very hard to change.

In the case of FICO, it became the standard due to lender adoption. There are 2 main reasons why this happened:

• FICO Score was simply a superior product compared to the lender’s own models at that time. So it allowed lenders to underwrite better loans.

• Regulatory changes such as the Equal Credit Opportunity Act (1974) convinced regulators that an algorithmic credit score is non-discriminatory.

Compound this first mover advantage with how difficult it is to change to a new standard. Lenders will only be willing to adopt a new standard only after others change first.

Either a very large lender moves, or a group of lenders move. But this won’t happen in the US because the mortgage industry alone is fragmented:

There is also high switching cost because FICO Score is integrated into multiple systems, workflows, and regulatory requirements. It is also present in secondary market securitization. Below are some of the areas implicated by FICO:

• Loan Origination Systems
  This are the systems that lenders have integrated with resellers to pull tri-merge credit reports.

• Automated Underwriting Systems
  Fannie & Freddie have systems called Desktop Underwriter and Loan Product Advisor. They require FICO Score as an input. No score means no loan approval.

• Loan Level Pricing Adjustments (LLPA)
  There are LLPA grids that determine basis points pricing for loans. These also require FICO Score input.

• Servicing & Loan Performance

  Mortgage servicers monitor borrower credit health post-origination using FICO Scores to make updated assessments of delinquency risk. This feeds into mortgage servicing rights and loss mitigation waterfalls.

• Mortgage Backed Securities (MBS)
  When GSEs securitize loans, the FICO Score is referenced as a risk metric.

• Regulatory Models
  Stress testing and capital frameworks use aggregated loan-level FICO distributions as key inputs to loss models.

• Mortgage Insurance
  Insurance premiums are calculated with FICO Score as one of the inputs.

Clearly, we should appreciate the entrenched nature of FICO within the mortgage industry alone. These disparate industry players have little incentives to make the investments required to support a second credit score beyond FICO. Needless to say, changing the industry standard is an extremely unlikely event.

Risk of Loan Repurchases

A lender is required to repurchase loans from the GSEs (Fannie, Freddie) if there is a breach in representations. The largest drivers for loan repurchases are missing key documents and/or inaccurate income calculations.

A former Vice Chairman of VantageScore said this:

If a loan goes into default owned/guaranteed by Fannie or Freddie, they will pull the loan and look at the documentation, and if there’s any discrepancy in the documentation, then they pull it out of the securitization and put it back to the lender.

That’s a big deal from a liquidity standpoint. They’re big dollar transactions as opposed to a credit card. The lender is deathly afraid of buybacks.

[…] They need to have a model that can accurately predict default because they don’t want to be in a buyback position. That’s a huge issue for the mortgage lenders.

The lender must pay the GSE $100 par value when they buyback the loan. So if there were loan buybacks occuring in 2024 when rates were 7%, compared to 3% when these loans were written, then the financial losses for lenders can be huge.

MBS Market

FICO is the preferred score in the securitization market. It has a long and proven history as the measure of credit risk for pools of mortgages. Therefore, FICO is a very sticky risk-based pricing product used by MBS investors.

MBS investors care primarily about cash flow predictability. This is driven by two things:

1. Charge-offs incurred when a loan defaults.

2. Pre-payments when the borrower pays off the loan earlier than expected, creating reinvestment risk for the lender. Note that loan pre-payments typically accelerate in a declining interest rate environment, meaning that the MBS holder is not only forgoing the future interest payments they were expecting, but will also generate less interest income on new issuance.

FICO Score is basically the uniform credit standard which gives investors confidence in these MBS. Without it, the secondary markets will be less liquid and likely result in higher mortgage rates for borrowers.

However, with the Federal Housing Finance Agency (FHFA) pushing for greater competition by allowing GSEs to accept VantageScore, we have to assess this threat.

VantageScore (VS)

This leads us to VS introduced in 2006 by the 3 CRAs to challenge FICO’s monopoly. However, after 20 years, VS impact has been only modest.

Why so?

Most financial institutions using FICO are sophisticated companies too. They have their own complicated risk models to manage. Calling in another scoring model would require them to recalibrate their models, given that it’s impossible to accurately and reliably map FICO to VS. The fact that VS has not been stress-tested in real-life economic downturns is another reason not to adopt.

In fact, the industry is very slow to adopt even FICO’s own updated scores. FICO 10T is the new release, but look at how many sectors are still on the old scores:

William Lansing pointed this reluctance to change in MS TMT Conference 2025:

So going from FICO 8 to FICO 9, it took four years before 50% of the market was on FICO 9. And I think you’ll see the same kind of thing. It’ll be four years again before FICO 10 is half the market. So it’s a very slow adoption rate.

Back in October 2022, the FHFA announced that a dual-score mandate would be implemented, whereby the GSEs would be required to use both FICO and VS on conforming mortgages. The FHFA then changed their minds after years of consultation in July 2025, announcing that it was moving to lender’s choice, whereby lenders can choose to use either FICO or VS!

Dual-score is just double the cost and complexity across the whole value chain. Consider these reasons and it makes no economic sense to have more than one scoring system:

• FICO Score cost is so low relative to a mortgage closing cost.

• FICO Score cost is passed to consumers.

• Change cost is massive for the mortgage industry with no proven incremental benefits.

However, it would be complacent to ignore the increased competition from VS. Let’s see which score is actually more performant.

There is a third-party study by Urban Institute that analyzed the predictive strength of Classic FICO (a 20 year old score currently being used by GSEs) and VS4.0.

They examined loans between 2013—2023 in Fannie Mae Single-Family Loan Performance database. The study looked at 90+ days past due data and plotted the cumulative default rates. This is the graph showing almost no difference between Classic FICO and VS4.0 ability to delineate between high/low-risk borrowers:

Now, FICO also did a separate study comparing their new FICO 10T model against Classic FICO. They noted better predictive powers with 18% relative improvement in defaulters captured in the lowest decile:

If this is correct, the implication is that FICO 10T should outperform VS4.0.

There are some benefits to VS4.0 as it is the first scoring model to use time-series data. This means that lenders using VS4.0 are likely able to write more loans. VS4.0 claims to be capable of scoring an additional 33 million consumers:

However, the ability to score more consumers comes from VS4.0 requiring only 1 month of credit history, compared to 6 months by FICO. This sounds like a less conservative underwriting practice to us. Moreover, FICO 10T also uses time-series data and it claims to improve mortgage approval rates by 5% relative to Classic FICO without adding incremental risk.

Put together, we don’t think VS will steal much market share from FICO as both as almost the same.

FICO Price Hikes

Historically, FICO has kept its royalty rates flat for the last 30 years, only increasing at the rate of inflation. Despite delivering so much value to the mortgage industry, scores growth has entirely come from volume rather than price.

This changed in 2012 after FICO renegotiated its license agreements with each of the CRAs. They reached an agreement that FICO would be able to adjust the base royalty rates once a year. Following this, FICO made the first adjustment for mortgage scores in 2018, pricing them at 50—60¢ per score.

Then in 2022, FICO told the CRAs that it was going on tier-based royalties. The rate for highest volume tier is unchanged, but rates increased for lower-volume tiers. After this change, FICO was collecting $0.60 to $2.75 per score. Some smaller industry players voiced concerns over this and FICO eliminated the tier-based structure in 2024.

Instead, they went for $3.50 per score wholesale royalty rate for 2024. A year later, prices increased to $4.95.

Although $4.95 per score doesn’t sound like a lot of money, it was the pace of price hikes that triggered regulators.

In March 2025, Bill Pulte was appointed Director of the Federal Housing Finance Agency (FHFA). In this role, Pulte oversees Fannie Mae, Freddie Mac, and the Federal Home Loan Banks (FHLB).

Before his appointment, Pulte had a career in homebuilding. In 2011, he founded Pulte Capital Partners LLC, an investment firm focused on building and housing products. He also served on the board of Pulte Homes, one of America’s largest homebuilders.

The FHFA’s authority over the GSEs gives it a direct relationship with FICO, given that the GSEs rely on FICO Score. But FHFA doesn’t have authority over FICO pricing.

Pulte started to put a big target board on the back of FICO, saying that the price increases are detrimental to home ownership.

To put bluntly, we don’t trust what politicians say. What they say only colours the public’s perceptions of FICO, and if that causes mispricing of the business, then it’s a good thing for investors.

So let’s go down the rabbit hole and see who is actually contributing to higher credit report costs.

Cost Breakdown

We can model a two person tri-merge report for mortgage lending. About ⅔ of home purchases involve a married couple, both people require a credit pull (FICO gets paid twice).

Both individuals will usually be checked multiple times during the loan approval process:

1. Pre-qualification: This is a quick, initial assessment of the amount that the person is able to borrow for a home purchase. During this stage there will be a “soft credit pull”, based on unverified information provided to the lender (income, debt, estimated down payment). More serious buyers typically skip the pre-qualification stage and move to pre-approval, which is a more involved process that results in a conditional commitment from the lender.

2. Pre-approval: Lender will do a “hard credit pull”. It requires a tri-merge report, where the lender formally access credit reports and FICO Scores from all three CRAs (Equifax, Experian, TransUnion). This results in 6 credit pulls (3 CRA*2 individuals). FICO collects $29.70 worth of royalties ($4.95*6). The tri-merge report is to ensure that the borrowers qualify for a conforming mortgage (620 FICO Score is the minimum requirement). The lender will take the middle score out of the tri-merge (for multiple borrowers the lowest of these middle scores is used).

3. Loan processing: A credit pull will occur during the loan processing stage to ensure that the borrowers’ financial status has not shifted, particularly if there are delays in the underwriting or home search process, where more than 90—120 days have elapsed since the initial credit check.

4. Final check: A soft pull will be performed just before closing the mortgage as a final check. FICO collects $9.90 ($4.95*2).

Note: Fannie & Freddie don’t accept 120 days old tri-merge reports after the initial pull. A new tri-merge must be pulled, if the process takes more than 120 days. The median time to close a mortgage is 55 days as of May 2026.

From this example, FICO will collect at least $29.7+$9.9 = $39.6 (from #2, #4).

In reality, lenders average 10 credit pulls per mortgage application. Assuming the other 3 pulls are “soft”, then the total royalties to FICO is $39.6+$9.9*3 = $69.3. This amount is very tiny compared to average mortgage closing cost (taxes, property agent commissions, title insurance…) of $6k to $15k.

Ok, this fact is already well-known, but the criticism isn’t the dollar amount… it’s the % increase.

Did you know that CRAs and resellers mark up the tri-merge report along the way? Of course, they have to make money too!

To show that FICO fees are actually the lowest % of total tri-merge cost, we can model the following between 2026 and 2025:

• Only FICO increased prices from $4.95 to $10 per credit pull.

• Third-party resellers & CRAs kept prices constant.

As we can see even if we assumed only FICO raised prices, it is still the lowest % of total tri-merge cost (yellow cells):

In reality, we can be sure that CRAs & resellers definitely raised their prices too!

Now, this exercise leads to the question of whether VS pricing can benefit the consumer. The answer is yes, but the CRAs will unlikely do it:

Using 2025 pricing, we can estimate that VS price per pull is $9.24 (+87% higher than FICO $4.95) for CRAs to earn the same revenues. Indeed, the tri-merge report will be 35% less costly.

However, if this happened the CRAs would risk triggering regulators by pricing at such a significant premium to FICO. In the current situation, FICO is taking most of the blame and the CRAs get to “quietly” raise prices and benefit themselves without the spotlight.

FICO Angered the CRAs

You must be quite confused now by all these pricing games they are playing. Well, to complicate further, we have the new FICO 10T which attempts to bypass CRAs markups by going direct-to-resellers. The CRAs still charge their “credit data fees” but lose selling the tri-merge report to resellers.

FICO 10T charges $0.99/pull upfront + $65 funding fee per borrower (only when the loan is successful).

The CRAs offer VS4.0 at $0.99/pull without funding fee. Very much cheaper than FICO.

This situation is in favor of VS4.0:

1. Lenders are always trying to find incremental pockets of consumers that they can underwrite loans — which VS4.0 enables.

2. The US is in a mortgage bear market, so lenders are struggling to maintain profit margins — VS4.0 is cheaper, every dollar matters.

So before FICO started to go direct-to-resellers, the entire ecosystem was one, big happy family. Each entity down the value chain would add on their markups and sell the credit report to consumers.

By launching direct-to-resellers, FICO basically angered the CRAs and turned them into competitors. Be reminded that FICO needs credit data which is sitting with the CRAs!

Why Lenders Complain

It’s not only the CRAs that are unhappy with FICO. Lenders, especially independent mortgage banks (IMBs), are also unhappy.

Don’t these lenders simply pass the cost down to consumers?

Not really, when you look at smaller lenders like IMBs. These are small businesses operating in local communities which use their own cash and short-term borrowings to fund their mortgage loan book.

IMBs play an important role in mortgage lending. They originate 83% of all single-family mortgages and 75% of GSE loans. But these lenders have thin margins, during the rate hike of 2022, IMBs as a group had 8 consecutive quarters of losses.

So what happens when a mortgage deal is unsuccessful, the IMBs have to bear the cost of multiple credit pulls. With such thin margins, the price hikes from FICO, CRAs, and resellers quickly stack up.

Ending Notes

To put everything together, the situation for FICO doesn’t look good.

1. High interest rates causing weak mortgage market. This has persisted for the last 4 years and we have not seen signs of recovery.

2. The credit bureaus with VS4.0 are going on a price war with FICO. How to think about valuations when the market goes from monopoly to duopoly?

3. Economics of FICO are still durable, but VS adoption in non-mortgage verticals is very rapid. For example, VS utilization in credit card sector is up +142% to 24.4 billion pulls in 2024. Also, they reported dramatic increases in telecoms, utilities and rental screening organizations.

4. FICO believes that they won’t lose market share for mortgage lending because the ecosystem is more complex. In this complexity, they think that it’s possible to get away with price hikes. However, we think that they pulled the pricing lever (from $0.6 to $10) too hard.

5. The two credit scores are not that different. Everyone on this value chain is trying to monetize on this transition between FICO and VS. In our mind, the brand of the credit score doesn’t matter.

Let's go back to somewhere around 2014. This was a period when smartphone usage proliferated, it was estimated that half of the world’s used a smartphone, and the trend was not stopping.

Search has been Google’s main product and the smartphone posed a major threat. Typing queries into a text box and clicking on links is a very desktop-suited experience, but on smartphones native apps offered a superior experience.

During that time, the argument was that if the mobile apps experience made more sense for users, then advertisers would follow, making mobile app advertising a disruptor to Google Search.

This wasn’t a far fetched idea. The tech industry saw it with Microsoft's failure to capture the operating system in smartphones.

So the threat was real, but how did Google evade and survive this challenge?

Knowledge Panels

Google came up with this thing called Knowledge Panels. See these small areas of information embedded within the search result? They effectively act like mini-apps:

This circumvented the need for users to download different apps. We had a chat with someone who worked on this project in 2014, and it was a fascinating lesson on how the correct incentives produce best results.

During that time, the VP of Engineering set ambitious targets on how much traffic Knowledge Panels should take from traditional Search.

One engineer asked if that would be detrimental to Google’s revenues as a whole.

Obviously, every high level executive had stock options whose valuation depended on how much money Google made. So shouldn’t they be worried about taking traffic away from Search?

Instead, the VP answer was literally “dont worry about the money”.

That is some powerful incentives alignment. His incentives was not tied to the sales performance of Google, rather it was based on the impact and effectiveness of Knowledge Panels.

In essence, the VP's job was to kill the blue-links in Search.

Does this ring a bell?

Yes, it's the same idea when Jeff Bezos asked Steve Kessel to kill the physical book business and invent a digital one in 2004. And thus Kindle was born.

There were also other initiatives like Voice Search, Web Answers, and PageSpeed. All these kept Google Search relevant in the age of smartphones.

Sundar Pichai took over as CEO from co-founder Larry Page in 2015. Before his promotion, Pichai's area of expertise was Search and his business acumen proved to be brilliant.

The monetization rate on smartphones is very low at 0.02%, compared to 1.8% on desktop. But the insight was that there are so many more queries users make through a device that’s always in their pocket.

Only 43.1% of mobile searches led to users leaving to an external website either through an organic or paid links. The remaining 56.9% of searches were free; Google works hard to satisfy our needs without making a cent!

This is analogous to how retailers like Costco use $1.50 hotdogs to keep customers loyal. It's a loss-leader strategy.

Google was correct that user loyalty results in long-term monetization that outweigh short-term pain, and so they were able to escape the Innovators’ Dilemma by disrupting themselves.

Round #2: AI

The interesting discussion today is whether Google is facing the same problem again — this time in the form of AI?

Well, Google is once again transforming Search. Today it’s called Generative Search Experience (GSE). We see features like:

• AI Overviews

• AI Mode

• Google Lens

• Circle to Search

All of these aim to keep Search relevant in the age of AI.

Now, we don’t know if the incentives are being set like in 2014 “don’t worry about the money”. But we do get insights from Elizabeth Reid (VP, Search) from her interview with Financial Times:

There are a lot of opportunities for ads. We show them both above and below in AI Overviews, but also within. Ads are relevant whenever users are going to make a choice that has some commercial aspect.

When a query is predominantly commercial intent, like we think you want to buy something, then we might often show ads. But sometimes we think you probably don’t want to see ads, and so we don’t want to give everyone ads.

But some people might want to buy something. If you search “how to clean a stain out of the couch” and the first thing we show is a bunch of ads, you’re like, “Whoa, I just wanted some advice.”

But if we’re giving you ideas and then we say, “if you’re having trouble you might want to consider a stain-remover product”, and then we give you some ads for stain-remover products, it feels natural and in context. And so, there are new opportunities.

Obviously, AI search cost more for Google but it seems like they have no issues subsidizing. Whether this works or not remains to be seen, but we are optimistic about how Google is willing to transform itself to keep customers loyal on Search.

For all we know, the $1.50 hotdog is coming out again!

In the midst of the dot-com bubble, Warren Buffett sat down with his friend Bill Gates on a stage in the University of Washington. The year was 1998, Microsoft had just destroyed Netscape in a high-tech browser war. Yet, the audience was interested in hearing the opinions of a 67 year old Buffett.

No surprises, he stuck to the things that he thought wouldn’t change. Explaining his contrarian views in the face of high flying internet stocks:

I look for businesses where I think I can see what they’re going to look like in 10, 15 or 20 years […] I don’t think the internet is going to affect how people chew gum.

[…] I don’t think the internet is going to affect the fact that Coke will be the drink of preference and will gain in per capita consumption around the world.

I don’t think it’ll change whether people shave, or how they shave, or the fact that they like a better shaving system, as the Mach 3 is going to be.

As we all know, the Mach 3 is the shaving blade brand of Gillette. It was a wildly successful product. The business locked-in customers by selling the shaver first, then selling the blades as they worn out.

It’s like selling a subscription service, the business strategy was pioneered by founder King C. Gillette. It adopted the name “razor-and-blade”.

Buffett was a big fan of Gillette. He famously explained that he could sleep well at night owning Gillette, knowing that “hair is growing on the faces of billions of men on Earth”.

Back in his 1996 shareholders letter, he called Gillette as one of “The Inevitables” alongside the great Coca-Cola.

No sensible observer — not even these companies' most vigorous competitors, assuming they are assessing the matter honestly — questions that Coke and Gillette will dominate their fields worldwide for an investment lifetime.

Berkshire shareholders letter (1996)

Quite the bold claim!

At the end of 1998, Buffett’s Berkshire Hathaway owned $4.6b worth of Gillette stock, it was the fifth largest at 12.5% of the equities portfolio.

The strategy of ignoring internet stocks turned out smart in hindsight. The old man seemed to have proved his critics wrong; Buffett wasn’t losing his touch.

As for Gillette, they launched Mach 3 in 1998 by simply adding a third blade. Their customers kept on shaving, sales for razor blades grew almost 50% for that year, quickly recouping the $750m of R&D expenses.

However, Buffett wasn’t complete correct in his assessment. It's true, people didn’t change the way they shaved — but the business of Gillette itself did change as a result of the internet.

And the changes were detrimental! In business, things rarely flip from night to day. More often, companies lose competitiveness until a point where they suddenly break down.

It would take another decade for the pieces to come together.

Dollar Shave Club

The market continued to roar higher after that 1998 lecture. In Berkshire’s 2000 annual meeting, there was a shareholder who said his internet stocks made up for more than the losses on his Berkshire stock. He asked Buffett and Munger if they would “please put a small sum to speculate on internet stocks”. That year, Berkshire lost -19.9% while the S&P500 gained +21%!

Then it all came crashing down in 2000. Another shareholder went up during Berkshire’s 2002 annual meeting to thank Buffett for not listening to that man 2 years back.

So Buffett continued to expound on the virtues of Gillette:

Gillette now has 71%, by value, of the blade and razor business in the world. Just think of that. I mean, 71%.

Here’s a product that everybody knows what it does, they know where it’s sold, they know that it’s a high-margin business. I mean, it isn’t like the capitalist world is unaware of the money that could be made if they could knock off Gillette.

But they can’t knock off Gillette. […] And I would say that 5 or 10 years from now I would be amazed if Gillette has lost market share.

He turned out to be half-correct, Gillette remained dominant until 2010 when it lost market share for the next 6 years, from 70% (2010) to 54% (2016).

What happened?

Well, an internet company distributing razors came up. It was named DollarShaveClub.com.

By the way, we have to commend the super cheap razors promo during America’s 250ᵗʰ birthday:

Founded by Mark Levine and Michael Dubin, the pair met at a party and spoke of their frustrations with the cost of razor blades. So they launched DollarShaveClub in January 2011.

The value proposition was great:

For a dollar per month, we will send high quality razors to your door.

Yeah! A dollar!

This was the silver bullet to kill Gillette’s business. A cheaper, good enough product.

But why would it take so many years for someone to breach this moat?

To answer that question, we must first examine Gillette’s competitive advantage. In 2005, P&G bought Gillette for $57b and leveraged enormous resources that led to repeated success:

1. Spend on R&D for new products.

2. Justify price increases from such new products.

3. Spend on advertising to take mind-share.

4. Spend on retail distribution.

This formula worked well in the pre-internet economy, allowing Gillette to keep market share even when they sold blade cartridges at $5 per pack.

For example, in 2006 Gillette launched Fusion with five blades, two more blades from Mach 3. They charged 40% more!

The market? Listen, we make the market. All we have to do is put her out there with a little jingle. It’s as easy as, “Hey, shaving with anything less than five blades is like scraping your beard off with a dull hatchet.”

The Fusion product flopped. Sales were slower than expected because 3 blades is good enough for many people.

In the meantime, there were two advertising businesses that were blooming off the back of the internet.

They were… Google and Facebook.

Both enabled global content distribution at a fraction of the cost of traditional advertising media.

These 2 factors combined was what killed Gillette:

1. People were satisfied with a good enough product.

2. Content distribution became cheap through Google and Facebook.

This enabled DollarShaveClub to compete at a lower cost through their direct-to-consumer model.

Innovator’s Dilemma

So how did the incumbent P&G react to these problems?

Former Gillette executive, Mike Norton, said this:

At Gillette’s headquarters in Boston, the reaction was a collective yawn […]

Our side-by-side tests would prove to you that you could take any Gillette razor, and it would be better than DollarShaveClub.

Billion-Dollar Brand Club (Lawrence Ingrassia, 2018)

Yes, the Innovator’s Dilemma in full display. Did you know even Buffett wasn’t immune to this psychological pitfall — he rejected the usefulness of Progressive’s telematics in pricing auto insurance risk, causing GEICO to lag behind for many years.

Gillette indeed had a product that was superior in quality, but they were expensive and people had to go to a store to buy one. DollarShaveClub’s lower prices and higher convenience was an attractive offer.

DollarShaveClub’s advertisements were bluntly good:

In 2014, Gillette introduced a copycat called Our Shave Club. But it required customers to subscribe through a preferred retailer. Why?

Again it’s because Gillette’s business had always been selling through physical retailers. If they went direct to consumers, Walmart is going to remove them from their stores. For a giant conglomerate like P&G, the risk was too much to take.

Gillette even filed a patent infringement lawsuit in 2015. The dispute spanned through various claim rulings in 2017, but was eventually settled for a confidential amount.

Lessons

Finally, in 2017, Gillette capitulated and cut prices by an average of 12%. Gillette had lost its pricing power from the internet’s disruption.

DollarShaveClub was eventually acquired by Unilever in 2016 for $1b. Dubin remains the CEO, while Levine went into design and real estate. They went from zero to $1b in just 6 years!

For Buffett, he swapped Berkshire’s stake in Gillette for P&G stock as part of that $57b merger (0.975 P&G shares for 1 Gillette share). He then held onto those P&G shares from 2005 to 2014, when he traded them back to P&G for their Duracell battery business.

The Gillette investment went from $600m in 1989 to $4.3b in 2005. Excluding dividends the price CAGR was +13% over 16 years — a good result despite the disruption.

Blind spots in investing are everywhere. Some are obvious, while others subtle. This case study of how the internet eventually disrupted Gillette is one of the subtler ones. Even Warren Buffett, a business analyst genius, couldn’t anticipate this disruption.

Fortunately, he bought those Gillette convertible preferred shares with enough margin of safety, paying only 13x forward earnings and getting 8.75% dividends. For context, P&G purchased Gillette for 25x earnings.

Today when you go to a supermarket, you pick up a basket and load up the goods you want to buy and pay at the cashier. This self-service grocery experience is nothing new in modern times, but did you know that this concept actually has a patent dated back to 1920?

It was submitted for patent by Clarence Saunders, the founder of Piggly Wiggly stores. This is his patented self-serving store layout:

Here’s the link for the full 5-pages: Patent #1357521.

Before this patent, shoppers handed a grocery list to a clerk behind a counter, who fetched the goods. Saunders’ patented arrangement drastically cut store labour costs and lowered food prices for consumers. More interestingly, it encouraged the psychology of “impulse buying” by putting products directly into the hands of the shopper. It also gave rise to the necessity for modern products packaging and brand advertising that we see today, since items now have to “sell themselves” on the shelves.

The picture below shows how Piggly Wiggly looked like in 1918. See those baskets? They even charged shoppers for using them:

But, as the title of this post suggests, the history we will be talking about is much more fascinating.

Piggly Wiggly wasn’t just an innovative grocery chain, it had some greedy financial engineering. This is the story of how trying to corner a market ended up badly for everyone.

Clarence Saunders

In 1881, Clarence Saunders was born into a poor family in Virginia. Unable to afford proper education, he went through menial jobs until 1916, at the age of 35, he set up America’s first self-service grocery store in Memphis, Tennessee.

When asked why did he call it Piggly Wiggly… Saunders would often answer:

So people like you can ask me that question.

According to Tennessee Historical Society, a journal written by Mike Freeman (vol. 51 #3, 1992), Saunders was boasting after just a few months of opening his first store:

One day Memphis shall be proud of Piggly Wiggly… And it shall be said by all men… That the Piggly Wigglies shall multiply and replenish the earth with more and cleaner things to eat.

Piggly Wiggly didn’t quite conquer the earth, but by 1922 it owned or franchised more than 1,200 stores throughout the US, each painted in brown, blue, and yellow.

It also started expanding into Canada.

Because of his innovative idea of self-service shopping, the stores were very low cost and had an average annual sales per location of $120,000. This was much higher than most grocery stores at that time.

The business model was a mix of wholly owned and franchised stores. The headquarters, Piggly Wiggly Stores, owned about 650 stores, while the rest were franchised and paid 0.5% royalties on gross sales to a separate holding company, Piggly Wiggly Corp, which in turn owned trademarks and the Patent #1357521.

As the business expanded, Saunders became rich and listed Piggly Wiggly Stores on the NYSE in February 1922. It went public with 150,000 shares outstanding issued at $43/share. The shares paid a dividend of $4/share each year.

Everything was going smoothly. Profits were $1.1m in 1922, a huge increase from prior year’s $712k.

Short Sellers

However, bad news would soon arrive in November 1922, a Piggly Wiggly franchisee operating in New York, New Jersey, and Connecticut went bankrupt. 35 stores were affected, but this was a franchisee and not the holding company itself.

This bad news coincided with an announcement that Piggly Wiggly would issue 50,000 more shares — not a particularly good combination.

During this time, Wall Street was booming and many short sellers (called “bears”) were betting that the market was overvalued. Securities regulations were weak back then, so these bears would borrow shares to sell them, hoping to buy them back later at lower prices.

Piggly Wiggly became a target board for bears and they drove the stock price from $50 to below $40.

Short Squeeze

Clarence Saunders was livid. He would recount that he felt like Wall Street was swooping down on Piggly Wiggly “like an eagle on a chicken”.

So he decided to hit back by announcing that he was going to buy Piggly Wiggly’s stock. But he carried this out to an extreme.

After using all the money at his company, Saunders went to borrow $10m from investors and banks to fund his shares purchases. In a week, he had bought 105,000 shares and the price quickly rose to $60 in January 1923.

A man on a mission, Saunders was bent on beating the bears into submission. He tried to force short sellers to buy back their borrowed shares at a much higher price. This is what we term today as a “short squeeze”.

But, Saunders had to also unwind his positions to repay the money he borrowed. Since he held so many shares, the act of him selling could crater the stock price, this is assuming if he even had any buyers at all.

Knowing this, the bears were encouraged to simply come back later.

Saunders had a brilliant plan. He sold his shares directly to the public by advertising in the newspapers and offering it at a discount:

He went further to try to shrink the public shares float because a smaller volume of shares would make the bears think twice. Saunders offered shares that he acquired to small investors on an installment plan, each priced at $55/share. $25 was paid upfront, the rest paid in ten $3 installments.

About 57,000 shares were sold this way, allowing Saunders to repay some of this debts without putting shares back on the open market.

By March 1923, Saunders had controlled virtually the entire public float of Piggly Wiggly.

The set up was done and on 20 March 1923, Saunders decided that his short squeeze plan was ready. That morning, he called for the return of 42,000 shares that the group of bears owed. They had 24 hours to comply.

The bears scrambled out of their caves and bidded Piggly Wiggly’s shares from $75 to $124 in the first few hours of trading.

The Piggly’s have slaughtered the Bears!

Or did they?

NYSE saw the nefarious plan to corner the market, and indefinitely suspended all trading of Piggly’s shares. On 22 March 1923, they delisted Piggly Wiggly. They even extended the time for delivery of the shorted shares until the following week!

Behold the power of regulators!

The stock fell back down to $82, and Clarence Saunders bemoaned:

Wall Street got licked and called its “mama”. Of course, “mama” heard the cry of her petted child.

Saunders was left with $5m in debt and 100,000 shares of his own company that he couldn’t sell.

There was still some glimmer of hope though, as there were still about 11,200 shares that needed to be repurchased by the short sellers. To poke the bears one last time, Saunders declared that he would only sell his shares at $250 a piece.

Remember those small investors he sold to at $55/share on installments? Because the shares were delisted, those people saw one last opportunity to profit from this debacle. They undercut Saunders’ price.

Betrayed by his own plan, Saunders relented and offered his shares at just $100 each. The small investors followed suit, offering an even lower price, enabling the bears to close their trades.

Yes, the bears definitely lost money. But Saunders was in deeper trouble. In a few months time, on 1 September 1923, he would need to repay half of his $5m debt, the other half was due 1 January 1924.

Desperate to save himself from bankruptcy, he offered his shares to the same group of small investors at the same $55. This time, nobody was interested.

To save their own kind, local businessmen in Memphis launched advertisements to find investors to bail out Piggly Wiggly’s founder. Alas, this was not enough to save Saunders.

The last resort led to Piggly Wiggly selling off assets… 17 stores in San Antonio were sold to franchisees for $300k in May. Another 97 in Chicago sold for $1m. That’s what you call a fire-sale, considering that annual sales was more than $100k per store.

In August 1923, with payments due next month, Saunders resigned as president of Piggly Wiggly and relinquished all his property — his stock, cars, even his house dubbed “Pink Palace” — to his creditors. Some of his shares were subsequently auctioned for $1. A year later, he filed for bankruptcy.

Anyways, his residence Pink Palace was built in 1923 using pink Georgian marble. Today, it’s a museum featuring exhibits ranging from archeology to chemistry, as well as a large theatre and planetarium. Of course, one exhibit features a replica of the original Piggly Wiggly store.

Second Wind

Clarence Saunders was bankrupt at age of 43, the bankruptcy judge reminded him that he owned absolutely nothing but his own name.

Now, this was insult to injury. Indeed, Saunders was determined to bounce back and launched a new grocery chain in 1928 called: Clarence Saunders, Sole Owner of My Name. A sarcastic punt to that judge.

Although eccentric, Saunders was also a grocery retail genius. He added meat departments and bakeries and his new shops became an instant hit, expanding to hundreds of locations.

Oh yes, Saunders a millionaire once again!

By 1929, he had 675 stores generating $60m in sales.

He also established a professional football team in Memphis, with an equally long name: Clarence Saunders Sole Owner of My Name Tigers.

Remarkably, the Tigers were a success, attracting up to 8,000 paying fans for a game against the NFL champion Green Bay Packers in December 1929. The Tigers, against all odds, defeated the Packers that day 20-6.

Life has its way around Saunders as his second fortune was short-lived. The Great Depression crushed the economy, and his Sole Owner chain went bankrupt.

Never Say Die

Oh no, Clarence Saunders wouldn’t simply lay down. Afterall, that’s not characteristic of American entrepreneurial spirit!

He wasn’t done with innovations in the grocery business. Learning from his past, he pivoted away from running retail stores and instead focused on building automated stores.

In 1937, he came up with Keedoozle. The name came from the phrase “key does all”.

Customers shopping in the Keedoozle would insert a key into the bottom of a glass display cabinet to select the item they wished to purchase. The item would then be deposited onto a conveyor belt by a stock clerk and the customer would be given a punched ticket tape. By the time the shopper got to the check-out, the selected items would be boxed, and the ticket fed into a machine would automatically tally the price.

This was actually the world’s first fully automated grocery store. In essence, it was like a giant vending machine.

Unfortunately, this idea would also fail… this time due to WW2.

After the war, at the end of his life, Saunders went for one final push. He wanted to make a store called Foodelectric.

This was described in a book titled TNT: The Power Within You.

It operates so automatically that the customer can collect her groceries herself, wrap them and act as her own cashier. It eliminates the checkout crush, cuts overhead expenses and enables a small staff to handle a tremendous volume.

Alas, death would claim him before his ideas took off. In 1953, Saunders died of a heart attack at 72 years old.

Piggly Wiggly is still operational today as small local supermarkets. Franchisees operate the vast majority of the 500+ stores across rural America. The logo has a cute pig mascot first introduced in 1949:

Two Great Lessons

The story of Piggly Wiggly brings us two important lessons for investors:

1. Focus on the business; NOT the stock price.

   Saunders was a good grocery business operator. He had a good business and a long runway of growth. However, he lost focus and turned a temporary fall in his stock price into a grudge against short sellers. He further stoked this fire with debt, blowing up a perfectly fine business along the way.

2. Never risk what you have for what you don’t need.

   Saunders was already a rich man until he decided that he should be rewarded for punishing the bears. Even though it was smart, legitimate financial engineering — he didn’t need this extra money. In the end, he went bankrupt because he took unnecessary risks.

On 18 September 1873, the first American investment bank, Jay Cooke & Company, went bankrupt and sparked The Panic of 1873.

The founder was a man named Jay Cooke, at that time he had the most trusted name in finance. He was even called “the financier of American freedom” after his involvment in civil war bonds. Yet, the crisis left his company bankrupt, and dragged the entire economy into a depression that lasted 6 years.

Civil War

To know the full story, we have to go back a decade before the panic happened. In 1861, the Civil War had just started and President Abraham Lincoln led the Union faction that fought against the Confederacy.

After the war began in April 1861, the Treasury faced a massive challenge funding the war efforts. Congress first imposed the first federal income tax in US history, creating a flat tax of 3% on incomes above $800, as the average urban worker made $600/year, this tax burden was mostly borne by the rich.

As military spending soared, the Congress passed more legislations like the Legal Tender Act, which printed $150m of fiat money. That’s the origin of “greenbacks” — the first banknotes issued by the federal government since the end of the American Revolution.

A new tax was also imposed on a long list of items: Liquor, beer, tobacco, cigars, lard, paper, soap, salt, flour, railroad passengers, ferries, trolleys, advertisements and carriages…

They also layered more taxes for businesses at 5% on earnings over $10,000.

One year later, tariffs were also raised to 37% in July 1862. Britian was the largest trading partner, and out of retaliation, the Brits funded a system of blockade runners to supply the Confederacy with food and munitions.

Despite these new measures, funding the war proved to be very difficult. The government continued to borrow and grew the national debt from $65m in 1860 to $2b in 1866.

It was estimated that the war cost $3m per day!

Nobody was interested in lending the government money. Considering that the Union did poorly at the beginning of the war, those wealthy elites who lent money had every reason to believe that their bonds were going to zero.

Enter Jay Cooke

At an age of 40, just months before the Civil War broke out, Cooke opened his private banking house Jay Cooke & Company in Philadelphia. At the start of the war, Cooke worked with the Treasury to secure loans from leading bankers in the Northern cities. He figured that regular Americans actually believed in this country. It was possible to sell war bonds to commoner. He started printing pamphlets, sending speakers, and deployed 2,500 salesmen across the North.

The most prominent was the 1865 pamphlet: How our National Debt can be Paid.

His message was clever:

Your duty to your country IS your duty to your wealth.

Buy these bonds, you’re supporting the troops AND earning a return.

Patriotic AND Profitable.

Patriotism worked extremely well. Cooke sold hundreds of millions of dollars in government bonds and earned his commissions. Ordinary citizens got their first taste of participating in large scale financial markets — you could say that Jay Cooke started retail investing.

By 1865, Jay Cooke was one of the most trusted names in finance. When Cooke spoke about money matters, people believed. However, his reputation would eventually become his greatest liability.

Summer of 1865, the Civil War ended with the Union as the victor, the Confederacy was dissolved and slavery was abolished, freeing 4 million African Americans. Cooke personally profited from war bond sales of between $700,000 and $1,000,000.

Capitalizing on his success, he started to construct a Victorian holiday mansion on Gibraltar Island in Lake Erie called Cooke Castle. Today, Ohio State University uses it as one of America’s oldest freshwater field research stations, the estate is categorized as a National Historic Landmark.

Halo Effect

Jay Cooke was not merely wealthy. He was famous and trusted by the public. Here is where the story becomes fascinating from a psychological standpoint — he began to believe his own legend. Afterall, he had solved an impossible problem during the war. He had moved markets through sheer persuasion.

What else could he not do?

So, by 1869, the government bond business had wound down and opportunities for profit appeared elsewhere. Cooke set his sights on building a railway from Lake Superior to Puget Sound, straight through the heart of America’s wilderness.

The government was handing out 47 million acres of land grants. Indeed if someone could build the railways, then he could profit from higher land values. Cooke saw this and envisioned cities thriving along the rails, agriculture sprouting across the lands, and commerce flowing from coast to coast.

He was right. All of that did happen… but not on his watch.

So that year he agreed to be the banker and agent for the Northern Pacific Railway Company. He believed this was another impossible problem that only he could solve.

To finance this project, he sold bonds the same way he did during the war. Leveraging on his reputation that anything with Jay Cooke’s name on it had to be safe… he deployed old methods of aggressive marketing appealing to patriotism and progress.

Here comes the concept called the halo effect. When we admire someone for one quality, we tend to assume they possess other positive qualities as well. A banker who financed a war must surely know which railroad investment is good, right?

The public trusted Cooke because of the war bonds.

Cooke trusted himself for the exact same reason.

But selling war bonds and financing speculative railroads through empty wilderness are fundamentally different activities. War bonds had the backing of the government, while railroad bonds had the backing of… optimism.

Despite Cooke's stellar reputation and an army of regional and local agents, the bonds proved very difficult to sell. Because of problems with climate and terrain, the railroad's projected cost per mile was much greater than its competitors. The bonds of other western railroads were selling at higher yields with less risk because they had been completed and were in operation. The Northern Pacific was sarcastically dubbed “Jay Cooke's Banana Belt” referring to the shape of the route on Cooke's map, and to misleading assertions that the climate along the line was temperate.

By the early 1870s, with bond sales faltering, the railroad was burning cash faster than Cooke could raise it. He repeatedly loaned money to the railroad to keep it afloat and maintain construction.

The Insiders Fled

In 1872, Cooke’s own partner Harris Fahnestock who was running the New York office blunted warned him:

…the present condition of the Northern Pacific, if it were understood by the public, would be fatal to the negotiation of its securities.

Cooke didn’t wake up to reality. The insiders anticipating the failure quietly fled, leaving depositors, ordinary investors and civil war veterans in the dark.

Eventually, one year later, on 18 September 1873, Jay Cooke & Company went belly-up, triggering a financial meltdown dubbed “The Panic of 1873”.

The NYSE shut down for 10 days.

Within a month many companies collapsed, including 89 out of 364 railroad companies and around 18,000 other businesses failed in the ensuing years.

Nationwide unemployment peaked at 8.25% in 1878. New York City had 25%. Former soldiers of the war lost their savings and ended up on the streets.

While the Northern Pacific railroad survived the first years of the Panic, it was also declared bankrupt in 1875, and construction was halted.

Remember Cooke Castle? He was forced to sell the island in 1874, but recovered some of his fortune and repurchased the property in 1880.

Systemic Risk

Why did one bad bet, by one banker, bring the entire economy down?

Although the term “systemic risk” was only coined in 1980 by William Cline following the Latin America debt crisis, this 1873 Panic is one of many examples in financial history that went down due to systemic risk.

• Other banks lent money to Cooke.

• Other companies bought his railroad bonds.

• These bonds were used as collateral for other borrowings.

• Businesses depended on the credit that flowed through Cooke.

Each bank that couldn’t collect from Cooke found itself unable to pay its own creditors. Systemic risk brings about interconnected balance sheets; one failure leads to another.

Jay Cooke Recovers

In 1875, two prospectors struck an extraordinarily rich vein of silver ore in a remote western region that would soon become the town of Frisco, Utah. While the mine possessed immense raw value, it lacked the capital, equipment, and logistical infrastructure required to transport heavy ore out of the desert to be refined.

Despite being bankrupt, Cooke retained his unique persuasive charm, financial acumen, and industrial connections. The owners of the struggling mine offered Cooke a 50% ownership stake in the venture under one condition: Secure the construction of a railroad to connect the isolated mine to the outside world.

Cooke tapped on his influence to gain financial backing from wealthy Eastern capitalists and prominent Mormon religious leaders. He struck a verbal deal with railroad tycoon Jay Gould and completed the railroad in just 5 months.

With the railroad operational, the Horn Silver Mine boomed into one of the highest-producing silver operations in the US. During its peak operational lifespan from 1870 to 1880, it produced ores that routinely contained 70—200 ounces of silver per ton (context: high grade mines yield 40—100oz/t).

In 1879, Cooke sold his 50% stake for over $1m ($33.5m in today’s money). With this windfall, Cooke cleared every single one of his remaining debts. He successfully bought back his beloved 52-rooms Cooke Castle and lived out the remainder of his life in financial comfort until his death in 1905.

Perhaps as a man of great ambitions, he might not have enjoyed his late comfortable life. He would see his peers make fortunes from his wreckage:

• J.P. Morgan bought up those bankrupt railroads and reorganized them. By 1900, he had control of ⅙ of all US railways including NY Central Railroad, Erie Railroad, Southern Railway, Northern Pacific.

• Andrew Carnegie capitalized on collapsing construction cost and finished his Edgar Thompson steelworks, which was the foundational mill of his steel empire.

• The Vanderbilts refused to bail out Jay Cooke, instead chose to hoard cash and bought up rival railroads at steep discounts, further consolidating their railway monopoly in NYC, Chicago, and the Midwest.

Lessons

This short history teaches us lessons that challenge our comfortable assumptions about how investing actually works:

1. Reputation is not the same as soundness. Jay Cooke was not a fraud; he had genuine reputation earned through real accomplishments. But this tells you nothing about the current quality of someone’s judgment.

2. Narrative is dangerous. We all love optimistic stories; society progressed because of optimism. But a compelling story and a good investment are two different things. In fact, the more compelling a story, the harder it becomes to evaluate an investment on its own merits.

3. Success is dangerous for a risk taker. Investors who succeed can have a hard time recognizing when conditions have changed, because every prior success reinforces confidence in your own judgment.

4. Timing is important. Cooke’s vision of a booming economy from railroads did eventually come into fruition, but it was not as early as he thought.

The Kirkland Signature brand is Costco’s private label, or what people call “in-house” brand. In 2025, Kirkland produced a massive $90b of revenues. Compare it to these consumer products giants:

• Coca-Cola: $48b

• Pepsi: $92b

• P&G: $84b

• Nestlé: $98b

• Unilever: $58b

Let’s go back in history to see how the Kirkland brand became a central piece of Costco today.

History

After WW2, national consumer brands (Gillette, Campbell’s, Heinz, Kellogg’s, P&G…) went on a prosperous streak off the back of several factors:

1. Demographic cohort born from 1946 to 1964. This was a generation of economic prosperity and rapid technological progress.

2. TV as a form of media expanded rapidly. Networks like ABC, NBC, CBS leveraged the TV media for advertisements.

3. Car ownership grew from 25 million to over 70 million vehicles, paving the way for suburban expansion and the Interstate Highway.

4. Construction of highways led to supermarkets with huge parking lots for the cars.

TV networks were the main medium for advertisements. So many great ads appeared in this era: Tony the Tiger, Jolly Green Giant, Ronald McDonald, Marlboro Man…

The business model for brands depended on economies of scale. They would spend on TV ads to drive sales and push out smaller competitors, then reinvest into more ads and spread the cost over an ever growing base.

Then came a long period of inflation in the 1970s which led to higher prices, until Paul Volcker raised interest rates to 20% and finally tamed the disaster. However, when prices started to subside in the mid-1980s, the national brands did not lower their prices.

This was came to be known as The Nifty Fifty era. At that time, the conventional wisdom was that investors could buy these stocks and be assured of continued growth.

The Nifty Fifty stocks were bidded up to absurd prices and experienced a decade of underperformance after their early 1970s peak.

The underlying companies were indeed good businesses, just that people were over optimistic at some point, so after the crash their shares began to recover. From 1981 to 1991, companies like Heinz, Hershey, Kellogg, Clorox, and Kimberly-Clark showed market gains that were multiples of the general market.

Consumer goods companies greatly widened their operating margins over this decade. Since 1981, for example, Heinz's operating margins increased from 11% to 18%, Ralston Purina's from 9% to 16%, Sara Lee's from 7% to 10%, and Clorox's from 11% to 18%.

Quaker Oats was selling their breakfast cereal at $0.73 for 500-gram box in 1980, then in 1991 it went up to $1.73, despite prices of oats fell from $1.7 to $1.2 per bushel.

It was in this environment that Jim Sinegal (co-founder) saw an opportunity for private labels. He decided that Costco could undercut the national brands while offering a better product, and what a better time to do this during the 1990 recession.

Sinegal had an experience with private labels back in his time at FedMart, which had an in-house product called FM Brand. The initial business model was to sell B2B, then they discovered that individual consumers were willing to shop in bulk at Costco.

Costco started to study private labels in the UK (Tesco, Sainsbury) and in Canada (President’s Choice). This led to the consolidation of multiple private labels into a single Kirkland Signature.

Kirkland was launched in 1995, it’s name came from Costco’s HQ which was in Kirkland, Washington at that time. It kept the name even after Costco moved to Seattle.

The first two Kirkland products were shampoo and vitamins.

Kirkland Quality Assurance

Costco typically spends years developing Kirkland products and will drop anything that doesn’t check all the boxes.

These are the challenges for trying to be a vendor under the Kirkland brand:

• Costco will trial a new product for 8-12 weeks.

• Since packaging is so important, Costco spends most of its time with the vendor making sure the packaging is impactful.

• The vendor’s manufacturing facility have to meet Costco’s exact quality requirements and the supplier is responsible for delivering the goods. Pallets have to exactly meet Costco’s requirements, or they will be rejected.

• Costco doesn’t provide vendor financing and will pay the supplier 15 days after delivery. This is where Costco gets their negative working capital.

• Gross margins are capped at 15% for Kirkland. When Costco rewards a vendor based on sales and not margins, it changes their behaviour; cost efficiency is prioritized.

• Costco will work with the vendor to have some product differentiation from national brands.

For example, it took 8 pitches for Costco to approve the Kirkland Signature gas station. Sinegal was very worried about trying to deliver a high-quality and low-price alternative to vertically-integrated O&G giants.

Costco now has 747 Kirkland Signature gas stations. On average, it sells gas at $0.30 cheaper per gallon and these customers end up going inside the warehouse buying more stuff, and eventually renewing their memberships.

Sinegal would famously require Kirkland vendors to have products to be at least “1% better than the market leading product”. Costco will identify data points in which the Kirkland product must be slightly better. As a supplier, you have no objections because the data is clear, and you’re certainly going to have more sales selling under Kirkland. So both your original product and Kirkland will be placed together on the shelf — the consumer has 2 choices, both of which are yours!

Even the breakfast sandwich was challenged on how the vendor could exceed the national brand on the bacon. They managed to put 40% more protein at the same price. Yes, maniacal quality control is what brings back customers!

We can't leave out their famous $1.50 hotdog combo, also under the Kirkland brand. This $1.50 price tag has been unchanged since 1984. CEO Ron Vachris, publicly vowed that the price is “sacrosanct” and will not change as long as he is with the company.

This Seattle Times interview with Sinegal in 2011 has by far the best reply:

Seattle Times: If the hotdog combo price ever goes up, what will it mean?

Sinegal: That I’m dead.

By the way, Sinegal is 90 years old now and still visits the office weekly on Tuesdays.

Here are some examples of confirmed big suppliers of Kirkland:

• Duracell makes AA and AAA batteries only for Kirkland. Their own Duracell batteries deliver an average 1.28V, while Kirkland 1.62V.

• First Quality makes diapers at half the cost of previous manufacturer Kimberly-Clark.

• Bumble Bee packs tuna cans for Kirkland for $20 (8*7oz) with thicker tuna chunks than their own Bumble brand at over $30.

  

The situation indeed says alot about who has more bargaining power. Costco offers big brands 2 slots on the shelves, 1 of them will be taken by Kirkland and they are forced to manufacture the Kirkland to be slightly higher quality than their own. Yet, Costco will sell Kirkland cheaper.

Nevertheless, vendors are willing to work with Costco because of its strong brand reputation and loyal customers.

Private Labels Here to Stay

National brands will not die out, but they no longer produce growth that was seen in the past. In 2025, total sales of national brands in America grew +1.2% to $1.05t, but private labels grew faster at +3.3% to $283b. From 2021 to 2025, private labels increased sales by $64.8b, while unit sales increased by 2.7 billion units.

Roughly 1-in-4 food and non-grocery products sold in US retail outlets are private labels.

Costco currently has 148.5 million members worldwide, of which 82.9 million are paid members, and the renewal rate on the annual membership is ~92%. The ratio of total/paid members is ~1.8 so the free supplementary household member of a Costco paid member probably has the same shopping habits, it will be hard to leave Costco's ecosystem.

Now that Kirkland Signature has climbed to 33% of total sales and the trend is not stopping… there will be this growing group of members who need Kirkland products because the alternative is simply more expensive and lower quality (usually).

That is a very strong business outlook to think about.

The value proposition of Kirkland is amazing — it acts like the “Price Police”. If a national brand tries to squeeze too much margins on a staple product, Costco will come in and drive prices down to the benefit of their members.

We can expect the future growth of Kirkland to be guaranteed. It produced $56b of sales 5 years ago, now it’s at $90b. There’s about 350 Kirkland products across total 4000 SKUs, so it’s bringing in 33% of total sales but only occupying 9% of total SKUs!

Given that the product mix is heavy on staples like toilet paper, diapers, bottled water and paper towels… we can confidently expect that revenue visibility is high.

Furthermore, Costco plans to make about 1000 SKUs variable, meaning that they are items that can be changed to keep the “treasure hunt” shopping experience. If so, we can also expect Kirkland to grow to above 500 products.

To forecast a very possible baseline financials in 2036 if Kirkland makes up 50% of sales:

Of course, Costco stock is expensive today trading at 41x EBIT. If this multiple doesn’t shrink, then based on the scenario above Costco will be worth $1t in 10 years time.

Bear Case: Demand Collapses?

We have written many articles about CPRT, refer here for a comprehensive explanation of business economics.

Ever since we have bought the stock at an average price of $41, it has fallen by over -20%. The stock price divergence with their financials is stark, as if suddenly the market thinks that CPRT is structural impaired:

In our Q3 FY2026 update, we have explained the near-term headwinds that CPRT is facing and why these are non-structural issues.

However, there is one structural bear case we want to address here. It attacks the demand side of CPRT’s business:

As cars become more complex and EV penetration increases, repairing cars will become more difficult due to lack of parts and skilled labour.

We might end up with a future where nobody has an incentive to repair salvaged cars leading to low demand.

To answer this case, we have to first see what type of auctions is CPRT doing today. We went to its auction page and downloaded 3 sets of data (sample size = 3000).

The dataset can be split into 3 types of salvage vehicles:

1. Non Repair: Very badly wrecked cars that cannot be repaired, most likely sold for scraps and dismantled for raw materials.

2. Salvage: Repairable cars, this is what the bear case is attacking.

3. Run & Drive: Car engines that can start on their own without intervention, supposedly able to buy and drive off (no gaurantees).

We refined the dataset so that “average bid amount” makes sense:

• Excluded bid amounts that are less than 50% of estimated retail price. This removes entries which are very early in the bidding process.

• Excluded Purple Wave vehicles as these are not supplied from insured accidents.

Firstly, based on the total bid amount, majority 83% of bids are for Run & Drive types. 14% are repairable (Salvage), while a tiny 3% are Non Repair:

So that’s one piece of good news: Auctions are mostly made up of cars that are “ready to drive”.

Baseline Model

Next, we look at the average bid amount and estimate average daily commissions CPRT receives from buyers. This will serve as our baseline:

As expected, Non Repair has very low bid amount, followed by Salvage. The average bid against retail value also makes sense:

• Non Repair are only worth 25% of their retail value.

• Salvage are also very low at 38% of retail value.

• Run & Drive is high at 78% of retail value. People are willing to bid for decent conditioned cars.

The average commissions are 12% for sales that are above $5,000. All of the categories exceed this amount.

CPRT states on their website that the number of cars auctioned at any given time is ~350,000. Then, we proportionally split it by the share of total bid amount in the above pie chart.

CPRT also says that it takes 7 to 10 days to sell a car after it is put on auction. We assume 10 days turnover:

Average daily comms = 12% * Est. Final Bid * Number of cars/10

Therefore, we estimate CPRT earns daily commissions of ~$91m.

Sense check:

• Gross $91m up to full 365 days for each 10-day cohort = $3.3b.

• Compare to FY2025 service revenues of $4b, the remainder $0.7b is reasonably made up of insurers and membership fees.

Scenario Model

Now we have a baseline, let’s simulate a scenario after 10 years where 50% of Salvage cars become Non Repair.

Our assumed changes from baseline:

1. Proportion of Non Repair increase +7% (from 3% to 10%), contributed by Salvage -7%.

2. Average retail price CAGR at +5.4%.

3. Average bid amount CAGR at +5% (slightly lower than retail price growth).

4. Commissions increase from 12% to 14%.

5. Number of cars auctioned increase from 350k to 400k (CAGR +1.3% represents total loss frequency trend).

6. Turnover unchanged at 10 days. Assume no operational improvements.

The average daily commissions works out to be $188m (CAGR +7.6% from baseline). This is essentially the sales growth rate we can expect if this scenario plays out.

Note: Vehicle sales segment (15% of revenues) assumed to grow at same +7.6%.

Finally, we can calculate our total return… To keep things simple, let’s assume:

• No further economies of scale and overseas transition to agency model: +2% margins improvement come from hiking commission rate from 12% to 14%.

• No more shares repurchases. Current 963 million shares outstanding unchanged.

• PE ratio 19x unchanged.

Based on these inputs total returns is ~10%.

We think this is a decent return for such a conservative scenario.

There are upsides that we have not considered which require no heroic expectations:

1. Margins improvement from international expansion as CPRT slowly move from principal to agency model.

2. Share count reduction. CPRT has zero debt and $5.1b in cash against the current market cap of $27.3b.

3. Car auction sales turnover can be improved. CPRT is investing in their towing network to reduce friction in the whole process.

At the 2017 annual meeting of the Daily Journal, Charlie Munger mentioned an investment that made him $80m in profit.

That investment was a bond and stock purchase of a company called Tenneco.

Tenneco was a high-risk high-reward investment. The stock/bonds were very cheap, but this wasn't an idea that you would put alot of money into. Considering that Charlie Munger was already a billionaire when he made the investment in 2001… $80m was insignificant relative to his net worth.

Munger claimed he spent less than two hours deciding to invest in Tenneco, calling it a “cigar butt”. Unlike Buffett, who thinks of “cigar butts” as businesses where you can shut down operations for one last “puff”, with the return inherent in the liquidation value exceeding the price paid for the entire company; Munger looked at it as an incredibly cheap stock, protected by predictable earnings.

In fact, Tenneco had no liquidation value. Tangible equity was negative due to large accumulated losses. It had $1.6b of debt with an average interest rate of 10.3% in 2001. More than $1b consisted of senior credit facilities, mostly bank loans, where Tenneco had already breached 2 of 3 covenants, and was close to breaching the third by the end of 2001.

They had renegotiated lower covenants three times in exchange for higher interest rates.

The remaining ~$500m of subordinated debt paid 11.625%. Its price went from $100 par to $35. Munger scooped up some of these bonds.

Operating income had dropped from $227m in 1998 to $92m in 2001. The dividend was suspended in 2001 and wasn't reinstated until 2017. With a fully diluted market cap of $51m and net debt at a market value of $1.2b, the stock was trading as low as $1.35/share. Clearly, the market was pricing for bankruptcy.

So how did Tenneco find itself in such a bad situation?

Restructuring

In the years prior to 2000, Tenneco spun off many businesses, turning the company from a conglomerate into a much smaller business.

In 1993, Tenneco was a large conglomerate in six unrelated businesses:

1. Automotive

2. Packaging

3. Farm equipment

4. Shipbuilding

5. Energy

6. Chemicals

In 1994/95, the chemicals and farm equipment businesses was spun out. In 1996, the shipbuilding and energy segments were gone. The remaining automotive and packaging segments were further broken up by forming a joint venture.

Eventually the packaging business was also spun off into a new company called Pactiv Corp.

In these 7 years, Tenneco went from being a large, diversified conglomerate with $13.2b revenues to a smaller automotive business with just $3.5b.

There were likely a bunch of shareholders who first owned shares in a diversified conglomerate and, by 2001, found themselves owning an undiversified automotive business in an industry suffering from a recession.

To make matters worse, Tenneco likely kept the debt at parent level when it spun off those parts. This is likely to make them look more attractive to buyers.

After all these spinoffs were finished, Tenneco was left with $1.5b of long-term debt. The net interest rate was 10%, draining $150m per year in interest payments.

Operating income was only $115m in 2000, and so Tenneco was losing money. Upcoming maturities from 2001 to 2003 amounted to total of $262m.

Looking at the financial situation in 2000, investors must have doubted if Tenneco could pay off its creditors.

The marcoeconomic environment added another blow. America was reeling back from the 9-11 terrorist attack and the bursting of dot-com bubble.

Meanwhile, aftermarket distribution channels were consolidating around companies like Autozone and Advanced Auto Parts.

Tenneco was competing in a matured industry with limited pricing power. And with the US in a recession, new vehicle sales declined by -1% two years in a row, disproportionately hitting auto parts manufacturers.

Given so many glaring problems, why did Munger even bother investing?

A Branded Cigar

Although most “cigar butt” type investments eventually realize their value through dismantling the company, they do come in different flavours; some are cheap relative to liquidation value, while others have a strong enough brand to keep afloat.

Tenneco was the latter.

This was written in their annual report:

In each of our operating segments, we manufacture market leading brand names. Monroe ride control products and Walker exhaust products are two of the most recognized brand names in the automotive parts industry...

In the market, we believe that we are among the top four suppliers in the world for both emissions control and ride control products and systems. In the aftermarket, we believe that we are the market share leader in the supply of both emissions control and ride control products in the world.

These claims were true:

• Monroe secured significant factory-fitted contracts. Globally, roughly 1 out of every 4 new passenger vehicles were engineered with Monroe shock absorbers.

• Tenneco’s list of original equipment manufacturers was large and diversified, including just about every major auto maker in the world. The largest automaker (GM) accounted for only 16.6% of sales.

• Even today, Monroe commands a 15% to 20% share in aftermarket parts.

Munger mentioned this reasoning too:

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.

Prisoner’s Dilemma

But how about their huge debt, surely brands can’t save it from bankruptcy?

Munger would've saw a very difficult financial situation, and a prolonged recession could very well destroy Tenneco.

The fact that Munger did not invest fully in Tenneco’s equity, and chose to split his investment into its bonds, showed that he realized this.

But the situation is not so simple. Prisoner's dilemma comes into play.

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. The total amount of debt was $989m plus $500m of subordinate bonds.

How are lenders going to recover $989m?

1. Force assets liquidation, lay off workers, shut down operations… and try to sell Monroe and Walker brands to someone?

2. Recapitalize the company, exchange debt for new equity and continue operating the business, eventually selling the equity?

Considering that EBIT was still profitable at around 3—4% margins, it would make more sense to allow Tenneco to continue operating. Furthermore, management was making big efforts to cut expenses:

Beginning in the fourth quarter of 2000, we undertook an aggressive cost-cutting initiative to reduce global salaried employment, followed by a second round of salaried employee reductions announced in the first quarter of 2001. We expect these combined actions to eliminate approximately 22% of our global salaried work force by early 2002. When this initiative is fully implemented, we anticipate realizing cost savings of $72 million annually.

Tenneco also had $357m left on their credit revolver. In a dire situation, this could buy some time.

Investment Results

Putting everything together, Charlie Munger invested in the stock and bonds under these conditions:

1. The stock and bond were priced for bankruptcy.

2. Tenneco had products that had strong brand value. Management committed to cut expenses.

3. Creditors not likely to force bankruptcy (prisoner’s dilemma).

We don’t know the proportion between stocks and bonds.

In 2004, Munger took $80m profits, we can work backwards assuming 50% stocks/bonds… he would’ve turned $14m into $88m (assume he sold both stock at $15 and bonds at $106, at the same time).

Rolling It Forward

A great outcome in just 3 years, but it was definitely a risky bet.

The story doesn’t quite end here…this $80m profits multiplied another 5x under the hands of Li Lu.

In 2004, Li Lu, a Chinese fund manager who escaped the Cultural Revolution, met Munger during a dinner. He explained the businesses he held and asked for advice on how to run a fund. Munger was very impressed by Li Lu's knowledge, so he passed the $80m over to him. Li Lu was the only person who ever managed money for Munger. Living up to his name, he turned $80m into at least $400m!

In the end Tenneco was acquired by Apollo Global at an enterprise value of $7.1b (debt $5b) in 2022. Taking profits after the deep value play succeeded was smart, because Tenneco is not a high quality business, if Munger simply held onto his shares, he would’ve just 5x his money from the shares.

Daniel Ek was born in 1983 in Stockholm, Sweden. From an early age, he showed both talent and passion for entrepreneurship. As a young 13 year old boy, he started building websites for small companies from his bedroom, charging $100 per site but quickly increased prices to $5,000 as demand grew.

By the age of 14, Daniel had learned C++ and HTML CGI programming languages, making money from software development. In fact, even in middle school, he used video games, iPods, and cell phones as bribes and incentives to get his classmates to do his schoolwork and code for his early web-design business.

By the age of 15, he was earning $50,000 a month creating websites. Later, his businesses grew to the point where he could employ 30 people!

When Daniel applied for a job at Google at the age of 16, he was rejected because he had not finished university. It’s funny that around this period, Daniel was already a taxpayer before graduating from university.

Thinking that money could buy happiness, Daniel retired very young at just 23 years old, after selling his online advertising startup Advertigo in 2006 for $1.25m. After becoming a millionaire, he decided to quit and enjoy his wealth. So he bought himself a Ferrari and hung out in Stockholm nightclubs where he got his friends expensive tables and wooed attractive women with fancy champagne.

However, after a year of this lavish lifestyle, Daniel had a realization… the women he spent time with were not very nice, and his friends were only using him for his money.

They were people who were there for the good times, but if it ever turned ugly they’d leave me in a heartbeat. I had always wanted to belong and I had been thinking that this was going to get solved when I had money, and instead I had no idea how I wanted to live my life.

Daniel Ek, 2014

Afterall, money cannot buy true happiness.

So he decided to move to a cabin near a suburb called Ragsved where his parents lived. It was in this place that he talked to Martin Lorentzon about creating a product that would allow people to have seamless access to all of the world’s music.

Both men came together and founded Spotify in the same year 2006.

Piracy was rampant in the early 2000s from websites like Napster and LimeWire. For those who lived their teenage years during this period would be familiar with the memories of illegal music sharing on a dial-up modem.

In fact, The Pirate Bay emerged in Sweden. Daniel Ek and Martin Lorentzon envisioned a platform that could offer instant access to any song, anywhere, without the need for downloading. The development of Spotify was driven by a commitment to legality, fairness to artists, and user experience. Just after inception, the founders negotiated with record labels and artists to ensure a vast catalog of music while developing a platform that was intuitive and user-friendly. Spotify officially launched to the public in October 2008 after roughly two years in the development phase, initially through an invite-only system in Europe.

Spotify quickly gained popularity, largely due to its superior user experience and freemium model: a free, ad-supported service alongside a premium subscription option. Their user base grew quickly, just 8 years after initial launch they reached 100 million users worldwide. Spotify entered the US market in 2011 and continued to grow its global footprint.

Today, Spotify is available in 184 countries with 751 million monthly active users (290 million paid users). Daniel Ek still receives zero salary, no long-term incentives, no stock options. Both founders are still the largest shareholders of Spotify with controlling stakes: Daniel owns 13.9%, Martin 9.2%.

Industry Background

We have to go back to 1979 when Sony launched the Walkman. It was a portable cassette player that allowed people to listen to music at any place and time. Before this, music was mostly confined inside homes or cars.

This sparked off the creation of audio CD (1980s) and the MP3 player (1990s). Then the smartphone era came and Apple invented the iPod. It was only until 2015, when music streaming finally scaled up. Like many other forms of entertainment, music has also gone through a few decades of transformation.

Following the rise of the internet, the music industry value chain has changed drastically. We started off with physical mediums, then downloading and piracy became rampant, in fact the whole music industry saw declining revenues until streaming took off in 2015.

Although we have nominal higher revenues, when adjusted for inflation it actually shows an industry with no real growth (in the US):

However, we need to focus on streaming because growth there is very real and durable. It’s estimated that people spend 4 hours per day listening to audio entertainment, the cost per hour of music entertainment is the lowest among different choices (Spotify has different premium plan prices for different markets, the comparison is likely to be valid):

Now that music has become even more pervasive with the rise of social media, this is adding growth opportunities for Spotify.

Value Chain

For the entire 20th century, music labels owned both the copyrights and hardware. Their business was essentially vertically integrated tech and manufacturing monopolies disguised as art companies. They controlled the entire means of music consumption:

• RCA (Radio Corporation of America) owned RCA records (Elvis Presley, David Bowie, John Denver…). They also invented the 45 RPM vinyl format and manufactured the physical record players you had to buy to listen to it.

• Philips owned PolyGram Records. They also invented the Compact Cassette tape in the 1960s.

• Sony owned CBS Records (Michael Jackson). But they also co-invented the Compact Disc in 1982, and manufactured the Sony Walkman that played them.

But today, the labels have lost this positioning.

Now the music industry is dependent on third-party platforms to distribute its products. These are very big players like Apple, Google, Amazon, and Spotify. They own the distribution, algorithms, hardware, and most importantly, the customer data.

Music labels continue to be the constant gatekeepers, but they lost the music value chain control. The top 3 labels: Universal (UMG), Warner (WMG), and Sony Music still control nearly 70% of digital/physical music market share today. However, their scope has been reduced to artists, while paying “rent” to platforms.

Music Labels

We need to segment music into 2 separate businesses:

1. Publishers: They work with songwriters and pitch songs to artists to perform. Publishing royalties are earned for the use of a song’s composition and paid to the songwriters and their publishers. They include performance and mechanical royalties, which are earned, tracked, and collected by societies around the world.

2. Recorded music: This is the area music labels operate. They discover and fund artists and make agreements to share the earnings from the “master record” at a pre-determined split. Labels get paid only based on the master record, so if an artist performs at a concert, the publisher gets paid and not the labels.

We will focus on labels since they have significant impact on Spotify's gross margins.

The economics of labels are similar to venture capital; they fund lots of artists but only a small handful make it big.

The power of labels lie in their control of music supply. Spotify’s suppliers are the music labels and such concentration of market share is usually a bad sign. You can quickly see Spotify gross margin is inferior to Netflix even though both of them have the same streaming business model:

Since it is the streaming platforms who are doing the hard work of growing the market, the labels get to enjoy a growing royalty fee.

But there’s a case of prisoner’s dilemma at work here. Since royalty fees are distributed based on market share of music streaming, if one label decides to leave Spotify, then the remaining labels will make more money because the whole pie will be distributed to whoever is left. The only way labels can hurt Spotify is for all of them to leave together.

We can illustrate the economics of Spotify vs Labels:

• Suppose Spotify earns $1000 from users who only listen to 2 singers: Bruno and Taylor, in a 3:1 proportion.

• Spotify keeps 30% of revenues before passing the remainder to labels (Spotify doesn’t pay the artists directly). So the labels will receive total of $700, split into Bruno and Taylor based on the stream-count share (3:1).

• Bruno’s label (WMG) gets $525 (75%), while Taylor’s label (UMG) gets $175 (25%).

These labels will take their cut and distribute the money further down the funnel to artists, songwriters and other parties involved according to the contracts they signed.

As you can see, artists under labels receive a smaller portion of the pie compared to independent artists who will get royalties after the distributor’s cut which is typically far lower at 10% to 25%.

But there is a good reason why famous artists sign with labels instead of going independent. In today’s world, there are between 100,000 to 150,000 new tracks uploaded to Spotify daily. That’s a huge number of audiences. Once physical distribution has lost it relevance, the world of streaming is much more complicated and independent artists will need a team of professionals to manage this work.

Given that most artists earn their money from concerts, marketing and merchandise, they don’t have incentives to do manage distribution and will delegate it to streaming platforms. Without a distribution platform, independent artists may enjoy a larger share of the pie, but the pie itself wouldn’t be large.

A special note is given for radio. In the US, traditional AM/FM radio stations are entirely exempt from paying performance royalties to record labels and performing artists. However, they still pay royalties to songwriters and music publishers.

Put together, the entire music value chain is an uneven playing field.

Business Model

Now that the labels lost control of the formats, music as a product has become commoditized. Most people reference a parallel to the video streaming wars between Netflix, HBO, Disney… this is a wrong comparison because each of them are differentiated. For example:

• Stranger Things is exclusive to Netflix.

• The Last of Us is only on HBO.

Music is really commoditized with Apple, Amazon, Tidal, Spotify… all of them offering the same library of hundreds of millions of songs.

Spotify can be described as a business that streams audio on demand. It operates a freemium model and reports in 2 segments:

1. Premium (Revenues: $15.4b; Gross profit: $5.2b)

2. Ad-Supported (Revenues: $1.8b; Gross profit: $0.3b)

Premium Segment

Like most platform businesses, the 2 most important drivers for Spotify’s business are monthly active users (MAU) and average revenue per user (ARPU).

Spotify’s MAU went from 124 million (2016) to 751 million (2025). MAU is effectively a leading indicator for Premium subscribers, which grew from 48 million to 290 million over the same period. The proportion of Premium users have been about 39% to 48%, this has been stable because Spotify prices their subscription depending on the affordability of the general population of the country.

The drivers of MAU is smartphone penetration. Because Spotify cannot gain share unless users have smartphones and internet access. Both factors have seen continued structural growth:

• In 2014, only 12% of the world’s population owned a smartphone, this number surged to 60% in 2024.

• Since 1991 internet penetration went from zero to over 60% today.

Since music’s appeal is universal, we expect Spotify’s MAU mix will shift further into markets outside the US and Europe. Take a look at the region mix changes:

• North America share: 34% (2016); 16% (2025).

• LATAM: 19% (2016); 21% (2025)

• Rest of World: 9% (2016); 37% (2025)

A key driver why Rest of World grew so much is because increasing smartphone penetration:

For ARPU, Spotify has many different Premium plans for every country. In the US, Spotify did not increase monthly prices since 2011 until a series of price hikes after 2022:

• 2022: $9.99

• 2023: $10.99

• 2024: $11.99

• 2026: $12.99

In emerging markets such as Bangladesh where purchasing power is lacking, Spotify has introduced daily/weekly plans. They also partnered with local telcos to let users buy pre-paid plans.

Due to the increasing number of MAUs from emerging markets, the ARPU has consistently decreased, until they increased prices recently:

Ad-Supported Segment

An Ad-Supported user is someone who is using Spotify for free, but listens or views advertisements between songs. The ARPU for ad-supported is very low at about €0.4 per MAU.

In 2025, there were 476 million ad-supported MAU. The growth is quite good; gaining 97 million users from 379 million since 2023.

Ad-Supported revenues are not a large piece, therefore Spotify have not focused on this segment:

Daniel Ek indicated before that he believes this segment should be 20% of revenues, and even more over the next 5 to 10 years.

Since the Premium segment is growing at +15% CAGR, Ad-Supported must do something special for Daniel’s statement to be true. There are a few possibilities:

1. Creating a two-sided marketplace where they sell ads and data insights to music labels and artists.

2. Taking radio advertising share. Radio remains the dominant medium for audio listening time, capturing ~62% of daily share. If we look only at ad-supported audio, 82% of audio time goes to radio and podcasts, while only 16% is spent with streaming music services. In terms of total ad dollars, radio’s share has shrunk as budgets have shifted to streaming.

The shift is better understood looking at how advertising spend inevitably shifts to time spent over time:

One of the reasons for the persistent share of radio is because people choose to listen to the radio when they are driving. Americans drive about 13,500 miles a year, if we assume an average speed of 60mph, then people spend 19 hours per month driving!

As cars become more technologically advanced, we can expect streaming platforms to eat into radio's share. Spotify tried to launch a hardware accessory called “Car Thing” for old cars to play Spotify in 2021, however it was quickly ended in July 2022 due to poor sales. There are also Apple CarPlay and Android Auto which are big competitors.

Rise of Podcast

There is a large number of podcast listeners worldwide estimated at over 500 million. 55% of the US population above 12 years old listen to a podcast at least once a month.

Below is an estimate of podcast revenues in the US:

Spotify used to significantly lag behind Apple for podcasts. Today it has grown into one of the top platforms for podcasts. They achieved this through acquiring exclusive content, copying the playbook of Netflix.

Podcasts escape the control of big labels and exclusive content not only creates more advertising opportunities but also converts free users into paid subscribers with a higher ARPU.

We have a growing body of evidence showing that there are significant benefits to engagement, retention, and conversion of users from Ad-Supported to Premium stemming from consumption of Podcast content.

We have seen benefits to retention on the order of several hundred basis points, which is a material change on a retention curve, for users that engage with spoken word content relative to those that haven’t, and early data indicates that these users are more likely to convert to Premium over time.

Daniel Ek, 2020

Acquisitions

Here’s a list of acquisitions Spotify made since 2019. The direction and ambitions of Spotify is quite clear: Empower artists to earn from their work on the platform.

Anchor

Consider their acquisition of Anchor (rebanded as “Spotify for Podcasters”) which is an introductory operating system for podcasting. It can be daunting for first-time podcasters to think about launching a show, Anchor simplifies and lower the barrier of entry.

Anchor has ~25% market share in the podcast hosting and publishing space. The other two major competitors are PodBean (34%) and Libsyn (10%).

Podz

We mentioned above that music content is a commodity, but music curation is not.

When you have an enormous number of content, effective curation is critical. For curation to be good, Spotify leverages on their large user database. Since podcasts are much longer than a typical song, matching podcasts with the relevant listener is a strong value proposition.

This is where Podz is useful. It uses machine learning to pull the best clips from podcasts and deliver them to listeners in personalized feeds (AI can make this even more effective).

The idea is similar to social media feeds/reels applied to podcasts.

Take a look at podcast download distribution. The median is a miserable 124 downloads — nobody makes money on that. To solve this issue, Spotify wants to find a way to increase the reach of smaller podcasters.

On October 2020, Podz filed a patent on “audio segment recommendation” algorithm. The core of Podz product is the ability to ingest large amounts of audio and pull the signal from the noise.

Currently, Podz algorithm is trained on inputs from professionals. Over time, it will continue to improve as it learns what listeners most resonate with across categories.

Megaphone

Megaphone is technology for publishers and advertisers seeking targeted slots on podcasts.

The Spotify Audience Network is the ad network where Spotify delivers on-demand ads based on their understanding of the customer. In this sense the ads are much more akin to social media ads than webpage ads.

The problem is the only place these ads worked were with Spotify-owned podcasts played within Spotify.

This is why the Megaphone acquisition is important to Spotify’s plans. By owning a podcast host Spotify gained infrastructure to start inserting ads into podcasts played on other platforms, and also podcasts streamed on their own. These ads won’t always be quite as good given that Spotify’s knowledge of the listener isn’t as accurate, but the more inventory the better.

When the target is the listener, then every podcast can be priced perfectly, regardless of how many fans they have. This is good for the smaller players, but it’s especially good for the market maker, because there is no longer a premium being paid to larger players simply for being worth the trouble.

Challenge #1: Relationship with Labels

Spotify’s financial model behaves differently from typical software platforms. Usually in SaaS business, profit margins increase exponentially as the number of users increase. We know this as operational leverage. The cost of serving an additional user is near zero, a larger user base spread across the same fixed cost leads to higher margins.

Streaming music is not like that. They pay out ~70% of every dollar earned back to labels and publishers, hence costs scale linearly with their user base.

This model works for non-pure play music streaming platforms like Apple, Amazon, and Google. They don’t need their music platform to be profitable:

• Amazon Music is a loss-leader to keep users paying for Amazon Prime subscription.

• Apple Music attaches itself to sell iPhones.

For pure-play Spotify, when margins are structurally capped, it is actually the artists that suffer. Although not explicitly said, it makes sense that Spotify doesn’t want artists to have a relationship with their fans.

Of course… Spotify wants their fans to have a relationship with Spotify!

As a result, Spotify have to guard listeners’ data because it’s the most important proprietary resource.

Fortunately, thanks to the rise of social media, independent artists (indies) have a much better chance to go direct to listeners. The rise of indies have taken share away from the top labels on Spotify.

The collective market share of the major labels has seen a steady decline, dropping from 87% in 2017 to 72% in 2025, as indies become more popular. This is a tailwind for margins.

Another quirk of the relationship between labels and Spotify is that subscription fees are not split between music and non-music content (see here for specifics).

Let’s say that you listen to 100 hours of audio streaming, split into 80 hours music, 20 hours podcasts (non-music). Your subscription fee goes entirely to the revenue pool to be distributed to labels/publishers/indies. So if you spend all your time on podcasts, your entire subscription fee still goes to the labels’ revenue pool.

If non-music content is growing rapidly, then we can see the vulnerability of the labels’ royalties. Why should labels receive royalties when you spend most of time on podcasts?

This is a touchy subject because the labels relationship is very much symbiotic and any large dispute can potentially hurt streaming growth for both parties. But we don’t think this equilibrium can last, the trend of non-music content will go up, and at some point Spotify should split revenue based on type of consumption.

In any case, Spotify has bargaining power simply due to the large user base.

Challenge #2: Big Tech

Competition is very strong with Google’s YouTube, Apple Music, and Amazon Music. All of them have deep pockets, low customer acquisition cost, and technological prowess.

There’s no doubt a challenge, but the details might prove otherwise.

Apple has been around the music industry even before iPhone was launched, but Apple Music was launched only in 2015. With 1.5 billion iPhone users today, Apple has an inherent advantage.

In June 2019, Apple Music mentioned that they had 60 million subscribers, today it’s estimated to be 95 million. During the same period, Spotify went from 108 million to 290 million paid users. It’s quite amazing that they added 182 million paid users, 5.2x the increase for Apple Music.

The reason is because iPhones market share is only ~20% globally although in the US their market share is much higher at ~60%. Android phones have a larger share in emerging markets, that’s why Apple Music loses out on the overseas growth.

Amazon Music launched in 2007 and charges a lower subscription fee than Spotify, but is only available in 50 countries. It is not designed to be a standalone product, and so it’s not really a serious competitor. It also has only 80 million users.

YouTube Music is probably the most dangerous competitor. Their strategy is to insert enough ads in the free version to incentivize upgrading to paid tier.

YouTube Music (and its ad-free bundle with YouTube Premium) has surpassed 125 million paid subscribers worldwide, including free trial users. Growth is significant, adding 25 million subscribers from 100 million in 2024. The potential of YouTube is their ability to buy exclusive non-music content, they can easily attack the podcast market and buy top content to become a big player.

However, video is far more profitable than audio for YouTube. So there’s a cannibalizing effect that YouTube has to think about. On the other hand, it’s easier for Spotify to move from audio to video (they launched video podcasting in 2020).

Indirect China Exposure

In 2017, Spotify and Tencent Music Group (TMG) swapped equity with each other.

Spotify received 9% share of TMG.

Tencent (parent) and TMG took 9% stake in Spotify too, which makes them Spotify’s 3ʳᵈ largest shareholder.

As far as we are aware, such equity swaps are uncommon. It made sense because China prohibits foreign investors from investing in companies engaged in audio-visual programs businesses.

We think there a several mutual benefits:

1. Spotify gains a lot from a closer relationship with TMG and learning from their playbook of integrating social experiences with music.

2. Both companies gain additional negotiating power with music labels. Tencent is extending this influence, in 2020, they bought 10% of UMG and 5.2% of WMG.

3. Spotify shows a novel approach to gain exposure to the Chinese market. Competing directly in China is never a good idea for Western companies.

These notes are not from us. It’s a compilation of conversations offline with people familiar with the topic, edited for reading.

Intro

For the last 2 years, the topic of AI compute was centered around GPUs. It is still true today, but there are developments in agentic AI (agents) that shifts the inference portion away from GPUs and into CPUs.

The reason is that agentic AI systems work on multiple processes like tool calls, API requests, memory lookups, orchestration logic, all of which are handled by the CPU. In this environment, CPU cores play a big role — more cores means more tasks, and memory hierarchy determine GPU utilization. These factors contribute to overall throughput and total cost of ownership of a compute rack.

Due to the way AI agents work, the demand will shift the CPU-to-GPU ratio in a compute rack towards CPUs. We can take the signals from Nvidia launching Vera CPU on 1 June 2026, selling it as standalone platform for agentic processing.

For those technically inclined, you can read this paper in November 2025 which estimated that tool processing on CPUs accounting for 50% to 90% of total latency in agentic workloads.

Inference vs. Agentic Workloads

The LLM (large language model) workflow that we’re all familiar with looks like that:

1. Human types a request.

2. CPU tokenize text (into numerical tokens) for the model to process.

3. GPU takes the tokens and run the model for inference.

4. CPU de-tokenize into text and produce the reply.

5. Human reads & processes the info, and acts on it.

This loop is called “pure inference”. It’s slow since there is a human processing the info. Notice that the CPU’s role is limited to getting the request to the GPU and fetching the output to the user.

All the compute happens on GPU, that’s why the GPU has dominated AI infrastructure, as the CPU is not the bottleneck.

However, in agentic workloads, the human is not in the loop. A person triggers a request, then agents read, plan, call for tools, and produces the result. The difference is that there can be many sub-agents per person, all running this loop simultaneously. So the faster this loop completes, the quicker agents finish their tasks.

More agents means that tasks must be performed in parallel, leading to the need for “orchestration of agents”, which all happens on the CPU.

In an agentic workload, the CPU performance dictates GPU utilization rate:

1. Human types a request.

2. CPU tokenizes it and generates an execution plan.

3. Multiple sub-agents run in parallel to carry out the plan. Each using different tools and having different runtime.

4. CPU manages the relationship between sub-agents. Some of them are independent, others are dependent on order of completion.

5. Sub-agents passes the tokens to GPU for calculation.

6. CPU de-tokenize output, collates and produce output.

Notice a few requirements here:

• CPU with high clock frequency can monitor tokens at low latency. This means it can pass tokens between sub-agents quickly.

• CPU with high core count allow for many sub-agents to run parallel.

• Low latency access to memory is important for long context windows.

Reasoning vs Action Oriented Workloads

The type of workload, whether reasoning or action, affects the CPU/GPU ratio. For sake of simplicity, we assume matrix operations are only handled by GPUs. So we don’t discuss features like Intel’s Advanced Matrix eXtention (AMX) where inference can occur on CPUs.

We can consider these 9 metrics below to show the relative reliance on the CPU:

(1) Per-core performance and clock speed

CPU needs to make orchestration decisions quickly to keep the GPU busy. Clock speed determines efficiency.

(2) Core count

Higher core counts means more agents can run in parallel. Single agents can also use multiple cores.

(3) CPU-xPU interconnect

xPU means whatever “x” computing architecture you’re using (GPU, TPU, NPU…). The better the interconnect, the faster data can be transferred.

(4) CPU memory

Higher memory bandwidth allows for larger context windows.

(5) Cache design and size

L3 cache is a high-speed memory bank located on the CPU. It acts as a staging area between the CPU cores and the slower main system RAM. More L3 cache means lower memory latency.

(6) Performance per watt

The standard measure of unit economics; how much energy is required for a given compute task.

(7) PCIe generation and lane count

Agents can do many actions, like API calls, read storage, go online… PCIe (Peripheral Component Interconnect Express) generations determine the raw speed of the data, while lane counts (the number of physical data pathways) dictate how much total data can travel at once.

(8) Instruction Set Architecture (ISA)

x86 or ARM determines what support exists for tool calls, and what ecosystem you use. For example:

• Grace (ARM) + Nvidia GPU

• x86 + Google TPU

(9) Non Uniform Memory Access (NUMA) domains

Modern server CPUs are composed of multiple chiplets to accommodate many cores. Depending on the interconnect technology between chiplets and physical layout, each CPU can have a non-uniform access latency to memory. This implies agents running on different cores can see different memory performance.

Now, we can compare a reasoning-oriented against action-oriented workflow across these 9 metrics.

Reasoning-Oriented

Consider this reasoning-oriented task:

Produce a comprehensive analysis of the memory chips market based on a large collection of research reports.

The agent plans a few research areas and assigns to sub-agents. Then, the sub-agents run inference loops, generating reasoning tokens, reads uploaded documents, goes online to search… Once all sub-agents finish, the CPU de-tokenize and produce output.

Nvidia’s Vera CPU is well suited for reasoning-oriented workloads.

• CPU-xPU interconnect uses NVLink-C2C at 1.8TB/s bidirectional. Coherent memory sharing lets their Rubin GPUs read directly from Vera CPUs memory. Applications can treat LPDDR5X and HBM4 as a single coherent pool, reducing data movement overhead and enabling techniques such as KV-cache offload and efficient multi-model execution.

• CPU memory is 1.5TB LPDDR5X across 8 SOCAMM (Small Outline Compression Attached Memory Modules) up to 1.2TB/s of bandwidth at low power.

Action-Oriented

Suppose you run a company and deploy agents for logistics:

There are thousands of customer requests.

Each customer has records from a database.

Logistics data comes from third-party database.

Agents need to generate an update for your customers and send an email to them.

This situation likely employs hundreds of agents running in parallel, with the GPU doing small inference. Most of the time would be spent on CPU orchestration.

We can see that CPU-related metrics matter more when it comes to action-oriented workloads. In this case, AMD’s Venice Dense CPU is suitable:

• High instructions per clock and 4MB L3 cache per core × 32 cores per CCD × 8 CCDs = 1 GB L3 cache. "Core per CCD" refers to how AMD organizes processor cores into individual, modular building blocks called CCDs (Core Complex Dies).

• 256 Zen6c cores, 512 threads with SMT (Simultaneous Multithreading). Many cores for many agents.

• x86 architecture.

AMD’s multi-chiplet architecture (which links multiple compute chiplets to a central I/O die) means that cores are separated into distinct Non-Uniform Memory Access (NUMA) domain. If a core needs to pull data in a memory bank on a different chiplet, it must route through the central I/O die, increasing latency.