This is the once-a-week free edition of The Diff, the newsletter about inflections in finance and technology. The free edition goes out to 11,102 subscribers, up 551 week-over-week. This week’s subscribers-only posts were part of a series looking at big tech companies and asking what could cause them to lose most of their value over the next ten years:
- Netflix: Product vs. Feature Competition looks at the Netflix streaming video bundle, and compares it to everyone else’s streaming-video-and-other-services bundle.
- Alphabet and Things That Don’t Scale: Alphabet has to outrun changes in form factor and competing vertical-specific search products. They can do the latter, but it goes against the company’s grain.
- Facebook and Superlinear Risk Scaling: The core of the bear thesis on Facebook is that the newsworthiness of nefarious activity on Facebook is a function of the company’s size, but the frequency of that behavior is, too. So platform abuse scales at the size of the network, squared.
- Microsoft and Costly Diversification: Microsoft is the big tech company that’s safest from antitrust, but at a cost. They have many more fronts to compete on.
This piece marks the close of the “Which FANMG Dies How?” series. The premise: if you look at the top companies by market cap in any given period, at least a few of them materially underperformed the market over the next ten years, sometimes losing most of their value. It would be odd if 2020 were the first exception.
The goal of this series is not to make specific predictions about which companies will fail, it’s to think about what risks a long investor is being paid to take.
In this issue:
- Amazon: A New Kind of Antitrust Risk
- PPE Nationalism, Vaccine Internationalism
- Macro Renaissance
- Tech Sees Like a State
- Technology-Governance Fit
- China’s Internal Immigration Reform
Amazon: A New Kind of Antitrust Risk
(Disclosure: I’m long AMZN.)
Amazon is the company most at risk from antitrust action, because Amazon doesn’t look like a conventional monopolist, and yet it feels like one. So if Amazon faces antitrust action, it will rest on a novel, untested theory—which means we have no precedents for determining the outcome.
It sounds a little ridiculous to say that Amazon “feels” like a monopoly, even though on paper it’s not. (Depending on how you count—a topic fraught with controversy—Amazon’s market share in e-commerce is somewhere between 5% and 35%.) Their best defense is that on paper, Amazon is a fast-growing business of unexceptional profitability, earning 5% operating margins last year and growing its topline at 20% per year. The stock trades at 120x earnings, though, because it feels like one of the best businesses of all time, and you have to invent a Theory of Amazon to justify owning it.
How Amazon Works
Every growth company has some kind of virtuous cycle: some chemical reaction where the reagents are cash, assets, and a competitive advantage, and the output is more cash and a deeper competitive advantage.
The Amazon cycle works like this: Amazon maximizes the number of economic actions it touches, and is willing to enter low-return businesses just to cozy up to more customers and suppliers. It overbuilds everything, so its incremental margins are always better than competitors even though its net margins are often far worse. So Amazon is always in a position to take its foot off the gas, and see revenue rise at a slightly slower pace while margins rocket, or it can continue to reinvest. All those economic touchpoints give it unparalleled information on where to invest, so it’s been in capital expenditure mode for a long time.
The last piece of the Amazon puzzle is cost of capital. Once investors accepted that Amazon could eventually show returns, and that the more losses it took early on the better its eventual results would be, they accorded the stock a high valuation, which meant that Amazon could lean in to capital-intensive logistics and cloud investments.
The simplest way to break Amazon down is into three segments, which don’t quite line up with the company’s own reporting:
Front-end retail is the classic Amazon business of operating a website that lists goods for sale. We can split this up further: Amazon started out selling first-party goods that were available elsewhere. Books, then electronics, eventually everything. They’ve expanded into a much more lucrative third-party business, where merchants bid to sell products to Amazon’s users. In that model, Amazon doesn’t have to source inventory itself, or choose products; it just has to vet sellers and kick bad actors off its platform. But that business led to a first-party private-label business, where Amazon identified successful product categories and then created its own in-house brands. This is a completely standard retail tactic, usually targeting high-margin products that aren’t special but, for whatever reason, don’t have much competition. My first experience with Amazon’s private-label goods was their iPhone charger cables. $7.50 on Amazon, $19 on Apple.com. (Edit: A reader notes that I linked to the wrong Amazon listing; the price gap is a bit closer for the comparable product, but still big.)
Amazon has added retail touchpoints over the years, both novel (apps, Alexa) and old-fashioned (Amazon Go, Whole Foods). But the core is still the same: a retail outlet where you can find basically anything, and get it for a reasonable-to-ultra-cheap price—after which the store will try to find you, hitting you with retargeted ads and nonstop product recommendation emails.
Their retail unit’s margins are a blend of two categories of products: emerging ones, where Amazon is losing money but building market share, and mature ones, where Amazon is harvesting profits. That’s completely obscured by the average numbers, pricing evidence suggests it’s true. This is also how retailers behave, on a different time scale: grocery stores know there are some products that get people in the door, and others that are high margin enough to make the visit worth it for the seller.
Logistics: Back-end retail: when Amazon reached their invisible asymptote and discovered the expensive and slow shipping was the key barrier to growth, they started to invest in warehouses and shipping. They never stopped. They now have 175 fulfillment centers with 150 million square feet of storage space, 60,000 trucks, and 51 planes. In-house delivery is not just a business for Amazon; it’s a BATNA. Every time they negotiate with other shipping partners, the subtext of the negotiation is “… and if that price doesn’t work for you, we’ll buy a few more trucks and planes and handle it ourselves.”
The Amazon logistics network has allowed them to offer Fulfillment by Amazon, where Amazon sellers ship their inventory to Amazon, and Amazon ships it to customers. And non-Amazon merchants can, too: Amazon is happy to store and ship goods that get ordered through Shopify, eBay, or an independent site.
Delivery networks have always had network effects: it’s more cost effective for a truck to make five stops on one block than one stop every five blocks, and more warehouses mean a smaller median distance between customers and warehouses. And Amazon has singlehandedly extended the network effects in delivery, by continuously spoiling customers with faster delivery times. The optimal network for two-day delivery is denser than the optimal network for 3- to 5-day delivery, so the original introduction of Prime meant that logistics could absorb more capital. And when Prime switched to same-day shipping, that meant more capital still. Amazon has continuously rearranged the logistics industry so companies with Amazon’s two advantages—low costs and cheap capital—have more room to grow.
AWS is Amazon’s crown jewel, a thought experiment that turned into the world’s most valuable SaaS company. I wrote about the beauty of the AWS model earlier this week, in the Microsoft piece:
AWS presents two challenges: scale and the tail. On scale, there are pure cost benefits to being bigger: AWS can open more datacenters, offer more regions, build in less slack to get the same redundancy, get better bulk deals from suppliers, and better amortize the cost of custom hardware. On the tail: with its larger customer base, AWS can better identify new ways to slice up its core product (system-resources-as-a-service, at some latency) into distinct products. Its userbase lets AWS profit from tiny differences in customer preference, like a willingness to accept high charges for uploading and downloading data in exchange for ultra-low prices for long-term storage, or uncertain spot prices versus certain reserved prices. Like with Google in search, it’s easier to solve for one-in-a-million edge cases if you have more millions around.
The AWS story reminds me of the line about meatpacking companies, that they could only turn a profit if they used “everything but the squeal.” And they did: Wilson meatpacking made sporting goods (literal pigskins) and before recombinant DNA it made insulin by crushing pig pancreases. AWS came into existence because Amazon needed enough computing power to handle the site’s peak load the week before Christmas, and that computing power wasn’t doing them any good the rest of the year. (Edit: a reader notes that this is not the origin story, and Amazon had other reasons to launch AWS.) They decided to package it up as a product, and rent it out, which lowered the risk to having a bit more capacity than they needed. AWS is in the very fortunate position that at first, it’s a cheap complement to an expensive, bottlenecked product (if you have an AWS bill, you have developers, and at first the devs are more expensive). Over time, AWS bills relentlessly rise, but this just makes users more dependent on AWS, and it doesn’t hurt that transferring data out is pricey. The nice thing about charging in tiny, usage-based increments, is that you can design your pricing so the maximum sticker shock is from trying to leave.
And AWS also benefits from an annuity-like trait: when programmers quit or get fired, they don’t necessarily leave behind a nice list of which AWS services are essential and which they just haven’t bothered to shut down. So AWS accumulates recurring revenue as a function of how many engineers have left an organization over time, and grows its revenue base according to how many are still there.
Even this breakdown doesn’t cover everything. I haven’t talked about ads, which are functionally part of their retail operation but spiritually part of AWS, since they, too, are a high fixed-cost effort to convert something Amazon already does—buy and sell pixels-minutes to shape demand—into a business. And I haven’t dug into their financing operation, both their ability to raise capital and their ability to strategically lend it to merchants and shoppers.
But what I’ve vaguely touched on, but not really emphasized, is that all of these businesses are functionally independent. Amazon Retail and Logistics are customers of AWS, but AWS sells to other businesses, including Amazon competitors. Retail sells goods that are fulfilled by Amazon, and goods delivered other ways. Amazon’s credit card works just fine when making purchases outside of Amazon. What Amazon has done is replicated the classic economic development strategy of “export discipline”: industrializing countries often subsidize a manufacturer early on, but force it to sell to goods externally, not just to the domestic market. They want to ensure that the products are competitive at scale, and the way to do that is to force them to compete globally.
The Antitrust Complexities
One long-term antitrust debate is over the concept of “predatory pricing”: selling goods below cost to drive competitors out of the market, then raising prices when they’re gone. This used to be an important part of antitrust law, and was a key element of the Standard Oil breakup. But Chicago School economists have pushed back on this: the trouble with using predatory pricing to eliminate a small competitor is that it’s expensive.
Suppose there are 1,000 widgets sold every year at $10/each, with a cost of $8. Red Widget has 50% market share, and so does Blue. Red decides to cut prices to $7 to eliminate Blue. So it loses $500/year. But if its market share rises to 80%, it’s now losing $800/year. At 100% market share, of course, it can raise its price, perhaps to $12, and start to recoup the cost—but if the profit per widget doubles, that attracts new entrants, so perhaps Green Widget will go into business, selling widgets at $11, and taking market share from Red, until Red cuts costs again—and goes through another expensive cycle of losing money at scale—until Green gives up. Meanwhile, widget buyers might notice that if they fund a small widget company that aims for 5% market share and sells its widgets at $7, they can trick Red into continuously providing cut-price widgets: the widget-buyers lose $50/year funding their competitor, and save $950/year buying widgets. And this is not hypothetical, either. In one of the great railroad price wars of the 19th century, Cornelius Vanderbilt and Daniel Drew kept cutting the price of cattle cars on competing railroads, until Drew decided to buy cattle and ship them on Vanderbilt’s line.
The irony of this antitrust critique is that it always sounded good in theory, but tricky in practice. Capital and management attention are not that fungible, suppliers can’t coordinate, so it always made more sense on paper than in practice. But, for the big tech companies currently in the antitrust crosshairs, the argument is basically true: all of them are constantly trying to identify places where their suppliers or buyers exercise monopoly power, and then commoditizing them.
More recent antitrust theory focuses on matters other than pricing. Consumer welfare could be hurt by less choice, or lower quality. It’s hard to make the counterfactual argument that we’d have more of a cornucopia of consumer goods if Amazon were smaller and other competitors sold the same products, and despite occasional negative headlines, Amazon is absolutely relentless at punishing sellers for perceived quality deficiencies.
Identifying instances of predatory pricing for Amazon is a fiendishly difficult task. Suppose Amazon sells me a book for $12, with free one-day shipping thanks to Amazon prime. It’s easy to account for Amazon’s cost to buy the inventory, but what about:
- The warehouse where the book was held
- The network of warehouses that allow Amazon to ensure they can ship one copy of a somewhat obscure book in one day, to anyone who subscribes to Amazon Prime
- The fixed cost for inventory-management systems that make the above project possible, and the nontrivial ongoing cost of estimating the supply and demand curves for every product Amazon sells, which as I’ve argued is a sufficiently big accomplishment that it reframes a longstanding debate in economics
- The marginal cost of running that pricing system—which is not cheap!
- The fixed and marginal costs of the delivery network
- The cost of all of Prime’s free goodies, some mix of which keeps each Prime subscriber renewing year after year
- The marketing cost of acquiring me as a customer, way back in 1999 or so, and the continuous cost of using their site, email marketing, an app, physical retail locations, and PR to keep me buying
- The cost of giving me 5% cash back on my Amazon Visa, relative to the expected future benefit of more Amazon spending in response to that benefit
Computing Amazon’s effective unit cost is a multidimensional nightmare. I’m sure it gives their accountants, data scientists, and double-barreled CPA + Stats PhD types no end of trouble. And that means there’s no way to tell if Amazon is selling a given product at a loss, which is the first step to figuring out if they’re selling it at a loss to grab market share. The decentralized approach makes this a little easier, because one Amazon division’s cost line is part of another’s revenue, but understanding the full picture is challenging. Amazon’s Escher drawing economics make standard antitrust tools irrelevant.
But sometimes courts exercise discretion. The best argument against antitrust action for Amazon is that they’ve lowered prices, increased selection, report low margins, and don’t have dominant market share in retail unless it’s defined very narrowly. The best argument in favor is that everyone understands why this video is darkly funny:
The deep risk here is that, since Amazon doesn’t violate current antitrust standards, we’ll need an entirely new set of them to actually go after the company. And who knows what that will entail?
There is one option available to Amazon, that the other big tech companies don’t have: preemptive spinoffs. If Amazon’s major divisions do business with each other as if they were independent companies, they could also function as independent companies. That would not be bad for valuations, either, especially for AWS. It would weaken one part of the model—that Amazon divisions are one another’s anchor customers. AWS benefited from Retail’s demand, and existed in part because Retail had excess capacity most of the year. Whole Foods helps Amazon promote last-mile delivery for more products. Logistics certainly gets a boost from how big Retail is, and how sensitive it is to delivery speed.
And in the very long term, a spinoff presents another risk: AWS, Amazon Retail, and Amazon Logistics will all be run by hyper-competitive Amazonians. At some point, one of them will look at all the money it’s paying to another, and think “We’d probably get a better deal if we had a backup option.” So they’d take it in-house: after Amazon splits itself up, Amazon Retail benefits from launching an in-house delivery unit to strong-arm Amazon Logistics.
In the end, this is the negative FANMG thesis I’m most confident in, but only by playing games with accounting. Ten years from now, it’s likely that Amazon.com will have a lower market cap than it does today—as long as you only count the original parent company, not the Amazon Logistics and AWS spinoffs.
Thanks to Andrew Walker for proposing the original thought experiment and collaborating on these writeups. For more on Amazon, The Everything Store is a good look at what we can now call “the early days,” through 2012 or so. If you want a lot on Amazon’s antitrust risk, and the history of antitrust thought, look at Amazon’s Antitrust Paradox. And stay tuned for a more detailed Diff writeup of the company.
Today at noon Eastern, I’ll be hosting a Zoom Q&A with paying subscribers.
Two podcasts this week: I joined the TechMeme ride home podcast to talk about Reliance and Jio, and Andrew Walker’s Yet Another Value Blog Podcast to discuss the FANMG RIP series, careers, and more. This one features a brief cameo from my two-year-old.
PPE Nationalism, Vaccine Internationalism
Canada is building local N95 mask production facilities. N95 masks are normally the kind of product that gets outsourced to low cost countries, because while it has sophisticated inputs, assembly doesn’t require as much specialized skill or equipment. Most of the time, that’s fine, though of course it didn’t work out that way this year. We may see more low-value-added manufacturing returning to rich countries, not because of tariffs or deregulation, but because of risk-averse subsidies.
One common concern about Covid is vaccine nationalism: that countries will run crash programs to develop vaccines, and distribute them to everyone in their country before distributing them to front-line healthcare workers elsewhere. A promising sign that this won’t necessarily happen: Russia is seeking a partnership with India to produce a Russian vaccine candidate. While vaccine nationalism is possible, vaccine diplomacy is a valuable tool.
A Macro Renaissance?
By process of elimination, macro investing should be a great field to be in right now. Most investment strategies implicit sell insurance against macro risk. Active investors who build a portfolio of stocks or bonds are usually predicting some version of the economic status quo (if you model a recession next year, or an oil crisis, or runaway inflation, it quickly dominates the outcomes you’re predicting—so most stock pickers, for example, will model the status quo and treat these events as exogenous risks). That, it turns out, is accurate ($):
Caxton Associates has returned 31 per cent this year, according to investors, while a fund run by the firm’s chief executive Andrew Law is up 42 per cent. Meanwhile, Louis Bacon’s Moore Capital, which last year decided to eject the remaining external investors from its flagship funds after a long barren stretch, notched up a 25 per cent gain in seven months through July.
In a healthy market, what you’d expect is for macro funds' returns to be fairly random, but widely-dispersed: if macro investors are thinking about low-probability, high-impact events, they’re only adding value if they’re all making variant bets.
Looking back at the heyday of macro, roughly the 70s through the 90s, the main reason behind macro investors’ alpha was that macro investors could track economic trends, and could assume that policymakers would be run over by economic forces. You can date the decline of macro from 1997 through 2012: from the start of the East Asian crisis, when some countries had to get rescued by the IMF but successfully restructured their economies, through the end of the Euro crisis, when Mario Draghi credibly committed to saving the system regardless of the cost. In the post-2012 period, central banks have been more interventionist, both at home and abroad; the US has been much more willing to provide dollar liquidity to overseas central banks, and as it turns out many of the profit-generating crises of the peak macro period were caused by acute dollar shortages.
Now, the challenge is not to predict when central banks or legislators will run out of firepower, but to predict exactly how they’ll react, and when political constraints will override governments' theoretical power to avert or mitigate depressions. That’s especially hard for traders, because the area governments are best able to intervene in is financial markets. The Fed was much faster than Congress. There’s still room for interesting macro maneuvers, though. As MMT advocates never hesitate to remind us, the constraint on government spending is not local currency, but real resources. The unhappy converse of that is that a shortage of real resources—whether it’s chips that can’t legally be exported to China, or oil tankers that can’t safely leave Saudi Arabia—still matter.
Tech Sees Like a State: Parties, Fires
Two new examples in the long-running Diff series on tech companies rationally choosing to act like governments:
- Google has enhanced its maps with real-time fire tracking. It’s a striking example of progress in bits rather than atoms; we’re better-equipped than ever to collect and distribute information on fires, but Covid and its consequences have left the state of California short-staffed for actually fighting fires.
- Airbnb has started to regulate users more, by capping group size and banning parties. Every government sets rules to minimize negative externalities by preemptively banning high-risk activities. Now Airbnb does it, too.
Pondering Durian has a great essay on which technologies are compatible with which regimes.
AI seems compatible with top down, authoritarian regimes such as China. AI is enabled by massive data-sets collated in centralized databases used to train predictive algorithms. The rise of the smart-phone and the rapid digitization of China create a teeming sandbox for even a mediocre data scientist, let alone some of the world’s best. The government is paying attention. For the first time in history, centralized decisions may actually scale to allocate resources as efficiently as a market economy. AI is an essential tool to craft Xi Jingping’s 21st Century China. Beijing at the core, plugging into centralized platforms with massive distribution and fine-tuned algorithms to incentivize behaviors on a mass scale in line with the party’s vision.
This is interesting, because the US has made plenty of progress in AI without an authoritarian system. We have many companies that collect staggering amounts of data, mostly to target ads, so our training sets are vast. And since ad dollars subsidize more data collection, it’s possible for a non-authoritarian country to end up with more data because, while collecting it isn’t a national priority, it’s a side effect of wealth and we can afford it.
That only holds true if the data collectors are allowed to use their data; it may turn out that there’s a tradeoff between protecting Internet users' privacy and winning the AI race. And that means the US and Western Europe still have centralized decisionmakers determining whether or not we’ll succeed in AI, just through different means. In one sense, any decentralized order requires a centralized substrate, and the more decentralized the approach is the more important it is that you can count on the underlying system. This is one of the left critiques of neoliberalism: it’s hypocritical to say that free trade and deregulation make the government less important, because they make a government that reliably enforces contracts far more systemically important.
China’s (Internal) Immigration Reform
The Economist has a great piece on China’s hukou system ($) of intra-country passports. Essentially, government services are tied to hometowns, not current residences, so migrant workers within China are basically guest workers. One important note:
[China’s cities] generally offer four routes to a local hukou, much like the pathways to immigration in Western countries: investing in a local business; buying a home; having a degree; or holding a qualified job.
This is very important to keep in mind when looking at China’s education system, and especially the real estate market. Buying a house in China isn’t just buying shelter, it’s also getting an entire package of services—education, healthcare, pensions. That makes price-to-rent ratios in Chinese cities misleading, but it also means the highly-levered Chinese property market is partly a bet on government spending on social programs.