In this issue:
Artificial Urgency as a Forcing Function
Frontend Omnichannel, Backend Specialization
Another Business Mafia
Cloud Computing: Cheap Because it's Expensive
As a rule, the more complicated a project is, the harder it is to create a remotely reasonable estimate for how long it will take. Dependencies accumulate, bottlenecks arise, delays at one stage ripple through the entire system. And that's for projects whose steps are known in advance: if a plan hinges on solving a problem that's not logically impossible, but not specifically complete, the timeline can be arbitrarily lengthy. Fermat's Last Theorem was posed in 1637, and the proof was completed three centuries later. A project the relies on a new proof might take months, or it might take generations.
So there should be a direct relationship between how ambitious a task is and how uncertain its completion date. Which makes it surprising that so many success stories involve completely arbitrary deadlines:
The Manhattan Project essentially had two artificial deadlines: at first, the goal of building nuclear weapons before Germany did (to physicists in the US, it was very big news indeed that Germany had banned uranium exports before the war). Later on, part of the goal was to demonstrate nuclear weapons in wartime before the war was won by conventional means—an idea deeply uncomfortable to many people working on the project.
JFK's 1962 speech at Rice University set a goal of landing someone on the moon "before this decade is out," which was a lot more euphonious than "first" or "soon," but which was only theoretically possible at the time.
In 1909, AT&T decided to perfect cross-continental phone calls within five years in order to use them as a PR stunt at the San Francisco Exhibition scheduled for 1914.
I, Claudius, one of the best novels of all time, was written in a few months to pay off a debt.
More prosaically, events like Google I/O and Apple's Worldwide Developers Conference have an artificial regular cadence, and regularly involve the announcement of products that were technologically or economically infeasible a year or two beforehand.
Uber's full-app rewrite in 2016 was scheduled so the app could be tested and globally rolled out before Apple stopped accepting app updates around Christmas.
Facebook finished FB Platform just a few hours before its launch at f8 in 2007, helped along, according to The Facebook Effect, with a gray-market alertness-enhancing drug more frequently used by the military. (40-hour periods of nonstop work offer certain competitive advantages!)
More frequently, in equity research there's a pattern where analysts quickly ramp up their expertise on a company and industry ahead of earnings. Earnings are not just an incremental puzzle piece in determining a company's valuation; sometimes, the exact numbers matter less than more open-ended data involving shifts in strategy, pricing, or marketing. It's very hard to make money being the first to react to a number with known importance, but there's a lot of money in being the first to integrate new qualitative data into a thesis. Sometimes, a stock will react one way based on earnings, and, over the next few minutes, hours, or days, end up moving in the opposite direction due to qualitative commentary. (Facebook, for example, moves instantaneously on news about revenue and margins, but on more of a delay when the company talks about usage metrics, or, especially, its strategy around what kinds of usage it targets.) A more robust mental model of a company's economic drivers will lead to a more accurate financial model of the next few years.
What all of these deadlines have in common is that 1) they have basically nothing to do with the nature of the underlying project, having been entirely imposed by external forces such as war and Apple's vacation schedule, and 2) they're associated with surprisingly successful outcomes.
It shouldn't work this way at all: an external constraint (do X and Y, and by the way subject to rule Z) necessarily eliminates options, and if successful management consists of choosing the best among the available options, then reducing the number of options should reduce the odds and magnitude of success.
Of course, that model only holds if the approach is to consider every option and choose the best one. In practice, there's a cost to considering options; it takes a nonzero amount of time to make choices, and given the usual uncertainty in estimating deadlines, there's limited utility in asking for 95% confidence intervals for a bunch of hypothetical unknown-unknowns. In practice, the Gantt charts and precise timelines only come into play for the Nth iteration of something, not the first version.
Artificial deadlines have two useful functions: first, they set a higher motivation baseline, and the effective morale of a team is disproportionately weighted to the most dispirited member. A fake deadline is a decent reason to quit, but if you get one and don't quit, meeting it is a matter of pride. Second, artificial deadlines for one project help prune the rest of the organization's decision tree. Since the artificial deadline is hard to negotiate, it makes the rest of the organization ultimately deferential to that goal. So it's a way to make it politically feasible to concentrate more resources towards one aim, instead of having a conventionally diversified portfolio of projects.
This motivational/political dimension requires a real deadline, not a completely fake one. This is one reason the successful outcomes tend to come from outside arbitrary pressure, rather than internal pressure. Internally, it's easy to see through artificial goals, and push back against them, but if the goals are set by someone who isn't an expert but is in charge, or who is also dealing with external constraints they can't affect, then the deadline becomes a premise rather than a conclusion, and everything else reorganizes around it.
So the upshot is not to randomly set artificial goals; that's a good recipe for burnout and mistrust throughout an organization. Instead, it's to identify cases where there's a plausible solution to a problem at hand, but that solution requires a fixed timeline. And then mark your calendar.
Frontend Omnichannel, Backend Specialization
Ben Evans offers a clever way to divide up the retail industry: instead of focusing on offline versus online, an increasingly blurry distinction, focus on shipping modes. One reason Amazon started with books is that the products are diverse (so searching, sorting, and ranking are valuable) but they're physically very similar, so it's possible to achieve economies of scale on shipping early on. Other products are more heterogeneous, like flowers (a delay that pushes delivery from February 14th to the 15th has a material impact on customer experience) or furniture (not only are the product sizes and weights variable, but the last smidgen of the Last Mile can involve extra complications). Among other things, this framing makes the Wayfair story clearer: they were competing with Amazon, but with a piece of Amazon that didn't have nearly the same scale advantages as the overall business.
Another Business Mafia
Business mafias are an important topic, because identifying one early is a scalable and repeatable way to make above-average investing and career decisions. I've written before about Paypal and Tiger Cubs, both of which shared a few common traits: high performance at the original company, a headcount close to Dunbar's Number, and a catalyst that led to a large number of employees leaving at the same time without suffering reputational harm. (Businesses that fail generally don't produce the same high-value alumni network.) There is, as it turns out, another such network in the railroad industry: six of the seven largest railroads in North America have senior executives who worked under the same CEO at Canadian National Railroad or Canadian Pacific ($, Globe and Mail). While railroads don't have the same low headcount as the other businesses mentioned, they did have a similar catalyst: that CEO, Hunter Harrison, died suddenly less than a year after taking the CEO job at CSX.
Business mafias are a partial answer to a broad market inefficiency: management best practices don't instantly transmit to other companies, even when they're widely known and are assumed to be good ideas. One possibility is that friendly, gregarious people are more likely to get promoted, and less likely to want to shake things up; a related possibility is that it requires some critical mass, since business culture norms only work when they're widely adopted. A cohort of executives who all come from the same company and share the same approach to business might be necessary to transmit the culture from one company to another; too few, and the corporate culture antibodies reject the transplant.
Cloud Computing: Cheap Because it's Expensive
A16Z argues that the cost of reliance on cloud computing is up to half a trillion dollars in aggregate market cap because of the pressure it puts on companies' incremental operating margins. In one sense, that's true: all else being equal, higher margins are better, and at scale the cost of renting rather than buying compute starts to add up. The same high incremental margins that make cloud companies' economics so attractive come right out of the economics of their customers.
But the big companies with the biggest cloud budgets are generally tech companies that pay their executives with a healthy amount of equity; if there's free money available, why not take it? That half-trillion in market cap improvement from higher margins can be inverted: it implies that the cost of diverting engineering resources away from product features and towards reimplementing the backend is even higher. Since companies add headcount before that headcount produces revenues, the general pattern is that growth companies will at some point overexpand, realize their addressable market isn't quite as big as they thought, and have to retrench; if they're overstaffed for the growth they thought they had ahead of them, they can focus those people on cutting costs instead. But markets are good at reading this kind of signal from management: a company for which the best use of talent is cutting compute costs rather than adding revenue-accretive features is a company that will offset lower valuation multiples with higher margins. One of the A16Z case studies demonstrates this: Dropbox was able to raise gross margins from 33% to 67% in two years by switching to their own infrastructure—but in the three years since then, annual sales growth has decelerated from +31% to +15%.
Above Avalon argues that Apple has a 5+ year lead in wearable computing, mostly due to better chips. This is always hard to judge; Apple has demonstrated to other hardware companies that they either need low prices or jaw-dropping launches, so the state of the art is either a few years behind (and cheaper) or a bit ahead (and not released yet). But taking it as a given, it's a testament to the value of Apple's increasingly full-stack approach: the more they can offer features that are limited by silicon, and also control the supply of that silicon—as big and reliable customers, they're unlikely to get bumped back in the queue when there's a fab shortage—Apple can commit to lengthier product roadmaps than other companies. And since new form factors often start out closely tied to old ones (the original iPod, for example, was Mac-only), Apple can use its existing install base to keep expanding to new markets.
One of the darker reports you can read online right now is this 2019 report assessing countries' preparedness for a pandemic, which ranks the United States at #1, although it notes that the UK would be better at "rapid response to and mitigation of the spread of an epidemic." Predicting the impact of extreme events is necessarily hard; all the one- or two-standard deviation stuff is being taken care of, so anything weird enough to matter has to be very weird indeed. As it turns out, preparedness is only measurable in retrospect.
This WSJ report ($) shows that nine of Texas' 13 black start generators, and six of fifteen backup generators, had problems during February's power loss. There doesn't seem to be an adequate solution to such risks other than fairly frequent drills. This is expensive, but cheaper than actually restarting the grid given a complete failure. Using continuous, artificially-induced failure in production systems is a good way to keep streaming movies reliable, and something similar might work for more important services.
ByteDance's lamp with built-in cameras to monitor kids' studying has been a hit in China ($, WSJ). Education tends to get pricey in countries that get rich, because it's been hard to scale. The converse of this is that in countries with small families, there are often two parents and four grandparents for every school-age child, which means products that are the complement to parents' and grandparents' time can be a big hit.