In this issue:
- FP&A as the Telemetry and Counterespionage—Good financial projections won't save a broken company and don't make a huge difference for a sufficiently good business that's hard to mess up. But there are particular kinds of companies that benefit from having a very keen sense of what a customer is worth, what it costs to get one, and how these numbers will evolve over time.
- 0.5x AI—A is probably off the table, but the biggest labs have indirectly coordinated to slow down the release of the most powerful models.
- Media Incentives—The Twitter algorithm competes with Twitter accounts that repost other people's material. It's very hard to beat them, but a good first step is to stop subsidizing them.
- Infrastructure—Building the infrastructure for well-behaved bots.
- Perks—The Willy Wonka Principle: if you have something unique enough that it's hard to measure the market price, make access a lottery where every economic interaction that's valuable to you constitutes another ticket.
- Externalities—Smart companies outsource tasks outside of their core competency. And that's why, when Meta does what a politician asks them to do, they make it very plain that this politician is the one they want defending the results.
Talk to this post on Read.Haus.
FP&A as the Telemetry and Counterespionage #
The two most common grades a company gets for managing its finances are an F and a gentleman's C. A company gets a failing grade if it had a reasonable business that could have turned a profit, but managed to finance it the wrong way so it ran out of cash. And there are many companies for which you can say all sorts of positive and negative things about their business, and where the best thing you can say about their capital allocation is that they didn't mess up too badly—they didn't do any ill-advised mergers, they gracefully responded to diminishing opportunities to invest by returning more capital to shareholders, they didn't structure some part of the business so GAAP profits were misleadingly front-loaded while later costs more than devoured those gains, and they weren't buying back stock at $30 and issuing it two years later at $5.[1]
And then there’s companies for whom granular financial analysis—at the level of products or end customers—is the defining feature of the business. There are many companies for which the gap between sounding good as a concept and actually working as a business comes down to these granular questions: what fraction of first-time customers convert into regulars, and how do their margins evolve over time? How long do people stay subscribed, what makes them leave, and how will they react when prices change? When this big piece of capital equipment gets sold, how much will the customer spend on maintenance, what margin will that produce, and how likely are they to switch to some cheaper provider? How is foot traffic likely to change over the duration of this lease—and, if whoever's getting that lease drives additional foot traffic to other nearby locations and increases the value of nearby real estate, how airtight can the case for this be?
In one sense, every company faces these problems, but can't necessarily unbundle these decisions from everyday management choices. Professional services companies of many types have unit economic constraints, but they don't have the sample size necessary to do especially sophisticated modeling, and sometimes other strategic imperatives make the choice for them. If an accounting firm decides it's not going to audit shakier companies, it's also choosing a longer sales cycle, better customer retention (in the two senses that big companies don't like to change auditors and in the sense that big companies are less likely to go bankrupt), access to better talent, etc. And if that accounting firm decides it's going to cross-sell consulting, that process is also going to be naturally fuzzier, with slower feedback, and subject to constraints that don't neatly fit into a spreadsheet. [2] These businesses can't be completely indifferent to unit economics, but they're usually not thinking "let's preserve our net dollar retention." They're thinking "uh-oh, we just lost the Walmart account."
Whereas if you visit Amazon's office the day after somebody canceled their Prime subscription, it's not so funereal. Customers come and go, on average, though of course Amazon wants them to stick around. But, at a deep level, Amazon's retail business can work backward from the question of: what does a new user spend, how does this spending ramp up, and what are the ways to make them start a larger fraction of their shopping efforts at Amazon and to add a few more things to the cart when they're done? Solving for this underwrites investments in new fulfillment centers, ads, new product lines, inference spending on their AI shopping assistant, and all the rest. (There are problems to solve on the merchant acquisition side, too, but every time customer count goes up by 1 or the net present value of the contribution margin from the typical customer goes up by $0.01, that problem gets easier to solve.)
In this wonderful Eugene Wei essay, which is about data visualization in roughly the same sense that Jurassic Park is a story about sometimes it's hard to take kids to the zoo, there's a wonderful line about this:
Late one night in 1997, a few days after I had started, and about my third or fourth time reading the most recent edition of the Analytics Package, a monthly report on every detail of Amazon's financial performance, mostly consisting of a few hundred graphs cover to back, I knew our hidden truth: all the naysaying about Amazon's profitless business model was a lie. Every dollar of our profit we didn't reinvest into the business, and every dollar we didn't raise from investors to add to that investment, would be just kneecapping ourselves. The only governor of our potential was the breadth of our ambition.
Financial markets in the 90s were capable of underwriting businesses that continuously reinvested in order to grow. This being the era of 15% annualized earnings per share growth targets and multiples far too high to make that achievable through buybacks alone, every growth company was reinvesting something. But if they were buying something tangible, it showed up in the P&L as depreciation. Amazon's investments were economically similar, and had different accounting treatment for good reason, but they meant that the economic engine of the company was described to investors qualitatively, and quantification involved repeat-order share rather than pure cohort math. And an explicit plan to aggressively invest in the business in order to earn profits in the future was a pitch made by many other companies that ultimately didn't do that. Amazon did tell investors this, but on a lag: for example, the headline for their Q4 1999 earnings release was Amazon.com Announces Profitability In U.S.-Based Book Sales, Financial Results for Fourth Quarter 1999; solve for solvency, optimize around cohorts, just as it should be. (But a cynic could have pointed out that a high fixed-cost retailer trumpeting the profitability of a subset of their business during the busiest shopping season of the year was a great way to downplay the fact that operating losses had risen from $18m to $175m year-over-year.)[3]
From Amazon's perspective, the GAAP-to-reality gap was a blessing for as long as investors believed in the long-term vision: they had to know that the cost of acquiring new customers would rise over time, that finding the right product for someone is easier if you know what they've bought before, that brands accumulate value over time, and brands that thoroughly solve a recurring problem tend to become verbs in a consumer’s lexicon. [4] Knowing all that, and knowing that investors were willing to underwrite the general model without asking too much about the specifics, it was perfectly sensible for them to keep things vague. There's no sense in sharing their master plan to capture just one market and risk inspiring other people to be the Amazon of DVDs, electronics, and other categories they entered. The people who figured this out were either trying to reverse-engineer Amazon's survival and growth after the dot-com collapse, or licking their wounds and trying to understand how Amazon had beaten them. This wasn't the only problem Amazon had to solve—they were an early enough dot-com that the list of ecommerce-related technical problems they had to deal with was pretty close to "all of them except inventing Unix, C, and Perl." But a keen awareness of the importance of cohort economics, and a ruthless commitment to growing the size of the new ones and the wallet share of the rest of them, was basically what kept Amazon's technology, logistics, and customer service engines humming at full efficiency for so long. A company is in a good position if they're able to spend incremental money on growth and the net present value of the resulting contribution profit is meaningfully higher than what they spent. They're in a great position if the way this shakes out in their financials is that they look like a money furnace, and people assume that competing with them means being better at snookering investors rather than doing whatever it is that the company does.[5]
All this is not to say that making decisions informed by detailed quantitative analysis is always strictly a good thing. When Bob Iger joined Disney by way of the ABC/Capital Cities acquisition, he was increasingly annoyed to find out that many of his decisions were second-guessed by Disney's planning team, which imposed a slow and uncertain veto on decisions that could have been made directly by business operators.
GE is a more pathological example, where planning aggressive targets around unrealistic quarterly goals encouraged the business to repeatedly make value-destructive decisions. Planning and finance are useful functions, but just like any other abstraction, they can swallow a company whole. GE's model looked amazing for a while: sell durable equipment, turn that into a predictable stream of maintenance revenue, turn the sale itself into recurring revenue by helping to finance it, etc. But once they had a big enough finance division, there was no end to the number of ideas that looked good on paper, and the returns from esoteric lending are more stable quarter to quarter than the returns from selling consumer durables or capital goods. But not every quarter is any given quarter, and GE ended up being an over-levered specialty credit hedge fund with a few manufacturing businesses bolted on.
FP&A is a kind of business telemetry tool, which, like any such tool, is a lot more valuable when it aggregates lots of backwards-looking information in order to inform future decisions. Which means it's one of those tricky information problems where the meta problem is figuring out how much you can be certain of. When it works well, it's partly a sign that a company has a good model that hasn't yet developed into a visibly good business. [6] So beyond the pass/fail binary, FP&A is particularly effective at helping stealthily great businesses maximize their lead before anyone catches on. It's a weird mix: a lot of finance, a fair amount of strategy, and, in the end, a company that gets an A+ for effective counterintelligence.
Disclosure: long AMZN
Technically this last one is situational, and you can imagine cases where it's the right call. A commodity company at the peak of the cycle might decide that the risk-adjusted return on investment from adding more capacity is low, but that cycles exist and they shouldn't commit to a dividend. So for them to do a buyback is not a market-timing signal, just a tax-efficient alternative to a special dividend. And, when that cycle turns, they might see a better opportunity set, particularly if their competitors are doing distressed liquidations. If you determine that the cost of issuing more equity capital is 15%, but that your return on capital is 20%, then you
shouldissue stock even if you previously bought it back at a higher price. But you'd also better be sure you're right, because the optics are terrible.↩︎Where, for example, do you put the line about
spending a few years publicly arguing about whether or not to split the business into separate consulting and auditing companies, before ultimately failing to do so? Auditing firms might be a special case; for risk-averse reasons that make a lot of sense to accountants, they wanted a wide distribution of veto power, butas various polities have discovered in the past, that means that a decision can't just be good for the firm, but has to be good for everyone who might veto it, in which means nothing gets done.↩︎Amazon in general has a habit of emphasizing tacit ten-year guidance over anything that would be especially helpful in tracking various KPIs and building to a quarterly number. And they used to absolutely
bury us in completely decontextualized information after Prime Day("US customers bought 1.2m pairs of sunglasses and 1m swimsuits? Holy cow, better update rows 13,832 and 48,397 in my model!"). Thosegradually got more vague, and this year it looks like they didn't do one at all.↩︎Unfortunately, "Amazoned" is a polymorphic verb that might mean: undercut on price to a competitor, subjected to bare-knuckles negotiating tactics to an acquisition target, or worked past the point of burnout as an employer. Maybe one reason is that the big brands-as-verbs imply an interpersonal context: "I'm Ubering there" matters to someone who wants to know why you're late, "I Googled it" means something so someone who wants to know how you know something, etc. Whereas in retail, Amazon dominates the categories where people generally don't say "Where did you get that?" unless it's some weird gadget with a hyper-specific use case and a brand name like "Mosptnspg," in which case you don't have to ask.
↩︎It's been nice for tech companies that some of the expenses that are fundamentally capex—like sales, marketing, some of their R&D spend, and the cost of integrating a product with customers' existing setup—all get booked as opex. AOL infamously hated this enough that they just capitalized their marketing costs, which was more accurate economically but which made their financials look better than competitors, and which didn’t follow the fairly strict accounting rules for when this was allowed. They ended up
writing off $385m of this in 1996, and paying a $3.5m fine in 2000.↩︎Amazon has lots of alumni from its finance function, including the CFOs of Expedia, Roku, Chewy, and Amazon itself. Google, with its higher margins and historically asset-light business, could think about things from more of a broad usage-maximizing perspective, knowing that incremental search volume more than solved their operating cash flow needs. They still do this, and Google's finance function has produced a few CFOs, albeit of smaller companies like Hims & Hers, but the big decisions seem less like the output of Excel and more the result of a
Sawzallquery showing the results of some A/B tests.↩︎ You're on the free list for * The Diff*. Last week, paying subscribers read about
crypto as a global bug bounty program, and how LLM-driven hacking affects this($); how AI discourse is driven by power users overestimating how much AI benefits everyone else, and AI opponents being less aware of how much it affects them($); and
how getting the most out of intelligence involves finding the maximum amount of stupidity you can get way with($). Upgrade today for full access.
Upgrade Today Diff Jobs
Companies in the Diff network are actively looking for talent. See a sampling of current open roles below:
- AI Transformation firm with an ambition to build an economic world model to run swathes of the private, unstructured economy is looking for a security engineering lead to ensure their transformation platform is secure, compliant, and governed. Experience hardening mission-critical software infrastructure in regulated industries (HIPAA, SOC2, etc.) a plus.
- Well-funded, frontier AI neolab working on video pretraining and computer action models as the path to general intelligence is looking for researchers who are excited about creating machines that learn from experience, not text. Ideally you have zero-to-one pre-training experience and/or are a high-slope generalist who’s frustrated that the big labs aren't doing this. (SF)
- High-growth startup building dev tools that help highly technical organizations autonomously test and debug complex codebases is looking for someone who can help scale and manage their rapidly expanding facilities footprint: everything from lease negotiations, to proactively anticipating / catalyzing new office acquisition, to making sure those offices are stocked with the best snacks. If you want a seat on a rocketship and enjoy fixing things that break–literally and figurately–this one is for you. (SF, DC, London)
- Lightspeed-backed team building the engineering services firm of the future is looking for founding members of technical staff excited about working alongside civil engineers to translate their domain expertise into the operating system that powers the next era of great American infrastructure. If you’re an engineer with strong product intuition, who's energized by access to users, and excited by the prospect of transforming how we design and construct our built world with frontier AI, this is for you. (NYC, SF or Remote)
- Ex-Anduril, Ex-Abnormal Security, Ex-Bridgewater, fast growing startup bringing agentic cybersecurity to 99% of businesses via MSPs is looking for platform and machine learning engineers. Startup experience preferred; what matters most is that you've grown in scope and handled ambiguity over the last few years. (SF)
Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up.
If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority. And: we're now actively deploying capital into early-stage companies through Anomaly. Our focus is on defense, logistics, robotics, and energy. If you'd like to chat, please reach out.
Elsewhere #
0.5x AI
The first big open letter arguing for a in AI development came out shortly after GPT-4, and argued against making anything more powerful than that. Since then, models have gotten vastly more capable, with surprisingly few incidents that rise above the background noise of either how new technologies experience hiccups in their development or how new kinds of media can be a dangerous tipping point for people with mental health issues. And the models have been getting more aligned in another sense: a great deal of capital has been sunk into GPUs, buildings to house them, equipment to power them and cool them, etc., and the providers of that capital want AI to be a net-helpful consumer product or enterprise tool, not something that buries their investment under a pile of lawsuits.
But recent software breaches—those being the de facto boolean eval for frontier models—have illustrated that these tools are smarter and more amoral than we necessarily want (or that the labs don't have adequate procedures for dealing with them). On the technical side, there isn't a precedent that combines the risk profile with scale, but on the organizational side, the nearest analog in kind if not severity is somewhere between the Wells Fargo upselling scandal and various war crimes committed by militaries that don't do that as a matter of course. Large organizations will sometimes have top-down mandates that, when executed to the letter with a modicum of moral flexibility, lead to disastrous unplanned outcomes. But now, anyone with access to a sufficiently powerful model is in the position of someone a few levels up in these organizations, who is setting high-level goals that get converted to intermediate projects and then into specific tasks.
So now, OpenAI is pausing some internal tests of its new Astra models until it's sure that they won't launch some disastrous hack as a shortcut. What's especially interesting about this is that the labs have slowly, indirectly, coordinated to moderate the pace of AI development, or at least ensure that it initially gets deployed in a defensive capacity—both in the sense of finding and fixing vulnerabilities before they're exploited, and in the sense of being applied under military or intelligence supervision. Usually, the evolution goes the other way: one resolution of satellite photos lets you look at a picture of your house from space, and higher-resolution ones for military use. But AI is an unusual technology, and if it's rewritten some rules entirely it's not crazy that it does some in reverse.
Media Incentives
Twitter (currently X) is revamping their revenue-sharing program to reward only original content. They have learned very well the lesson that if users get rewarded for attracting attention to themselves, they'll act as a replacement for the algorithm that's supposed to surface interesting content—but one with a higher discount rate than the company would like. So, they're trying to pay people for creating content, and then to get paid for having a feed that recommends good content interspersed with ads. They've also set a higher usage threshold for participating, which is both a way to ensure that they're mostly rewarding people who are already fairly popular on Twitter, but also a way to lower that discount rate by forcing people to invest effort into their account before they can convert it into money.
Infrastructure
Cloudflare's role as an institution is to look for the missing or mis-specified protocols, and then write good ones instead. One recent example of this: they have a new browser, Kitesurf, that's designed to be used by AI agents. Right now, agents have mixed incentives to behave: the big labs aren't quite at a point that they're directly delivering material revenue to the publishers and retailers they link to, though their share is growing all the time. If chatbots are a low share of some site's revenue, but a high share of its usage, there's an incentive to cut them off. And agents may not care at all. Managing this kind of issue is exactly the kind of problem Cloudflare has taken upon itself to solve. So if they can build a browser that's more efficient at collecting public data than the typical agent, they can also encourage usage that keeps Kitesurf whitelisted by default even if competing alternatives all get detected and blacklisted.
Perks
The market for restaurant reservations is getting more efficient, and, as is the case any time you think of some category without using the phrase "the market for..." that efficiency takes the annoying form of making the most desirable table/time tuples unavailable except to power users ($, WSJ), and, increasingly, people with lots of spend on credit cards that have some kind of connection to reservation platforms. Call it the Willy Wonka Principle: if you have something truly scarce, and you want to price discriminate but you can't imagine how much price discrimination you can get away with, what you should do is make access random, but weighted by propensity to do something economically valuable to you.
Externalities
Meta is trying to address the unpopularity of datacenters by just covering the costs of most of the externalities people are worried about. This is a very Coasian approach: if datacenters use energy, let them subsidize other energy users so it doesn't affect them; if they use water, let them pay for it so nobody else needs to adjust their behavior; if they aren't connected to the local economy, connect them by subsidizing jobs like teachers, firemen, cops—any job that adheres to the Richard Scarry rule ($, Economist). This also encourages datacenters to be more efficient, since they capture the upside directly instead of reducing a moderate nuisance for someone else. Meta is very much framing this as a politial decision: they mention Texas governor Greg Abbott four times by name in the announcement, always favorably. Which is a nice PR move for them, because the message to Abbott is: we gave you everything you asked for, and it's your job to make sure the voters appreciate that.
Disclosure: long META.