The Diff is a newsletter exploring the technologies, companies, and trends that are making the future high-variance. Posts range from in-depth company profiles, applied financial theory, strategy breakdowns and macroeconomics.
## P&L Beta: How Much You Keep out of How Much You Make
If you've ever had a job where you're directly responsible for revenue, and you see how much of it you make—whether that's working in sales, being the person the sales team brings along to explain the product, managing revenue and costs in a subset of a business, contributing to investment decisions—you might be inclined from time to time to calculate your P&L Beta, i.e. for every dollar of revenue that you personally cause your employer to earn, how many pennies flow directly to you?
As a rule, if your compensation package doesn't explicitly include a number that corresponds to this—sales commissions, a negotiated P&L split, a bonus structure with quantifiable hurdles—then the number will be disappointingly low. Which makes it sound like a classic instance of companies using their negotiating leverage against the median employee to redistribute upside to themselves, except that it's also happening in plenty of intra-company competition. Heinz cares about the quality of the tomatoes that go into its ketchup, but the Heinz label is doing a lot of the work, and whoever provides the raw materials doesn't have any room to extract a markup just because their product is being used to make a recognizable branded product rather than a discounted white-label one. (In fact, mass-market consumer food companies have an incentive to go a bit downmarket for their inputs: if there really is a best tomato producer, and Heinz can charge even more by using them, the problem for Heinz is that this upcharge is directly attributable to exactly one vendor, who will then try to charge a premium equivalent to whatever price premium they enable.)
One place where this shows up is in investment research, particularly in which kinds of investors can rationally buy lots of data, and build out the infrastructure to analyze it. There are many opportunities of the form "Spend $x on additional data, make some amount greater than $x in incremental P&L." But what does the hierarchy look like?
- For extremely short-term oriented traders, this is a straightforward optimization problem. You can, for example, calculate what fraction of the time you come in second place trying to make some specific trade, estimate how much money you'd have to spend to get to first place, and decide if buying those FPGAs or hiring that C++ wizard is really worth it. And once you put this into action, you know pretty quickly whether or not you were right.
- For a pod shop, there's also a short turnaround time and the whole model is designed to isolate the results of skill from the results of luck. They can't quite do the same fine-grained estimates that work at lower frequency, but they can justify substantial budgets for research and data, especially for real-time signals of fundamentals.
- A long-only has a harder time doing attribution, because in any given year the vast majority of the variance in their returns will come from how the general market is doing. Since the excess returns they're betting on are typically realized over longer periods, there's also short-term variance in how much alpha they produce. If your performance is a point or two about the market, plus or minus 10% each year, and the asset class you're betting on has some return +/- 20% each year, it takes a long time to put a good estimate on how much alpha you're producing, and it's that much harder to attribute returns to any given source.
- For venture, the calculation is even harder; the legal life of a fund is probably longer than the useful life of the median investment-informing dataset, so even if you're fully convinced that whatever specific dataset you relied on materially contributed to making whatever decision ended up returning the fund, by the time you try to use it again either a) it won't exist any more, or b) it'll be something that everyone else looks at, too.
This is pretty annoying! If there's some way to gather new information in order to make the relative pricing of S&P 500 ETFs and S&P 500 futures 0.01% more cost-effective, it's fairly easy to underwrite the decision to invest in this. But if there's a source of data that might identify one of the handful of generational companies that keep global economic growth humming along, create tens or hundreds of thousands of jobs, and justify the existence of venture as an asset class, it has to be justified on more whimsical grounds.
You can also think about this problem by reversing it, and looking at careers where compensation is less directly linked to revenue generated. For example, here are some ways a software engineer could maximize the amount of revenue that can be reliably attributed to their actions:
- Don't write any documentation whatsoever. Just code.
- Use whatever coding standards work for you, whether or not this makes it hard for other people to maintain what you've written.
- Under no circumstances should you devote any time to training people more junior than you. (But do feel free to ask a colleague for help when you need it.)
- Avoid spending any time on email or meetings that involve telling other people what you're working on.
- Push everything to prod as soon as it looks good to you.
In any given week, doing whatever of these you can get away with will be a huge productivity boost. Over time, it inevitably means taking down some important product and needing a long time to get it back up and running. Which is part of the incentive problem that compensation is trying to solve—in jobs where the upside can be directly and immediately linked to one person's performance, but the downside link is more tenuous, it's a better idea to focus on standards for inputs than to directly focus on outputs. That's the only way to credibly avoid funding the asset of working code with the off-balance-sheet liability of tech debt—and if you aren't measuring tech debt at all, you'll inevitably rely on it to excess.
The corollary to all of this is that if you're optimizing for value capture, you have a few steps you can take: target customers who are better at measuring their results, try to sell to people who have P&L responsibility, and adjust pricing down when these conditions don't hold. There are many products for which that won't apply, and some industries where pricing power doesn't really exist—if there are futures for whatever you're selling, your margin advantage has to exist entirely on the cost side. But that, too, is fine; plenty of the economy functions perfectly well with indirect attribution. But that's not something to bet a business on forever. In many industries, the pace of turnover in the list of leading companies is a function of how well results are getting measured, and which businesses have adapted to it. In investment banking, the shift from mostly low-risk intermediation in the mid twentieth century to risk-taking from the 70s through the financial crisis was tough to navigate for the banks that built their business on hiring bankers who went to the same boarding schools as the CEOs of their clients. It was an easier transition for companies whose competitive advantage was being willing and able to bid a quarter point higher than the competition for a particular bond. Advertising had its own transition like this: any time measurement got better for some category of direct response, the companies that insisted on metrics-free brand advertising had their business squeezed. There's a lot of room for error if your job or industry has low P&L beta, but that isn't a permanent trait for either.
## A Word From Our Sponsors
Runway is life for startups, so why do most banks chip away at it with fees, minimums, and delays? Founders deserve better — that’s why Brex built a banking solution that helps startups take every dollar further. With Brex, you get the best of checking, treasury, and FDIC protection in one account. Send and receive money fast worldwide. Earn industry-leading yield from your first dollar — while being able to access your funds anytime. And protect your cash with up to $6M in FDIC coverage through program banks. Ready to join the 1 in 3 US startups using Brex and take every dollar further?
Learn more at Brex.com
## Elsewhere
### Narratives and Volatility
Bear markets aren't made in a day, but it only takes a few days to produce a bear market narratives. US markets are a bit off their high, Japanese stocks are having their worst day in percentage terms since 1987, crypto's also having a bad day, etc. Looking at recent history, a low-rates/low-growth environment is pretty good for big tech valuations, and big tech remains a big chunk of the S&P. But it's important to note that one effect of AI is to shorten the duration of these companies' cash flows: to the extent that their earnings are relatively less dependent on recurring income from monopolistic traditional software businesses, and more dependent on getting maximum returns from investments in rapidly-depreciating GPU hardware, the present value of those cash flows becomes more sensitive to overall economic growth and less sensitive to lower rates.
### The New Acqui-hire Model
There have been a few AI deals that take the form of hiring senior members of the team, possibly partly cashing out investors, and letting the remaining staff run whatever's left: Inflection, Adept, and the proposed acqui-hire of the OpenAI team all fit this model. The latest instance of this is Character.ai's founder, Noam Shazeer, returning to Google along with several other members of the team. Per Bloomberg, they're getting cashed out at a premium to where Character last raised funds. Meanwhile, Character itself sees "an advantage in making greater use of third-party LLMs alongside our own," i.e. becoming a wrapper company.
Wrappers can work, sometimes quite well; Google did just fine as a thin layer wrapped around the Internet itself, though it helped that they found the exact right thing to wrap. These deals often get cast as a workaround for antitrust restrictions on acquiring big competitors outright, but that's always going to be speculative: to the extent that it's the justification, it was never discussed in any legally discoverable medium, so we'll all be arguing about circumstantial evidence forever. Meanwhile, there is a decent reason to expect deals like this, given the structure of AI companies: they're capital-intensive, both because they're hiring expensive talent and because they're using pricey hardware. But there's a lag between when model training starts and when the new model is available for testing, so for any given participant it's never clear if they've calibrated their investment to be 150% of what they needed, or 95%, i.e. not so much better that it's a competitive advantage. In that environment, a company's capital and staffing needs can change wildly once they conclude that existing models can be tweaked to serve the purpose their models are built for, and that they have product-market fit for how they present models rather than exactly how well those models work, it suddenly makes much less sense to have all that hardware and all those research staffers. There is still a viable business, just not the one they thought there was. It's not an especially uncommon corporate origin story—Harley Davidson became an independent company when some employees of its corporate parent, American Machine and Foundry, decided to buy it out from them. Sometimes a company is, for sometimes arbitrary reasons, no longer the optimal owner of some of its assets or the ideal employer for some of its people.
### A Good Trade
Warren Buffett started buying Apple in early 2016, with a cost basis of around $34/share. In the last quarter, Berkshire sold nearly half its stake, after cutting 13% in Q1 (Apple averaged $180 during the quarter). It was a pretty good trade; earning 20%+ compounded, and in dollar terms it makes Buffett one of the most successful tech investors of all time. Part of what underwrote that investment was that Apple was becoming a more legible company: it sold a consumer product, could earn a price premium for similar hardware because of its brand name, and had lots of levers to pull to keep margins high. But that evolution towards the conventional means that it was also going to be subject to more conventional analysis, and at some point a portfolio that's massively overweight a low-growth, high-ish multiple stock—even in an extremely well-run company with valuable intangible assets—isn't the safest setup.
What this also highlights is that one source of Buffett's alpha is his flexible mandate. The Buffett portfolio is much more cash-heavy than it was a few quarters ago, and permanent capital vehicles don't have to worry about redemptions, whether they're from irate investors annoyed by a big cash position in a bull market or by investors worried about high exposure during a downturn. It's easier to maximize absolute dollar returns if you have a choice about how fully invested to be.
### Terms
Being a credit investor is partly a game of analyzing the fundamentals of companies, and getting some sense of whether or not they'll be able to pay their bills. But it's also a question of knowing the details of how debts are structured—a bond is a simple concept with a complex implementation, and some of the implementation details can have surprising consequences. As a result, credit investing is increasingly a business of reading covenants rather than balance sheets and cash flow statements, and about finding ways that current bonds (or refinancings) can advantage one set of creditors over another. And one reason this is getting more common is that, while default rates remain low, if you add in distressed reorganizations you get an annual default rate above 2%, or around where it was during the pandemic. Those aren't quite defaults, and some of them are probably less distressed than others, but they're still a sign of stress in credit—which is a lot more stressful when the outcome hinges on legal minutia rather than fundamentals.
### Celebrity AI Economics
Meta is paying millions of dollars for the rights to use famous people's voices for AI products. This is part of a recent theme in The Diff: usually, as celebrities age they get worse at reprising their old roles, but as technology improves they start getting better again, and probably won't stop for a while. For Meta, there are some interesting options: the product started out asa way for people to interact with their real-world friends, and is increasingly a way to interact with influencers. Somewhere in between those models is having real-time conversations with AI instantiations of famous or semi-famous people.
That raises the question: why start with independently famous people instead of home-grown influencers? There are execution risks with this kind of product, and there's also a fixed cost to training models on every member of the long tail. Licensing for a small-scale test might be much cheaper than an immediate broad rollout, especially if it turns out that people don't want to interact with celebrities at all. When AI spending is high, existing fame is the cheap substitute for training a bunch of celebrity-specific models for people with followings in the tens of thousands rather than tens of millions.
Disclosure: Long META.
## Diff Jobs