Is the Business of AI More Like Steel or VBA?
Some say that sufficiently transformative businesses don’t have ready analogies for what is to come. And yet:
- A global communications network that reduces the cost of transmitting data between arbitrary third-parties to roughly zero compared to where it was before, creating not only new businesses but a fundamental rearchitecting of the structure of existing ones—the Internet is a big deal, but I'm actually talking about the telegraph, which, among other things, led to the first modern corporations. And even before that, reliable ways to send letters also had a profound impact, and somewhat enabled the rise of new kinds of businesses. (Using data rails to transmit money has in one sense been an important project for PayPal, Stripe, etc. to pursue, but it's also a decent description of the Medici.)
- Cars led to a restructuring of cities, and to a massive burst of capital spending in order to make them ubiquitous, as well as a surge in the consumption of complementary goods—but so did railroads, and, for some cities, container shipping and earlier advances in ocean freight.
- The smartphone created a platform on which new software companies could build novel products and achieve massive distribution; on the backend, cloud computing ensured that as long as they had some kind of economic model (or as long as their investors trusted that they'd find one), computing itself was a mostly financial rather than operational problem. But this, of course, is an apt description of the PC boom a generation earlier—you don't have to think much about compute or storage if your users are all running things on their local machine!
So it's useful, at least as a tentative exercise, to ask what the shape of AI's economic impact might be. This doesn't require some kind of maximalist view that AI will replace all knowledge workers once GPT-5 ships, because that’s fundamentally unpredictable ("But of that day and hour no one knows, not even me, a large language model trained by OpenAI..."). All we really need is to remember this general rule: when some category of product gets radically cheaper, the use cases become surprisingly abundant. When homes first got wired for electricity, electric appliances looked like absurd power hogs compared to lighting, but scaling power plants to handle orders of magnitude more houses turned out to reduce the cost of power enough to make things like toasters and irons, and later much larger appliances, a viable option.
It's easy to imagine a universe of natural language computer interactions that's an order of magnitude larger than the current universe of constrained-language interactions; a CRM that can handle cases like "I'm ten minutes early for a meeting—which other client should I check in with right now, and, by the way, what should I say to them?" can get more usage than Salesforce, and an ERP system that can answer questions like "Which of our costs has been subtly drifting up over the last few years, and what's a good checklist for figuring out why?" will see more interactions than SAP.
Let's start with a somewhat dark case for AI businesses. What if they're the next steel industry? Steel is a useful and ubiquitous product; this post opening an upcoming series on the steel industry notes that "Nearly every product of industrial civilization relies on steel, either as a component or as part of the equipment used to produce it."—but that doesn’t make it a great business.
For example, US Steel has returned 1.8% compounded over 32 years, ArcelorMittal has lost 2.1% annualized since 1997, and Nippon Steel shareholders roughly broke even over the last 31 years. Of course, there are some exceptions; Korean national champion POSCO produced a solid 11.9% annualized return since the early 90s. And steel can generate fortunes early in a country’s development, when rising demand is more important than the industry’s steady-state dynamics. But regardless, it's a tough industry when the biggest companies rarely produce good returns for shareholders. And there are a few reasons for this:
- Steel is a cyclical, capital-intensive business. When demand is rising, firms have to invest more to retain enough scale to be relevant, which, as with every other cyclical business, means that they're always at peak capacity when demand rolls over.
- It still has marginal costs, but since fixed costs are so high, firms basically have to keep producing through a recession. In some cases, deeply indebted cyclicals have to produce even more when prices decline, because they still get incremental profits that are needed to service debts.
- They're tied to supply and demand constraints they can't really control. If the steel industry is optimistic and the iron mining industry isn't, then there's a global bidding war for iron, which further compresses margins.
- As with other asset-heavy businesses, workers have leverage to negotiate better wages when times are good. And it's politically difficult to negotiate wage packages down. (This is especially true as corporate disclosure has improved. Since executives' compensation is announced on a lag, if the market is good in one year and gets worse in the next, an executive who just publicly got a massive compensation package will be in the awkward position of asking workers to economize. It doesn't go well.)
- Many countries view steel as a strategic industry. And they have good reasons to! A lot of steel investment and a touch of IP theft made the US steel industry dominant over Britain's in the late 19th century, and that played an important role in making the US a bigger and more important economy. Steel was an early part of Japan's growth. Korea's early economic strategy closely followed Japan, with some war reparations taking the form of assistance to POSCO. And China's steel production is absolutely staggering: since 1967, global steel production outside of China has risen a total of 90%. China's steel production is up 73x over that time period, and is now half the world total.
This is an entirely plausible future for the AI industry. Models, while not a traditional physical asset, keep getting more expensive to maintain. And the bigger their datasets become, the more these datasets must converge: the biggest imaginable model uses all the public data in the world, which means every instantiation of that biggest possible model is using the same underlying data. There are some differences in cost based on proprietary hardware (more on this in a future post!), but the general story of recent AI development has been relentlessly bigger models with higher fixed costs and longer training times. Those training times start to introduce worrisome lags into the business: training a model means predicting what future demand will look like, but also means guessing what competitors are cooking up.
The bidding war for AI talent doesn't seem to have abated much (anecdotal reports from SF indicate that we're rapidly approaching a world in which every single person in the Bay either works for OpenAI or doesn't have a job, though this may be a slight exaggeration.) But as with other tech companies, AI employees have more levers to pull than just asking for higher compensation; they can also influence the nature of the products their company creates. And since many people in the AI community are deeply concerned with the risks of AI—fairly immediate ones about spam, propaganda, and asking ChatGPT how to produce illegal drugs, or more distant and apocalyptic worries about AI ending human civilization. An employee base where a substantial number of people firmly believe that a) they are saving the world, and that b) misguided competitors may literally end the world, is a group of employees highly motivated to play an active role in company strategy. Free cash flow per share may not be their top concern.
And that's just at the level of worker-company relationships. There are also company-government relationships. One rough heuristic for guessing what sectors the US and Europe will consider strategically essential in a few years is to look at what the Chinese government is currently dumping massive subsidies into. Whether this is just because governments chase superficially emerging threats, because technocratic governments can identify promising new technology fields, or some combination thereof is irrelevant: in industries that produce exports, some governments lead and many others eventually follow. This hurts AI companies in two ways. First, by limiting who they can sell to and what they can sell, and second, by ensuring that some of their competitors will be subsidized by the government and won't need to turn a profit. (It's hard to find old data on this, but Nippon Steel did not need to be a great investment on its own to be a worthwhile deal; it made more sense as a subsidy to the auto, construction, machine tool, etc. industries.)
So that's the pessimistic view for investors: AI will be as important and ubiquitous as a product, like steel, but AI companies will be relatively minor players in the economy they prop up. They might end up in the same position as social-dependent companies like Zynga or search-dependent ones like Demand Media, except that the AI companies won't even know which platform risk is the biggest one.
And then there's the optimistic case: AI as VBA. This case is compelling because large language models are a nice natural language glue between a) software products that don't have good APIs, or b) mixed software-and-human processes that are tricky to fully automate. A fair amount of internal software at some companies is basically glue connecting different processes: dump data from X, Y, and Z, and put it in a dashboard; query this table and produce a regular email report or an occasional Slack alert; make the online marketing budget over here respond to the customer churn rate over there; tell Finance what to do when Sales has done its thing, and update numerous spreadsheets accordingly. These processes support innumerable white collar jobs and a few people who have automated enough of their job to essentially retire with a full salary pension.
But why focus on just VBA? Why talk about VBA and not bash, perl, or a more recent scripting language, all of which have been used as this kind of glue? For two reasons: First, because the venn diagram of good prompt engineers and good engineers is not a perfect circle. In fact, AI's impact on white collar workers partly comes from the fact that it offers something close to software-style talent leverage without requiring as much time to get reasonably productive. And it seems to require a different set of skills. In fact, there's an active tradeoff between being good at using an AI text interface and coding, since the former rewards finding clever ways to accomplish a specific thing while with the latter leads to unmaintainable code. VBA was historically a way for non-programmers to automate a little of their work, so long as it lived in the Microsoft ecosystem. Since it can grow directly out of manual processes, and can keep growing for a while before it gets difficult to use, VBA has been a persistent force for a long time. And while plenty of programmers deride it (it's Stack Overflow's "most dreaded" language), plenty of smart people have used it, including Jane Street.
The world's many companies running some form of legacy software, with idiosyncratic levels of automation and organizations partly built around where they choose to have humans in the loop, will benefit from AI tools that connect these systems together. And what most of these businesses almost certainly have in common is that they're almost certainly running Microsoft software. It's definitely not a coincidence that Microsoft is looking at ways to add GPT to its office suite ($, The Information). This is, to be clear, on the just on the human-to-software interaction side—suggesting content, being smarter about recognizing the context of terms used within an organization, etc.—but it's a step in the direction of integrating proprietary tools with existing proprietary software to make it even harder to dislodge.
Will LLM-based content suggesters and coding assistants create jobs or destroy them? As a shorthand, there are two answers to the job-creation question for any new technology:
- In the short term, new technology is a net job creator, because the deployment stage requires lots of upfront expense compared to the steady state. Cars are a good example of this: at average annual mileage, modern cars have a useful life of roughly 13 years. So going from a 50% to a 51% rate of car ownership means about 20% higher car sales than just staying at 50% penetration and replacing cars as they fall apart. (In the 50s, US car penetration was rising at about two points per year, i.e. the industry had to make and sell 40% more cars than were required to maintain the steady state. It was a good time to be selling cars, or making them for that matter.)
- Once that deployment phase wraps up, labor-saving technologies tend to destroy jobs for a while, both because the industry shrinks after a period of extraordinary demand and because the efficiencies of new technologies replace workers.
- In the even longer-term, though, these technologies are job-creating because they make the world richer, so there's more to spend on other goods and services. People and countries spend more on services as they make more money, and services spending creates more jobs in the long run than goods spending because efficiency gains are rarer.
But that model is somewhat threatened because AI tools seem better at replacing services than products. (Though there are certainly some promising developments on the latter, too.) If the economy loses the main way to redeploy human capital that's freed up by rising levels of physical capital and increasing technological progress, it might become a social problem.
This is, naturally, pretty distant from the question of what AI businesses' economics will look like. But it's a question that will need an answer, and one whose answer will partly be revealed by what the experience of AI deployment within existing software is.
One important possibility, which doesn't do much for the job-destruction argument, is that AI makes it much easier to do tedious tasks than completely new ones. ChatGPT is plenty creative, but it's hard to give it exactly the right constraints without lurching into surreal interactions in one direction or railroaded ones in another. And more deeply, it's much easier to escape a local maximum in a system you can fully reason about than one whose behavior is hard to explain. This applies in many different fields, of course: in programming the languages with more powerful abstractions can produce more powerful results; in finance, the assets with fundamentals and cash flows are less bubble-prone than the ones that have those loosely (currencies, precious metals) or in no meaningful sense (some cryptos, fine art, meme stocks); in economics, going back to the basic questions about incentives, information, and thinking on the margin is always a good sanity check.
So AI-enhanced tools may end up speeding up things that 95% (or 99.95%) of white-collar workers do, making software relatively more valuable and humans relatively less, while having a weaker effect on the lucky few who can rigorously reason about extremely complex systems.
Which creates a fascinating dynamic for the AI companies. It's always good to sell a high-value complement to a product that's continuously dropping in price, whether that product is software for ever-cheaper hardware (like Microsoft and the first generation of software companies) or organizing tools for ever-more-abundant data (like Google and Meta). The possible world where humans are a necessary complement to AI tools, but are an ever cheaper one, is a world where AI companies are very valuable indeed.
Disclosure: Long MSFT, META.
There's already evidence that AI research is increasingly being conducted by the biggest companies and most elite universities, and that this is partly a function of their access to compute and to unique data. (On the other hand, costs may come down.) ↩︎
Though this brings up yet another caveat: will reproducibility become a big concern with commercial uses of AI? If the model is changing all the time, and getting more complex all the time, not only can you not guarantee that it will give you the same answer to the same query in the future, but it can't tell you why that is.) The second reason for Visual Basic as an analogy is that it's owned by a single company, which uses it to a) link between different products, improving customer productivity, and not incidentally b) create a large codebase that's both essential to the business and pretty much unmaintainable. ↩︎
From a maximally cynical perspective, a thriving middle class is just a temporary and expensive substitute for a finishing deployment of a labor-saving technology. This is a dynamic that the middle class can survive, given a continuing supply of new technologies that take time to deploy. ↩︎
A Word From Our Sponsors
Discover investment opportunities in deep tech and fintech startups.
What if you were able to unleash technological innovations, grow businesses, and truly impact the way we live? And what if you could do it while potentially making investment returns? Our goal is to make that vision a reality. Propel(x) is an investment platform that gives accredited investors the power to potentially change the world. By funding breakthrough technology startups, you may empower companies that are finding potential cures for debilitating diseases, creating abundant energy for all, and even taking us into space.
Formalizing an Economy
The number of people registering for pension plans and the number of small businesses registered at all in India is rising rapidly ($, Economist). In a very poor country, a large informal sector can be an economic stabilizer of sorts, because informal jobs are much less likely to have fixed compensation, so workers can have their pay rate or hours cut, which means companies don't go through cycles of accumulating workers and then doing wholesale layoffs when the macro situation gets worse. But such an economy has plenty of other constraints: those companies and workers can't borrow cheaply, are hard to tax, and are hard to fit into a legal framework since they're operating outside of one. A rising formal economy share is a good sign when a government has high state capacity, and a neutral or negative one otherwise.
Walgreens has dialed back spending on theft prevention ($, WSJ), both because of the return relative to costs and because they feel they may have exaggerated the growth of the problem. Since shrinkage dropped from 3.5% last year to 2.5% now, this implies that their spending on preventing it was on the order of 1% of retail revenue, or about $1.1bn. It's entirely possible that this spending was based on extrapolating their future loss prevention costs forward on the assumption that theft would continue to increase, and that the problem was more cyclical than structural. Making upfront investments despite an uncertain long-term outlook doesn't just apply to revenue-generating decisions like opening a new store; it also applies to the upfront price of reducing a rising category of costs.
One of my favorite details from the book Capital Returns is that when it was written in the early 2010s, the author was casting about for an example of why revenue was hard to predict, and settled on "number of international flights in 2020" as one of those numbers that's fundamentally hard to guess. Quite right! But industry disruptions, even generational ones, don't permanently change the direction of long-term trends, and now, three years post-Covid, the airline industry faces a shortage rather than a glut of planes. Some of this, like so much of the modern economy, is because of uneven recoveries from Covid and uneven impacts of Covid-offsetting policies: planes are scarce in part because of lingering supply chain problems, which tend to be worst for the companies with the most complex supply chains (modern commercial aircraft have a few million discrete parts). This, too, is temporary: when lead times are longer, it means that orders are being placed with ever more uncertainty about what the demand environment will look like when the planes finally arrive.
Mastodon, which quintupled in size shortly after Elon Musk took over Twitter, has since lost almost a third of its active users. The main upshot of this story is not that Mastodon suddenly got less popular, but that its surge in popularity was smaller than it looked: day to day, active user numbers are a good proxy for, well, how many active users there are on a site, and the higher the frequency of measurement, the more likely those users are to stick around (daily actives are clearly addicted, monthly actives are much less engaged). When there's a sudden burst of signups, it makes these higher-frequency numbers look amazingly good, but also means that much of the newly-joining cohort will end up being infrequent users, or will never use the service again. It's rare but not unheard of for a product to be rescued specifically because people are mad at the biggest competitor (Lyft's S-1 shows an amazing metrics improvement following the #deleteuber situation, for example), but a more common pattern is for a service to show steady growth, with occasional spikes as people defect en masse from a competitor and then don't come back.
The Mystery of the UAE Holding Company
A holding company that's literally called "International Holding Company" has risen in value by 42,000% since 2019 through a series of acquisitions of UAE-based businesses at basically zero cost ($, FT). The stock is now a third of the UAE's market benchmark. Its chart shows oddly steady returns, with some periods of flat share prices and some of high growth (the biggest drawdown in the company's history was a 12% drop from peak to trough in 2020, and in the last year shares have always traded within 3% of an all-time high. This is a very weird situation where there isn't a legitimate explanation for the company's growth or the share price, but there also isn't a good explanation for what the endgame is. Sprawling holding companies have been built out of basically nothing (there was a rash of them in China, including the infamous HNA, which at one point owned multibillion-dollar stakes in Uber, Deutsche Bank, and Hilton, and which went bankrupt in 2021), or National Student Marketing in the US. But usually the endgame is to buy some legitimate cash flow-generating assets and use them as the core of a real business. In this case, the buyer is getting a great deal—buying companies for free—and the sellers' motives are a complete mystery. It's possible that this is the world's most colossal wash trade, where the sellers are shareholders and their underpriced sales create a narrative around infinite growth. But at a $240bn market cap, it's a pretty big company for such a simple scam.
Companies in the Diff network are actively seeking talent! If you're interested in exploring growth opportunities at unique companies, please reach out. Some top current roles:
- A well funded early stage startup founded by two SpaceX engineers is building the software stack for hardware companies. They're looking for a backend engineer who can build services that quickly process large amounts of data. (Los Angeles)
- A firm using NLP and other ML tools to give retail and institutional investors access to custom-taylored portfolios is looking for a data engineer. (NYC)
- A company bringing machine learning tools to everyone is looking for experienced ML engineers with strong product sense. (Remote)
- A company building ML-powered tools to accelerate developer productivity is looking for a mathematician. (Washington DC area)
- A new service that's trolling the dating market with a better product and better monetization is looking for a full-stack founding engineer. (Los Angeles)
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.