## Life with Nonzero Interest Rates

Interest rates are finance's answer to a fundamental force like gravity. They put the economy in an elliptical orbit around its sustainable growth rate, dragging different things in different directions depending on how far from optimal we are. Rates, overall, reflect the market's consensus view about the costs and rewards of deferred consumption. Because they're a single-dimension compression of a many-dimensional phenomenon, you can tell plenty of stories about a given change in rates, and because so many of these forces offset each other, you can plausibly explain 100% of the net change in rates by only looking at a specific cause. From the perspective of households, rates are a way to think about the supply and demand of savings: high rates mean that people want to spend rather than save, and if that what you want to spend on is some durable asset—usually a house, but cars and other big durables count—then you're paying a lot to tie up capital for so long. Low rates are a way for savers to basically beg people to find something, anything, useful to do with money. "You can't find a way to make money if your cost of capital is 4%? Fine, anything you can do at 3.5?" For companies, it's a similar signal, about whether to retain or return capital: if you can mechanically grow earnings per share at, say, 10% a year merely by buying back stock, why bother doing anything else? And if the market says that doing that retires 2% of your shares each year instead, the market is also saying that you should invest more. Rates tell the government when to get its fiscal house in order by reducing spending and controlling inflation, or when the real problem is slow growth and a substantial deficit can be tolerated for a long time.

But rates go deeper than that. A few weeks after ChatGPT came out, The Diff wrote about how technologies you can describe as "a solution in search of a problem" are a low-rates phenomenon, specifically because low interest rates mean there's a smaller difference between an investment that starts producing returns in five years or in fifteen. This encourages exploratory research, like spending money on statistical models of text that show promise in fields like spellcheck and translation—maybe in a few years they'd be useful for something else, like autocomplete! 

Attention is All You Need was published in June 2017, when the ten-year yielded 2.3%. It hasn't had a two-handle since mid-2022, and in the last few months, rates have shot up higher. It's annoyingly difficult to trace this rise in rates directly to AI capex. PIMCO doesn't see AI-related borrowing crowding out other kinds of funding; treasuries aren't selling off because everyone's dumping them to invest in the latest Nebius convertible issue or whatever. AI-related borrowing was about 40% of total corporate debt issuance in the first half of this year, compared to 20% for last year. But that makes it about 6% of total borrowing—pretty meaningful if you work in fixed income, but not something that can trivially explain the rise in rates. You can extend this a bit and get more compelling numbers, because prior to the burst of capex, one of the criticisms of companies now investing heavily in AI was how much lending they were doing: they accumulated large balances of cash and cash equivalents, with "and equivalents" meaning mostly short-term debt. If they're drawing down cash balances to fund AI investment, that's additional net borrowing, but that doesn't quite double the sum they explicitly borrowed. Meanwhile, if you look at net nonfinancial borrowing, AI-linked companies were about 9% of it and the US government was 44%.

So the rates situation looks like an old-fashioned problem—government spending as a share of GDP hitting a level in a boom that was previously associated with recessions: the deficit was 5.7% of GDP last year, a little higher than in 2012 (6.6%) but otherwise a level that, prior to the financial crisis, had only been achieved in wartime. New Deal-era deficits peaked at around 5% of GDP, the lows in 1983—a mix of tax cuts, higher defense spending, and a recession pushed it to 5.7%—and is otherwise unprecedented for the US. If the government is spending money without recapturing that in tax revenue, unemployment is low, and the biggest growth industries are bottlenecked, the natural result is both higher real rates (because there are more investments competing for capital) and higher inflation. 

Higher real rates may or may not be directly caused by AI, whether directly or through second-order effects. But it's an appropriate circumstance, particularly if inflation is a big component. Because some costs are fixed through borrow and long-term contracts, and since companies and workers are averse to pay cuts rather than layoffs, higher inflation tends to accelerate the reshuffling of resources in the economy, by reducing the real income of everyone outside of the sectors where there's enough growth that employees can negotiate raises. It basically puts the economy on a threshold dose of psychedelics, causing everyone to take a step back and get more perspective. What the non-inflation component of rates tells you is that now is a time for choosing deployment over research, because the timing of the payoff is more certain. High rates are terrible for the present value of long-term research and development, and do a lot less to the expected return of some cost-cutting project that's going to pay for itself in a few months. Capabilities will continue to advance, of course, and we'll find uses for whatever new models come out. But in the immediate future, more capital can shift to figuring out what to do with current models. Interest rates are the price of time, and sometimes the future is cheap from a distance but expensive once it gets here.


## Elsewhere

### Stimulus

China plans to spend about 3% of its GDP through 2030 on some big infrastructure projects. This has been a part of China's economic toolkit for a long time, to the point that they've worked their way very far down the list of viable infrastructure projects. It's usually led to overinvestment and low returns. On the other hand, this displays some optimism from China about what global economic growth looks like: if they're going to increase their capacity to produce more goods, the world will need to absorb them. Since at least some GDP growth tail risks come from China's policies, it's a way to bet on a positive slant to the status quo.

### Ads

Apple has been blocking connections to some advertisers in iOS, and has set itself up to add more to the list. Apple's behavior here does keep users' information more private, but it actually strengthens their most serious strategic threats, because it makes scale and vertical integration more valuable. The big platforms have a lot of data about their users, and can monetize it plenty on services they own. It's the publishers who rely on third-party ads, and the companies that advertise with them, who do worse from this. It's a dangerous time for Apple to give the big platforms an advantage, because the question of which form factor will dominate computing is more up in the air than it was a few years ago. 

### Gambling

NPR has a piece on problem gamblers who discovered Kalshi and Polymarket, with their central case being someone who put himself on sportsbooks' self-exclusion lists but had to badger Kalshi support to delete his account and ban him from the site. The market he ultimately chose is one of the purest gambling markets on the platform, in which traders bet on whether Bitcoin will rise or fall over the next fifteen minutes. This is, naturally, something that can be hedged, albeit imperfectly, by trading Bitcoin. And once there are hedgers, the modeling will get more complicated, traders will start faking out and otherwise exploiting one another, and the result is that the whole thing is a random number generator that will occasionally produce patterns, but only temporarily and by accident. That's what slot machine designers go for, and if that's where these sites' users end up, then it's clear what they're there for.

### Acqui-hires

Two engineers from the surviving entity of Groq, much of which was acqui-hired by Nvidia in a deal announced late last year, are suing. They note that there aren't good precedents around these deals, which makes the lawsuit a healthy one. It is, in principle, possible to construct a transaction like this, where IP and/or key employees leave, investors get cashed out, and the sum of the value of that deal plus what's left is more valuable than the original company would have been. But it's also very possible for these deals to be one-sided and asymmetric; they're similar in spirit to corporate spinoffs, and for a while spinoffs were an incredibly attractive place for fundamental investors to focus their time and attention: they tended to be misunderstood and thus mispriced, with sellers who weren't always economically rational. That changed over time, and now the spinoff world is better-understood, and thus harder to take advantage of. Spinoffs, though, involve public markets, where there's more transparency and it's easier to judge the results. Acqui-hires will take a long time to replicate that, and setting a few legal precedents about which deals stand and what makes them fair is a good thing.

Disclosure: long NVDA.

### IPOs

Australian AI company Firmus is going public, but expects existing investors to take about half the deal. As much attentio as there is to AI infrastructure's voracious demand for cash, it's worth noting that there's still a large supply of cash looking for more AI exposure, too. The market is moderating a lively debate about the terminal value of GPUs, and the return on incremental watts, and there are still big participants with healthy balance sheets who want more of this exposure.