The AI bull case for bonds

Treasury investors were focused on the FOMC today, parsing — as they always do — every word of the committee’s statement and the Chair’s press conference for clues about where interest rates might go next.

Hanno Lustig thinks they should be parsing the latest AI models instead.

“If you own a 10-year Treasury,” he advises in a blog post, “you are, whether you know it or not, making a sizable bet on AI.”

Some Treasury investors do know it.

Lustig cites a paper showing that Treasury yields fall when AI labs release new frontier models: by 6.6 basis points on average in the three days around the release and 9 basis points over 15 days. 

The paper’s authors assume that lower Treasury yields, as they usually do, implies something negative: investors must be pricing in lower growth in a post-scarcity economy, they suggest, or higher odds of an AI apocalypse. 

Lustig, however, says it implies something distinctly positive: a productivity shock to economic growth.

Treasury yields falling in response to advances in AI, he says, “might be exactly what you should see if investors are long growth through their Treasury holdings and are revising growth expectations modestly upward.” 

We don’t usually think of Treasuries that way.

Treasuries are bonds, after all, so we really just want our money back — with interest and in dollars that haven’t lost too much of their purchasing power.

But Lustig says they’re also an asset whose present value is a function of future budget surpluses — surpluses that could get a lot larger, a lot sooner, if AI raises economic productivity even a little. 

Productivity leads to faster growth and faster growth widens the gap between what the government collects and what it spends. The difference flows directly into larger primary surpluses.

This gives government debt a “large positive productivity beta,” Lustig explains. “Its fundamental value seems highly sensitive to the long-run productivity growth outlook” — which AI should raise.

For example, if AI added just a tenth of a percentage point to productivity growth, Lustig estimates it would raise the fundamental value of US government debt by $1.3 trillion (4.2% of its market value).

Half a percentage point of extra growth would raise the value of the debt by $6.5 trillion. 

Why so large?

“The key insight is that revenue and spending respond very differently to GDP growth,” Lustig writes.

On the revenue side: as incomes rise, taxpayers are pushed into higher brackets, causing government revenue to grow a little faster than GDP.

On the spending side: nothing really changes — Medicare and Social Security rise with inflation, but not GDP.

The US budget has shifted dramatically away from discretionary spending: from 50% in the 1980s to 30% now, and falling.

This is normally a cause of grave concern: how do we get the budget deficit under control if more than 70% of spending is mandatory?

AI, however, turns this into a cause for hope: the falling share of discretionary spending turns Treasuries into a leveraged bet on productivity growth: the growing difference between government revenue and spending flows straight to primary surpluses.

Over time, the difference can really add up. 

Lustig cites CBO data estimating that, for every dollar of revenue the government gains from economic growth, spending rises by just eight cents.

The result is that, in the scenario where AI adds 0.5% to productivity growth, Lustig estimates government revenue would rise to 20.1% of GDP by 2055 while spending falls to 18.9%. 

The two lines — revenue and spending — appear to cross around 2043. That would be the moment when the government flips from running a primary deficit to running a surplus.

(In the text, Lustig says the lines cross in 2035, but I’m conservatively going with the chart because I think everyone would be more than happy with budget surpluses from 2043.) 

All this assumes the government doesn’t mess things up, of course.

Lustig’s forecasts assume that the current tax code remains in place, which it won’t. It usually changes for the worse (from the perspective of bond holders, at least).

It’s easy to imagine how it could change for the much worse: if AI productivity generates large fiscal windfalls, politicians will be keen to ensure reelection by giving them away in tax cuts.

“Bondholders betting on AI growth are therefore also betting that Congress will not fully dissipate the windfall through tax cuts,” Lustig writes.

That is not a bet I would normally make.

But what if AI raises productivity faster than Congress can cut taxes?

It’s worth asking because, as hopeful as Lustig’s estimates are, they might not be hopeful enough.

Economists at Goldman Sachs, for example, estimate that AI might boost productivity growth by as much as three percentage points. 

If Lustig’s 0.5% increase in productivity growth would raise the present value of Treasuries by $6.5 trillion, a naive extrapolation of Goldman’s high-end estimate of 3% would raise it by $39 trillion.

That’s assuming all things are otherwise equal, which they won’t be — and that the effect is linear, which it probably isn’t.

The current value of US government debt is $32 trillion, so my naive estimate of $39 trillion is likely too good to be true.

But as long as we’re dreaming, why not dream big? 

The research group Epoch.ai has speculated that a rapidly expanding stock of “digital workers” could lead to explosive economic growth: adding digital workers to the economy, their paper says, could raise annual GDP growth to 30%.

Economists might not attribute this to “productivity,” however. The digital workers are a form of labor (as the researchers see it), and adding labor to an economy makes the economy bigger, but not necessarily more productive.

But the effect would be the same: government revenue would soar and spending would not.

(Assuming robots do not become eligible for Medicare and Social Security, that is).

I don’t expect Treasury investors will be pricing in a scenario of explosive growth any time soon. 

But perhaps they should at least consider it?

Less time thinking about the FOMC and more time thinking about AI would surely be… productive.

— Byron Gilliam