The Longest Duration

Four signals this week. OpenAI is raising money at a $1.2 trillion valuation. The Treasury Secretary wants AI labs to lose their liability shield. Nvidia's CEO says AI needs no new laws. And somewhere, a central bank raised rates.

Three of those signals are about AI. One is about the price of money. Read as separate stories, they say nothing. Read together, they are the same story.

The three AI signals disagree — but not about what AI is worth. They disagree about when it pays.

OpenAI at $1.2 trillion is a bet on the decade. A number like that only holds together if the cash flows keep arriving a decade out. No company's current revenue justifies $1.2 trillion, so the number is, by construction, a statement about 2035 pulled back to today.

The Treasury Secretary is pricing a different horizon. Liability, open-source mandates, the shape of regulation — these are questions that get answered on a political clock. An election is two years away. A bill takes eighteen months to move. The policy price of AI sits inside the next four years.

And Nvidia reports every ninety days. A company that discloses its numbers quarterly is priced quarterly. There is no impatience in that — it is simply the time scale the structure imposes.

Three parties. Ten years, four years, ninety days. They are not arguing about the value of AI. They are pricing three different maturities of the same name.

Finance already has a word for "when the money comes back." It is duration.

Duration measures how far into the future an asset's cash flows actually sit. It is the weighted average of when you get paid. A one-year bond has a duration near one. A company whose profits are ten years out has a duration near ten. The longer the wait, the longer the duration — and the more an asset reacts when the price of time moves.

AI is the extreme case. Its current profits are thin. Its promised profits are vast, and they are far away. A $1.2 trillion valuation with cash flows concentrated on the far side of 2030 carries a duration of fifteen to twenty years.

15–20 years

Duration has a deeper form underneath it. The present value of AI is not one number. It is a sum over every future year, discounted back to today:

V = Σ CF_t / (1 + r + ρ)^t

CF(t) is the flow of AI's future earnings. r is the interest rate. ρ is the risk premium — the price of uncertainty. Together, r + ρ, they are the discount.

The whole argument sits in that one line. A rate hike does not touch CF(t), the flow. It only touches the discount — the e^−(r+ρ)t. Same future, different weight.

But then the honest question arrives: what is F(t)?

If AI's earnings were flat every year, F(t) would be a constant, and the valuation would fall gently as rates rose. If they grew forever, F(t) would rise without bound, and no rate hike could ever kill the number. If they rose, peaked, and then faded — the shape t^k e^−βt — the value would have a finite half-life, a tail that decays toward nothing.

Most of the market is pricing the second: growth without end. The honest model is the third: rise, peak, fade. And the difference between them is not opinion. It is the number β.

Now the hike's arithmetic becomes precise. Duration turns a rate move into a price move: the percentage fall is roughly duration times the rate change. An eighteen-year duration, hit with a one-percentage-point hike, falls about eighteen percent.

$1.2T → $984B

Read that carefully. The company, the technology, the 2035 cash flows — none of it moved. Only the denominator moved. A rate hike does not touch AI's future cash flows. It changes what those flows are worth now.

There is a recent precedent worth measuring against.

In 2022, central banks hiked quickly, and the longest-duration assets — companies priced almost entirely on distant promises — fell the furthest. Some corners of the market lost sixty, seventy percent. Duration did exactly what the formula says: when the price of time rose, the assets carrying the most time in them were repriced the hardest.

But 2026 is not 2022. In 2022, the long-duration companies had little or no earnings. Their valuations were pure duration — all future, no present. In 2026, the AI leaders book real revenue, and some book profit. The duration is still long. But there is now a present underneath it.

And this is where the shape of F(t) decides everything. A company whose earnings rise forever can carry almost any valuation through any hike. A company whose earnings rise, peak, and fade — whose flow is t^k e^−βt — has a valuation with a real, computable ceiling. The question the $1.2 trillion raises is not whether the hike kills AI. It is how much of that number is the flow, and how much is the discount.

So the $1.2 trillion is not one number. It is two things multiplied: the future flow F(t), and the discount e^−rt. The hike only touches the second. The market has not yet answered the first.

That is what the four signals, read together, actually ask. Capital bids on a flow that grows without end. Policy prices a risk that arrives within a term. Industry prices a quarter. And the rate hike quietly re-prices the discount on all three — without ever touching the flow.

The longest asset in the room is the one most exposed to the price of time. But the deeper question is not the duration. It is the shape of the flow underneath it — whether AI's earnings rise forever, or rise, peak, and fade.

A hike does not change AI the technology. It changes the discount — the price of time and the price of uncertainty — and so it forces the market to reprice AI's timing, its certainty, and its present value. Which of those three dominates is computable. It lives in β and ρ, not in headlines.