a16z Infographic Trends: AI costs halved this year while usage doubled, American 30-year-olds' life milestones are being pushed back across the board

Author: a16z New Media

Translation: Deep潮 TechFlow

Original link:

Deep潮 Guide: This week’s chart weekly report from a16z covers four topics, each worth a separate article: the decline in AI costs triggering the Jevons effect, the true scale of tech giants’ capital expenditures, Kalshi’s prediction markets outperforming professional forecasting agencies, and the widespread delay in life milestones for 30-year-olds in the U.S. The data sources are solid, and the perspectives are calm and restrained, making it a high-quality reference for understanding the intersection of current tech and macro trends.

DExit… Is it a real trend or just an illusion?

Delaware remains the preferred state for company registration in the U.S., but this position is quietly weakening:

According to Ramp data, Delaware’s share of new company registrations has been steadily declining since 2023, with a drop of about 10% in Q3 2025.

History doesn’t simply repeat itself, but it often rhymes… maybe.

Delaware has not always been the holy land for business registration.

About a century ago, Delaware replaced New Jersey—the original “Trust Mother”—as the top state for company registration. New Jersey lost its advantage because then-Governor Woodrow Wilson tried to curb corporate abuses, which severely worsened the business environment there. Delaware’s corporate law was modeled after New Jersey laws before Wilson’s era, naturally welcoming companies fleeing other states. Over the next nearly 100 years, Delaware’s Chancery Court, combined with this legal foundation, built a reputation as a fair and mature forum for resolving corporate and investor disputes.

However, what took a century to establish has begun to shake in just a few years. Regardless of right or wrong, Delaware’s Chancery Court has recently adopted a more lenient stance on shareholder lawsuits (notably in several high-profile cases, including but not limited to Tesla), prompting companies to seriously consider relocating their registration elsewhere. Farewell, good luck, Delaware.

This is at least the mainstream narrative, but other data suggest the situation is more complex.

First, even the founding myth of Delaware isn’t entirely accurate.

It wasn’t until the 1980s (about 60 years after Wilson’s governorship) that Delaware truly surpassed New Jersey to become the state with the most business registrations in the U.S.:

New Jersey’s dominance lasted longer than mainstream narratives suggest. The catalyst for Delaware’s eventual overtaking was likely its passage of a series of laws related to director responsibilities, which made it especially attractive to public companies, combined with network effects that reinforced this inertia.

Second, regardless of what’s happening with high-profile public companies (and those listed in Ramp data), Delaware overall still performs well—actually, it’s doing better than just “well”:

According to data from Harvard Law School’s Forum on Corporate Governance, from late 2024 to 2025, Delaware’s share of the total number of U.S. companies actually increased significantly.

In fact, if you’re looking for a clear “DExit” case, it’s probably this one, which has nothing to do with Tesla but involves a specific corporate form:

Wyoming LLCs began to grow rapidly around 2015.

Why? This is likely related to Wyoming’s LLC laws, which emphasize asset protection and privacy, and the state promotes this corporate structure as a “cowboy cocktail.”

In summary, the point isn’t that DExit isn’t happening (since some data clearly show it is—at least some high-profile companies are moving out), but the reality is definitely more complex than the mainstream narrative suggests.

The truth is, Delaware still enjoys the advantage of being the default choice, along with all the network effects tied to it, which are hard to shake.

We previously released an early version of this chart, but as data increased, the effect became even more striking.

Token costs decline, token consumption rises:

Since the beginning of this year, paid token prices have dropped from about $0.90 per million tokens to $0.50, while the number of tokens processed has nearly doubled—from about 6,000 to 12,000.

This is a classic Jevons effect. The cheaper AI becomes, the more we use it. It’s encouraging.

Remember when someone said that once better GPUs hit the market, old GPUs would become worthless?

It seems that’s not quite the case:

According to Silicon Data, the rental prices for Nvidia H100 and A100 have increased this year.

There’s no sign of an oversupply of computing power; in fact, demand seems to be barely scratched.

This comparison isn’t perfect, but if history offers any reference, it may take some more time before we can truly understand what an “AI-driven” economy looks like:

From Faraday and Henry’s initial discussions of electricity to the industrial productivity boom of the early 20th century, it took about 100 years.

Since the 1820s, technological iteration cycles have indeed sped up, but a platform-level shift involves many variables.

Roy Amara’s famous quote: “We tend to overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten.”

Capital expenditure, viewed on a graph

Here’s a set of data that never gets old: AI capital expenditure is enormous.

Compare the following:

Projected AI capital expenditure in 2026 is close to the total net new loans made by U.S. banks in 2025:

Capital expenditure is about 33% higher than total U.S. corporate income tax revenue, roughly three times the amount of tariffs:

It’s six times the military budget of any G7 non-U.S. country:

So yes, the scale of capital expenditure is truly massive.

Kalshi enters the macro prediction space

Federal Reserve researchers believe prediction markets are quite effective.

At least for one indicator, Kalshi’s forecast of the federal funds rate has already outperformed professional forecasting agencies:

For the forecast of the federal funds rate 150 days ahead (after three FOMC meetings), Kalshi’s average absolute error is very close to that of professional forecasters. Unlike surveys that provide only a snapshot every six weeks, Kalshi offers continuous updates of the full probability distribution… We found that Kalshi’s median and mode predictions perfectly recorded the rate forecast the day before the FOMC meeting, a statistically significant improvement over federal funds futures forecasts.

In other words, although all predictors start from similar points, Kalshi’s “continuous updating” forecasts keep improving over time, ultimately achieving “perfect prediction records” the day before the rate is officially announced. Moreover, Kalshi’s performance surpasses that of futures markets.

Kalshi’s advantage isn’t limited to the federal funds rate. As the Fed researchers point out, because macro indicators like inflation, growth, and unemployment lack other options markets, Kalshi is the only place providing “high-frequency, continuously updated, rich probability distribution benchmarks” to reflect the “public’s” expectations of these economic indicators.

Sounds quite important.

Delays in reaching adulthood

This is a thought-provoking chart, with some commentary:

The proportion of 30-year-olds reaching major life milestones has been declining quite steeply since at least the 1980s.

Among 30-year-olds, fewer and fewer are:

Living independently;

Getting married;

Living with children;

Owning their own home.

The only exception is college enrollment—the proportion of 30-year-olds with a bachelor’s degree has nearly doubled since 1995.

So, is college worth it?

Milestones? More like a grindstone for the neck, right?!

Maybe yes, maybe no, but the feeling of “buyer’s remorse” seems to be in the air.