AI is supposed to be deflationary. Here's what the macro data suggests.

Rent, home prices, electricity, food and streaming since 2022

Rent on Zillow's index has gone up 2.8% a year since the cutoff and Case-Shiller home prices 3.6% a year. The average residential electricity price that EIA collects from utilities is up 4.8% a year. FAO's world food index is down 0.5% a year, but meat is up 3.1% a year, vegetable oils 6.6%, and the whole index is still a third above 2019.

I expected streaming to get cheaper first, since it is almost all software and the cost of delivering video keeps falling. I built an equal-weight index of eight plans: Netflix, Disney+, Hulu, Max, Spotify, YouTube Premium, Apple TV+ and Amazon Prime. It is up 32% on December 2019, and 14 of the 21 price increases since 2020 came after November 2022.

The World Bank energy index is crude oil, natural gas and coal. The BLS producer price line is here for comparison only, and none of the argument depends on a government price index.

Token prices fell 80 to 90%, and software list prices went up

OpenAI charged $60 per million output tokens for GPT-4 in March 2023. It charges $10 for GPT-5 and $8 for GPT-4.1. Google cut Gemini 1.5 Flash from $1.05 to $0.30 in three months. No other series in the repository fell anywhere near that fast.

Tokens are a cost for software companies, and those companies raised what they charge. Adobe Creative Cloud All Apps went from $49.99 to $69.99 a month, Salesforce Sales Cloud Enterprise from $125 to $175 a seat, Slack Pro from $6.67 to $8.75 and Microsoft 365 Business Standard from $12.50 to $15. Microsoft charges $30 a seat extra for Copilot, and Notion sells its AI features as a separate plan.

Eight of the 12 SaaS plans I tracked changed price after the cutoff, and all eight went up. The only cut in the list is GitHub Team, from $9 to $4 a seat in April 2020.

Software vendors' margins went up after 2022

I expected AI to make software competition more expensive, with more R&D to keep up, more marketing to sell new features, and thinner margins. The 10-Ks of 32 listed software companies show the opposite. Sales and marketing fell from 20.0% of revenue in fiscal 2022 to 17.3% in fiscal 2025, and R&D from 15.7% to 15.1%. Gross margin rose from 70.9% to 72.1%, and operating margin from 28.9% to 34.2%.

There are two explanations and this data cannot choose between them. AI may have made software cheaper to build and sell while vendors kept their prices up, or the 2023 layoffs cut costs and AI had little to do with it. Either way, my guess that competition would squeeze profits was wrong for these 32 companies. Microsoft and Oracle carry a lot of the revenue weight, and the panel without them moves the same way; that version is its own CSV in the repository.

Status-page incidents at Cloudflare, Google Cloud, GitHub and Datadog

Cloudflare posted 379 incidents a year from 2019 to 2022 and 699 a year from 2023 on. Google Cloud went from 121 to 258 a year and GitHub from 84 to 162. Datadog went from 58 to 32 a year, but total incident hours rose slightly, from 100 to 114, so each incident lasted longer.

A status page lists what a company decides to disclose, which is different from measured uptime. Cloudflare and GitHub both launched products and regions after 2022, and a company that starts reporting more openly will look worse here. AWS is in the repository, but its history before 2023 comes from Internet Archive snapshots with too many gaps to use. What the table shows is that three of these four companies reported more incidents and more incident hours after 2022, and none reported fewer of both.

I read this as a sign that AI is making software less reliable. Google and Microsoft, which owns GitHub, cut about 22,000 jobs in early 2023, and both now say AI tools write more than a quarter to 30% of their new code. The doubling at Google Cloud and GitHub starts that same year. Status-page counts cannot prove AI caused it, and Datadog went the other way.

The data center buildout is 0.7% of GDP, and goods prices would need to fall 3.4% to offset it

Amazon, Alphabet, Microsoft, Meta and Oracle spent $379bn on property and equipment in fiscal 2025 and $155bn in fiscal 2022. I use the $224bn difference as a rough measure of AI capex. It equals 0.73% of US GDP and 3.4% of what Americans spent on goods. For that spending to be offset in the same year, goods would have to cost 3.4% less than they otherwise would, and goods prices went up.

The $224bn is wrong in both directions. It leaves out NVIDIA's other customers, and NVIDIA had $130bn of revenue in fiscal 2025, much of it from outside these five. It also leaves out utilities and construction. It includes some ordinary growth in warehouses and cloud regions. My range is 0.7% to 1.3% of GDP. For comparison, US telecom capex peaked near 1.2% of GDP in 2000, shale drilling near 1% in 2014 and railroads at 2 to 3% in the 1880s, each spread across hundreds of companies.

Output per hour rose in industries that use AI and stayed flat in ones that do not

Interest rates, tariffs and the pandemic affected every industry, so comparing industries that use AI heavily with ones that barely can removes some of that. BLS publishes output per hour for detailed industries. Five heavy users have data through 2025: software publishers, commercial banking, engineering services, publishing and travel agencies. With 2019 set to 100, their median went from 103 in 2022 to 129 in 2025. Ten light users, including trucking, restaurants, hotels, grocery stores and electric utilities, were at 99 in both years.

Two other comparisons agree. Producer prices in the heavy-user industries, divided by prices in the light-user ones, fell from 97.6 in 2022 to 96.4 in 2026, so the heavy users raised prices about a point a year more slowly. US GDP per hour is up 12% on 2019, against 2% for the euro area and the UK and 4% for Japan, and the US is where almost all of the data center spending is.

This is the strongest evidence against my view, and it has three problems. Output per hour goes up when a company lays people off and keeps its revenue, which software and banking both did in 2023. Five industries is a small sample, because BLS has not yet published 2025 figures for the rest. The US was already ahead of Europe in 2021, before ChatGPT. And prices rising a point a year more slowly in a few industries does not make rent or groceries cheaper.

What I conclude

AI has not made anything a household buys cheaper yet. So far, buyers are paying more for software that breaks more often: list prices rose 7 to 40%, and three of the four status pages show about twice as many incidents. I was wrong about the vendors. I expected AI competition to raise their costs and cut their margins, and instead they spent less on sales and marketing, raised list prices, and increased operating margin by five points.

The long-run version of the deflation argument could still be right. Computers took about a decade to show up in US productivity in the 1990s, and output per hour in the heavy-user industries is already rising. Three years after ChatGPT, the five biggest buyers are spending at the scale of the 2000 telecom buildout. I think calling AI obviously deflationary today is wrong, because the price that fell is for tokens, software companies buy most of them, and those companies charged their customers more.

At a glance

Question

Is there evidence, outside the CPI, that AI has lowered prices?

Data

Zillow ZHVI and ZORI, Case-Shiller, FAO and World Bank food and commodity indexes, EIA electricity, 8 streaming and 12 SaaS list-price histories, LLM token prices, 32 software 10-Ks via SEC XBRL, five status pages back to 2019, BLS industry productivity, OECD GDP per hour

Cutoff

November 2022, ChatGPT's public release. Before is 2019 to 2022, after is 2023 on

Cost

$0 in data. Everything is a public download