How we measure
Breaking out
An item breaks out when it grows faster than 90% of similar items over the same window, and by at least 10%. Growth is the figure now divided by the figure at the start of the window: 2× means it doubled. Rankings start at the median size of each group, so a tiny item that went from 1 to 10 does not lead.
What each figure counts
How many AI tokens an app uses in a week.
Tokens are usage, not users.
How many AI tokens a model serves in a week, through one large AI model marketplace.
Tokens through one marketplace, not all use of a model. Free, paid and dated versions of a model count together.
Each AI lab's share of all AI tokens served in a week, through one large AI model marketplace.
A share of one marketplace, not of all AI use. A lab is the sum of its models; unnamed stealth models count in the whole but belong to no lab.
How often developers download each AI lab's own SDKs and coding agents.
Downloads of packages a lab publishes itself, not developers; build systems download too. Compare a lab with itself over time.
Open jobs where AI labs pay experts to train and test their models, by field.
Open jobs on one large AI work marketplace, not all hiring. One job can seek many people.
How often an open AI model is downloaded.
Downloads, not users. Models released in the last year; copies and converted versions are left out.
How many people install an extension or add-on.
Installs, not active users. Browser extensions are measured by their reviews.
How many times a week developers install a skill for their AI agents.
Installs, not users. Skills installed together as a pack count once.
How many times a week developers download the code behind an MCP server, a tool an AI agent can use.
Downloads, not users; servers run only from a hosted address are not counted.
How many developers install a company's packages.
Downloads, not developers.
How widely a piece of software is adopted by the companies on one large business-spend panel, month by month.
The panel's own adoption score, mostly venture-backed small and mid-size US companies. Not a share of all businesses.
How many websites run a technology.
Websites that run it, not companies that pay for it.
How many open roles a company lists on its own careers page.
Roles on the company's own careers page, read about once a week. Not every company lists every role.
Points an organization's links earned on one large tech news forum over the last 12 months.
A rolling year of points on one forum; a jump means a new story about the company took off.
How a company's website ranks among all websites by visits.
A rank, not a visitor count. Lower is better; growth means times closer to the top.
The share of new AI research papers about each idea, this quarter against the one before.
A share of papers, not a count: it rises only when an idea gains on the rest.
Companies
A product counts toward the company whose web address it belongs to. A company breaks out on several fronts when two or more of its measures break out at once; web reach supports a story but never counts as a front.
When it updates
Every night. The figures on each page are from one snapshot, now Oct 10, 2026, and each page says its date. A group with a gap in collection uses a shorter window until the gap is behind it.
How to cite
Quote any figure with its window and date, and link the page it came from:
Vela Breakout (Vela Partners), Oct 10, 2026, breakout.vela.partners/…
Agents can read the same figures through our MCP server.