Daily index · Updated 11 November 2026, 10:00 UTC
2026 Global AI Momentum Index
Ten themes, ranked by how much attention they are drawing today. The score is an editorial composite built from release notes, pricing changes, procurement language and developer chatter — not a market cap, and not a forecast.
Composite of the four sub-indices below.
Funding, filings and secondary activity.
Model releases, benchmark movement, task completion.
Deployment signals, hiring, procurement language.
The board
Ordered by 24-hour momentum. Status reflects how settled the signal is, not how important it is.
-
NO.01
Agent tooling moves from chat windows to browser control
The centre of gravity in the agent category is shifting from conversation to controlled browsing. Teams that spent last year wiring chat to a database are now wiring a browser to a permission model, and the hard part has moved from prompt design to session handling, credential scope and audit trails.
2720 momentum peak 3019 -
NO.02
Inference pricing splits into two tiers, and buyers notice
Frontier pricing is holding while small-model pricing keeps falling, so procurement teams are being asked to route work by task value rather than by vendor. The practical result is a two-tier stack in the same company: an expensive model for judgement, a cheap one for volume.
1914 momentum peak 2050 -
NO.03
Open-weight models close the gap on structured output
The argument that open weights are only good for chat is getting harder to defend. On schema-constrained extraction and classification, the gap now shows up mainly in edge cases and long documents, which is exactly where a human review step already sits.
1602 momentum peak 1688 -
NO.04
Enterprise buyers ask for audit trails before autonomy
The procurement question has changed. It is no longer whether an agent can do the task, but whether the company can reconstruct what it did six weeks later. Vendors that ship a readable action log are clearing security review faster than vendors with a better demo.
1288 momentum peak 1402 -
NO.05
Voice becomes a default input on the desktop
Dictation stopped being a accessibility feature and became a normal way to draft. The interesting metric is not transcription accuracy but how much text survives the edit, and teams report the first draft is now often faster to speak than to type.
1104 momentum peak 1104 -
NO.06
Small models win the edge deployment argument
When the data cannot leave the building, a smaller model on local hardware stops being a compromise and becomes the only option. The trade-off has moved from quality to maintenance: someone has to own the quantised build, the GPU and the fallback path.
946 momentum peak 1012 -
NO.07
Evaluation shifts from benchmarks to task completion
Leaderboard numbers are being replaced by completion rates on the buyer's own work. It is slower and less quotable, but it is the only number that survives contact with a real workflow, and vendors are starting to publish the harness rather than the score.
812 momentum peak 860 -
NO.08
Data residency turns into a procurement blocker
The conversation moved from 'is it secure' to 'where exactly is it processed'. Teams that had a policy in theory are discovering it does not map onto a vendor's inference topology, and the deals that stall are stalling on paperwork rather than capability.
688 momentum peak 902 -
NO.09
Design tools absorb generation into the canvas
The separate 'AI image tool' is being folded into the design surface. The useful change is not image quality but that a revision now propagates across a set of assets, which is the part that used to be manual file management.
604 momentum peak 640 -
NO.10
Coding agents shift from completion to review
The value is moving downstream. Writing the line is cheap; deciding whether the change is safe is not, and the tools getting adopted are the ones that read the whole repository and argue about the diff rather than autocompleting it.
522 momentum peak 560
Sector heat
Relative attention across the eight categories we track, 0–100, with the 24-hour change.
Frontier signals
Five things worth carrying into next week, written by the desk rather than generated from a feed.
Two-tier routing is becoming the default architecture in mid-size companies: frontier model for judgement, small model for volume, with a router in between that nobody has audited yet.
Security questionnaires now ask for an action log and a deletion path before they ask for accuracy. Vendors without either are losing deals they would have won a year ago.
The interesting number is completion rate on the buyer's own tasks. Everything else is a proxy that flatters the seller.
The gap is narrowing fastest where the output has a schema. Free-form judgement still belongs to the frontier models, for now.
Every team that shipped an agent this quarter is now hiring someone to watch it. That role did not exist in any org chart last year.
How the index works
Each theme is scored on four inputs: the number of independent sources reporting it, the number of distinct reports, how fast the topic is spreading between them, and how much of the discussion comes from operators rather than vendors.
The composite index is a weighted average of the four sub-indices, rebased to 1000. A reading of 742 means the desk judges current momentum to be roughly three-quarters of the strongest week this year. It is a relative measure, not an absolute one.
These are editorial estimates, not market data, and nothing here is investment advice. Where a number cannot be verified from a primary source, we say so in the brief instead of printing it.
Weights
- Capital — funding, filings, secondary activity30%
- Capability — releases, benchmarks, task completion30%
- Adoption — deployments, hiring, procurement25%
- Policy — regulation, standards, litigation15%