The AI Frontier Report
A structural analysis of artificial intelligence at the close of its fourth compounding year — compute, capital, capability, labour, sovereignty and risk.
Abstract
The AI Frontier Report is Frontier Labs' annual compendium: the load-bearing numbers of the AI economy assembled into one place, at one moment, with the sourcing attached to each page so a reader can check the work rather than trust the author. This first edition measures artificial intelligence across eleven dimensions and arrives at one structural conclusion — four years into the compounding, the technology is not the constraint. Substations, credit spreads, liability doctrine and organisational change now set the pace; model capability does not. The report establishes unprecedented reach, compounding token demand, the largest privately financed infrastructure build in history and a thousandfold fall in the price of inference. It leaves four questions open: cluster utilisation, audited margins, long-horizon reliability, and whether the enterprise pilot gap closes. Each chapter ends with the observations that would prove it wrong, and a scored review is published with the next edition.
Key findings
- 01The constraint on what happens next is physical, financial and institutional rather than technical. Substations, credit spreads, liability doctrine and organisational change set the pace; model capability does not.
- 02$725bn of big-tech AI capital expenditure is committed for 2026 — the largest privately financed infrastructure build in history, on current guidance.
- 03Global data-centre power demand is forecast at 132 GW for 2026, which is why electricity, not silicon, is the binding input.
- 04The price of GPT-4-class output has fallen by approximately 1,000× since 2023, while frontier revenue grew faster than prices fell.
- 05Approximately 900 million weekly users on the largest assistant — reported and unaudited, which is precisely the problem with the headline figures in this field.
- 0631 per cent of organisations report an agent in production, against a median 20-step agent success rate of 41 per cent on benchmark. Deployment is running ahead of measured reliability.
- 07Chinese open-weight models account for an estimated 61 per cent of open-weight token share, a structural shift in where diffusion originates.
- 0886 per cent of US venture dollars went into AI in the first half of 2026, a concentration without precedent in the sector's history.
- 09Of the nine figures carrying the most analytical weight in this report, three are audited, two are Frontier Labs estimates, and four are company statements that nobody has verified.
Four years in, the technology is not the constraint. The limits on what happens next are physical, financial and institutional.
Artificial intelligence has now spent four consecutive years compounding on three axes at once — capability, capital and usage. Each of those axes is separately measurable. Almost nobody measures them together, and the ones who do tend to publish a single slice: a model leaderboard, a capex table, an adoption survey.
The consequence is a discourse in which two people can hold opposite views on whether AI is working and both cite real data. One is looking at inference cost curves. The other is looking at enterprise deployment rates. They are describing different layers of the same machine.
The AI Frontier Report exists to put those layers on the same table. It is a structural compendium: the load-bearing numbers of the AI economy assembled into one place, at one moment, with the sourcing attached to each page so a reader can check the work rather than trust the author.
Executive summary
This report measures artificial intelligence across eleven dimensions and arrives at one structural conclusion: the limits on what happens next are physical, financial and institutional. Substations, credit spreads, liability doctrine and organisational change now set the pace. Model capability does not.
What is established
Unprecedented reach, compounding token demand, the largest privately financed infrastructure build in history, and a thousandfold fall in the price of inference. These are not contested. They are visible in filings, in guidance, in grid interconnection records and in published price schedules.
What is not
Cluster utilisation, audited margins, long-horizon reliability, and whether the enterprise pilot gap closes. Four unknowns, each decisive. Every confident projection currently in circulation resolves, on inspection, into an assumption about one of these four.
What follows
Eleven chapters, each ending with the observations that would prove it wrong, and a scored review published with the next edition. We think a report that cannot be marked is not a report; it is an opinion with charts.
The report in nine numbers
These are the figures that carry the most analytical weight across the eleven chapters. Their provenance matters as much as their magnitude, which is why it is stated alongside each one.
| Figure | Measure | Status | Chapter |
|---|---|---|---|
| ~900M | Weekly users, largest assistant | Reported, unaudited | 01 |
| $725B | 2026 big-tech AI capital expenditure | Guidance | 02 |
| 132 GW | Global data-centre power demand | Forecast | 02 |
| ~1,000× | Fall in GPT-4-class output price | Estimate | 04 |
| ~61% | Chinese open-weight token share | Measured | 05 |
| 31% | Organisations with an agent in production | Survey | 01 |
| 41% | Median 20-step agent success rate | Benchmark | 03 |
| 86% | AI share of US venture funding, H1 2026 | Reported | 09 |
| 2 Aug 2026 | EU enforcement powers commence | Statutory | 10 |
Three of these nine are audited. Two are Frontier Labs estimates. The remaining four are company statements that nobody has verified. That distribution is the single most important thing to carry out of this table.
The eleven chapters
Each chapter is self-contained, carries its own source line on every page, and closes with the observations that would falsify its thesis.
| No. | Chapter | Subject |
|---|---|---|
| 01 | The Acceleration | Diffusion time has collapsed by an order of magnitude. Reach, usage and the pilot-to-production gap. |
| 02 | Compute & Capital | The build: capital expenditure, data-centre power demand, and the physical inputs that gate expansion. |
| 03 | Models & Capability | What improved, what did not, and the distance between benchmark performance and long-horizon reliability. |
| 04 | The Economics of Inference | The thousandfold price decline, its composition, and what it implies for margin. |
| 05 | Open Weights | Where openly released capability now originates, and what the token-share data shows. |
| 06 | Sovereign AI | National programmes, committed capital, and delivered versus announced capacity. |
| 07 | The Physical World | Robotics, manipulation reliability, and the distance from demonstration to duty cycle. |
| 08 | Work & the Firm | Task composition, organisational change, and what the labour data does and does not support. |
| 09 | Money | Venture concentration, financing structures, and the credit dimension of the build. |
| 10 | Risk & Governance | Statutory instruments in force, enforcement timelines, and liability doctrine. |
| 11 | The Frontier Ahead | What would have to be true for each reading of the next two years to hold. |
Method and evidence standard
Figures in this report are selected, reconciled, labelled and sourced under a fixed procedure. Primary sources first: filings, transcripts, regulator texts, official statistics and laboratory technical reports. Where two credible sources conflict, both are recorded and the spread is carried forward rather than resolved by preference.
- 01Retrieve — primary first: filings, transcripts, regulator texts, official statistics, lab technical reports.
- 02Reconcile — where two credible sources conflict, both are recorded and the spread is carried forward.
- 03Label — actual, estimate (E) or forecast (F), marked on the exhibit itself and not buried in an endnote.
- 04Attribute — every page carries its own source line. No exhibit is orphaned from its evidence.
Hierarchy of evidence
- Company filings and audited statements.
- Direct company guidance and transcripts.
- Official statistics and regulator publications.
- Specialist trackers with published methodology.
- Reputable press with named sourcing.
- Frontier Labs estimate, explicitly marked.
What gets excluded
Single-source vendor marketing claims. Numbers whose methodology is not stated anywhere. Aggregator sites that recirculate a figure without an origin. Any statistic whose provenance the desk could not trace to a first publisher.
Roughly one exhibit in six contains a Frontier Labs estimate. These are constructed from disclosed anchors and stated assumptions, and carry an E marker. They exist because the alternative — leaving a structural gap unquantified — is less useful, not because the number is precise.
Units, currency and dates
All figures in US dollars unless marked. Power in gigawatts for capacity and terawatt-hours for consumption. Token prices per million tokens. Fiscal years are labelled where they diverge from calendar years. The data cut-off for this edition is 8 August 2026; where a figure is older, the page states its vintage.
How to read the exhibits
The visual grammar is consistent throughout the printed edition. Every exhibit page carries a kicker naming the sub-theme, a headline stating the finding as a sentence rather than a label, a deck defining the measure and period, the exhibit itself, a takeaway line stating what the page establishes, and a source line giving publisher, publication date and access date.
One page, one idea, one source line. Full-bleed dark pages carry argument rather than data and are marked as opinion. E denotes a Frontier Labs estimate; F denotes a published forecast by the named source; unmarked figures are reported actuals.
What this report is not
It is not a forecast product, a vendor comparison, an investment recommendation, or a position on whether AI is good. Where the evidence supports two readings, both are stated and the tension is left visible.
The intent is a single coherent point of view rather than a survey of consensus. Every chapter thesis, chart selection and interpretation in this document is the author's editorial responsibility.
Front matter, About this report
The complete first edition runs to 500 pages across eleven chapters, with every exhibit sourced on its own page. It is available in full as a PDF below.
© 2026 Frontier Labs. All rights reserved. No part of this publication may be reproduced for commercial purposes without written permission. Charts may be reproduced with attribution to The AI Frontier Report. This document is research, not investment, legal or tax advice. It contains estimates and forward-looking statements that are uncertain by construction.
Sources and references
- [1]Company filings, audited statements and direct guidance.First tier of the report's evidence hierarchy. Used wherever a figure exists in a filed or audited form.
- [2]Official statistics and regulator publications.Including national statistical agencies, energy regulators and the statutory instruments underpinning Chapter 10.
- [3]International Energy Agency; national grid operators.Data-centre power demand and capacity series used in Chapter 02.
- [4]Epoch AI; Stanford HAI; OpenRouter; Hugging Face.Specialist trackers with published methodology, used for compute, capability and open-weight series.
- [5]Crunchbase; KPMG; Gartner.Venture, enterprise adoption and deployment survey series.
- [6]Frontier Labs research desk estimates, explicitly marked.Approximately one exhibit in six. Constructed from disclosed anchors with stated assumptions and carrying an E marker on the exhibit itself.
Source assessments, provenance records and the full data appendix are described in the methodology. Corrections to this publication are recorded on the corrections page.
Keywords
- artificial intelligence
- compute
- inference
- capital expenditure
- robotics
- governance
- sovereign AI
- open weights
Full edition
Read the complete report.
The full edition is released to named readers. Tell us who you are and it opens straight away.
- Extent
- 500 pages
- File
- PDF · 1.6 MB
- Price
- Free
Cite this publication
Smit S. Joshi (2026). The AI Frontier Report. Frontier Labs. https://frontierlabs.research/research/ai-frontier-report-2026
Research infrastructure powered by AGI-S.