Independent research institution
Research since 2022
First Edition · 2026
Frontier Labs
Researching Tomorrow, Today.
Independent research focused on artificial intelligence, computing, economics and emerging technologies. Every finding is published with its methodology, its sources and its uncertainty.
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.
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 …
Recent papers and briefs.
Browse the full archive- 01
Sovereign Compute
National AI infrastructure programmes: stated objectives, committed capital and delivered capacity.
ForthcomingAreaPublic PolicyPublished15 SEPT 2026Paper · 20 min read - 02
Evaluation Integrity
A contamination and construct-validity audit of twenty-eight public benchmarks.
In reviewAreaArtificial IntelligencePublished22 JUL 2026Paper · 17 min read - 03
Software Production After Assistants
Throughput, defect density and maintenance burden in machine-assisted engineering.
PDFAreaSoftwarePublished02 JUN 2026Paper · 18 min read - 04
The Inference Economy
Reconstructing the unit cost of serving large models, 2023 to 2026.
PDFAreaEconomicsPublished28 MAY 2026Paper · 20 min read - 05
Compute Concentration
Mapping the global distribution of installed AI training capacity across 214 facilities.
PDFAreaComputingPublished12 MAY 2026Paper · 22 min read - 06
Compute Thresholds in Regulation
An assessment of training-compute triggers as a regulatory instrument.
PDFAreaPublic PolicyPublished06 MAY 2026Brief · 14 min read
Research papers
Peer-reviewed internally and externally refereed
Published reports
Long-form assessments, including annual editions
Primary data sources
Filings, registers, records and datasets under active tracking
Charts produced
Each traceable to its underlying series
Countries covered
Across infrastructure, policy and labour datasets
Years analysed
Longest continuous series begins in 2000
Why Frontier Labs exists.
The most consequential technologies of this decade are being described almost entirely by the people building them and the people selling them.
That is not a conspiracy; it is a structural consequence of where the information sits. But it produces a public record that is confident where it should be uncertain, precise where the evidence is thin, and silent about the assumptions that carry the conclusions.
Frontier Labs was founded to publish a different kind of record: one assembled from primary sources, stated with its uncertainty attached, and open to being checked by anyone who wants to check it. We would rather publish a narrow finding we can defend than a broad one we cannot.
Evidence before opinion
We do not publish a view we cannot support with material a reader could check. Where the evidence is thin, we say the evidence is thin. Where we hold a view that the evidence does not yet establish, we mark it as a judgement and separate it from the findings.
Independence
Frontier Labs does not accept funding contingent on findings, does not offer pre-publication review to subjects of research, and does not publish commissioned work under its research imprint. Our funding structure is disclosed and our conflicts register is public.
Transparent methodology
Every publication carries its methodology in full: sources, assumptions, intervals, and the analytical choices that could have gone another way. A reader who disagrees with our conclusion should be able to locate precisely where they disagree.
Long-term thinking
We choose questions by their durability rather than their currency. Most of what we study will still matter in a decade. We would rather be slow and right than fast and cited.
Neutrality
We describe rather than advocate. On contested questions we set out the strongest version of each position and identify the empirical claims on which the disagreement actually turns. We do not have a policy programme.
Correction as practice
We are wrong sometimes. When we are, we correct it with the same prominence as the original claim, we record it permanently, and we examine why the error occurred. Published work is never silently revised or withdrawn.
Six stages, applied without exception.
Read the full methodology- 01
Research
A question is specified precisely enough that it could be answered wrongly.
Work begins with a written question, a statement of what evidence would answer it, and an explicit account of what the answer would change. Questions that cannot be stated in a form that admits of being wrong do not enter the programme. At this stage we also record the prior view of the research group, so that the distance travelled by the evidence is visible in the final publication.
Outputs- Question specification
- Evidence plan
- Recorded prior
- 02
Verification
Every quantitative claim is traced to a primary source and the source is assessed.
We work from primary material: regulatory filings, statutory instruments, interconnection registers, customs records, planning submissions, published microdata and disclosed technical documentation. Secondary sources are used to locate primary material, never as evidence in themselves. Each source is assessed for provenance, collection method, known bias and revision history, and that assessment is recorded alongside the datum.
Outputs- Source register
- Provenance assessment
- Revision history
- 03
Cross validation
Independent methods are used to check whether they reach the same answer.
Wherever a quantity can be estimated by more than one route, we estimate it by more than one route. Compute capacity is derived independently from electrical records, from construction footprint and from hardware trade flows. Where methods disagree beyond their stated intervals, the disagreement is investigated and, if it cannot be resolved, published as an open discrepancy rather than averaged away.
Outputs- Independent estimates
- Reconciliation record
- Open discrepancies
- 04
Analysis
Assumptions are made explicit and their influence on the conclusion is quantified.
Every model we publish carries a sensitivity analysis identifying which assumptions the conclusion actually depends on. Where a conclusion is fragile to a single assumption, we say so in the body of the text rather than in an appendix. Analysis is conducted by a team that did not assemble the data, and a designated reader argues the opposing case before any finding is accepted.
Outputs- Sensitivity analysis
- Assumption register
- Adversarial review note
- 05
Visualisation
Charts are constructed to be checkable, not to be persuasive.
Every chart is generated directly from the underlying series, never redrawn by hand. Axes begin at zero unless there is a stated reason otherwise, in which case the reason appears in the caption. Intervals are shown where they exist. Any chart that could not be reproduced by a reader holding the published data is not published.
Outputs- Reproducible chart specification
- Underlying series
- 06
Publication
The finding, its uncertainty and the conditions that would overturn it are published together.
Publication includes the finding, the interval around it, the methodology in full, the source register, and an explicit statement of what evidence would change the assessment. Corrections are published with the same prominence as the original claim and are recorded permanently in the change log. We do not remove or silently revise published work.
Outputs- Publication
- Methodology appendix
- Change log
- Falsification conditions
Research notes, published as we go.
All insightsThe utilisation number nobody publishes
Cluster size is reported. Achieved utilisation is not. The second number is the one that matters.
Computing28 July 2026 · 5 min readHow to read an AI capex disclosure
Four questions that separate what a filing says from what it means.
Economics11 July 2026 · 6 min readVerification cost explains most of the capability distribution
Where checking an answer is cheap, systems improve fast. Where it is expensive, they do not.
Artificial Intelligence30 June 2026 · 5 min readAnnounced, contracted, energised, consumed
Four different numbers, one word, and most of the confusion in datacentre reporting.
Energy12 June 2026 · 4 min read
Monthly frontier intelligence.
One email each month. New publications, corrections to existing work, and the datasets we have released. No marketing, no forwarding of your address, no spam.
You can unsubscribe from any email. We explain exactly what we store in the privacy notice.