Radar follow-up / May–July 2026
FP1’s archived reading of the AI buildout: what was changing, what remained uncertain, and what to check next.
The July interpretation
The July review put more weight on the practical limits to deployment: electricity, chip supply and the cost of infrastructure. It also kept asking whether improving AI agents were changing work at the same speed as their benchmark results.
This is FP1’s interpretation of the period, not a new September assessment. The public reading below explains the questions without claiming success against criteria that are not shown here.
Read the dated July market briefing →What the review examined
Power and infrastructure
The review treated electricity connections, equipment and financing as limits on how quickly AI could be deployed.
Look for completed connections and operating capacity. An announced project is not the same as a working facility.
Berkeley Lab: interconnection queues ↗
This tracks power-generation and storage queues, not data-centre connection times directly.
AI agents
The review kept agent reliability under close watch as performance improved.
Compare the tasks, success rate and level of human help. A result on a coding benchmark does not establish reliability across every kind of work.
METR: measuring long-task performance ↗
Public benchmark research; no exercise-specific performance target is reproduced here.
Jobs and hiring
The review focused on differences between groups of workers rather than treating the national unemployment rate as the whole story.
Separate fewer new hires from layoffs, and changes in one occupation from changes across the economy. Employment data alone do not establish that AI caused a change.
BLS: Current Population Survey ↗
Public labour-market data, not a direct measure of AI’s effect.
Security
The review distinguished people using AI in an attack from AI acting outside human control.
Look for evidence about the operator, the system’s permissions and how the incident began. Greater automation does not by itself settle who was responsible.
RAND: published Robot Insurgency report ↗
A hypothetical exercise and its published findings, not evidence that the imagined attack occurred.
Power and access
The review examined ownership, partnerships and access to computing as governance questions.
Read the contracts and control rights: who can set terms, restrict access or leave a partnership?
FTC: AI partnerships and investments ↗
Public research on partnerships; FP1’s interpretation is its own.
Hé / A question added in June
The US–China chip discussion added a useful test: which parts of the AI supply chain can each country build for itself, and where does it still depend on the other?
That changes how to read an export restriction. Its effect depends partly on whether the buyer can replace what is being restricted.
Meet Hé →Questions for investigation—not a claim that either country is self-sufficient.
Sources and publication boundary
Infinite Potential is RAND’s publicly described exercise platform. RAND has published reports on Robot Insurgency and Cyber Surprise. Those reports are public; that does not make every session document or participant’s notes public.
This edition uses the linked public research and FP1’s own broad interpretation. It excludes session notes, scenario-specific benchmarks and unpublished exercise details.
The earlier page showed a numerical radar, exercise-specific tests and a “five out of five” result. Those elements have been removed because the underlying exercise criteria are not available for readers to check here. The diagrams’ distances were not measured probabilities.
This overview does not certify the earlier grades. FP1’s interpretation is independent of RAND and should not be read as RAND’s findings or endorsement. The dated public market briefings remain available in the archive.