Capex pullback
Revenue lags the buildout and investors cut AI capex.
Cuts off 79 boxes, 41 of 53 stages · slows 41 · Show on the map →
Hits directly
Signs
The IMF and Bank of England warned in October 2025 of a "sharp market correction" if AI enthusiasm fades, naming slower adoption as a trigger. Bain puts the revenue needed to justify the buildout at about $6tn a year, against about $105bn of run rate at the two largest labs.
What would soften it
Revenue keeps compounding at 2026 rates, and cheaper inference keeps widening demand.
Cut off 79
- Cognitive abundance
- Compute price indices and futures
- State backstop for the AI buildout
- AI-designed medicines
- Strategic AI competition
- AI answers replace search
- AI companions
- AI in public services
- AI tutoring at scale
- Embodied agents
- Low-price AI for emerging markets
- Military AI and autonomy
- World models and simulation
- AI cyber defence
- Ambient AI wearables
- Capex payback gap
- Clinical AI in practice
- Firm-owned post-trained models
- Home robots
- Opaque reasoning
- Self-replicating robot factories
- Virtual cell models
- AGI: top-expert-level AI
- AI-native medicine
- AI-run cyber operations
- Always-on background agents
- Knowledge-rich agents
- Licensed generative media
- Misbehaviour cascades in agent swarms
- Model weight security
- Offshore services displacement
- Trace capture
- AGI declared
- AI R&D uplift
- AI-generated trading signals
- AI-native entertainment
- Audit and attribution
- Automated coder
- Compute deterrence
- Computer use
- Continual-learning agents
- Deskilling from AI use
- Expertise pipeline
- Firms shrink to a few operators
- Government AGI project
- Software that maintains itself
- Verification of AI agreements
- Agent-native organisations
- AI as an electoral issue
- AI in research mathematics
- AI-run capital allocation
- Drop-in remote worker
- Institutions for agent societies
- Legitimate governance
- Liability and insurance
- Measured productivity gains
- Multi-agent coordination
- Output verification at scale
- Reputation and credentials
- Rollback
- Scalable oversight
- Superhuman AI researcher
- Verified slowdown
- Agent economies
- Agent-run discovery
- Agents designing agents
- AI-enabled power concentration
- Alignment
- Civilisation-scale knowledge
- Divergent-copy reconciliation
- Explosive economic growth
- Intelligence explosion
- Internal deployment oversight
- Long-horizon value stability
- Superintelligence
- Human-agent symbiosis
- Long-latency civilisation
- Choosing the objective
- Extended mind
Slowed 41
- China critical-mineral export controls
- US ban on foreign-made robots and drones
- Workflow redesign
- Commodity memory crowded out by HBM
- Compute supply
- Datacentre local backlash
- Macro dependence on AI capex
- White-collar hiring slowdown
- Lab automation
- On-prem and air-gapped frontier AI
- Pretraining at scale
- Protein structure and design
- Conversational assistants
- Materials discovery
- Multimodal perception
- Music generation
- Open-weight models
- RL on checkable tasks
- Test-time compute
- AI research tools in investing
- AI writes most new code
- Image generation
- On-device models
- Perturbation data at scale
- Reasoning models
- Robotaxis at scale
- Sycophancy
- Voice cloning and real-time speech
- AI-enabled persuasion
- Ambient documentation beyond medicine
- Long-horizon agents
- Narrowing of scientific inquiry
- Video generation with sound
- Attention and curation under content glut
- Coding agents as products
- Synthetic media at scale
- Agents that trade for investors
- AI-native application revenue
- Professional-services agents
- Public AI scepticism
- Experimental validation capacity