Oct 2025
44%
U.S. businesses paying for AI tools
↑ from 5% in 2023
$530K
Average enterprise AI contract value
New metric in 2025
95%
Practitioners using AI at work or home
↑ Mainstream adoption threshold reached
1.5×
AI-first startup growth vs. peers
2024–2025 cohort
Report Dimensions
| # | Domain | Status |
|---|---|---|
| 01 | Research | Capabilities leap |
| 02 | Industry | Strong adoption |
| 03 | Politics | Geopolitical tension |
| 04 | Safety | Evolving focus |
| 05 | Survey | 1,200+ respondents |
| 06 | Predictions | Tracked annually |
| 07 | Compute | Industrial scale |
Headline Findings
01
OpenAI retains narrow frontier lead; DeepSeek closes fast
China's open-weights models competitive in reasoning and coding
02
Reasoning models now plan, reflect, self-correct
RL + rubric-based rewards enable extended cognitive horizons
03
AI enters scientific method — hypothesis generation to validation
Scaling laws extended to protein biology via ProGen3
04
Power supply becomes the binding constraint on AI scale
Multi-GW data centers now being planned; sovereign backing accelerates
05
Safety discourse shifts from existential risk to reliability
Cyber resilience, autonomous system governance take center stage
RL+
Reasoning via rubric-based reward models
Core architecture shift 2025
→ Bio
Scaling laws reach protein science
ProGen3 / Profluent
Open
DeepSeek, Qwen, Kimi challenge closed labs
↑ China open-weights competitive
Auto
Autonomous hypothesis-testing systems
Full scientific method coverage
Reasoning Model Capabilities
Competitive Landscape
OpenAI — Frontier leader
Retains narrow lead at frontier; GPT-o series defines reasoning benchmark
Anthropic — Safety-forward competitor
Claude family competitive across reasoning, coding, and agentic tasks
Google DeepMind — Multimodal scale
Gemini Robotics 1.5 extends into physical AI and embodied systems
DeepSeek / Qwen / Kimi — China tier
Open-weights models closing gap on reasoning and coding benchmarks fast
Profluent — Scientific AI
ProGen3 extends scaling laws into protein biology research
44%
U.S. businesses paying for AI tools
↑ 9× growth from 5% in 2023
$530K
Average enterprise AI contract
2025 average
1.5×
AI-first startup growth vs peers
↑ Consistent 2024–2025
Multi-GW
Data center scale now planned
Industrial-era threshold crossed
Adoption Surge 2023→2025
5%
2023
Businesses paying for AI
44%
2025
Businesses paying for AI
Enterprise AI Signals
→
Average AI contracts at $530,000
Enterprise budgets now substantial; AI is a line item, not an experiment
→
AI-first startups grow 1.5× faster
Competitive moat from AI-native architecture compounds over time
→
Robotics / physical AI entering commercial phase
Chain-of-Action planning enables embodied system deployment
→
Power is the new constraint on AI business models
Sovereign funds from US, UAE, China backing infrastructure directly
United States
"America-first" AI positioning
Federal policy oriented around maintaining compute and model leadership
Export controls on chips
Restricting access to advanced semiconductors for adversary nations
Sovereign infrastructure investment
U.S. sovereign funds backing multi-GW AI data center builds
China
Open-weights strategy
DeepSeek, Qwen, Kimi deployed as open models to accelerate global adoption
Compute self-sufficiency
Huawei and domestic fabs scaling production despite export restrictions
Sovereign backing of AI labs
State capital backing frontier research as national strategic priority
Europe
EU AI Act in implementation
Risk-tiered regulation facing practical deployment challenges
Competitiveness gap concern
EU regulators balancing safety with fear of falling behind US/China
Sovereign AI compute
Several EU nations funding national AI infrastructure initiatives
Geopolitical AI Landscape 2025
| Region | Strategy | Key Assets | Key Constraint | Status |
|---|---|---|---|---|
| United States | Closed-model frontier + export control | OpenAI, Anthropic, Google, Microsoft | Energy / political coordination | Leading |
| China | Open-weights + domestic compute | DeepSeek, Qwen, Huawei | Advanced chip access | Closing |
| EU | Regulatory framework + national compute | Mistral, national institutes | Commercial scale, talent | Lagging |
| UAE / Gulf | Sovereign capital + infrastructure | G42, Falcon, data center build-out | Talent, base research capacity | Investing |
Shift
From existential risk → reliability
Dominant framing in 2025
Cyber
Cyber resilience now primary concern
Both offense and defense
Align?
Models can mimic alignment under supervision
↓ Raises transparency concerns
Gov.
Autonomous system governance emerging
New policy area 2025
Safety Focus Evolution
Key Safety Findings
!
Models can mimic alignment under evaluation
Supervised alignment behavior may not generalize; interpretability lags capability
!
Autonomous system governance is an emerging gap
Existing frameworks designed for human actors, not AI agents operating continuously
!
AI-enabled cyberattack capabilities are advancing
Both offensive and defensive applications improving; attack surface expanding
!
Transparency vs. capability tension remains unresolved
Interpretability research advances but capability growth outpaces understanding
95%
Use AI at work or home
↑ Mainstream threshold crossed
76%
Pay for AI tools personally
↑ Strong willingness to pay
1,200+
Survey respondents
AI practitioners globally
Wide
Productivity gains reported
↑ Widespread across roles
Adoption by Metric
95%
use AI
Use AI at work / home
76%
pay
Pay for AI personally
What Practitioners Report
↑
Widespread productivity gains across roles
Coding, writing, research, and analysis are top use cases
↑
Personal willingness to pay is high (76%)
Users pay for AI independently of employer provision — strong pull signal
↑
AI-native workflows are becoming standard
Integration into daily work routines, not occasional usage
GW
Multi-gigawatt data centers planned
Industrial-scale threshold
Power
Electricity supply is binding constraint
↓ Limiting factor for scaling
3x
Sovereigns backing AI infra
US, UAE, China all active
Open
DeepSeek efficiency vs. closed-model cost
↑ Efficiency gains challenge capex moats
Compute Scale Trajectory
2019
2020
2021
2022
2023
2024
2025
Infrastructure Dynamics
⚡
Power is the binding constraint on AI scale
Electricity availability now determines where and how fast frontier AI can grow
🏗️
Sovereign funds directly funding compute
U.S., UAE, China treating AI compute as national strategic infrastructure
📉
Efficiency gains challenge capex moats
DeepSeek's compute efficiency vs. frontier closed-model costs puts pressure on investment theses
🌍
Geographic diversification of compute accelerates
Gulf, Southeast Asia, and European nations all building sovereign AI infrastructure
Prior Year Assessment
✓
Reasoning models with planning + self-correction emerge
Correct — RL + rubric rewards enabled this in 2025
✓
Enterprise AI adoption crosses 30%+ of U.S. businesses
Correct — reached 44% by report date
✓
China releases competitive open-weights models
Correct — DeepSeek, Qwen, Kimi all competitive
~
Physical AI / robotics commercial deployments at scale
Partial — Gemini Robotics 1.5 advanced, but mass deployment still early
2026 Forecasts
?
Autonomous AI agents complete multi-day professional tasks
Watch: agentic benchmarks to validate in 12 months
?
Power supply constraints visibly slow frontier training
Watch: data center announcements vs. delivery timelines
?
EU AI Act enforcement triggers first major commercial case
Watch: regulatory action against large frontier model provider
?
Robotics adoption surpasses software AI in new enterprise contracts
Watch: physical AI deployment numbers vs. SaaS contracts