BLUE SHIFT REPORT
AI'S Hidden Dependencies
Unpacking AI's resource strain and systemic vulnerabilities
Download the full report
On 20 October 2025, a massive AWS outage exposed the fragility of our digital infrastructure, echoing René Barjavel’s 1943 novel Ravage, in which a civilization entirely built on electricity collapses the moment power fails.
On 18 November 2025, a major Cloudflare outage — a key web infrastructure provider — made many sites like X and ChatGPT unavailable for hours!
AI'S DEPENDENCIES
The next chapter of our global exploration of AI’s impact
SIGN UP TO OUR NEWSLETTER
Sign up
Find your nearest office
“The future is not what will happen, but what we will do.”
As organizations scale up adoption of more complex AI applications & commercial solutions built on agentic architectures, their overall environmental footprint could expand dramatically without their awareness — depending on how green their energy provider is
BROWSERS
OPERATING SYSTEMS
EMBEDDED USAGE WILL DRIVE FUTURE ADOPTION
Embedded AI will drive the next wave of growth
What are the “no-regret” moves businesses can make?
What are the “no-regret” moves businesses can make?
What does it mean for businesses?
What does it mean for businesses?
What is the real impact of AI on energy and water consumption?
What is the real impact of AI on energy and water consumption?
AI adoption is exploding
ChatGPT weekly active users (millions)
0
100
200
300
400
500
600
700
800
May-25
Jan-25
Sep-24
May-24
Jan-24
Sep-23
May-23
Jan-23
Sep-22
8x increase in 18 months
100
400
800
Source: Arthur D. Little, SemiAnalysis
THE NUMBER OF AI USERS
IS EXPLODING
As tasks grow in number and complexity
AI'S ENERGY DEMAND WILL
INCREASE EXPONENTIALLY
But inference dominates AI’s carbon footprint
MOST THINK TRAINING
IS THE PROBLEM
Complex queries increase energy needs
Text
Video
~ avg Google search
~ iPhone full charge
~1h of laptop
5 min video ~ Tesla full charge
0
5
10
15
20
25
30
35
40
800
950 Wh
5 second 16 FPS video with CogVideoX
GPT-4o document review (75k words)
Avg. Chat GPT-5 100 word query
Avg. GPT-4 100 word query
8hr human work day
800 Wh
0.2-0.34 Wh
18
40
950 Wh
~ 90x
~ 120x
~ 2,800x
Source: Arthur D. Little, EpochAI, MIT Technology Review, expert interviews
Embodied impacts
Operational impacts
End-of-life
Compute-level Impact of AI across OpenAI GPT-4 lifecycle
~0.3 - 1.5
Material extraction and chip production
~0.05-0.15
Downstream distribution of GPUs for training GPT-4
10-15
GPT-4 model training on Microsoft Iowa data center
~600-850
Annual inference emissions from GPT-4
0.01-0.05
E-waste emits low levels of GHGs
Emission impact
(KtCO2e)
AI has an environmental footprint across its entire lifecycle, although as usage scales, inference could account for up to 90% of the total
Source: Arthur D. Little; Jegham, Nidhal, et al. “How Hungry Is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference.” arXiv preprint, 14 May 2025; Falk, Sophia, et al. “More Than Carbon: Cradle-to-Grave Environmental Impacts of GenAI Training on the Nvidia A100 GPU.” arXiv preprint, 27 August 2025.
End-of-life
Model Inferemce
Model Training
Product Assembly
Raw materials
Negligible
Moderate
High
Very High
Key sources of pricing uncertainty
Faster hardware turnover — GPUs now depreciate in <1 year (vs. 4-5 years before), undermining ROI
Infrastructure lag — data centers take 2-3 years to build, but energy hookups can exceed 7 years, slowing deployment
Broken business models — freemium AI services keep prices artificially low while costs surge
Disappointing returns — MIT Media Lab study found 95% of AI pilots generated no measurable business benefit
Lock-in is increasing across models, chips, and cloud platforms
GLOBAL AI DEPENDENCIES MAKE SWITCHING
OR DIVERSIFYING NEARLY IMPOSSIBLE
They shape cost predictability, sustainability credibility, and strategic autonomy
FOR EXECUTIVES
AI'S RESOURCES ARE STRATEGIC
What are the “no-regret” moves businesses can make?
What are the “no-regret” moves businesses can make?
What does it mean for businesses?
What does it mean for businesses?
What is the real impact of AI on energy and water consumption?
What is the real impact of AI on energy and water consumption?
And hiding significant future costs
AI INFRASTRUCTURE SPENDING
IS OUTPACING RETURNS
As AI usage scales across systems and suppliers
COMPANIES ARE LOOSING VISIBILITY
ON AI'S TRUE ENVIROMENTAL FOOTPRINT
AI infrastructure spending is outpacing returns and hiding future costs
Massive CAPEX race among the Big Six tech players…
0
100
200
300
400
500
600
CAPEX of Big Six tech1 ($Bn)
30
25
24
23
22
21
20
19
18
17
16
15
14
500 - 600+
400
210
130
125
95
65
60
65
46
42
35
30
2x
2x
2x
1.5x
Note (1): Big Six includes Apple, Nvidia, Microsoft, Alphabet/Google, Amazon (only AWS, excludes Amazon Retail), and Meta
Source: Arthur D. Little, Bond Capital
… is driving cost instability & uncertain platform pricing
Transparency has collapsed — fewer than 3% of new AI models now disclose environmental data, down from 10% in 2023
Voluntary reporting is unreliable — lifecycle metrics are inconsistent, partial & often unverifiable
A new baseline is needed — a minimum viable transparency standard combining model- & facility-level data could realign claims with physical reality
North America controls most of the AI value chain, from chips to cloud
Europe & APAC remain dependent on US technologies
Taiwan holds a critical position as the world’s dominant hub for advanced chip fabrication
AI model transparency peaked in 2023, then collapsed in 2024
Low environmental disclosure
Increasing disclosure
Transparency
collapse?
0
30
60
90
120
Note: 1. Family rollouts often counted as 1 or 2 releases and not nominally (e.g., Anthropic: Claude 3 Opus, Claude 3 Sonnet, Claude 3.5; OpenAI: GPT-4o, GPT-4o mini, o1, o1-mini, o1-preview)
Source: Arthur D. Little; Epoch AI; Luccioni, Sasha, et al. “Misinformation by Omission: The Need for More Environmental Transparency in AI.” arXiv preprint, 18 June 2025
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
10
11
13
19
33
32
48
55
36
35
5
3
10
9.5
12
16
28
27
41
47
26
29
15
25
34
39
33
5
7
8
10
3
67
31
3
49
50
3
78
51
6
91
62
12
113
28
2
63
Direct disclosures
No disclosures
Indirect disclosures
AI Applications
Foundation
Models
Training Data
Cloud Computing
Chip fabrication
Lithography
CHIP DESIGN
Dominant, global leadership
Dominant, regional leadership
Competitive in some niches
Weak or emerging
Absent
US
EU
APAC
# AI models
Resilience in AI means managing physical dependencies as strategic assets rather than hidden risks
NO REGRET MOVES
What are the “no-regret” moves businesses can make?
What are the “no-regret” moves businesses can make?
What does it mean for businesses?
What does it mean for businesses?
What is the real impact of AI on energy and water consumption?
What is the real impact of AI on energy and water consumption?
Strategic move 1
Restore environmental credibility
Strategic move 2
Anticipate the real cost of AI
Strategic move 3
Build strategic resilience
Objective: Gain visibility and control over the real footprint of your AI use
Key actions
• Require transparency -
make environmental disclosure (energy, water, materials) a non-negotiable criterion in provider contracts.
• Push standardization -
align suppliers on comparable, auditable metrics to avoid inconsistent or partial reporting.
• Prepare for regulation -
establish internal tracking frameworks now; when rules arrive, you’ll already be compliant.
Objective: Keep AI costs predictable and aligned with real business value
Key actions
• Quantify total cost of ownership - include compute, energy, and water in financial planning as today’s prices are artificially low.
• Prioritize efficiency -
invest in model and workload optimization now to reduce future exposure.
• Renegotiate contracts -
secure long-term pricing and transparency clauses before providers pass on real energy costs.
Objective: Maintain the freedom to move and adapt across providers and jurisdictions
Key actions
• Design for portability -
use architectures and formats that can migrate between clouds, vendors, or regions.
• Protect your data rights -
ensure contracts guarantee ownership, portability, and prevent model training on your data.
• Diversify providers -
distribute workloads across multiple suppliers; combine global scale with sovereign or on-prem options.
FIND OUT HOW WE CAN HELP
Contact our team
Get in touch
Get in touch
Find your nearest office
Find your nearest office
“The future is not what will happen, but what we will do.”
— attributed to Henri Bergson, early 20th century
1