Macro Markets and Machines_The Economic and Market Transformation Driven by AI_GWM report - Flipbook - Page 24
Macro, Markets, and Machines
November 2025
There is one potential offset, however. Our discussion of “access” is being defined in terms of compute power,
semiconductors (and their quality), and skilled labour – areas where advanced economies have the edge. The
United States in particular, enjoys unparalleled access to Nvidia’s most advanced GPUs, an advantage that China
does not have due to U.S. export controls. However, models like DeepSeek show that AI can be trained more
efficiently, reducing upfront costs and potentially broadening access. Exhibit 18 shows DeepSeek achieves
performance comparable to models from OpenAI, Google, and Anthropic, with a noticeably lower cost. That
said, inference costs (i.e., the cost to run the model once it has been trained) must also be considered given
that they are a contributor to the overall cost of AI systems and a key determinant in how scalable these
solutions are. Over time, improvements in chip design and model optimization may improve energy efficiency
and compute utilization, suggesting that training and inference costs should fall, enabling more costeffective solutions.
On the subject of energy, it is emerging as one of the most important factors as computationally intensive AI
models run on sophisticated infrastructure – from GPUs to cooling systems – that require significant energy
input. Data centres currently consume about 1.5% of total global electricity, but that figure is expected to roughly
double to 3% by 2030 as model training and data storage requirements rise, according to the IEA. As such, the
AI race may be influenced not only by who has access to the most advanced chips or data but also by who has
access to affordable and reliable power to sustain long-term adoption without sacrificing cost-efficiency.
Exhibit 18 – Barriers Related to High Energy Consumption and Operational Costs Are Less Daunting, with
Relatively Inexpensive Open-Source Solutions
70
GPT-5 (high)
Artificial analysis intelligence index
60
Claude 4.5 Sonnet
50
Gemini 2.5 Pro
DeepSeek V3.2Exp
Best quadrant
40
30
Llama 4 Maverick
20
10
0
0
1
2
3
4
5
6
7
Price: USD per 1M tokens - lower is better
Source: Datacentermap; Scotia Wealth Management.
Scotia Wealth Management
23
