The AI Trifecta: How Tokens, Power, and Data Define AI’s Limits

Token definition

A token is the basic unit of information processed by AI. In Chinese a token corresponds to roughly 0.6–0.8 characters. For example, the phrase "人工智能" can be tokenized as two tokens ("人工" and "智能") or as four tokens ("人", "工", "智", "能") depending on the tokenization method.

Why tokens matter

A prompt of 1,000 tokens followed by a 2,000‑token response consumes 3,000 tokens.

Current large‑model context windows are about 128 K tokens , roughly 100 k Chinese characters.

GPT‑4‑level models cost a few to a few dozen dollars per million tokens.

Tokens are to AI what bytes are to computers—the fundamental measurement of information processing.

Electricity consumption

Training a large model

Requires tens of thousands of GPUs running simultaneously.

Runs for several months.

Total energy consumption is roughly 50–500 GWh . For comparison, the Three Gorges Dam generates about 2.5 GWh per day , so a single training run needs the dam’s output for 20–200 days .

Inference power

Each GPT‑4‑level conversation consumes about 3 Wh of electricity—comparable to an LED lamp lit for half an hour. Billions of such interactions occur daily worldwide.

Data‑center electricity share

Microsoft announced it will restart the Three Mile Island nuclear plant in 2025 to power AI.

Google and Amazon are competing globally for data‑center locations, making electricity cost a core competitive factor.

By 2030 AI and data‑center electricity could account for 3%–5% of global power use , roughly the total consumption of Japan.

Training data

Sources

Public internet text: webpages, Wikipedia, Reddit, GitHub, etc.

Books and academic papers.

Specialized datasets: code, formulas, scientific data.

Synthetic data generated by AI.

High‑quality data ceiling

Researchers estimate that high‑quality public text on the internet totals 5–35 trillion tokens , while a single GPT‑4‑scale training run consumes tens of trillions of tokens , suggesting an approaching data ceiling.

Resulting trends:

Synthetic data is rising to supplement training.

Companies are fiercely competing for exclusive data sources (e.g., Reddit licenses, book copyrights, domain‑specific databases).

Three‑way equation: Tokens × Power × Data

Token ↔ Power

More tokens require larger models, which need more GPUs and thus more electricity; longer context windows increase compute load and power consumption.

AI’s “intelligence” and its electricity consumption are almost proportional.

Token ↔ Data

Richer data improves the “value” of each token, leading to higher output quality; poor data yields low‑quality results regardless of token count.

Data ↔ Power

Processing larger datasets demands additional compute for cleaning, tokenization, and training, which raises power consumption.

Trade‑off triangle

Goal: larger model (more tokens) → Cost: more power + more data.

Goal: stronger capability (better data) → Cost: more power + larger model.

Goal: lower cost (less power) → Cost: model optimization + data reduction.

Future outlook

Energy as a moat

Cheap, abundant electricity may become a stronger competitive edge than algorithms. Nvidia CEO Jensen Huang stated that “AI’s next bottleneck is energy.” Regions with abundant solar, oil, nuclear, or geothermal resources are attracting AI data‑centers.

Data quality

When data volume ceases to be the bottleneck, exclusive high‑quality datasets become decisive, and synthetic‑data quality assessment will be a key technology.

Model efficiency revolution

Small models (7 B–70 B parameters) with efficient architectures are approaching large‑model performance.

Sparse Mixture‑of‑Experts (MoE) models activate only needed parameters, dramatically cutting inference cost.

Edge AI running on phones or PCs reduces reliance on cloud power.