LLM Inference Energy Efficiency: How Many Tokens Can One kWh of Electricity Generate?

As large language model services move from the “can it work” stage to the “can it be afforded” stage, inference cost is becoming a core variable that determines whether a business model is viable. Among these costs, electricity is a fundamental item that cannot be avoided. One question that developers often overlook—but that is highly worth analyzing—is: how many tokens can a single kilowatt-hour (kWh) of electricity produce? This metric reflects infrastructure efficiency and directly affects the long-term sustainability of API pricing.

This article starts with an energy consumption estimate, derives the token output per unit of electricity, converts that output into revenue at typical list prices, and discusses what these numbers imply for the model-serving landscape.

1. Why “Tokens per kWh” Deserves a Serious Calculation

For developers, tokens are the billing unit; for infrastructure, tokens are the product of computation and energy consumption. Connecting the two yields a very intuitive efficiency metric:

• Energy efficiency: how many output tokens can be generated per joule of energy;
• Cost mapping: the token volume corresponding to one kWh, multiplied by the unit price, gives the “revenue ceiling” of the electricity investment;
• Business model assessment: when the revenue per kWh is far higher than the electricity cost, pricing has considerable headroom; otherwise, the model is fragile.

Note that the following estimates are based on typical assumptions for general-purpose inference scenarios, not measured data from any specific product. Different hardware, batching strategies, model architectures, and data center conditions can all produce significant differences, so the conclusions should be treated as an order-of-magnitude reference rather than precise values.

2. Step One: How Many Output Tokens per kWh

2.1 Key Assumptions

Consider a large-batch inference scenario for frontier LLMs running on mainstream high-power accelerator clusters, with the following assumptions:

• Per-GPU power draw of roughly 700W;
• With data center cooling, networking, CPU, and other overheads included, using a PUE of about 1.3, the effective power draw is roughly 900W;
• Under large-batch inference, aggregate output throughput of about 1,000 tokens per second per GPU (a conservative range of 500–3,000).

2.2 Energy per Token

The energy consumed per output token is:

900 joules/second ÷ 1,000 tokens/second = 0.9 joules/token

One kWh equals 3.6 million joules, so:

3,600,000 ÷ 0.9 ≈ 4 million output tokens per kWh

In other words, under fairly ideal batch-inference conditions, one kWh of electricity can produce roughly 4 million output tokens. If throughput drops to 500 tokens per second, this figure halves; if throughput rises to 3,000 tokens per second, it can reach the tens of millions.

2.3 Why the Input Side Is Far Higher Than the Output Side

Prefill on the input side is parallel computation with much higher throughput. At an estimated 10,000 tokens per second per GPU, the energy per token is about:

900 joules/second ÷ 10,000 tokens/second = 0.09 joules/token

This corresponds to roughly 40 million input tokens per kWh, about 10 times the output side. This explains why input and output tokens are usually priced differently.

3. Step Two: Converting Token Output into Revenue

3.1 Output Side

Assume an output price of $50 per million tokens:

4 million tokens × $50 per million ≈ $200 per kWh

3.2 Input Side

Assume an input price of $10 per million tokens:

40 million tokens × $10 per million ≈ $400 per kWh

3.3 Mixed Workload

In real-world workloads, input and output tokens are usually combined in a certain ratio, with a typical input-to-output ratio of about 3:1. After blended conversion, the revenue per kWh is roughly:

$250–300 per kWh

Given the uncertainty in throughput, a reasonable range is approximately $100–800 per kWh.

4. What These Numbers Tell Us

4.1 Electricity Is a Small Share of Revenue

Even with conservative throughput estimates, the revenue per kWh is significantly higher than the procurement cost of the electricity itself. This means that in mature large-batch inference scenarios, electricity is not the main pricing bottleneck; hardware depreciation, R&D amortization, networking, and operations are the more significant cost items.

4.2 Throughput Is the Core Lever of Energy Efficiency

From 500 to 3,000 tokens per second, token output per unit of electricity varies by a factor of 6. Ways to improve throughput include:

• Larger batch sizes;
• More efficient attention implementations and KV cache management;
• Model compression techniques such as quantization and distillation;
• Request scheduling and continuous batching optimizations.

These optimizations do not require changing model capabilities, yet they directly improve token output per unit of electricity.

4.3 Implications for Service Models

When the revenue per kWh is high enough, providers have room to lower prices and expand the market; conversely, if a service model’s energy efficiency remains low over the long term, its pricing headroom will keep shrinking. In other words, energy efficiency is becoming a fundamental variable in the competitiveness of model serving. Below is a simplified energy-efficiency estimation example:

def tokens_per_kwh(power_w, pue, throughput_tps):
    effective_power = power_w * pue
    joules_per_token = effective_power / throughput_tps
    return 3_600_000 / joules_per_token
# Example: 700W per GPU, PUE 1.3, throughput of 1,000 tokens/s
print(tokens_per_kwh(700, 1.3, 1000))  # approximately 4 million

5. Conclusion

One kWh of electricity produces roughly 4 million output tokens, worth about $200 at list prices, or roughly $250–300 per kWh in a mixed input-output workload. These figures show that in large-batch inference scenarios, electricity costs still have considerable buffer relative to revenue. What truly determines long-term competitiveness is the ability to optimize throughput, hardware utilization, and overall engineering efficiency. For developers and enterprises, understanding this energy-efficiency logic helps in more rationally evaluating the cost structure and sustainability of different model services.