What is a Megawatt Worth?

Power is one of the biggest constraints on token generation. Data-centre projects are often delayed because a site cannot get enough power or connect to the grid quickly enough. Data-centre projects are therefore often described by the power available to them, such as a 100 MW site. Builders speak in megawatts, while token factories speak in tokens. What is the exchange rate between the two?

This article estimates how much revenue one megawatt could generate if it were used to run GPUs serving open-source language models. It uses the SemiAnalysis AgentX benchmark to estimate the token throughput of models running agent workloads, then values those tokens at market prices.

Value of a Megawatt

Agent Throughput Per Rack

To calcualte the value of a Megawatt we are going to figure out the token throughput that we can generate with a GB300 NVL72 rack per hour. The SemiAnalysis AgentX benchmark provides estimates for the token throughput of popular open source models, running agent-style benchmarks, on GB300 NVL72 racks. The benchmark uses real agent traces captured in the wild, so they are representative of production token counts and caching behaviour.

Megawatt per Rack

The other half of the equation is the number of such racks we can fit in a 1MW power envelope. A GB300 NVL72 requires approx. 143 kW of installed power. A one-megawatt power envelope can therefore support about seven racks. We will use this estimate, understanding that if in reality you can fit more or less in your power envelope that the total generated value of the tokens will be lower by the same factor.

Tokens Per Megawatt

Combining the throughput per rack with the number of racks gives an estimate of the tokens one megawatt can produce in a year. The calculation separates output tokens, cached input tokens, and uncached input tokens because each category has a different market price. We are using real numbers measured in production inference engines so the token productions are not theoretical. They do not rely on roofline estimates or anything like that, so these are very much achievable values.

What is assumed is that you have the user demand for all these tokens evenly over the course of a year. Of course in reality your system is sometimes over utilized (so you rate limit) or underutilized and you produce fewer tokens than your system can achieve possible. We are really bottlenecked by the ability to sell all these tokens, rather than the ability to generate them.

The total revenue is then calculated as:

$P = P_{\text{input-uncached}} \times T_{\text{input-uncached}} + P_{\text{input-cached}} \times T_{\text{input-cached}} + P_{\text{output}} \times T_{\text{output}}$

How Much Can You Make

In the interactive chart below we take the numbers from SemiAnalysis and use the market rate of the input and output tokens as seen on OpenRouter to estimate the total revenue for running a model on 1 MW worth of GB300s for a year. Typical returns are between 3-10x higher than for market rentals of GB300s which come in around $6/gpu/hour.

Appendix: Cost of a Megawatt

We have included in the chart above the cost of owning a MW of GPUs. We have used the SemiAnalysis GPU TCO Model. This takes into account the cost of the GPUs, and estimates of additional costs like energy, land, and networking. They amortize this over the expected useful lifetime of the GPUs.