EN
22 October 2026 · Jakub M.
An hour at full load on the Mac Studio M5 Ultra: what a million tokens costs
We kept our Mac Studio generating text with sixteen AI models at once for a full hour and logged speed, power and temperature every ten seconds. It produced 980,000 tokens without slowing down, and a million tokens took 0.074 kWh of electricity.
How much does it actually cost to run AI locally? We ran an hour-long test on our Mac Studio with the M5 Ultra and 256 GB of memory, the same machine we rent out, and recorded speed, chip power and chip temperature every ten seconds.
How we measured
- Hardware: Mac Studio with Apple M5 Ultra (36-core CPU, 80-core GPU), 256 GB of unified memory, macOS 27.0.1. Nothing else ran on the machine during the test.
- Model: OpenAI’s gpt-oss-20b in the MXFP4-Q8 version for Apple’s MLX (MLX 0.32.4, mlx-lm 0.32.2). It is a mixture-of-experts model with 21 billion parameters, of which about 3.6 billion are used for each token. It takes roughly 12 GB of memory.
- Load: sixteen separate programs generated text at the same time for 60 minutes, each with its own copy of the model. That simulates sixteen users or AI agents. Every generation was capped at 2,000 tokens and then started again with the same task, a long technical article about how a car engine works.
- Measurement: every 10 seconds we logged the number of generated tokens, chip power and the average CPU and GPU temperature, read from macOS with macmon 0.10.0. Energy is average power multiplied by time.
Results
| Result | |
|---|---|
| Test duration | 60 minutes |
| Parallel streams | 16 |
| Generated | 980,000 tokens |
| Speed | 273 tokens per second |
| Speed, first / last sixth | 272 / 272 tok/s |
| Chip power, average (max) | 73 W (76 W) |
| CPU / GPU temperature after settling | 63 / 66 °C |
| Peak CPU / GPU temperature | 66 / 70 °C |
| Energy per million tokens | 0.074 kWh |
The limit is memory speed, not compute
When generating text, the Mac is held back by memory bandwidth rather than compute. For each token the GPU has to read the relevant part of the model from memory and spends most of the time waiting for data. That is why the chip drew only 73 W even with sixteen streams.
The sixteen programs share the GPU and each reads its own copy of the model. Together they produced 273 tokens per second, about 17 per stream. That is slow for one person chatting, but enough for a group of agents working in the background. Inference servers handle this differently: they process requests from many users in one batch and read the model once for all of them, which gets more tokens per second out of the same Mac. We did not measure that in this test.
Temperatures and stability
The GPU temperature rose to 70 °C in the fourth minute, then dropped and held at 66 °C from the fifth minute on. The CPU settled at 63 °C. The fans kept up without the chip having to slow down: in both the first and the last sixth of the test the Mac generated 272 tokens per second.
What the electricity costs
The hour at full load used 0.073 kWh. Running like this for a whole day would take about 1.75 kWh, and one million tokens needs 0.074 kWh. At €0.25 per kWh that is under 2 cents per million tokens and about 44 cents per day. For local AI the electricity hardly matters; the price of the computer is what counts.
What the test does not show
- Power of the whole computer. We measured only the chip (CPU, GPU and Neural Engine). Fans, SSD, network and power supply losses are not included, so the draw at the wall is somewhat higher.
- Batched serving. Sixteen separate programs are not the same as a server that handles many users in one batch.
- Large models. gpt-oss-20b is small and fast. A dense model with 70 billion parameters would load and power the Mac differently.
- Output quality. We measured speed and power only, not what the model wrote.
These are results measured under the conditions above, not a guarantee: real speed and power depend on the model, quantisation, context length, software and workload.
Try it yourself
If you want to see how your own model behaves under load, you can rent this Mac Studio M5 Ultra with 256 GB from one day. More measurements are in our benchmark of six models.