Sovereign AI is growing: our first operating milestones
Sovereign AI is growing: our first operating milestones
Argyll Data Development’s sovereign AI inference platform is moving from launch into regular use. Five contracted users are putting it to work on code development and agentic workflows—giving us an early picture of demand, performance and energy use in a live environment.
Our first benchmark was one billion tokens processed in a day. On 3 October, the platform processed roughly 3.7 billion. Across the first ten days of October 2026, it processed about 11.2 billion tokens. The total includes both input tokens sent to models and output tokens generated in response.
These figures reflect more than individual questions and answers. Agentic workflows can make repeated requests as they work through code, context and successive tasks. The growing volume shows how our first users are beginning to put sovereign inference to practical use.
What happened on the busiest day
The 3 October peak gave us an opportunity to see how the service behaved under heavier demand. For MiniMax M3, which handled most of the token volume, the daily p90 total response time was around 15 seconds. That was lower than on the days immediately before and after the peak.
It is an encouraging early result: the highest-volume day did not bring a corresponding rise in that measure of response time. We are also examining the requests that took much longer. Reducing those longest response times is an important part of preparing the service for more users and workloads.
Measuring energy as well as throughput
Across 1–10 October, the cluster used 12.1 MWh while processing about 11.2 billion tokens. During the busier 2–4 October period, it processed 4.95 billion tokens using 3.70 MWh—around 45% more tokens per MWh than the ten-day average.
This gives us an operating baseline to improve against. It is not a direct comparison with another AI platform: energy per token depends on the models, the mix of input and output, and how fully the infrastructure is used. We intend to track those factors as demand grows, so that greater capacity also means better use of the power available.
The system uses no water for cooling. We plan to begin introducing green energy in 2027. Alongside work to reduce energy per token, that should lower the operational emissions associated with serving each workload.
Building the next stage
Sovereign AI has to be useful in practice. Days like this are especially useful. They show us how real workloads behave as demand rises and where response times can stretch. We’re studying the longest requests and using what we learn to tune the service, helping us maintain dependable delivery as more users and workloads come online. For us, that means giving organisations a controlled environment for real work while being open about what the platform delivers and what we are still improving.
These are early milestones, achieved with a small group of contracted users. It is a small but important step. Each real workload gives us evidence to improve the service and build the dependable support customers need as they grow. We will keep measuring performance and energy use, and sharing what we learn.
Become part of the story. To explore how sovereign AI could help your business, contact Argyll Data Development.
