Executive Summary

• Sarah Rees, founder of data intelligence firm Signwl.com, explores how real-time GPU pricing fluctuations could serve as the missing leading indicator for predicting power grid strain across the UK and Europe.
• Across Europe, the newer generations of GPU (H200 and GB200) have shown little movement. Both on-demand and spot discount pricing have held steady since January 2026. The older H100 is currently varying from region to region and month-to-month which makes it a more useful signal for reading regional AI activity currently.
• AI usage is not uniform across Europe. In London, Frankfurt and Amsterdam, capacity appears to be loosening, which suggests AI is unlikely to be the summer grid challenge this year that some expect.
• As AI growth surges across the UK and Europe, the regions where operators, grid planners and cloud compute providers share utilisation forecasts and trends will almost certainly be the regions that manage to keep the lights on.

Cloud GPUs account for a growing share of data centre electricity demand and as a result of
load on the grid more broadly. Power draw varies considerably between GPU types, so tracking
where usage is rising for the highest-draw GPUs can give an early signal of where the grid is
most likely to come under pressure.

Table 1: GPU power draw by GPU

Source: Manufacturer specifications

Usage trends can be inferred from pricing. Unsold GPU capacity is offered at a discount to standard rates, a pricing product known as a ‘spot price’. The difference between the standard price and the spot price is termed the ‘spot discount’. When the discount narrows, it indicates that available capacity is tightening and that usage is rising.

At Signwl.com we have analysed first-half pricing movements across major hyperscalers in Europe to identify where the grid could come under the most pressure this summer. Across Europe, the newer generations (H200 and GB200) have shown little movement. Both on-demand and spot discount pricing have held steady since January 2026. However, the older H100 is currently varying from region to region and month-to-month which makes it a more useful signal for reading regional AI activity currently.

Observation 1: Madrid is seeing consistent utilisation increases of the power-hungry H100 training GPU, suggesting potential increased grid draw in Summer 2026

Chart 1 shows that spot discounts on the H100 in Madrid have narrowed steadily in recent months, indicating a sustained rise in training workloads across both regions. This is notable because the H100 is an older-generation training GPU. In Madrid, the newer H200 has shown no spot discount since January 2026, which suggests little of its capacity is reaching the open spot market, leaving the H100 to absorb the region’s flexible training demand illustrated by the narrowing discount in Chart 1. As demand on the grid rises through June, July and August the data suggests Madrid will need careful planning and collaboration between relevant parties.

Chart 1: Madrid H100 Spot Discount January – June 2026

Source: Signwl.com

Observation 2: London, Frankfurt and Amsterdam are seeing reduced utilisation of the power hungry H100, which may be bad news for AI training, but good news for the grid

As can be observed in Chart 2, unlike Madrid London experienced a sustained increase in spot
discount in H1 suggesting reduction in usage of the H100 in the region. This may be good news
for the grid over the difficult summer months, however bad news for AI training in the region if it
is indicating a broad deceleration of training models locally.

Chart 2: London H100 Spot Discount Jan-June 2026

Source: Signwl.com

Similarly, although even more notably, in Amsterdam and Frankfurt the actual on-demand price
has moved. Chart 3 demonstrates material reductions in on-demand pricing in both Amsterdam
and Frankfurt suggesting materially reduced usage of the H100. Again, potentially good news
for the grid this summer although not good news for AI training velocity in the region.

Chart 3: Frankfurt & Amsterdam H100 On Demand GPU $/hr Jan-June 2026

Source: Signwl.com

Closing Thoughts

Evidently AI usage is not uniform across Europe. In London, Frankfurt and Amsterdam, capacity appears to be loosening, which suggests AI is unlikely to be the summer grid challenge this year that some expect.

Madrid is the clear exception, tightening just as the summer demand on the grid builds, and that is the data worth watching closely over the coming months. For the relevant parties in Spain, if not already happening it would certainly make sense for data centre operators to share compute-demand forecasts with grid operators so that plans can be made accordingly.

And as the AI buildout continues in Europe, the regions where operators, grid planners and cloud compute providers share utilisation forecasts and trends will almost certainly be the regions that manage to keep the lights on.