We aren’t delivering energy fast enough to meet the UK’s artificial intelligence (AI) growth aspirations. London, Europe’s largest data centre market, has an electricity grid already at capacity, partly due to high power needs from the data centres already connected.
Hi Paul - the UK isn't developing much new hydropower but certainly relies on it for some of its energy now. I think the UK is betting on nuclear for the future rather than more hydro, given the high cost, long build times, and disruption to the environment.
Great article! We’re looking at CBE at the viability of these types of flexible data center models in Africa, too. Though lack of fibre capacity limits land available for co-located project development.
A single cloud services provider may not use a single physical building for JUST AI model training. Meaning uptime requirements would default to the highest minimum standard acceptable for other services running. How would we find out?
Interesting on your Africa fibre constraints, in the UK the grid is the big constraint and fibre is comparatively easier/quicker to build because it's privately owned and less regulated.
My observation of uptime for training was more about how the process can handle an outage or intermittency without it ruining the model training entirely or disrupting daily economic life and internet usage. Uptime itself for the asset would be a trade-off that I haven't analysed in depth - e.g. lower utilisation of the data centre might not be worth it for the cost gain in reducing power bills, conversely getting online a few years earlier at half the capacity might be beneficial versus waiting only until the grid connection comes. There's also another nuance discussed at a conference this week, that the large models like ChatGPT and Anthropic are in a race against each other to innovate, so for them uptime matters not because of disruption to daily data services but because it sets them back relative to competitors. For them, an intermittent model is unlikely to be worth the reduced energy cost benefit, but off-grid could be if it means they can get power faster.
Hi Lucy,
Do you see hydro power , whether stored or run of the river having any scope to power off-grid data centres in UK please?
Regards
Paul
Hi Paul - the UK isn't developing much new hydropower but certainly relies on it for some of its energy now. I think the UK is betting on nuclear for the future rather than more hydro, given the high cost, long build times, and disruption to the environment.
Great article! We’re looking at CBE at the viability of these types of flexible data center models in Africa, too. Though lack of fibre capacity limits land available for co-located project development.
Can you say more about the observation that data centers training AI models don’t require the same level of uptime? 🤔 Not sure exactly how space in data centers is owned / rented, but for redundancy purposes instances of a single component are frequently run across multiple physical devices: https://learn.microsoft.com/en-us/azure/reliability/concept-redundancy-replication-backup#physical-locations-in-the-cloud
A single cloud services provider may not use a single physical building for JUST AI model training. Meaning uptime requirements would default to the highest minimum standard acceptable for other services running. How would we find out?
Interesting on your Africa fibre constraints, in the UK the grid is the big constraint and fibre is comparatively easier/quicker to build because it's privately owned and less regulated.
My observation of uptime for training was more about how the process can handle an outage or intermittency without it ruining the model training entirely or disrupting daily economic life and internet usage. Uptime itself for the asset would be a trade-off that I haven't analysed in depth - e.g. lower utilisation of the data centre might not be worth it for the cost gain in reducing power bills, conversely getting online a few years earlier at half the capacity might be beneficial versus waiting only until the grid connection comes. There's also another nuance discussed at a conference this week, that the large models like ChatGPT and Anthropic are in a race against each other to innovate, so for them uptime matters not because of disruption to daily data services but because it sets them back relative to competitors. For them, an intermittent model is unlikely to be worth the reduced energy cost benefit, but off-grid could be if it means they can get power faster.