[graypegg]

> Sovereign AI compute, owned and/or allocated by the public sector, will enable the UK to quickly and independently allocate compute to national priorities. [...] Sovereign AI compute will almost certainly be the smallest component of the UK’s overall compute portfolio.

> NB: this review has not considered the requirements of non-AI high-performance computing, for which there is already a well-established case, including the need to deliver an exascale capability. Government should seek to resolve this as soon as possible, noting that these systems will play a crucial role in supporting AI science and research.

What an interesting addendum to add to that. The government, or at least who ever is writing this report, seems to imagine "AI" as a uniform and distinct process, where you turn on the AI machine, electricity goes in and AI comes out. "Can't use a non-AI machine. It doesn't make AI." The government should already be investing in computational resources for a huge number of government projects and services. The N.B. makes a good point.

The fact someone needed to add this note, and it made it into the final public document, written in that tone, seems like this isn't really a plan, more so just a vibes-based promise for innovation from 1 team in the government.

[dkdbejwi383]

> The government, or at least who ever is writing this report, seems to imagine "AI" as a uniform and distinct process, where you turn on the AI machine, electricity goes in and AI comes out

This is how a large chunk of the general public think of "AI" as well. They're just waiting for someone to flip the AI switch or add a bunch of `from ai import magically_do_everything` to products and services they use.

[fredley]

Don't blame them, this is how products are marketed and sold now, e.g. AI washing machines, AI microwaves, AI features in cars, AI home lighting. Of course, most of these things don't have anything a software engineer or computer scientist might reasonably call AI in - they're just using sensor data and (conventional) algorithms, the same way they have for decades, but AI is the buzzword now.

[netdevphoenix]

Can't blame them. OpenAI, Alphabet and co have been pushing this AI magic marketing in the last 3 years or so. This AI magic seems to be also fed to their investors if Sam Altman's statements are anything to go by. At some point, the AI musical chair game will end and many players will lose everything

[segasaturn]

> seems like this isn't really a plan, more so just a vibes-based promise for innovation

To be fair, that description lines up with a lot of what I hear from other AI startups.

[elric]

> The government should already be investing in computational resources for a huge number of government projects and services

From what I see, governments are massively shifting to Azure and AWS (moreso the former).

[DrScientist]

As long as you have decent utilisation then buying and running your own hardware is typically a lot cheaper than using AWS etc al.

Cloud shines for dynamic loads, global resilience etc - but if you just want to run a lot of computational jobs for a long time, then running your own kit is likely to be much much cheaper.

The other advantage of your own kit is you can guarantee it will be available when you want it, and you can optimise the stack for your own applications.

There is a reason something like 3/4 of Amazons profits come from AWS and not selling books et al.

[elric]

> As long as you have decent utilisation then buying and running your own hardware is typically a lot cheaper than using AWS etc al.

The cynic in me suspects that some government consultant (I used to be one) is making a lot of money from peddling AWS/Azure.

[graypegg]

It does make me wonder if some nationalized utility like "British Compute" or something would be a good idea. They could over-allocate and sell the unused resources to the private sector, AWS style. Could bring long term costs down even further. (With bonus political levers to pull, like giving British owned companies special rates)

[DrScientist]

Are you suggesting an offshoot of British Energy - as a lot of the costs are energy costs anyway?

Note the same logic as above applies for centralised services whether it be British Compute or AWS versus having your own - any central service is going to have overheads around managing the shared resources - virtualisation, scheduling, availability, security, billing etc.

As in the real world - sometimes a renting model sense, but you pay a premium for the convenience.

[graypegg]

That is my thinking, yes. Plus also simply the government has access to a lot of capital and land.

Overhead from running your own datacentres is a factor, but (taking this report at face-value) it seems like there will be some modicum of government owned computer resources built. I feel like the cost difference between standing up a small datacentre capable of what they are pitching, versus standing up an over-allocated datacentre is mostly centred on initial 1-time cost. Once things are running, the ongoing cost may not be much different. Even if the government was the only client at a small datacentre, they would still have to have support, maintenance, and management. If they were instead selling over-allocated resources back to the private sector, there would be sales staff, but otherwise the other expenditures would remain nearly the same. They wouldn't even have to be that competitive on anything other than price, functioning much like a utility. Fancy PaaS features can be fully the responsibility of the private sector to develop on top of the true-blue british metal + electrons the government is offering up.

[DrScientist]

Shared resources add complexity - perhaps a middle ground would be a government run data centre ( modular nuclear reactor etc, reuse of wate heat etc ) which various organisations can host their own hardware.

This already exists in the private sector ( but without the dedicated nuclear reactors ).

[milliams]

They're already investing in non-cloud GPU-heavy supercomputers designed for AI research (i.e. into- and using-AI). For example the Isambard-AI supercomputer at Bristol University (https://www.bristol.ac.uk/news/2023/november/supercomputer-a...) and this report seems to be talking about doing a lot more of that - in addition to non-"AI".

Disclaimer - I work at Bristol on Isambard-AI