Hello Jim

> "My feeling is that there are certainly some data centres being built that won't be needed for their currently intended purposes."

Putting on my REIT hat but also my AI professional hat & my IT/HPC sysadmin professional hat. So there are a few ways to do data centers.

And you're absolutely right, some, probably most datacenters that are planned, will not be needed or used for their currently intended purpose.

I'd break down usage models like this:

1. AI/SAASCorp. Buy/lease land and/or building and power grid connection, build a Google Datacenter, fill with Google stuff. Sell Google web services.

2. CloudCorp. Buy/lease land and/or building and power grid connection, fill it with (network gear, cooling gear, servers). Sell hardware capacity.

3. DatacenterCorp. Buy/lease land and/or building and power grid connection, fill it with (network gear, cooling gear, NOT servers). Sell rackspace and connectivity and power.

4. PoweredShellCorp. Buy land, build a big concrete box, connect power. Lease it out to a DataCenterCorp, CloudCorp or AICorp to do what they like with it.


Why does this matter?

Well, if you own a datacenter, as a box with power and expensive PC gear, it can often be hard to imagine profitable alternative uses for the expensive PC gear.

But if you own a datacenter, as merely a box with power, then it's not hard to see profitable alternative uses for it and alternative tenants, if AI demand doesn't pan out.

Primarily: Robotic factories, robotic logistics warehouses, or any industry with high energy intensity in manufacturing.

The question I feel is more about, who is going to lose money on the expensive AI gear that is currently filling some datacenter buildouts.

Or who has slightly overpaid for their powered shell or grid connection.

Many unbuilt datacenters, without GPU racks filling them up, have vast industrial and commercial logistic potential that can be pivoted towards, next year or the year after.

You probably know most/all of this already; I'm likely talking to the audience here.


> "Intelligence per Watt: Measuring Intelligence Efficiency of Local AI"

With a solar panel, the energy cost efficiency of local AI becomes unbeatable.

Where you can re-use hardware assets you already have, the hardware cost/space efficiency of local AI becomes unbeatable.

Yet in my view, (as a professional in the field of AI for decades), the primary advantages of local AI are not in efficiency or skill but in 'control'.

1. You actually know for sure which version & model you are using. How it is configured. How much effort it can use. This is kind of a big deal because it's infuriating to have to adapt work processes on a monthly/weekly/hourly basis as model behaviour and personality changes continually behind the scenes at cloud providers. It's like the difference between having an enthusiastic, knowledgeable yet foolish intern; versus having a continuous stream of interns swapping out every day with different personalities, varying willingness to help, different knowledge and skills.

2. You know what kind of safety / security framework you are using it within your organisation.

3. The ideas and data you put in, are not stored forever, even potentially used against your own business later by competition.

4. Data locality. If I run AI locally, I can point it at a 10TB local dataset without issue. It's much trickier to copy/paste such a dataset into ChatGPT using Chrome browser.

5. Fine-tuning.

6. Other things; sorry, I'm physically/mentally exhausted today - I'll try and finish more on this post later, I just wanted to drop it into the discussion for now with what I had.

Hope this is of interest to someone.

TRS