Aardvark Daily

Please visit the sponsor!

Brilliance or insanity?

Love it or loath it, artifical intelligence (AI) looks like it is here to stay.

Right now we're still in the genesis stage of this technology and that makes it exciting because things are changing on an almost daily basis. Capabilities grow, flaws are remedied and we're still coming up with ways to apply its jaw-dropping capabilities to new problems.

The biggest problem however, is capacity.

Quite a few of the more popular AI systems are already straining under the load, something that has seen prices rise of late -- especially when you get into areas such as video generation and coding agents.

The capacity is limited by silicon availability, datacenter capacities and energy availability.

Those companies who previously have manufactured consumer/commercial-grade DRAM, NVME storage and GPUs are already pivoting away from that small margin market and dedicating most, if not all, of their capacity to server-grade silicon for AI datacentres. This may soon provide some relief for that constraint.

However, building datacentres takes a lot of time -- often several years from the time the decision is made until all the permissions have been obtained, concrete has been poured to when racking and servers are installed before commissioning. There aren't many ways to speed up this process.

Energy supply is also a big problem right now.

Even if sufficient generation capacity is available it can be very difficult and/or expensive to make sure that enough power is available at the location where a datacentre is being built. Once the reticulation of energy is in place there's still the issue of backup power systems and the bottleneck there, at least for larger installations, is that the manufacture of generators with sufficient output has a leadtime measured in years.

Clearly, if AI development is to continue and rollout is to be ramped up, some novel approach to solving the above problems has to be implemented.

Enter an unlikely player, a company called PulteGroup, a US housing company that has hatched a plan to build mini AI datacentres in the back yard of its new home builds.

This article gives a rundown on the concept and what work is being done to turn it into a reality.

Apparently, the company "can install 8,000 XFRA units about six times faster and at five times lower cost than the construction of a typical centralized 100 megawatt data center of the same size."

Sounds good on paper!

Now you might be wondering just how useful such a small micro-data-centre unit could be when it comes to running a half-decent AI agent.

You might be surprised, very surprised.

To give you some insight, here's a fascinating video showing how little CPU and RAM it actually takes to run some AI models, such as the Chinese Deepseek agentic system. It's fascinating that a humble 4GB Raspberry Pi can produce real-world AI results whilst actually only using 2GB of its limited memory.

Could the future see every household equipped with its own datacentre running a powerful AI system that optimises its internal systems such as solar generation, battery storage, heating, cooling, lighting and almost everything else -- while perhaps also providing a networked resource that supports far larger SuperAI systems with a power and scope we have yet to even imagine?

Scary or exciting?

Carpe Diem folks!

Please visit the sponsor!

Here is a PERMANENT link to this column

Rank This Aardvark Page

Change Font