John Carmack argues residential mini data centers running Nvidia hardware are economically nonsensical compared to centralized hyperscalers

@ID_AA_Carmack I could probably get 1B step/second small-model RL on a fridge sized box though! Maybe they'll deliver me one and forget about it when they go bankrupt?

@ID_AA_Carmack Sounds similar to what Tesla/SpaceX are doing at Superchargers.

@ID_AA_Carmack The communication would suffer of latency and tons of logistical problems right

@ID_AA_Carmack Well it’s a hell of a way to circumvent data center zoning restrictions

That’s probably 98% of it, especially since the home builder Pulte Homes is Atlanta based, a city with a slew of data center restrictions since Sept 2024

@ID_AA_Carmack local models

@ID_AA_Carmack They may deploy some as a PR move as a way to literally invest people in the idea of datacenters.

Kind of a victory garden sort of thing.

Cynical I know but I take nothing for granted these days when it comes to propaganda, even if I agree with the goal.

@ID_AA_Carmack Once people research tax laws and find out they're going to need to claim this as income, it'll put a damper on the whole situation. Not a bad idea in theory though.

@ID_AA_Carmack Maybe not to individual households, but why not businesses that might have space on their roof and already have physical security, power, and internet?

They just repurposed a half-abandoned mall as a pediatric clinic. It saved a lot of money not having to build it from scratch.

@ID_AA_Carmack seems like a crazy extension of the folding@home project where instead of computing proteins you're computing...everything else!

@ID_AA_Carmack the pitch makes sense until u think about installation and support at scale

gonna be a rough rollout

@ID_AA_Carmack Except inference doesn't need efficiencies of scale.

A home data center is more than enough to handle one request.

@ID_AA_Carmack Building large datacenters is quickly becoming illegal. So this is the alternative.

Backyard DCs is indeed nonsense, and they need to have regulatory capture around the large DCs - They have to bring their own power or add to grid 1:1 - They have to have soundproofing as part of building codes - No evap towers - use closed loop or heat pumps Outside of that, @satyanadella brings up a point - the US is training SOTA models that are really large and inefficient instead of focusing on maximally useful models. Smaller models are in many cases _BETTER_ and use far less resources.

AI needs far fewer resources than many think.

@PiedViper @ID_AA_Carmack They'd have to be paying enough for a business to bother with instaling it, and they'd have all the losses of shipping discrete components, all the enclosures and idiot-proofing, dealing with loses (failed hardware, stolen hardware), obviously the increased latency, etc.

@ID_AA_Carmack i said throw them in teslas and pay people to own their vehicles (or offset costs of maintenance), where they can either get gpu grid access or contribute.

@ID_AA_Carmack data centers efficient until the power bill goes public and vc realizes training one model costs more than their last 3 exits combined — seen 4 stealth gpu plays since 2021, all pivoted to "inference infrastructure" afte

@ID_AA_Carmack They don’t want facts. They want to hate datacenters … on the internet

@ID_AA_Carmack the sheer scale of data centers is honestly wild to think about.

@ID_AA_Carmack Nobody's going to park $250k in gear in their front yard. Crackheads will be all over this.

@ID_AA_Carmack at-home math fridges are what happens when the planning board becomes the hottest component in the cluster