Another AI is Possible

Rob Lach

When it comes to PFAS-based fluids and data center immersion cooling we’re usually talking about two-phase cooling systems where an operator can leverage the fluid phase-change for additional cooling efficiency. 

How these fluids work is that they are non-conductive (so an operator can fill a tank and submerge the computing hardware directly), have a low boiling point (so hot chips easily boil the fluid around them, pulling heat away through the liquid-to-gas phase change), and consequently condense at a low temperature (so the vapor readily turns back to liquid on a water-cooled radiator above and drips back into the pool). The primary benefits are that heat is moved off the chips without plumbing coolant to each one, operating temperatures are generally lower, and water use is lower than in a typical water-cooled system (because condensing the vapor takes less cooling than carrying heat directly off the chips with water).

The negatives are the many opportunities for PFAS-based fluids to escape these cooling systems and contaminate the environments where these types of data centers would be living. Maintenance of the computing hardware requires lifting the equipment out of these fluid baths to allow as much of the fluid to drip off as possible. Then the equipment is often submerged in a second bath of rinsing chemicals that washes off the remaining coolant and produces a waste product. The riskiest part is that containing gaseous PFAS fluid is difficult and an operator would need to engineer redundant capture systems within the facility to capture leaks during typical operation.

From a cost perspective an operator would be incentivized to keep these fluids contained, but even in the marketing materials of these chemicals they allude to “lowered” fluid maintenance costs assuming they’ll have to be topped off. A simple way to measure how much is leaving containment is by just metering how much they’re adding back to the system.  

There is also a world of immersion cooling that doesn’t use these fluorocarbon-based fluids. A good portion of our electrical grid infrastructure is immersion cooled with mineral oil for example. These don’t leverage the phase-change effects so it often makes less sense than water cooling the chips directly.  

Preston Rhea

Speaking only for public municipal scale compute – once a city has made the decision that it is worth having this resource and maintaining some control and expertise over it, rather than doing it all externally, I think the key is that the resource can start very small and scale without a lot of fuss. It’s unlikely that suddenly a city would need to go from one rack to a whole newly built building for anything it needs for compute. The capital outlay is much smaller than for something like a municipal broadband network. So you don’t have to start with something big, but can start with the smallest possible unit. The worst outcome is just that you have to go back to cloud contracts, not that you are left with a ton of debt.

Rob Lach

Definitely. 

The compute is the closest to a commodity that exists in our computing ecosystem and easily adapted for current and future use cases. For example, a recent trend is measuring computing capacity in Trillion Operations Per Second, or TOPS, purely because AI models are generally agnostic to what kind of computers they are running on and knowing the computing throughput will reveal how a range of AI models will perform on those systems.

The AI models themselves vary in how many TOPS they need to accomplish equivalent tasks so model builders are competing to make their systems more efficient. Taking ownership in a particular model or company would be a bet on a particular approach that could be entirely obsolete by the next innovation. 

Access to compute capacity is currently positioned as the most effective moat the large AI labs have created for themselves. If the public were to take a stake in anything, a stake in the underlying computing infrastructure, be it collectively or building capacity individually amongst ourselves, not only creates a more broadly accessible market for AI builders , but also provides the public some agency towards aligning digital services with our needs (which includes more control over how the data we create is utilized).

Andre Kudra

This is very much needed and worked on by a global community for many years. The individual-hosted/-controlled data space, as promoted by Tim Berners-Lee with Solid, is fantastic but has not really taken off yet. The tides may change now.

For organizational identity, there’s lots going on but focus in many jurisdictions is still on natural person identity. The EU is about to launch a Digital Identity Wallet for citizens as an offer for them to store data artifacts in a decentralized fashion. Going live at the end of the year. The expectation is that the EU will pass a similar law pertaining to organizations, under the label EU Business Wallet. Will take some time, then – from passed law to into force is 24 months.

What’s already there and could do the trick globally: Legal Entity Identifiers (LEI) governed by the Global LEI Foundation (GLEIF). 

Such an LEI could be used to attribute who’s doing, owning and mandating what, also in AI. There’s a cryptographically verifiable variant as well, vLEI.