Search on TFTC
The Convergence of AI Data Centers and Bitcoin Mining

The Convergence of AI Data Centers and Bitcoin Mining

Sep 19, 2023
Marty's Ƀent

The Convergence of AI Data Centers and Bitcoin Mining

I've been falling down the AI rabbit hole pretty hard over the last six months and my journey has led me to the base layer of the industry; the data centers powering the GPUs that train models and allow people to run inferences on those models. The more I pull on this particular thread, the more convinced I become that the future of AI data centers will be alongside bitcoin mining operations.

At the current moment, the AI industry is experiencing a problem that the bitcoin mining world has run into in the past; insane demand for computers with a supply that is lagging materially. This has pushed up the price of GPUs to astronomical levels and pushed companies to begin hoarding machines to ensure that their competitors don't get an edge. This problem is being exacerbated by another issue that bitcoin miners are familiar with - it is hard to find sufficient rack space to plug the GPUs in. The data center infrastructure simply isn't where it needs to be at the moment to service the demand from the AI industry.

I think it's fair to assume that there is a lot of hype within AI right now, but it is also safe to say that this is an industry that is likely to be around into perpetuity and the hype will fade as valuable products come to market that provide real utility for end users. As that happens, the market for GPUs will find an equilibrium. Chip producers will push out more GPUs, prices will correct, and AI compute will become more commoditized over time. When the market reaches this point, AI companies won't be competing on access to machines, but will instead be competing for cheap energy. This is where the convergence of AI data centers and bitcoin mining operations becomes clearer.

Once we reach the point where AI companies are competing on electricity price instead of access to machines the fight for cheap power contracts will become vicious and AI companies will need to have a bitcoin mining strategy to compete. Especially if they plan on training models, which demand very high uptime. To get access to the cheapest power contracts, data centers will have to be able to participate in demand response programs that allow them to lock in lower pricing by agreeing to divert a portion of their power back to the grid during peak demand periods. Since it is not feasible for AI companies that are training models or are trying to compete on UX by having quick response times on inference calls to turn off their machines to supply the grid with electricity when it is needed most, AI data centers will need to have bitcoin mining operation co-located with their GPUs that can check the demand response box.

We are still in the early days of commercially viable AI products finding their fit in the market, and even earlier when it comes to the infrastructure side of the industry finding an equilibrium, but I think it is safe to say that the future of the infrastructure layer will be dependent on bitcoin mining if AI companies want to produce good margins. Bitcoin miners are uniquely suited to provide the AI industry with the demand response capabilities that will be demanded of these data centers if they want to lock in low power prices.

In the future and at scale, it wouldn't surprise me if the industry matures to the point that these GPU + ASIC data centers that are plugged into demand response programs develop algorithms that allow them to achieve the most profitable outcomes possible. The algos will determine what the optimal mix of AI compute/bitcoin mining/selling power to the grid is at any given point in time and calibrate the data center appropraitely. It's going to get pretty heady.

At the very least, the AI industry is joining us on the front lines against the Malthusian-energy-hating-climate-hysterics whether they realize it or not.

Final thought...

There are too many cooking competition shows.


Current Block Height

Current Mempool Size

Current Difficulty