I can't figure out how there's both too little supply (so a dramatic need for more data centers) but also too little demand (so labs subsidize inference).
There isn't too little demand. There is massive demand and many competing companies trying to capture that demand, so they are attempting to make better offers than their competition. Hence subsidy.
- Every competitor is planning for the demand to be much higher in a few years than it is now, and aiming to capture as much of that as they can, which starts by getting companies hooked on their models now
- The data center capacity will get used no matter who captures the most demand
I can somewhat understand companies getting users depentant on their harnesses or workflow, but model vendors as in this deepseek case, I have absolutely 0 model loyalty when it's a simple config change away, and will always optimize for either capability or price (or whatever !/$ metric you can determine).
Depends what you’re doing. For example, Gemini is somehow still your only option if you need a model that can natively understand video and reference timestamps in its response.