Fantastic observation and insight! I have been thinking about this a LOT. I will add another wrinkle that has not pemeated the zeitgeist that much. We can all agree the cost per token usage going down has resulted in an explosion of usage and as you say here, the joules are joules, they dont care.
Here is the kicker: Historically these companies that offered SaaS pricing assuming the unit cost economics went down over time and were fixed, not its joules. So the most successful your customer, the more costly it is for you to afford them. So, rationinng of token usage which has already started will only continue and take an extreme shape. The logical cosequence of that will be people will use frontier models less and less and the Enterprises will shift to open source models that they manage themselves and can control the economics with upfront capex and not be hostage to surprises in opex cost. Uber and Microsoft are just a small sample what is likely to follow. My two cents.
great wrinkle, the inverted SaaS idea is a good one. I think rationing has already started. Lots of talk about sovereign AI, and the Palantir guy just went off about it on TV. Hah. If enterprises shift to self-managed open source to control economics with upfront capex, the tokens still get generated and the electrons still get consumed. And an enterprise that just sank capital into its own GPU fleet has every incentive to run it flat out regardless of power prices. The sunk capex you describe as the escape from opex surprises is precisely what makes that load even less price-sensitive on the grid side. IMO
They don't need to shift to "self-managed" open source to control the economics. The hyperscalers are the arms dealers in this, have all the models. It can still be opex with open-source models hosted by hyperscalers.
The economics today is 500:1 between frontier and open source models that are close to frontier models. If you get away from frontier, it is 1000:1. Pretty soon it will be 5000:1.. That is enough to give SaaS like economics.
Ultimately for SaaS to work, you need unit economics to work.
Fantastic observation and insight! I have been thinking about this a LOT. I will add another wrinkle that has not pemeated the zeitgeist that much. We can all agree the cost per token usage going down has resulted in an explosion of usage and as you say here, the joules are joules, they dont care.
Here is the kicker: Historically these companies that offered SaaS pricing assuming the unit cost economics went down over time and were fixed, not its joules. So the most successful your customer, the more costly it is for you to afford them. So, rationinng of token usage which has already started will only continue and take an extreme shape. The logical cosequence of that will be people will use frontier models less and less and the Enterprises will shift to open source models that they manage themselves and can control the economics with upfront capex and not be hostage to surprises in opex cost. Uber and Microsoft are just a small sample what is likely to follow. My two cents.
great wrinkle, the inverted SaaS idea is a good one. I think rationing has already started. Lots of talk about sovereign AI, and the Palantir guy just went off about it on TV. Hah. If enterprises shift to self-managed open source to control economics with upfront capex, the tokens still get generated and the electrons still get consumed. And an enterprise that just sank capital into its own GPU fleet has every incentive to run it flat out regardless of power prices. The sunk capex you describe as the escape from opex surprises is precisely what makes that load even less price-sensitive on the grid side. IMO
They don't need to shift to "self-managed" open source to control the economics. The hyperscalers are the arms dealers in this, have all the models. It can still be opex with open-source models hosted by hyperscalers.
The economics today is 500:1 between frontier and open source models that are close to frontier models. If you get away from frontier, it is 1000:1. Pretty soon it will be 5000:1.. That is enough to give SaaS like economics.
Ultimately for SaaS to work, you need unit economics to work.
And the Token Price war will intensify
New model on the block
The most intelligence per dollar.
Amazing results from @pathway_com team on ARC-AGI-1 : $0.0007 at 29.5%
~11x cheaper than GPT 5.6 Luna
Paper:
https://huggingface.co/papers/2608.09888