For the last couple of months I’d been dumping lots of strategy and product research for my company into various sessions of AI. And a couple of days ago I opened up the sessions thinking that I’d left all of that information in them. No. It appears that all of that work had been deleted. I never was asked if I wanted to keep it. I was never even warned that it was going to be deleted.
That was the moment when I realized I wasn’t using my AI but renting someone else’s.
There are so many different AI systems becoming available for access to cheap tokens, to free “tiers” of functionality, to increasing sizes of histories that can be accessed, even to brand new models every month or so. All this AI for so little money. Wonderful you might think ;)
None of this is yours.
Two letters from the AI landlords
Google introduced new pricing for Gemini 4 Argon output and Gemini 4 Argon input a few days ago. As of December 1, Gemini 4 Argon output will cost $20 per million tokens (previously $10 per million tokens) and Gemini 4 Argon input will cost $4 per million tokens (previously $2 per million tokens). Note that these price increases will affect all paid tiers so moving to a higher tier won’t bring any relief.
Liquid AI open-weight model now runs at 184 tokens/s on a laptop chip. Looks very interesting. Here’s license for this model: “Liquid AI grants you a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license.” Note that commercial rights cease to exist when your company reaches $10M in revenue. And its Termination Clause (Section 11): “This License will terminate automatically and immediately if You fail to comply with any of its terms and conditions. Upon termination, You must cease all use of the Work and any Derivative Works and delete all copies in Your possession.”
That’s tenancy. Access without ownership.
The lord’s mill
We’ve seen this before. On medieval manors you had to have your grain ground at the lord’s mill, and pay the lord’s toll. Even worse, you weren’t allowed to grind your grain at home with a hand mill.
The town of St Albans is famous for a spot where the monks took the townspeople’s millstones and set them into the floor of its cloister, where the monks could stroll above them daily. In June 1381, during the Peasants’ Revolt, the townsmen went into the cloister, pulled up the millstones and broke them up into small pieces which they then distributed amongst themselves.
The hand mill is back: Epoch AI puts open models about four months behind the best closed ones. But a model that runs on one consumer graphics card matches what the frontier could do 6 to 12 months earlier.
What owning AI actually means
Owning AI means you can run it, inspect it, modify it and rely on it. This generally refers to four things:
The model weights (i.e. the AI model itself).
The compute it’s running on (e.g. your laptop, your data center, your cloud instance).
The data it was trained on and is working with.
The interface you interact with (i.e. the chat window, and anything on top of it).
A “license” to use AI software is the equivalent of a “deed” to own a house. Can you sum up the major terms of any license to use AI software in one short sentence?
If not, i.e. if there are “revenue caps,” “usage restrictions,” and “terms and conditions that change”, then you are leasing not owning AI software.
Why this matters more than the price
This is not about saving a few dollars for tokens to use AI to get some work done.
I note also that current AI work enables much more than just answering questions (e.g. writing, booking travel, paying bills etc.). Let alone deciding for you. I don’t believe that most people would hand their working lives to rented intelligence just because it’s cheap, even if they were offered it.
No one appears to link big launch announcements on sustainability to the ceaseless race to the biggest AI model of all (IEA expects data centers to more than double electricity use by 2030) that will require the largest datacenter to power it and hence even more concentration. In reality, much of the most useful work will require nothing like the biggest AI model of all and can be done very well with a good-enough AI model that is under the full control of the people doing the work and is located as close to the work as possible.
Everyone’s asking the wrong question
And I’d rather ask the hard questions a tenant would ask a landlord:
How much would you still be able to use if you got evicted by the landlord?
Would you still be able to run it if the landlord refused to serve you?
Think about what will run if the provider says no (more on this soon).
It’s time to stop wondering which model will be great for now and to start thinking about what you can still run tomorrow if the provider of AI refuses to serve you.
Sources
OneHill Foundation, Christoph Welsbach: https://onehill.org/own2026/
Gemini 4 Argon pricing, Google AI for Developers: https://ai.google.dev/gemini-api/docs/pricing
Liquid AI, LFM2.5-24B-A2B release, Hugging Face, 13 Oct 2026: https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b
LFM Open License v1.0: https://liquid.ai/lfm-license
St Albans 1381 and the millstones: https://www.cdamm.org/articles/william-grindecobbe
Epoch AI, open vs closed gap (29 May 2026): https://epoch.ai/data-insights/open-closed-eci-gap
Epoch AI, consumer GPU vs frontier: https://epoch.ai/data-insights/consumer-gpu-model-gap
IEA, Energy and AI: https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works