Startups and researchers are starting to do something I think is very smart.
Instead of depending completely on OpenAI, Anthropic or any other frontier AI company, they are slowly building their own custom AI models on top of open-weight models.
Bloomberg put it very simply: “soaring artificial intelligence costs have nudged it to rethink its dependence on the AI giants.”
I keep thinking about the early internet.
At first, servers and hosting looked like boring infrastructure. Then businesses moved online, and suddenly whoever controlled the infrastructure underneath the internet controlled something extremely valuable.
AI may be going in the same direction.
Imagine building a $20 million Canadian company around one American AI API. Or an American company quietly running important work on a Chinese model. Everything works beautifully until the price changes, the model disappears, API rules change, access gets restricted, or relations between companies or countries become ugly.
Your business may still belong to you.
But its intelligence does not.
That is why I think learning to own even a small part of your AI stack now is a very smart early move.
Building your own AI model does not mean building another ChatGPT.
That would be insane for almost everyone.
The practical version is much simpler.
Take an already intelligent open-weight model.
Give it examples of your work.
Teach it one job.
Then run that model yourself.
Harvey is doing a very advanced version of this in legal AI. Its Tenet model starts from Moonshot AI’s open-weight Kimi K3 and is post-trained for legal work rather than being built from zero. Harvey says one of its goals is eventually letting firms build specialized models and “own their intelligence.”
You can do the same basic thing on a much smaller scale.
Maybe your model learns:
research → newsletter draft
or:
customer email → company-approved response
or:
meeting notes → weekly report
or:
lead information → sales analysis
or:
bug description → code written in your company's style
This is where custom AI becomes useful.
Not when it knows everything.
When it becomes unusually good at your one repeated job.
Opinion AI hit 2K paid subscribers, so yesterday we opened 40% off our annual plan as a small celebration gift.
Only a few hours are left before it expires. Grab it while it’s still live.
Below is the full guide on how to build a custom AI model that learns your work, runs on your terms, and becomes part of your real workflow.
This can save you a lot of time.
Before training anything, try solving the task in three simpler ways.
