Hi there,
The theme this week: AI is getting government restricted, and will soon be really expensive to use.
The future of work as we know it will be driven by AI, so it’s going to be pretty important to figure out how we keep control over our ability to do work and use its intelligence.
Lots going on, let’s get into it.
What I’m Seeing in AI
Fable is back
Remember that crazy Claude model that got taken down by the US government?
Recap: Anthropic released Fable 5 on 9 June, and three days later a US government decree took it down worldwide.
Yesterday it finally came back for everyone, but only for 5 days, before it moves to be cost-restricted - Anthropic’s announcement.
The insight: Frontier intelligence can be switched off by one government, overnight, for the whole world. OpenAI is now also restricting its newest model to only ~20 pre-approved organisations after the government stepped in there too (Axios).
The alternative to these models is ‘Open Source’ AI models.
Open Source Models in 30s
Claude and ChatGPT are closed models: you pay every time you use them, either through a subscription, or an API.
As users, we aren’t allowed to see their inner workings, all we see is the text we send in, and the answer we get back.
An open-source model is the opposite: the weights (how the model decides what words to answer with) are published on the internet, and anyone can download them and run them on their own computers (usually servers).
If you know how to set this up yourself, all you have to pay for is the electricity to keep the servers running and you can use the AI for free.
The problem is, Open Source obviously doesn’t have the same money-making capacity behind it that models like ChatGPT do, because they are free. Hence they aren’t at the same ability-level as closed models, but they’re catching up (mainly through Chinese models like DeepSeek). Eventually, closed models will have to increase their prices, and lots of people will start using open source ones (many people already are).
Open Source models consistently lag a few months behind closed ones, but soon they’ll be good enough to use day-to-day
Everyone wants to own their AI
I’m not the only one thinking about this.
Palantir’s CEO Alex Karp on CNBC this week said that the way labs sell tokens has “gone completely wrong”, and that enterprises are “livid” about their IP flowing into models they don’t own (CNBC).
Palantir and Nvidia now sell “sovereign AI” built on open source models: you train on your own data, and you own the resulting weights.
DeepSeek models (Chinese open source) are 20-50x cheaper per token than the frontier labs (pricing comparison), and MiniMax (another open source model) gives frontier-level coding at 5-10% of the cost (VentureBeat).
What I’m Building
What AI actually costs to run
The cost of running AI came up in three separate client conversations, each with a different arrangement, so we’re converging on a standard.
One of our systems was heading toward $900 this month in API credits. We changed the code to make it more cost-efficient, but it led us to change our setup.
Clients bring their own API key seems to be the way forward. This way, their AI runs on their existing infrastructure (better cost and security), and aligns with the general move to open source models - they can use one if they have it set up.
A near-equivalent open model would cut that bill to $45, but running your own comes with its own costs: hardware, upkeep, know-how.
Coinbase just showed what this looks like at scale: their CEO says they’ve cut AI spend nearly in half while their token usage keeps growing, largely by making open-source models the default for their engineers (the story).
Coinbase Spend vs Tokens
AI Business Package
I had a random thought this week.
We’re landing on an ideal customer segment and a repeatable system to deliver (i.e. a real business 😆), but that also means we have to say no to a lot of things clients want.
A lot of these are super custom, specific AI platforms that would make a great business in themselves. So I’m thinking about how we can pair these opportunities (and the first paying customer) with people early in their career, so they essentially would run their own startup.
It really aligns with what I want to do: drive myself and others constantly to a more interesting, self-driven path.
Any thoughts? Know someone who wants to run a business?
That’s all she wrote.
See you next Friday,
Finlay