The Unexpected Expected Rises in AI Costs (How Founders Stay Ahead)
Something major shifted in the AI space. More specifically in the economics.
It starts with NVIDIA’s VP stating that AI is more expensive than human employees. Not because human output is gutted by AI but because the computational effort and the employees using it makes it more expensive compared to human labor itself . Entrepreneur Article
Next, Microsoft quietly cancelled Claude Code licenses for thousands of their engineering team. Not necessarily because the AI output failed but because it was too effective and their internal teams used it so much costs spiraled out of control. TheStreet Article
Uber deployed AI coding tools to its team of 5,000 engineers. They proceed to burn through its entire $3.4 Billion AI budget in 4 months. Business Insider Article
Here’s the thing all of these companies are enterprise. They are the most well-resourced technology companies in the world. They got caught off guard.
If that’s happening at enterprise scale, what does it mean for a founder or small business running on a much smaller budget?
It means the “add AI just because” is over.
Why AI economics is tricky for small businesses
AI tooling is priced currently at a flat rate.
This is changing very fast. Just recently, Anthropic announced that starting June 15th 2026 programatic usage of claude (arguably the most valuable use case of claude code) moves to a separate metered credit pool billed at full API rates. What felt like a predictable monthly cost is becoming a variable bill that scales with how much you use it. Anthropic news
Usage-based pricing is even trickier. The better the tool works for you, the more you use it, and the higher your bill grows. There’s no ceiling unless you set one yourself.
For a large enterprise with a dedicated AI budget, this is manageable. (Well perhaps not as much) For a founder bootstrapping growth, an unexpected AI bill is a serious problem.
Patterns I observe among founders
After advising multiple founders on their AI tooling I’ve noticed they fall into one of three categories:
The Overspender: They are subscribed to ChatGPT, Claude, Gemini, and three other AI tools they found on a podcast. Paying for all of them. Actively using maybe one. The rest are running in the background doing nothing while quietly billing every month.
→ Fix: audit your AI subscriptions right now. Pick one primary tool that covers 80% of your use cases. Cancel the rest.
The Free Tier Maximizer: They use the free version of everything. Getting frustrated when they hit limits. Spending time working around restrictions instead of working on their business.
→ Fix: identify the one AI tool that saves you the most time and pay for it. One paid subscription used well beats five free tiers stitched together.
The Set and Forget: They connected AI to their workflows, automated a few things, and stopped thinking about it. Their usage grows quietly in the background. Then the bill arrives.
→ Fix: set a monthly spending cap on every AI tool that charges on usage. Most platforms support this. Most founders never set it.
Running AI locally actually makes sense
You don’t have to use a cloud AI service at all.
Local AI models let you run artificial intelligence directly on your own machine. No cloud. No subscription. No usage bill. The model lives on your computer and processes everything locally.
I’ve personally experimented this using Ollama to run Deepseek locally, specifically to boost my own coding efficiency. The setup takes about 30 minutes. Once it’s running there’s no bill. Ever.
What’s the catch? The tradeoff is local models are less powerful than frontier models like Claude or GPT-5. For tasks like drafting emails, summarizing documents, generating first drafts, or writing code snippets, they perform surprisingly well. For complex reasoning or highly nuanced tasks, frontier models still win.
So when does running locally make sense for a small business?
When you have repetitive, predictable AI tasks that don’t require frontier-level intelligence.
When you’re processing sensitive business data you’d rather not send to a third party server.
When your AI usage is high enough that subscription costs are becoming a real line item.
When you have a technical person on your team who can handle the setup.
Tools worth knowing:
Ollama: the simplest way to run local models on Mac, Windows, or Linux. Download it, pick a model, run it. That’s it.
LM Studio: a desktop app with a cleaner interface for running local models. Good for non-technical users who want to experiment.
Deepseek: one of the strongest open source models available right now. Runs well locally and punches above its weight on coding tasks.
The practical framework
Before your next AI purchase ask yourself four questions:
Do I actually have a repeatable process this AI will plug into or am I hoping AI creates the process for me?
Am I paying for this tool because it saves me time or because everyone else has it?
Have I set a spending cap on every usage-based AI tool I’m running?
Is this task complex enough to need a frontier model or could a local model handle it for free?
If you can’t answer question one clearly, don’t buy the tool yet.
The bottom line
AI is not going away. The costs are not going down anytime soon. The founders who win with AI in the next 12 months won’t be the ones using it most. they’ll be the ones using it most deliberately.
One well-placed automation beats five subscriptions you don’t fully understand.
I hope this was helpful on what to do with your AI needs.
Thanks for reading.
- Andres