The AI Brief #43 proprietary AI AI ROI fine-tuning token costs SMB AI infrastructure

Building your own AI capability vs. renting models: the debate that changes everything

Rodrigue Le Gall | | 3 min read

Satya Nadella just reminded a truth that many SMBs forgot while jumping on ChatGPT and Claude: the model is only 20% of the problem. The rest? It’s infrastructure, data, and especially the learning loop around the model.

The context is straightforward. For the past two years, companies have adopted AI in “API rental” mode: you pay per token, you use OpenAI’s or Anthropic’s cloud service, and you hope it works. Except Nadella reminds us what serious CTOs already know: competitive advantage doesn’t come from accessing the same model as everyone else. It comes from what you do with it. How you refine the output. The intellectual property you build around it.

For SMBs, this is the signal that “deploying AI” is no longer enough. Anyone can build a chatbot. But those who win? They’re the ones building an in-house database, proprietary workflows, and feedback loops that make their model more useful every day.

The catch? It costs more upfront. And it’s more complex than saying “we use Claude.”

What this means for your business

For an SMB, this means two possible paths. First: keep renting, but with a clear plan for internal deployment in 18-24 months. Second: accept right now that in-house AI (even fine-tuned, even with a simple structured database) creates value that generic APIs will never deliver. Concretely, this means: document your real use cases, start centralizing your business data, and test fine-tuning on an open-source model (Llama) alongside your API usage. The initial cost increases by 30-40%, but after 18 months, you won’t be dependent on token pricing anymore.


In brief

Uber burned its annual AI budget in a few months

Large enterprises are discovering that “tokenmaxxing” (pushing AI to the limit without counting costs) drains wallets fast. Uber, like others, had to scale back ambitions after exceeding budget forecasts. The actual ROI of AI is now being questioned seriously.

Read source

Snap shuts down its AI video unit (too expensive) and spins it into a startup

Snapchat creates Dotmo, a new company made up of its AI video team. The reason? Costs. This shows that even giants are reassessing whether in-house AI makes sense or if it’s a separate business that needs its own path to profitability.

Read source

Adobe Firefly can now remember your previous creations

Adobe’s AI gains persistent contextual memory. For agencies and creative SMBs, this means less re-explaining and more consistency across generations. A concrete example of the kind of incremental improvement that makes a real difference.

Read source

63% of Americans think AI is advancing too fast

According to Pew Research, while 49% regularly use chatbots, two-thirds believe the pace is too rapid. For SMBs, this reflects a real tension: external pressure to “do AI” often outpaces the actual capacity to integrate these tools in a stable way.

Read source

Get The AI Brief in your inbox

3x per week, the essentials of AI decoded for business leaders.

Subscribe

Take action

Ready to automate your repetitive tasks?

Discover what AI can concretely change in your business. In 2 hours, we identify your automation opportunities.

Free AI Checklist

10 processes to automate in your business

Download PDF