AI’s cheap seats are filling up: open-weight and small models take over the machine workload

What worries me about the economics of AI is that as models advance, there will be more and more tasks that can be completed by cheap, open source models. For instance, once AI can reliably complete the tax filings of 98% of people, then you won't need a frontier model to do it. That line of thinking is going to work through a whole series of workloads and Scotia has an interesting report today that's somewhat of a case study on a single healthcare company that's aggressively adopting AI.The healthcare technology firm Scotion spoke to ($2 billion in revenue, 5,000 employees) is running a two-tier model strategy, and the bottom tier is growing fastest. Machine-driven workloads, now about 70% of token usage, are split 50/50 between closed frontier models and open-weight models like Qwen and Kimi. The exec says that could go to 100% open-weight and he credits the switch with major cost savings while arguing that performance hasn’t suffered.Some notes:Open-weight models cost about 1/5th of frontier models all-inSmall language models cost about 1/100thSLMs trained on internal data can be more accurate than large models on repetitive tasks like classifying PDFsSLM adoption is just beginningThat’s the tax-return problem in real life. Classifying a document or summarizing internal research was a frontier task two years ago. Now it’s solved, and a solved task goes to the lowest-cost producer. Every advance at the top of the stack eventually pushes another set of workloads to the bottom.The frontier labs aren’t losing business here, at least not yet. Human users still default to Claude, ChatGPT and Gemini, and the exec doesn’t expect closed-model spend to plateau in 2027 because new workflows keep coming.Notably, the company's cloud spend doubled in 2026 overall, which Scotia says is good for compute providers like MSFT, GOOGL and AMZN because they're largely agnostic to which models are being used. More concerning are the frontier labs, where Scotia sees a "mixed outlook" which I take as a diplomatic way of saying "be cautious"."Token spend per developer is flat lining," the report says, at around $100-$500 per month.An Anthropic IPO is expected later this year with valuations now seen in the $1.4-$1.5 trillion range. This article was written by Adam Button at investinglive.com.