Open-weight and small models dominate as AI workloads shift to cheaper options

The concern in AI economics is that advancing models will enable a growing number of tasks to be handled by inexpensive open-source versions. For example, once AI can reliably file taxes for 98% of individuals, frontier models won't be necessary for that task. This logic applies across many workloads, and Scotia released a report today focusing on a healthcare company that is rapidly embracing AI.

The healthcare technology company that Scotion interviewed — with $2 billion in revenue and 5,000 employees — employs a two-tier model approach, and the lower tier is expanding most quickly. Machine-run workloads currently account for roughly 70% of token consumption and are evenly divided between closed frontier models and open-weight ones like Qwen and Kimi. According to an executive, that split could eventually reach 100% open-weight; he attributes this shift to significant cost reductions, while maintaining that performance remains unaffected.

A few notable points:

• The all-in cost of open-weight models is roughly one-fifth that of frontier models.
• Small language models come in at about one-hundredth of the cost.
• When trained on proprietary data, small language models can outperform larger ones on repetitive tasks such as PDF classification.
• Adoption of small language models is still in its early stages.

This mirrors the tax-return scenario in practice. Two years ago, classifying documents or summarizing internal research required frontier-level models. Today those tasks are considered solved, and solved tasks naturally migrate to the cheapest provider. Each improvement at the top of the stack ultimately moves another category of workloads downward.

Frontier labs are not yet losing business in this area. Human users continue to rely on Claude, ChatGPT, and Gemini, and the executive does not anticipate closed-model spending leveling off in 2027, as new workflows continue to emerge.

A noteworthy point: the company's overall cloud expenditure doubled in 2026. Scotia views this as positive for compute providers such as MSFT, GOOGL, and AMZN, since these companies are mostly indifferent to which models are deployed. The frontier labs are more worrying; Scotia describes their outlook as "mixed" — a polite term suggesting caution, according to the author.

The report states that "Token spend per developer is flat lining," at approximately $100 to $500 per month.

An initial public offering for Anthropic is anticipated later this year, with valuations currently estimated between $1.4 trillion and $1.5 trillion.