Business is leaving ChatGPT for cheaper Chinese AI models for savings, Microsoft saves with software

In June 2026, businesses began massively switching from branded language models like those released by OpenAI to Chinese alternatives — and, as noted by the author of the ML digest on Habr, the matter is not about fashion, but about cold calculation on saving API costs. Microsoft, meanwhile, is attempting to retain corporate clients' interest in AI through software rather than through the models themselves alone.

Why Companies Are Moving Away from Familiar Brands

Savings on API calls are only the top layer of a deeper shift across the entire AI stack, writes the digest author. A year or two ago, the choice in favor of OpenAI models or another large Western vendor was almost automatic: the brand guaranteed quality, and price was secondary against the backdrop of experiments and pilot projects. By mid-2026, many companies had gone through the pilot stage and moved to large-scale industrial exploitation of AI — and at such volumes, the difference in token cost between a branded model and a cheaper alternative turns into a noticeable line item in expenses.

Chinese laboratories over the past couple of years have systematically increased the quality of open and semi-open models, while simultaneously offering significantly lower prices per token compared to flagship products from Western laboratories. For companies that count the unit economics of their AI features, such a choice increasingly looks rational rather than risky.

What Microsoft Is Doing to Retain Customers

Significantly, Microsoft operates within this same logic — but from the opposite side. Instead of competing solely on the quality of the base model, the company relies on deep integration of AI into its software: Copilot is built into Windows, Office, and a suite of enterprise tools, which creates switching costs for clients that go beyond the price of a single API request. This is a strategy of retention through ecosystem lock-in, not through the cheapest or most powerful model on the market.

• June 2026 — according to the digest author, the month when the strategy shift became "the main driver of change in the industry"
• The reason for switching to Chinese models — cost savings on API expenses during industrial exploitation, not a change in preferences
• Microsoft's response — bet on software and deep integration of AI into its own products, rather than just on the model itself

What Lies Behind Discussions of an AI Bubble

The term "AI bubble" in recent years has been used to describe the gap between enormous investments in models and infrastructure and the still-unclear profitability of these investments for end business. The digest author proposes looking at the situation not as a question of belief in technology, but as ordinary cost optimization: as soon as a company moves from a pilot AI project to tens or hundreds of millions of model calls per month, savings of a few cents per thousand tokens turns into a noticeable budget line. In such logic, the choice of a model vendor becomes an ordinary procurement decision, not a matter of brand loyalty.

This also explains why the author recommends looking at "the entire AI stack from top to bottom," rather than just at the price of a specific API call: changes are happening simultaneously at the level of models, cloud infrastructure, and application software, and the winner is the one who can optimize all three layers at once, rather than just one of them.

What This Means

Conversations about "AI bubbles" are increasingly shifting from the question "will these models be used at all" to something far more practical — "which vendor has the cheaper option at large volumes." For the market, this means intensified price competition between branded and Chinese models, and for companies like Microsoft, it means an incentive to earn not from the model itself, but from the software around it.