Why private models could win

Clayton Christensen described this pattern in 1997. A cheaper, worse product enters from below. The market leader ignores it. The cheaper product improves. At some point it crosses a threshold: good enough for most customers. By the time the leader reacts, the market has moved.

He called it disruptive innovation. He was writing about disk drives and steel mills. He was also, without knowing it, describing what is happening right now to the AI model market.

The numbers are already there

In May 2026, models from DeepSeek and Qwen match or outperform GPT-4o on most published benchmarks. On coding specifically, DeepSeek V4-Pro outperforms GPT-4o on competitive programming benchmarks by a significant margin. The cost difference is roughly nine to one in favor of the open alternative.

On OpenRouter, the largest AI model API aggregation platform, Chinese models went from under 2% of traffic in early 2025 to over 60% by early 2026.

That is not a trend. That is a result.

The leading AI providers are in a difficult position

OpenAI is projected to lose 14 billion dollars in 2026 alone, with profitability not expected before 2030. Anthropic has projected its first profitable quarter but warned that sustained profitability through the rest of the year is uncertain. The price per token has come down, but the total cost of agentic workflows has gone up. Uber burned through its entire 2026 AI budget in four months.

The economics are unstable. When companies under this kind of pressure need to monetize, they find a way. Starting June 15, 2026, Anthropic moves programmatic usage out of the flat subscription and into a separate credit pool billed at full API rates. Claude Code and agent workflows that were included in your $100/month plan are now metered separately. The subscription price did not change. The bill did.

The privacy argument is not getting enough air

Banks, law firms, healthcare providers are sending sensitive data to external APIs without much friction. The justification is usually that enterprise contracts include confidentiality clauses and data does not train the models.

That may be true today. It is a contractual promise from a company that is not yet profitable and is racing to find a sustainable business model.

Running a capable model on your own infrastructure removes that dependency entirely. Two years ago that meant accepting a meaningful performance penalty. Today it does not.

Christensen’s pattern has one more piece

The incumbent always has the same reaction when the cheaper alternative appears. They say it does not compete with them. They are right, for a while. The cheaper product serves a different customer, a simpler use case, the bottom of the market. Then it improves. It does not stop.

The leading AI providers still have real advantages in the hardest reasoning tasks, the longest context windows, the most complex agentic workflows. That edge is real and it matters for a slice of use cases.

But most use cases are not that slice.

The disruption may not be coming. It may already be here. The signals are hard to ignore, and the cost structure is not moving in the leading AI providers’ favor.

Christensen called it in 1997. The playbook has not changed.