Most corporate AI strategies are stuck in a spot with no real competitive advantage, caught between models that are neither the best available nor the cheapest, according to a Fortune essay by Mark Minevich, a strategic partner at Mayfield. He calls that middle ground the AI death zone, and argues the market has split into two very different races, one over raw capability and one over distribution, that most companies are failing to pick between.
American labs still lead on capability, with frontier systems like GPT 5.5, Claude Fable 5, and Gemini 3.x pushing the ceiling of what these models can do. But distribution increasingly belongs to China. Minevich points to OpenRouter data showing Chinese models held 60 percent of traffic on the platform by July 2026, up sharply from roughly 30 percent a year earlier, while the US share fell from about 70 percent over the same period.
Names like DeepSeek, Xiaomi's MiMo, Alibaba's Qwen, and Moonshot's Kimi now top OpenRouter's rankings, and Qwen has overtaken Meta's Llama as the most downloaded open model family in the world after crossing one billion cumulative downloads. Much of that pull comes down to price. Minevich notes DeepSeek's V4 Flash costs $0.14 per million input tokens compared with $5.00 for GPT 5.5, a gap that exists even though Chinese labs built much of that efficiency after export controls limited their access to advanced GPU clusters.
Not every US player is caught in the squeeze. Minevich points to Anthropic as an example of successful premium positioning, capturing roughly half of all OpenRouter spending despite holding only about 12 percent of token share, evidence that charging more can still work if the capability gap is real and clear to buyers.
His advice for companies stuck in the middle comes down to four moves: build hybrid routing that mixes frontier models with cheaper open alternatives depending on the task, treat inference efficiency as a core part of strategy rather than an afterthought, differentiate through proprietary data and the application layer instead of the base model itself, and stop trying to compete in the middle at all, choosing instead to compete clearly on either capability or cost and openness.
On policy, Minevich argues against simply banning Chinese models, saying Washington's better move is to compete with credible US and allied open weight models released on a regular cadence. As the next generation of global software gets built, he writes, the download numbers already show whose models are winning that race, and it is not the answer America wants to hear.