The Real AI Race Is No Longer About Frontier Models

For weeks this summer, the AI world was fixated on Anthropic’s newest frontier releases and Washington’s debates over who should get access to them. But while the spotlight stayed on those headline‑making models, developers were quietly building something else, and they weren’t waiting for permission from the biggest names in the industry.

Data shows a striking shift is already underway:

• Chinese open‑weight models made up 41% of all downloads on Hugging Face this spring, surpassing U.S. models for the first time.

• On OpenRouter, the top six most popular models are all open releases from Chinese firms, including Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai. Anthropic’s Claude Opus 4.7 follows in seventh place.

• On Vercel, open models handled nearly one‑third of all AI requests in June, taking on most high‑volume workloads while closed models served as the higher‑cost premium layer.

These numbers don’t capture usage hosted directly by major labs, which likely makes up most of OpenAI and Anthropic’s traffic, but they raise a critical question: If most real‑world AI eventually runs on cheaper, customizable alternatives, how much do frontier models actually matter?

Frontier Models: For Specialists, Not Mass Production

Many industry leaders now believe tomorrow’s most advanced models will be reserved for narrow, high‑value work while almost everything else runs on open or in‑house systems.

“Maybe in a few years, frontier models will be for experimenting and very specific high‑value tasks,” said Hugging Face CEO Clem Delangue on the Equity podcast. “Most production workloads will be powered either by private models inside companies or by open source models.”

This shift is driven by more than just cost. After receiving steep bills for scaling closed APIs, businesses are increasingly wary of outsourcing their core capabilities to black boxes they cannot control, inspect, or own.

“If you’re building an AI business, you don’t want to hand your core advantage to someone else to an API you can’t see into and can’t keep if terms change,” Delangue explained.

That momentum is visible across Hugging Face: a new repository is created every seven seconds, with nearly three million public models and one million datasets now hosted on the platform. Half of all Fortune 500 companies use the service to deploy their own private or open models, moving away from the idea of “one model to rule them all” toward a future of many specialized tools.

Chinese Releases Are Redefining the Cost Curve

Fueling this transition is a steady stream of powerful open‑weight models from Chinese AI labs, which often match or approach frontier performance at a fraction of the cost. Most recently, Beijing‑based Z.ai launched GLM‑5.2, an open model built for agentic coding that rivals Anthropic’s latest systems in spotting security vulnerabilities.

These releases are reshaping the economics that U.S. firms have invested billions to build, making it harder to justify paying premium prices for capabilities that are increasingly available openly.

Independence Over Lock‑In

Microsoft CEO Satya Nadella has also warned against relying too heavily on a single provider, arguing that data ownership and control must come first.

“Ironically, we fight for fair use to train models on public data, then turn around and impose strict limits on how customers can use what they build,” Nadella said. “If learning only flows one way, all the value ends up with the infrastructure owners, not the people who create the knowledge. We need to spread that capability so every business can run its own learning loop.”

Safety: Open vs. Closed

The rise of open models has also sharpened a fierce debate over risk. Anthropic’s Dario Amodei has argued that releasing powerful model weights makes them impossible to control, putting them at risk of misuse for disinformation, cyberattacks, or biothreats.

Delangue offers a different view:

“The biggest risk in AI is the concentration of power. We make the world safer by leveling the playing field and being transparent so we can see and fix vulnerabilities before bad actors do.”

He notes that closed models are not immune to theft or misuse either: weights can be leaked, and API guardrails can be bypassed. Restricting access simply concentrates power in a handful of companies, creating dangerous imbalances rather than eliminating risk.