Speaking at a panel organised by lender DBS in Singapore on Monday, October 12, Intel CEO Lip-Bu Tan said the token costs of Chinese frontier AI models have been lower and that their adoption has been "very fast".

He named DeepSeek, Kimi and GLM as Chinese models whose performances are "good enough", and argued that markets, not gatekeepers, should decide which models fit which use cases. The remarks were reported by Reuters and picked up by AsiaOne, The Indian Express and Euronext.

It is striking for the head of a major Western chipmaker to declare rival Chinese models effectively at the frontier, given how much of the AI narrative has centred on US laboratories. Tan's framing runs against the closed-model consensus among US frontier developers.

It also lands amid an ongoing fight over openness: PPC Land reports that a July 24, 2026 letter signed by 77 organisations, including Nvidia, Google, Meta and OpenAI, defended open-weight models, while Anthropic, which did not sign, called for mandatory pre-release testing of all sufficiently capable models.

"I'm a strong believer of open source, that's the best way to drive efficiency," Tan said, according to Reuters via Euronext. His emphasis on efficiency echoes a wider industry debate over whether compute thresholds still capture where capability lives.

PPC Land reports that NVIDIA researchers argued in June 2025 that models under 10 billion parameters can handle 60% to 80% of agentic tasks now given to far larger systems, meaning fixed compute lines may miss capable models while catching bloated ones.

The term Tan invoked is itself contested. According to PPC Land's explainer, a frontier model is one whose capabilities match or exceed the most advanced models available at the time, and no single definition governs the phrase.

The Frontier Model Forum uses a relative capability test, while legislators prefer training compute: the EU AI Act presumes systemic risk above 10^25 FLOP, and California's SB 53 and New York's RAISE Act both draw the line at 10^26 FLOP for developers with over $500 million in annual revenue.

Epoch AI counted 33 publicly disclosed models at or above 10^25 FLOP from 12 developers as of June 2025.

If Chinese open models are genuinely good enough at lower token cost, buyers can route workloads away from expensive closed frontier systems, reshaping where inference revenue and cloud spend flow.

PPC Land notes that most AI features advertisers buy run on frontier models or their derivatives, and that training at frontier scale requires hyperscaler data centres — with UBS estimates suggesting OpenAI and Anthropic could account for 48% of Google Cloud revenue in 2027.

Cheaper open alternatives pressure that concentration, and regulatory status is commercial status for whoever sits near the line.

Tan also said markets will decide which models benefit different use cases the most, and predicted quantum computing could become mainstream in the next three to five years, with software playing a role in using the required infrastructure efficiently.

The broader frontier debate is still moving: New York's RAISE Act takes effect January 1, 2027, and California's governor has ordered recommendations by November 16, 2026 on whether to require kill switches for frontier models and onsite verifiers in large laboratories, per PPC Land.