Musk sees Chinese AI catching Anthropic's Fable by Q1 2027

The gap between Chinese and US frontier AI models has narrowed to months, not years, as open-weight systems from Zhipu and DeepSeek post benchmark scores within striking distance of the best closed models.

Elon Musk predicted Chinese large language models could reach parity with Anthropic's Fable by the first quarter of 2027, responding to a social media post about Zhipu AI's GLM-5.2 narrowing the gap. Google DeepMind Chief Executive Demis Hassabis has also said Chinese AI models may be "only months behind" their overseas counterparts, according to previous remarks cited by Chinese state media.

"The rate of improvement is what stands out," said Rachel Kim, an analyst at Edgen who tracks AI infrastructure. "Chinese labs are compressing what used to take years into quarters, and they're doing it on domestic silicon."

Zhipu AI released GLM-5.2 on June 16 under an MIT license, making it freely available for commercial use. The model scores 81.0 on Terminal-Bench 2.1, up from 62.0 for GLM-5.1 — a 31 percent jump in a single point release. On SWE-bench Pro, it scores 62.1, edging past GPT-5.5, and trails Anthropic's Opus 4.8 by a single point on FrontierSWE. The model carries a 1-million-token context window and costs roughly one-sixth of what leading US closed models charge per token.

DeepSeek's V4-Pro, a 1.6-trillion-parameter mixture-of-experts model that activates 49 billion parameters per token, posts 80.6 percent on SWE-bench Verified. At about 87 cents per million output tokens, it costs roughly one-thirtieth of frontier pricing. The weights are open. Alibaba's Qwen family crossed 1 billion downloads on Hugging Face in January, surpassing Meta's Llama as the most-downloaded open model family globally.

Three Releases, Four Months

The cadence of Chinese model releases shows the pace. GLM-5 arrived in February. GLM-5.1 followed in March, lifting its internal coding score from 35.4 to 45.3 — a 28 percent improvement. GLM-5.2 arrived in June, nearly doubling the Terminal-Bench result again. Each step was trained on Chinese silicon, with some evidence suggesting the entire stack is now Nvidia-free.

In 2023, open models trailed the closed frontier by two years. In 2024, that gap shrank to one year. In 2025, six months. Today, on the benchmarks that matter for engineering work, the gap is measured in weeks.

Where the Value Goes Next

As model weights approach commodity pricing, the economics shift to inference and infrastructure. Inference now consumes roughly two-thirds of all AI compute, up from one-third in 2023, according to industry estimates. Nebius Group reports one customer cut inference costs by 26 times using open models on its platform. Cloudflare now serves more than 70 models from its edge network.

Microsoft Chief Executive Satya Nadella framed the shift in a June 14 essay, arguing that companies must build both "human capital" and "token capital" — the AI capability they own rather than rent. His warning to staff: avoid routing every task through an expensive frontier model when a cheaper specialized one would do.

For investors, the narrowing gap raises questions about the $176 billion in potential understated depreciation across the data center industry that Michael Burry has flagged. If frontier-grade models run on a $4,700 DGX Spark desktop — Nvidia's Grace Blackwell machine with 128 gigabytes of unified memory — the centralized inference demand curve that underwrites five-year depreciation schedules may grow slower than the spreadsheets assume. Roughly half of US data centers planned for 2026 already face delay or cancellation, and prediction markets put the odds of a federal moratorium on large data center incentives before 2027 at about one in three.

This article is for informational purposes only and does not constitute investment advice.