← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Sourceupdated Sat Jun 27 2026 08:00:00 GMT+0800 (Philippine Standard Time)

China Is Having Another AI Moment (Economist)

economistchinaaizhipuglmdeepseekbenchmarkshierarchy-of-access

China Is Having Another AI Moment

Section: China Edition: 2026-06-27 Print headline: Roaring back once again? URL: https://www.economist.com/china/2026/06/21/china-is-having-another-ai-moment

The 2026 sequel to the January 2025 DeepSeek moment — but with three structural changes since. Direct extension of Hierarchy of Access (the June 12 Fable 5 export ban is the reason this moment matters) and Frontier AI Ecosystem (the Chinese labs are positioning as the "radical openness" alternative).

The event

June 13, 2026, 5:21pm Beijing time — Beijing-based lab Zhipu (Z.ai) released GLM 5.2. One day after the Trump administration ordered Anthropic to remove Fable 5 from access to non-Americans.

Zhipu tagline: "a step closer to frontier intelligence for everyone."

The timing is the story. GLM 5.2 was positioned as the open, allies-neutral alternative to Fable 5 the same week the US made frontier access politically conditioned.

The DeepSeek precedent (the pattern to match against)

  • January 2025 DeepSeek R1 release → wiped ~$1trn from US capital markets. Nvidia dropped 17% intraday, Nasdaq -3.1% in a day.
  • What shocked markets: not just that Chinese AI was good, but that it was free and open-weight.
  • Then the moment faded — markets rationalised the capability and re-priced the "AI is inevitable" bet.

In 2026, US markets have so far shown little reaction to GLM 5.2. The Economist reads this as a positive signal for pricing power, not for capability: the DeepSeek moment already reset expectations, so a second Chinese frontier-adjacent release doesn't shock.

Three axes of competition (the article's framework)

The Economist frames GLM 5.2 as competing across the three axes American labs have all along:

1. Capability

  • Artificial Analysis: GLM 5.2 ranks as the most intelligent open-source model on the market.
  • On its overall list, GLM 5.2 sits 4th — behind OpenAI ChatGPT 5.5, ahead of Google Gemini.
  • Fable 5 is ~17% cleverer than GLM 5.2 on average benchmark tasks.
  • Elon Musk on X (post-release): expects China to match the frontier by early next year. Zhipu co-founder Tang Jie shot back: "Won't take that long."

2. Cost (the caveat matters)

  • DeepSeek v4: $0.87 per 1M output tokens. Anthropic Fable 5: $50 per 1M output tokens. Fifty-plus times cheaper per token.
  • Ramp reports a sharp June rise in American firms paying for DeepSeek.
  • Microsoft is reportedly considering DeepSeek in Copilot — the enterprise switching signal.

But: Chinese models use many more tokens to reach the same answer.

  • Du Zheng (Georgia Tech) et al., updated this month: DeepSeek used 23× more tokens than an OpenAI rival to achieve basically the same result on the same tasks.
  • Correct comparison metric = total cost of tokens used, not price per token.
  • On a software-engineering benchmark, GLM 5.2 ended up costing more than systems from Anthropic and OpenAI.

This is the buried lede for enterprise-IT: the DeepSeek/Zhipu cost advantage may be a per-token illusion when total-cost accounting is applied. Directly extends the Token Scarcity "Token Reckoning" arc from the 2026-06-20 issue.

3. Reliability (the new axis — and the reason for the timing)

  • Zhipu / Tang: "Our attitude is one of radical openness"; attacked "external blockades" like the Fable 5 ban, calling them mechanisms that make AI "subject to revocation at any moment."
  • 80% of Anthropic's consumer use is overseas (from the 2026-06-20 cover Leader). The pool of non-US customers now looking at Chinese alternatives on political-risk grounds is enormous.
  • Two US congressional committees are investigating American firms using Chinese models — the counter-move that could close this window.
  • Chinese labs face their own reliability limits: compute-shortage-driven service interruptions during traffic spikes.

Why America's lead is bigger than four months

The single most valuable methodological point in the article — worth quoting in any conversation about Chinese-vs-US model gaps.

  • Naive read: GLM 5.2 today ≈ Western model from four months ago (February 2026).
  • Havard Tveit Ihle (Norwegian Defence Research Establishment) analysis, pre-GLM 5.2: Chinese models were 4–6 months behind on public benchmarks, but 8–10 months behind on private benchmarks — nearly double.
  • Mechanism: "Chinese labs appear, possibly unwittingly, to 'teach to the test'" on published benchmark questions.
  • Corroborating US government study (May 2026): similar gap.

GLM 5.2 on private benchmarks:

  • ~7 months behind on Weirdml (unusual ML tasks requiring careful reasoning).
  • ~1 year behind on SimpleBench (common-sense trap questions).

The countervailing datapoint:

  • Artificial Analysis's new office-worker exam (released June 19, on messy files + conflicting information) — GLM 5.2 couldn't have trained for it and it outperformed ChatGPT 5.5 (two months old).

Tveit Ihle's read: lead steady, gap not widening as some had expected.

Why Chinese models are good at some things and bad at others (the structural constraint)

  • Chinese models excel where answers are clear-right-or-wrong — maths, coding.
  • They fall down on open-ended / sustained-judgment tasks.
  • Root cause: export controls on advanced chips → Chinese labs are compute-short. They compensate with heavier post-training — fine-tuning, RLHF on preferred outputs, and (allegedly) distillation from American systems.
  • Post-training is fine for narrow-benchmark performance; it's the wrong tool for open-ended judgment.

This is the compute-moat argument from Hierarchy of Access with the specific mechanism spelled out. The moat isn't just chip count; it's the shape of capability the chip-poor labs can build.

The last-paragraph tell

"That Chinese models are not, for now, facing similar regulatory risk suggests China's government is not yet alarmed enough to act. That may be some of the clearest evidence that they remain behind their rivals."

I.e. the fact that the US export-banned Fable 5 (and not GLM 5.2) is itself evidence of the direction of the lead. Regulatory attention as a capability signal.

Connections

  • Hierarchy of Access — this article is the empirical mirror of the concept: the day after the June-12 ban, a Chinese lab positioned itself as the allies-neutral alternative.
  • Frontier AI Ecosystem — Satya's ecosystem argument; Zhipu's "radical openness" is the anti-lock-in pitch.
  • Zhipu (Z.ai) — new entity page seeded by this article.
  • GLM 5.2 — new tool/model page seeded here.
  • Anthropic — Fable 5 pricing ($50/1M output tokens) confirms + timing of the export ban.
  • DeepSeek — cost anchor + the 23× token efficiency finding (Du Zheng / Georgia Tech).
  • Token Scarcity / Companies Are Scrambling to Curtail Soaring AI Costs (Economist) — the "total cost, not per-token cost" reframing extends the 2026-06-20 Token Reckoning arc.
  • AI Capex Supercycle — compute-moat mechanism corroborates the 5×-capacity buildout thesis.