The China AI Manifesto Scaring the West — And Its Own Investors
Zhipu AI’s open letter is neither a geopolitical threat nor a compelling defence of the LLM business.
Grüezi!
1 One Letter, Two Readings
In the frenzied geopolitical contest over artificial intelligence, messages get easily distorted – and this weekend provided a spectacular example. The man behind GLM-5.2, China’s latest open source LLM sensation, wrote something that quickly got translated and shared in English.
Tang Jie (唐杰), the man behind Zhipu AI (智谱) – GLM’s creator – published an open letter headlined: “The Great Wave Has Arrived” (巨浪已来).
In the English-speaking world, readers seized on the threat from a Chinese frontier lab announcing autonomous cyber-hunting capabilities, self-replicating agents, and open-source distribution with no user restrictions – all wrapped up in Anthropic-style safety language.
In China, it landed rather differently. Chinese commentators noted a company with negative net assets, a stock that had just fallen more than 19% on intraday trading, and a founder trying to stave off critics by penning an inspirational letter the same evening that his first post-IPO lock-up expired.
What each ecosystem read into the text reveals more about its own prior anxieties than about the document itself.
2 Western AI nerves
In English-language coverage, the letter arrived at a moment of self-doubt for the bullish AI community. The US Commerce Department had weeks earlier ordered Anthropic to cut foreign access to its two most capable models, Fable-5 and Mythos-5, citing national security.
On the same day that order took effect, Zhipu had released GLM-5.2 under an MIT licence with no regional restrictions. The letter framed this as merely the opening salvo of a “Touch High (摸高)” strategy – Zhipu would open source a model with world-class cyber security expertise whilst the United States was denying its allies models that were capable of exploiting software vulnerabilities.
Tang’s letter even borrowed from Anthropic’s vocabulary: superalignment (keeping AI systems in line with human values once they exceed human capability); and mechanistic interpretability (机械可解释性) – establishing that AI models are safe by unpacking their internal computational processes rather than inferring it from how they behaved once deployed.
The letter also included a veiled reference to overseas labs delaying model releases as their leaders warned that AI would reshape the global balance of power.
Western commentary read Tang’s post as a competitive positioning move – China’s “Anthropic” claiming the safety mantle for the rest of the world whilst upending the expensive licensing model.
There’s some truth to that, but that’s assuming the letter is chiefly a maverick entrepreneur’s geopolitical warning shot, written on behalf of a precocious AI superpower. It isn’t.
And what that interpretation misses is the domestic pressure the letter was actually answering – it was a defensive play, and that was exactly how it went down at home.
3 China’s “Price-to-dream ratio”
The Chinese tech press – LatePost 晚点, 36氪, 钛媒体, and comment threads on Snowball 雪球, China’s main investor platform – saw the letter very differently.
Zhipu had listed on the Hong Kong Stock Exchange on 8 January 2026 at HK$116.2. It was oversubscribed more than 1,159 times at retail, and by late June had run to an intraday high of around HK$2,980 – over 24 times the issue price, a market cap that hovered briefly above HK$1.3 trillion.
The company’s 2025 results showed revenue of RMB 724 million, a net loss of RMB 4.7 billion, and negative shareholder equity. Wry Chinese observers called it the “price-to-dream ratio” 市梦率.
Zhipu’s share structure made the position even more fragile. Only around 2.5% of stock was freely tradable. When the first tranche of cornerstone investors became eligible to sell in early July, the stock initially held up – nearly 70% of cornerstones pledged to stay – then it fell sharply.
Meanwhile, Zhipu was launching a placement targeting up to HK$31 billion in fresh capital, the largest Hong Kong refinancing of 2026 and more than six times the original IPO.
Tang’s letter appeared that same evening, 11 July 2026.
The domestic commentariat’s recurring reference point was the SenseTime (商汤科技) crash of early 2022, when a high-valuation, low-float Chinese AI listing fell nearly 47% on its lock-up day.
On Xueqiu, retail commentary was brutal: no LLM company has a real moat, and Tang’s letter read as a desperate attempt to manage the narrative of a stock whose multiples were getting squeezed.
Zhipu’s third-party compute costs had grown from RMB 15 million in 2022 to more than RMB 1.5 billion in 2024, consuming more than one yuan for every yuan of revenue. Chinese tech writers called it a “compute tax” 算力税 with no obvious path for Zhipu to get software-style economies of scale as context windows continued to expand.
But neither Western geopolitical nail biting, nor Chinese bubble anxiety addresses Tang’s actual arguments.
4 A bet on the ceiling
The “Touch High” strategy, stripped of its founder-fluff rhetoric, is a thesis about competitive dynamics. The dominant strategy in the global LLM market is racing toward monetisation – build applications, expand enterprise sales, compound ARR.
Zhipu is betting that this is the wrong way to compete, and that whoever pushes the capability ceiling will capture everything downstream when the ceiling shifts.
The letter pushes four priorities over the next two years: long-horizon task capability (planning and executing over months rather than milliseconds); autonomous agent systems (self-organising networks of specialist agents); fully self-training models (synthetic data factories, self-play from scratch, models that can rewrite their own code in sandboxes); and “extreme safety governance” centred on mechanistic interpretability.
The quid pro quo for investors? A two-year “revenue-deprioritisation” pledge. It is a very bold move from the founder of a just-listed company to tell the market he’s just raised billions from that he intends to spend the next 24 months not focused on making money.
The logic is a little clearer if you look at the domestic competitive context. Up against DeepSeek, Moonshot/Kimi, Alibaba’s Qwen, and MiniMax, Zhipu’s only real distinguishing claim is its “pure-play frontier lab” identity – academically rooted, Tang’s own Tsinghua pedigree, and a DeepMind-like penchant for putting research primacy ahead of profits.
Investors might take comfort from the fact that Tang published a similar letter on IPO day back in January, committing to return to base-model research over application revenue.
GLM-5.2 dropped five months later.
5 Open vs closed: the big argument
The bigger question behind the letter is about where the AI industry’s centre of gravity is moving, and whether the US export-control strategy is producing the outcome it intends.
The American gambit, as embodied in the Fable-5/Mythos-5 order, is closed-and-restricted: keep the most capable models behind commercial and political gates, extract rent from access, use the capability gap as a coercive geopolitical instrument. One model to bind them.
The Chinese bet, as embodied in GLM-5.2’s MIT release and the sovereign-AI deployments that preceded it, is open-and-ubiquitous: distribute freely, compete on price and adoption, build moats at the infrastructure and sovereign-integration layer rather than the model layer.
The clearest example of what this looks like in practice is Malaysia choosing GLM as its national model, with the Z·UM AI Lab – a joint national AI institute established by Zhipu and the University of Malaya – built on Z.ai open weights and optimised for Malay and Chinese.
Deployments in parts of the Gulf and East Africa follow the same logic, typically pairing GLM with Huawei compute for states that cannot access Nvidia. For these buyers, the choice is not between GLM-5.2 and Claude’s cutting edge models – it is between GLM-5.2 and nothing, since the US has now made that decision for them.
GLM-5.2 trails the closed frontier by around seven percentage points on the hardest coding benchmarks while costing roughly one-fifth as much. When Elon Musk estimated on X that Chinese models were around six months from parity with Anthropic’s Fable, Tang replied simply that it wouldn’t take that long.
For most of the world’s AI buyers, the bleeding edge of frontier capability is less commercially valuable than access, price, and the legal right to self-host.
If that holds, the US strategy may succeed in protecting its technology frontier at the cost of accelerating adoption of Chinese alternatives across the Global South.
6 The safety claim and its limits
The safety pillar in the letter is the most Anthropic-shaped element and the least substantiated.
Zhipu and its affiliated THUDM research group – Tsinghua University’s NLP lab – do have real peer-reviewed behavioural safety and alignment work to their name: AlignBench (a Chinese-language alignment benchmark), SafetyBench (a content-safety evaluation framework), and BPO (Black-box Prompt Optimisation).
What Zhipu does not have is any published mechanistic-interpretability research, any named interpretability team, or any structured safety framework committing to specific evaluations before each capability increase.
The Future of Life Institute’s Summer 2026 AI Safety Index graded Z.ai at D-, noting that the company had “indicated it was developing an existential-risk plan” – which isn’t the same as actually having one.
D- might sound bad, but no company in the FLI index got over C- on existential safety, and the Index concluded that even mainstream interpretability approaches are questioned by researchers because “detection is not prevention.” The D- is nothing to brag about, but failure on this metric is industry-wide.
What the letter’s safety language is actually aiming for is to allow Zhipu to compete for the safety-conscious enterprise market and the governance-conscious sovereign-AI market simultaneously. That’s why it self-consciously uses terms that Western policymakers recognise, whilst not being bound by the commitments the terminology implies in a Western lab.
7 What to watch for
What’s needed for the interpretability pledge to become meaningful is a real Zhipu interpretability programme, a published technical output, or a concrete existential-risk framework. If you don’t see those showing up by the end of 2026, then Tang’s “tens of billions of RMB” commitment is just posturing.
The sovereign-AI deployment count also matters, but what to watch is the infrastructure pairing: whether the next national deployments – Saudi Arabia, Indonesia, and Kenya are the most frequently-cited candidates – replicate Malaysia’s GLM model on Huawei compute, which would give Tang a genuine moat rather than a one-off government contract.
GLM-5.5, the model Tang positioned as Fable-level by year-end, and its independent benchmark performance will be the clearest signal on his capability bet.
GLM-5.2 reportedly consumes several times the compute of peers at comparable task-completion rates – a real adoption constraint for China’s own developer communities. The letter doesn’t address this.
The larger January 2027 lock-up expiry, when a substantially bigger tranche of pre-IPO shares becomes tradable, is the moment of truth for Zhipu’s share-price narrative. Sell-side consensus for 2026 sits around RMB 7 billion, implying a price-to-sales multiple above 100x, at which point, as domestic analysts note, there is zero tolerance for deceleration.
This is the AI story in a nutshell.
American AI entrepreneurs fear being undercut by cheap Chinese models.
So too do China’s AI entrepreneurs…
Thanks for reading!
Bis bald,
Adrian