Zhipu, China’s most valuable AI startup, has lost 66% of its market value in two months — roughly HK$800 billion erased since its June peak. The striking part is not the number but the admission behind it: the company’s own chief scientist now publicly calls the trillion-parameter race “a detour.” The signal for the whole industry is that the moat has left the parameter table — it now lives in post-training efficiency, inference cost, and the ability to turn a model into revenue.
What Happened to Zhipu
On August 21, Zhipu shares closed at HK$1,014, down 0.88%, for a market value of about HK$472 billion. That is a 66% collapse from the intraday peak of HK$2,980 on June 22. At its highest, the “world’s first large-model stock” was worth more than HK$1.3 trillion; today it sits below HK$500 billion.
The trigger sequence matters. July brought the first unlock of locked-up shares. Then on July 16, Moonshot AI open-sourced Kimi K3, a 2.8-trillion-parameter model that topped the Code Arena leaderboard — punching a hole in Zhipu’s “technology leader” narrative. The next day Zhipu plunged 28.49%, losing over HK$200 billion in a single session. On July 20 it fell another 19.56%. Two trading days erased more than HK$300 billion.
The Chief Scientist Admits the Detour
On August 19, Zhipu co-founder and chief scientist Tang Jie published a long post re-examining scaling laws. His blunt line: “Looking back, the trillion-parameter round was a detour the whole field walked together — and then turned back together.”
He is not being philosophical; he is describing a documented history. In 2020, Kaplan and colleagues argued that parameters should grow roughly 2.7x faster than data, sending GPT-3, Gopher, and MT-NLG racing toward a trillion parameters. In 2022, DeepMind’s Chinchilla, built on about 400 model runs, reached nearly the opposite conclusion: the compute-optimal allocation is close to 20 tokens per parameter, and a 70B-parameter Chinchilla beat a 280B Gopher on several benchmarks.
Tang’s point is that the biggest models of that era were the most compute-misallocated. Worse, Chinchilla optimized for the old “train once, evaluate once” setting. Today models are called billions of times a day and inference dominates lifecycle cost, so the optimum slides toward “smaller models, trained longer.” He compares scaling to turning multiple knobs, adding: “This does not mean the other knobs have hit their limit.”
Getting Stronger Without Adding Parameters
The theory has a concrete exhibit: GLM-5.3, released early on August 19. It shares the same base and the same 744B total / 40B active parameters as GLM-5.2, with the only variable being about a month of extra long-horizon environment work and reinforcement-learning post-training. Zhipu reports GLM-5.3 improved roughly 50% over GLM-5.2 on its internal Z.ai Code Bench. No new parameters, yet meaningfully more capability. GLM-5.3 also shipped a notable security leap — see our breakdown of how open-source models became the defender.
Why the Market Is Still Wary
JPMorgan raised its target from HK$1,600 to HK$1,800 on August 16, keeping an overweight rating, citing GLM-5.3’s post-training gains and DeepSeek’s API price increases since August 17 as reasons to redraw the industry’s “capability–cost” frontier. It also raised 2026–2030 revenue estimates by 6–10%.
The market has not followed the analysts. Zhipu fell 2.97% to HK$1,012 on August 19 and rose a meager 1.09% the next day. Valuation is no longer tracking a single model release; it is pricing the harder question of whether the company can convert capability into revenue. Meanwhile Zhipu is pushing its A+H dual listing: it filed to list on Shanghai’s STAR Market on June 1, seeking 15 billion yuan (12 billion for its general-purpose base-model project), and completed a HK$31.4 billion placement in July to fund base-model R&D, compute infrastructure, and commercialization.
The Moat Has Left the Parameter Table
Tang’s reflection, whatever its effect on the stock, is a structural signal for the entire industry. When “bigger is always better” loses its authority, three things get repriced: post-training efficiency (what you can squeeze out of an existing base), inference cost (what it costs to serve a model at billions of calls a day), and the revenue engine (who can actually sell the capability). This is the same logic behind DeepSeek’s pricing moves, the widening gap between AI revenue and AI spending at big companies, and the same pattern behind Anthropic’s first profitable quarter.
For model companies, the differentiation is no longer “we have a bigger model.” It is “we can make the same model cheaper and smarter over time, and we can find buyers for it.”
What Teams Should Do
- Re-benchmark “capability per dollar” instead of leaderboard rank — the same base with better post-training can beat a bigger model at lower cost.
- Price inference into every architecture decision; at billions of daily calls, inference cost dominates the lifecycle.
- Watch post-training releases closely — GLM-5.3 proves the gain is now in the training loop, not the parameter sheet.
Frequently Asked Questions
Why did Zhipu stock crash so fast? Shares fell 66% in two months (from HK$2,980 to about HK$1,014), erasing over HK$800 billion in market value, after the first share unlock and Kimi K3’s open-source release broke its technology-leader narrative.
Did GLM-5.3 really improve without more parameters? Yes. GLM-5.3 shares the 744B-total / 40B-active base with GLM-5.2, and Zhipu says about a month of post-training lifted it roughly 50% on its internal coding benchmark.
What does “the trillion-parameter race was a detour” mean? Chief scientist Tang Jie argues the field overshot on parameters after Kaplan’s 2020 scaling rule; Chinchilla (2022) showed compute-optimal training favors more tokens per parameter, and today’s inference-heavy usage pushes toward smaller models trained longer.
