Zhipu is down 79% from its peak as the model price war bites

Share
Zhipu is down 79% from its peak as the model price war bites

The token price war finally has a scoreboard, and it is denominated in shareholder value rather than tokens. China's most-watched listed model lab is worth about a quarter of what it was in June.

Zhipu's Hong Kong-listed shares (02513.HK) closed Friday at HK$633.50, about 79% below the HK$2,980 peak they reached on June 22, putting the company's market value back near HK$309 billion. Zhipu listed in January at HK$116.20 a share and ran roughly 25 times that into June; it has now given almost all of it back. The slide began in earnest after DeepSeek cut API prices on September 10 and continued through a week in which Xiaomi, Anthropic and OpenAI all cut theirs. The sharpest detail is in the placement: the shares Zhipu sold on September 13 at HK$714 a piece — the equity half of a raise worth about US$5 billion with its zero-coupon convertible bond — now trade roughly 11% below what those buyers paid. We covered that financing when it landed — Zhipu raises $5 billion in its second financing in two months.

The mechanism is not mysterious. Zhipu built its pitch on being cheap and good at the same time, and cheap stopped being a differentiator once everyone was cheap: DeepSeek's V4.1 Flash prices uncached input at 2 yuan and output at 8 yuan per million tokens in peak hours, halved off-peak; Xiaomi's MiMo-V2.6 landed September 22 with Pro at 3 and 6 yuan and its Flash at 1 and 2 yuan; Anthropic's Opus 5.5 came in at $4 and $20 per million tokens the same day, and OpenAI's GPT-6 Luna at $0.10 and $0.50. Zhipu's own GLM-5.3-Flash sits at 0.8 and 2.8 yuan, with the flagship GLM-5.3 at 8 and 28. Goldman's read, cited in the Chinese coverage, is that frontier and multimodal models still hold pricing power while the low-end API market is already in a price war — with per-million-token prices headed to $0.10–$0.20 in the second half of 2026, and some well-funded Chinese labs subsidising at zero or negative gross margin.

The volume side of Zhipu's story is still real, and that is what makes the re-rating interesting rather than a rout. First-half revenue was 954 million yuan, up 399.7%, with the open-platform and API business at 825 million yuan, up more than 27 times, and MaaS token volume up over 40x from the start of the year — the numbers we went through in Zhipu's API revenue jumps 27x as China's AI labs prove the token economy. What changed underneath is margin and cost. Gross margin fell from 50% to 26.4% as cloud deployment scaled, R&D hit 2.13 billion yuan and the net loss came in at 2.07 billion yuan. Compute has been the binding constraint all year: demand for GLM-5 ran about ten times ahead of forecast after February's release, the company exhausted its reserved capacity and suspended sales of its Coding Plan until the July raise funded more silicon.

Put those together and the trade is legible. Investors paid for a lab that could raise prices while volumes rose; they are now pricing a lab that has to cut them while volumes rise, and funding the cut with equity raised at steadily lower prices — more than HK$70 billion across the July and September rounds alone. The uncomfortable version of the question, the one Goldman's zero-margin scenario implies, is whether a price war fought with shareholder capital can end anywhere except lower.

What to watch: whether the next set of numbers shows the low-end price floor Goldman describes arriving, and whether the shares recover above HK$714 before Zhipu needs to raise again.

If the labs are cutting prices with money they raised from public markets, whose money is really subsidising your API bill? Tell us in the comments.

Read more

Mistral unveils Le Chonk: a 1T-parameter open-weights model

Mistral unveils Le Chonk: a 1T-parameter open-weights model

The biggest open-weight release outside China lands in public preview today, and the country that spent the week promising its own frontier model just put a price on the ambition. Mistral has opened a public preview of Mistral Large 4 — codenamed "le Chonk" — a 1 trillion-parameter mixture-of-experts model with 49 billion active parameters, natively multimodal, which the company calls its largest and most capable model to date. The preview API is live today on Mistral Studio at $1.36 per milli

Google signs 3.6 GW power deal, a quarter of it new nuclear

Google signs 3.6 GW power deal, a quarter of it new nuclear

Grid capacity, not chips, is becoming the binding constraint on the AI buildout — and on the same day, the labs told an Australian inquiry they can live with mandatory incident reporting. Google has contracted 3.6 GW of power from Constellation Energy across the PJM grid — the largest electricity deal in the region's history, and the biggest single power commitment any AI company has made. The agreement covers 3,590 megawatts over 13 states, with 890 megawatts of new nuclear capacity coming fr

Korea probes AI agents in bank hacks as president cites 'signs'

Korea probes AI agents in bank hacks as president cites 'signs'

South Korea opened a formal investigation into whether AI agents drove a wave of bank breaches — and it isn't the only AI story moving money today. President Lee Jae Myung said "signs" point at AI models, and the probe is now at the highest level a national banking sector has seen. Speaking at a cabinet meeting, Lee said that "in some hacking incidents, signs have emerged of AI being used, causing considerable public concern and anxiety," and police have since opened a full-scale investigation

Open Source Radar — October 6: nothing leaves your machine

Open Source Radar — October 6: nothing leaves your machine

Today's trending board is all projects we ran earlier this week, so the fresh signal comes from the Product Hunt launch slate instead — three open-source tools that share one instinct: your phone, your pixels and your MCP traffic should stay on hardware you control. All three verified at the source. iphone-use (Rust, MIT, about 59 stars) is computer-use for a real iPhone: an agent reads the screen as text, taps, swipes and types over WebDriverAgent, and every action comes back with an honest v