JPMorgan Cuts Target Prices for Zhipu and MiniMax: Post-AI Coding Demand Gap Longer Than Expected
nashnova research
JPMorgan slashed Zhipu's target from HK$2,000 to HK$1,200 and MiniMax's from HK$300 to HK$200, compressing the valuation multiple from 20× to 12× 2030E P/E — the next wave of enterprise AI demand after coding is taking longer than expected, but both ratings hold.
Why is AI demand stuck between coding and white-collar work?
JPMorgan frames the AI-model market as a labor market: enterprises adopt at scale only when a model can reliably finish an entire job category — demand climbs in staircases, not slopes.
The coding staircase is already proven. Cursor's annualized revenue jumped from ~$1 bn to over $4 bn in eight months; legal-tech firm Harvey grew ARR roughly 8× in under two years.
This means → the next staircase is "broader white-collar work," but today's state resembles coding circa 2024 — early signals, no breakout yet. The market sits in the gap between two stairs.
Independent evaluator METR shows the best models can finish a 3-hour task roughly 80% of the time; the length of tasks a model can handle independently doubles roughly every 105 days. JPMorgan calls the next wave "a matter of when," but the timing is later than previously assumed.
Token prices keep falling — can AI firms still make money?
Prices are indeed falling: per the Ramp AI Index, the blended per-million-token price from the largest US providers has halved since August 2025. In China, the 2024 price war cut API list prices by over 90%.
But volume is rising faster: US enterprise token spend is up roughly 15× because consumption surged roughly 36×; China's daily token calls grew from ~100 bn in early 2024 to over 500 trillion by June 2026.
This means → "cheaper" and "profitable" can coexist — what matters is whether you sit at the frontier. Over the past two months, seven leading labs across the US and China released over ten new models; more than half priced below their predecessors, yet OpenAI's GPT-6 Astra and Google's Gemini 4 Argon — the two top-tier launches — both raised prices.
In plain terms = commodity models are in a price war, but the strongest models charge more than ever — pricing power lives at the very top of the pyramid.
Why is Zhipu still JPMorgan's top pick?
Zhipu's ARR grew from ~$35 mn in December 2025 to $1.8 bn by August 2026 — roughly 46× in eight months. Management then raised the year-end guide from $2.4 bn to $3.0 bn.
GLM-5.3's API price doubled versus its predecessor, while a cheaper GLM-5.3-Flash retains price-sensitive users. The open-platform and API business swung from a −0.4% gross margin to 24.6% in one half.
This means → Zhipu is running a dual track — frontier-model premium plus low-cost volume — and the margin flip shows the model is starting to work.
Near-term catalyst: GLM-5.5 / GLM-6, expected in November, with an estimated 2.5–3.5 trillion parameters. JPMorgan will watch whether the new model narrows the gap with US frontier models and whether enterprises begin using it for longer, multi-step agentic workflows.
Why did MiniMax get only a "Neutral"?
MiniMax competed on value this year — M3 and M3.1 Flash offer competitive models at the lowest price tier among leading labs — yet the company did not raise its ARR guidance all year.
In plain terms = cheap and capable, but enterprise customers haven't scaled up orders — "affordable" hasn't converted into "big contracts."
The key near-term watch: M3 Pro, expected in November. JPMorgan will look for it to close the gap on multi-step agentic work versus leading US and Chinese models — and to price at frontier levels rather than continuing the value play. The report states: "If it does, our view could turn more constructive."
What exactly is JPMorgan worried about after such deep target cuts?
The valuation multiple dropped from 20× to 12× 2030E P/E, stripping out most of the premium Zhipu previously carried over China's tier-one internet names. Zhipu's 2026 revenue forecast edged down from RMB 6.1 bn to RMB 5.7 bn, but the 2027 forecast actually rose from RMB 17.8 bn to RMB 23.9 bn.
This means → JPMorgan isn't cutting its revenue conviction — it's cutting how much premium the market will pay for a standalone AI-model company while the next demand wave refuses to arrive. Investor patience is wearing thin.
Key downside risks: fast-followers closing the gap and shrinking the premium window, platform giants subsidizing models at a loss, no new job category "graduating" after coding, persistent compute-supply constraints, and equity dilution from fundraising.
This reflects a clear focal point: whether Zhipu's GLM-5.5/6 and MiniMax's M3 Pro can deliver a frontier-capability leap in November — that is the make-or-break moment for validating whether the current de-rating has gone too far.
市场有风险，内容仅供研究参考，不构成投资建议。