Zhipu AI (智谱) has raised roughly $5 billion to bankroll its next GLM models and a self-training R&D pipeline.
The AMW Read
A $5B raise explicitly earmarked for a fully self-training R&D pipeline extends Zhipu's known GLM cadence and sits inside the broader China compute-stockpiling capital wave, giving it segment-plus-structural weight without overturning an existing debate.
Zhipu AI (智谱) has raised roughly $5 billion to bankroll its next GLM models and a self-training R&D pipeline.
On September 13, Zhipu disclosed a $5 billion financing package — about $2 billion from a share placement and roughly $3 billion from zero-coupon convertible bonds, netting close to HK$39.3 billion. Around 60%, roughly HK$23.5 billion, is earmarked for next-generation GLM models, a "fully self-training" R&D system, and large-scale training, inference, and compute infrastructure. First-half revenue reached RMB 954 million, up 399.7% year over year, with API revenue at RMB 825 million (up 2,736%) now 86.5% of total revenue versus 26.3% at the end of 2025 — a shift from project-based delivery to usage-billed API income.
The raise sits inside a broader compute-stockpiling wave among Chinese tech majors: ByteDance's roughly $29.6 billion data-center loan, Alibaba's HK$80 billion placement for global compute, Tencent's roughly $4.7 billion bond, and SoftBank's $40 billion bridge financing for OpenAI, against forecasts of over $886.7 billion in 2026 cloud capex globally. Zhipu has shipped GLM-5 through GLM-5.3 roughly every two months since February, lifting its Artificial Analysis intelligence score from 32 to 60 in eleven months. Its next target, "fully self-training" — models generating and cross-checking their own training data, building task environments, and tuning inference infrastructure, already visible in GLM-5.3's infra agent — extends the "Touch High" R&D-first strategy Zhipu outlined in July, per AMW's earlier coverage.
The convertible bonds' zero-coupon, back-loaded structure signals investors are underwriting a multi-year R&D runway over near-term returns. The open question the source raises: whether self-generated training data degrades model quality at scale — an unresolved risk that will decide whether Zhipu's iteration pace keeps compounding or plateaus within the next year.
