Zhipu Raises $5B for Next-Gen GLM and Domestic-Chip Infrastructure as AI Capital Race Intensifies

Zhipu's announcement said net proceeds from the share placement and convertible bonds total about HK$39.3 billion. About 60%, or HK$23.5 billion, will go to the next-generation GLM and a fully self-training system, as well as large-scale training, inference and compute infrastructure. A securities analyst tracking AI told Leiphone.com that the convertible bond-heavy structure with a premium suggests long-term money. A zero coupon means the company does not pay immediate interest, while conversion value is realized later, which suits companies with long R&D cycles.

The financing follows a rapid release cadence. From February to August 2026, Zhipu released GLM-5, GLM-5.1, GLM-5.2 and GLM-5.3, averaging a major upgrade every two months, according to Leiphone.com. Third-party data from Artificial Analysis showed its flagship model's composite intelligence index rose from 32 to 60 over the past 11 months. In the first half of 2026, Zhipu revenue was RMB 954 million, up 399.7% year on year. Open platform and API revenue was RMB 825 million, up 2,736%, accounting for 86.5% of total revenue, up from 26.3% at the end of 2025.

Zhipu is also pushing RSI into infrastructure. On Sept. 17, Tang Jie posted about RSI, saying Zhipu had built a GLM-5.3-driven inference infrastructure. GLM-5.3-Flash, the anonymous model known as Ox-Alpha, handled more than 62 trillion tokens in six days, all on 100,000 domestic chips, according to Leiphone.com. Tang said switching all real production traffic took two weeks, and end-to-end throughput rose 3.2 times. The company used software techniques including ReplaySSM, W8A8 quantization and mixed-precision KV cache compression, and a dense feedback system, according to the report.

On a Sept. 16 investor call, Zhipu executives said demand surged 10-fold after GLM-5 launched in February, exhausting compute reserves and forcing it to suspend sales of its Coding Plan. After $4 billion in financing in July, it restarted sales and saw volume rise more than 15 times. Zhipu estimated that if it invested RMB 30 billion in compute, with 40% for training and 60% for inference, and if the inference capacity were fully utilized at an 80% gross margin, annual revenue could reach RMB 40 billion. Cloud revenue-sharing agreements for GLM models will begin contributing revenue in October. About 100 security companies have connected to GLM, it said.

But AI cloud sales teams at large Chinese internet companies are under pressure. Leiphone.com reported that MaaS revenue targets have reached more than RMB 10 million per salesperson, up from tens of thousands to about RMB 1 million last year. In March, OpenClaw's popularity in China boosted enterprise AI Coding demand. The Coding and multimodal markets were largely captured by Zhipu and Seedance, respectively. After Seedance 2.0 launched on Feb. 12, some film and short-drama clients offered to pay RMB 3 million in advance for whitelist access, according to the report.

Cloud vendors adjusted incentives to let third-party model resale count fully toward sales performance, but price wars intensified. Sales offered discounts on Zhipu GLM and DeepSeek models, sometimes using cloud product vouchers to offset costs. Codex's cheaper reset mechanism also diverted some budgets, according to the report. Internal coordination with discount approval, presales and delivery teams often proved as difficult as external sales, and one regional team saw more than a dozen salespeople resign en masse, which was internally labeled a management incident.

In agriculture, a Fiji farmer, Cody, traveled nearly 10,000 kilometers to Hangzhou for the 2026 Global Agri-Innovator Competition, hosted by the UN Food and Agriculture Organization, Zhejiang University and Pinduoduo. According to QbitAI, Cody, 31, from Kadavu Island, spent more than 2,000 hours learning programming with AI and built Teivaka, a system for smallholder record-keeping, learning, product listing and buyer connections. Its AI assistant uses local soil, crop, season and farming rules. The system is testing with 100 users, half farmers and the rest buyers, suppliers and institutions. Fiji has about 83,000 farmers, 84% smallholders, and imports FJ$1.1 billion in food annually while exporting only FJ$200 million to FJ$300 million, according to the report.

The competition received 191 applications from 49 countries, with 137 eligible and 13 finalists, including five from small island developing states. Another finalist, Berikan Protein from Indonesia, uses enzyme hydrolysis to turn low-value fish into protein ingredients. Over three years it processed 720,000 kg of low-value fish, benefited 6,479 people and generated more than $1 million in revenue, according to QbitAI. It trained 30 fisherwomen in production and sales, then shifted digital marketing to younger community members. China's smallholders account for more than 98% of agricultural operating entities, and Pinduoduo's 'farmland cloud pooling' model aggregates demand to connect them with markets.

In consumer hardware, Hypershell unveiled Halo at IFA Berlin on Sept. 3, a hip-knee integrated exoskeleton with four motors. Its founder edition is priced at RMB 15,999, limited to 200 units, with shipping expected from November 2026. Leiphone.com reported that Hypershell delivered about 30,000 units in 2025 and is raising a new round at a $1 billion valuation. The company claims Halo is an industry-first hip-knee integrated architecture, but Vastnaut, founded by former DJI core members, launched a Kickstarter for its four-joint Vastnaut One in April and raised $1.35 million. Hypershell's early dual-motor product had a yield rate of about 46% and current overall return rate of about 38%, according to the report.

Also in hardware, LUMISTAR completed multiple financing rounds totaling tens of millions of US dollars from Oasis Capital and Monolith, with a new round under discussion. The company, incubated by Thick Snow Capital, focuses on AI sports training hardware. Its tennis ball machine TERO launched in April 2026, began domestic shipping in May, priced at RMB 5,899, and ranked first in its domestic category during the 618 promotion, with about 2,000 units shipped in one cycle and a second-half target of 10,000 units. Its basketball training robot CARRY raised $1.34 million on Kickstarter in July and is expected to ship from November, with a price of $2,399 and weight of 35 kg. The team of about 100 comes from DJI, ByteDance and Xiaohongshu, with more than half in product and R&D.