Tang Jie tells staff Zhipu will keep spending on AGI
Zhipu AI founder Tang Jie told employees in an internal letter on July 11 that the company will not shift its resources toward short-term monetization over the next two years, even after going public. Instead, it will continue to spend on basic research areas with high costs and no clear near-term payoff.
The letter, titled “The Giant Wave Has Arrived,” did not directly address the company’s share price, lock-up expiry pressure, or valuation debate. It focused on Zhipu’s longer-term direction and framed the IPO as a new starting point rather than an endpoint.
Tang summarized the company’s method as “essence, contrarian thinking, and focus.” He pointed to three milestones mentioned in the report: a 2006 academic search effort tied to long-term research on how disciplines evolve, the team’s 2021-2022 decision to pursue a 100-billion-parameter model that became GLM-130B, and Jan. 8, 2026, when Zhipu listed in Hong Kong and, in his telling, returned to foundation model research instead of treating the bell-ringing as a finish line.
How Tang described the path to AGI
Tang wrote that the largest lesson of the past two decades is that major business opportunities do not come from minor tweaks to products or business models. They emerge when the ceiling of machine intelligence is raised.
He described AI progress as a path from perceptual intelligence to cognitive intelligence and then to artificial general intelligence, or AGI. In the letter, AGI was defined in strict terms: not the intelligence of a single genius, but the combined wisdom of humanity, capable of producing original knowledge on the level of relativity.
He named three barriers on that path. The first is long-horizon task capability, where AI moves from instant answers to planning and carrying out projects that may last weeks, months, or years. The second is a fully autonomous agent system, moving beyond the idea of a “one-person company” to what he described as an “unmanned company” made up of collaborating specialist agents. The third is self-evolution, with AI writing code, generating and cleaning data, and training itself.
The report also said Tang referred to a Google DeepMind paper arguing that even if a single model never exceeds human-level ability, the number of AGI instances could theoretically reach 100 million within five years if it grows tenfold each year. Shared base intelligence, high-efficiency collaboration, and near-zero-cost knowledge replication could amount to artificial superintelligence at the group level, according to that framing.
The four engines in Zhipu’s “reach higher” plan
Tang said Zhipu’s strategy for the next two years centers on what he called a “reach higher” plan, aimed at the next frontier of AGI rather than short-term application revenue. The letter listed four engines.
- Long-horizon tasks: building a new memory architecture so models can learn, act, and accumulate knowledge across the full life cycle of a project. The example in the report was breaking the goal of designing an anti-cancer drug molecule into thousands of executable subtasks.
- Autonomous agent systems: moving from AI assistants to “digital employees,” with hundreds or thousands of agents that can debate, collaborate, review one another’s code, and allocate resources on their own.
- Fully self-training AI: building synthetic data factories as high-quality human data becomes scarce, and allowing systems in secure sandboxes to reconstruct their own code.
- Strict safety governance: encoding human ethics, social norms, and national laws and regulations into the model’s value function instead of relying on patchwork safeguards.
On safety, Tang said Zhipu plans to commit resources at the scale of tens of billions to mechanistic interpretability research, with the goal of understanding the neural logic behind each model decision and making black-box systems more transparent. He also said work on superintelligence and “super alignment” must move forward together.
Open source sits alongside the AGI push
The letter treated open source as the other side of the same strategy. Tang argued that real safety is not created through technical closure and barriers, but through broad participation, sharing, co-development, and public oversight.
As an example, the report pointed to GLM-5.2, which it described as Zhipu’s most capable open-source model so far. It supports a million-token context window and is released under the MIT license, allowing download, deployment, and commercial use without restrictions based on user type or organizational nature.
Tang’s message was that frontier intelligence should not belong only to a small group, nor should access be subject to recall by a small set of rule-makers. In the letter’s framing, one hand reaches upward toward the limits of intelligence, while the other lays a road for wider use.
Letter came after a sharp stock swing
The timing mattered. According to the report, Zhipu listed on the Hong Kong Stock Exchange on Jan. 8 this year at HK$116.2 per share, becoming the “first large-model stock.” As its flagship models kept iterating, the share price at one point rose to HK$2,980, up more than 24 times from the IPO price, and its market value briefly exceeded HK$1.3 trillion, above Xiaomi and close to three times Baidu.
Before and around the release of the letter, the first batch of locked-up shares became available for trading, and the stock at one point fell by more than 19%. That sparked heavier discussion in the market over whether the valuation could hold and whether the company had become a bubble.
Revenue growth and industry backdrop cited in the report
The report said Tang’s push into coding capability began after DeepSeek R1 was released in early 2025. At that point, he judged that exploration of the dialogue paradigm had largely peaked and shifted resources toward coding and reasoning. The direct result, according to the article, was a 60-fold increase in annual recurring revenue for Zhipu’s MaaS platform over the past year. It also said GLM-5.2 had entered the global top-three tier on international benchmark rankings.
The report compared Zhipu’s direction with moves by OpenAI, Anthropic, Google, and Meta. OpenAI has kept launching agent products, Anthropic has emphasized coding through Claude Code, Google has pushed Gemini Agent, and Meta has expanded its personal AI assistant efforts. In that context, Tang’s “unmanned company” concept fits into the same race around autonomous agents.
Closing line in the letter
Tang used the closing section of the letter to answer why a newly listed company would keep putting core resources into the most uncertain direction. His answer was: “Those who truly reach the summit will turn the mountain into a road.”
He also referred to the earlier Wudao large-model project and wrote that its shared conviction among hundreds of scientists later became infrastructure that a new generation of founders could build on. The letter ended with a sharper line: “If we do not reach the top, that is failure.”