An Internal Letter
July 11. Zhipu founder Tang Jie sent an internal letter named Touch High — the Reach-High Plan.
Its weight isn't in slogans but in laying out Zhipu's two-year direction clearly: no short-term monetization chase, aim straight at the next AGI frontier.
Four Engines
The plan bets on four directions:
- Long-horizon tasks: AI moves from instant Q&A to large-scale engineering, spanning entire project lifecycles — learning, doing, remembering. It can auto-decompose a grand goal into thousands of subtasks.
- Autonomous agent systems: from assistant to digital employee, building societies of tens of thousands of specialized agents that debate, collaborate, review code, and schedule resources on their own.
- Fully self-training: human high-quality data is running out. Synthetic data factories and AI-vs-AI adversarial play let the system evolve itself, even rewrite its own code.
- Extreme safety governance: investing billions-scale resources in mechanistic interpretability, turning black boxes transparent.
LLM OS: Redefining the Operating System
The pivotal line: in the future you won't open a computer to today's OS — you'll see an LLM OS, all functions generated on demand.
This isn't casual talk. Tang frames it as a challenge to the von Neumann architecture that's run for 80 years. Office software, search, apps might cease to exist — every capability generated dynamically by agents.
If that happens, today's API pricing debates will look like arguing over Windows installation fees twenty years ago — meaningless. What matters is who becomes the entry point.
Building Its Own Chips
The other thread: July 8, The Information reported Zhipu is considering self-designed AI chips.
Two reasons. One: GLM-5.2, released last month, saw daily token calls surge 27x — the fastest-growing model on the Vercel aggregation platform. Two: US export controls keep tightening; compute supply has shifted from a cost issue to a hard bottleneck.
Zhipu currently runs Huawei chips, other domestic chips, and some NVIDIA. Because it's on the US entity list, it can only look for domestic design partners. The full chip R&D cycle is two-plus years; it's still early, no design partner chosen yet.
How to Read This
Put both threads together and Zhipu's ambition is clear: not just models, but the foundational platform of the AI era.
The model is the foundation, LLM OS is the operating system, self-designed chips are the hardware base. This is the same path OpenAI, Google, and Anthropic envision — Zhipu just faces harder constraints, with compute restricted, forcing it to solve the problem itself.
Of course, Touch High is strategy, not product. Whether it delivers in two years depends on three things: compute, data, and engineering execution. But at least the direction is clear — not fighting price wars at the application layer.