Zhipu's Wager: When the Chat Paradigm Is Exhausted, Coding Is the Next Peak

The Event

Zhipu AI founder Tang Jie revealed in an internal letter that the company predicted "the Chat paradigm's exploratory phase essentially concluded" a year ago, and subsequently concentrated resources on the model's code generation (Coding) and reasoning (Reasoning) capabilities. The bet paid off—over the past six months, Zhipu's valuation increased 10-fold, entering the "trillion Hong Kong dollar club," surpassing Xiaomi and reaching three times Baidu's market cap. After the first wave of stock unlocking, the share price held firm.

On the surface, this looks like a stroke of genius, but the deeper phenomenon is: Zhipu identified a trend inflection point that DeepSeek hadn't fully exposed—when DeepSeek squeezed chatbots' logical reasoning capacity to the limit using chain-of-thought, chat itself was no longer the battlefield; the next peak lies in code generation and intelligent agent (Agent) collaboration.

Why This Matters

In the technology industry, track switching is often the dividing line between survivors and the eliminated. Three levels explain this phenomenon:

Level 1: Identifying Boundaries vs. Linear Extrapolation

Where lies the boundary of the chat paradigm? From a technical perspective, it faces these constraints:

• Reasoning depth: How many reasoning steps can a user's single query execute? The competition between ChatGPT, Claude, and DeepSeek R1 is essentially an "arms race of reasoning steps." But this race has physical limits—when reasoning ability is pushed to 95% accuracy, spending 100 times more resources only yields 99% accuracy, with diminishing marginal returns.
• Use case ceiling: Chat's most natural scenario is "human-machine dialogue." But AI's true commercial value lies in "AI replacing human workflows," which requires AI to autonomously write and execute code, integrate with systems, not merely answer questions.

Zhipu's insight lies here: while competitors still compete for "reasoning rankings in Chat," the next arena has already shifted to "code generation + autonomous agents." This isn't a quantitative difference in reasoning ability, but a shift in application boundaries.

Level 2: Organizational Inertia vs. Strategic Flexibility

Why didn't OpenAI and Baidu pivot first? Likely not because they failed to see it, but because:

1. Sunk costs: They've already invested enormous training resources, product lines, and customer expectations in the Chat paradigm. A sudden pivot means admitting past investments were "inadequate preparation." 2. Organizational habits: Product teams, sales teams, and marketing departments all target "better Chat." Changing direction requires comprehensive coordination, and executives typically choose to "push harder on the old track" rather than bet on a new direction. 3. Short-term earnings pressure: Chat already has clear revenue models (token-based pricing), while code generation's business model (SaaS, enterprise agent fees?) remains uncertain.

Zhipu had no such baggage—as a later entrant, it possessed strategic flexibility instead. The resource reallocation in early 2025 didn't disrupt vested interests, yet allowed it to claim first-mover advantage on the next battlefield.

Level 3: The Ghost of Timing

Clayton Christensen's disruption theory points out: disruptors often succeed not through superior technology, but by appearing at the moment new rules are being defined. Zhipu's success follows the same logic:

• DeepSeek R1 (late 2024) launched when industry conversation still centered on "reasoning performance."
• By early 2025, perceptive observers began asking: are reasoning paradigm marginal returns diminishing? What's the next functional boundary?
• Zhipu's wager occurred in this brief "cognitive vacuum"—they pivoted while competitors still optimized reasoning accuracy.
• By mid-2026, the market validated that "code generation + reasoning" indeed became the new value driver, and Zhipu had already established first-mover advantage.

The Deeper Question

But there's a trap in this story: timing identification itself is uncertain. Tang Jie said "the Chat paradigm essentially concluded," but what if he miscalculated? What if chatbots still had 3-5 years of rapid growth ahead? Then Zhipu's resource shift becomes "premature exit," potentially surpassed by OpenAI or Baidu's linear push.

We see Zhipu winning now and easily say with hindsight that "Tang Jie had exceptional foresight." But at the moment of decision, this was a high-risk wager with perhaps only 40% success probability. It's precisely because they wagered on that 40% correctly that today's story exists.

Transferable Insights

This atom's core isn't "Zhipu won," but:

Every rapidly-growing track has signals of "boundary proximity." Identifying these signals matters more than running faster on the old track.

Signals include:

1. Diminishing marginal returns: 10x resource investment yields only 10% performance gains 2. Homogenized competitor behavior: All players optimize along the same dimension 3. Emerging new use cases being overlooked: Such as the shift from "chat" to "process automation" needs 4. Time window: Typically 6-18 months—too short to judge, too long misses the opportunity

The insights for investors, entrepreneurs, and managers all lie here.