Summary

Qu Kai describes the current moment as the sharpest divergence in years: model momentum is at its strongest in years, while applications are at rock bottom and near death. Under the prevailing view that models will eat everything, domestic applications have still failed to deliver sufficiently strong revenue results; the funding market is even asking, “Who is still looking at applications?” and declaring that “the application category is already dead.”
The gap between US and Chinese application companies is especially stark: the US already has many companies with $100M-plus AR, while China can barely name a few with $10M-plus AR, and representatives such as Mynas and Janspark still rely mainly on overseas markets. Qu Kai’s criticism is blunt: many teams have turned going abroad into a financing script, ending up merely posting English promotional videos on the domestic internet; fewer than 1 in 10 are genuinely committed to overseas markets.
The model war is far from over, and the competitive order may reset every 6 months. Zhipu went from laggard to No. 1; Google went from being reviled to feted, its stock surging, and then abandoned by the market. Anthropic could overtake OpenAI and trail it again 6 months later. Chat has given way to Coding, and the next cycle is already visible in long-horizon tasks and autonomous agents; today’s lead is not stable.
Zhipu’s win reflects strategy, talent density and Tang Jie’s foresight, but may also owe something to luck. Qu Kai asks whether Zhipu could have stayed so focused on Coding and Reasoning if it had done better in multimodal or consumer applications. OpenAI, DeepSeek and Doubao may win the Chat phase, while Anthropic and Zhipu currently lead Coding; whether Tencent Hunyuan and WorkBuddy can overtake in a turn, how long Zhipu can stay ahead, and what role DeepSeek will play are all too early to call.
For application founders, Qu Kai offers a counter-consensus view: today is actually the best time in history. Competition and the bubble are gone, allowing teams to return to user value, products and first principles. The model-information arbitrage enjoyed by big tech’s high-P talent is being squeezed out—“there is an obvious glut of hammers, but not enough nails”—so the next opportunities must come from real use cases and demand.
The survival strategy is not to chase every hot theme, but to “ride the model wave and follow the flow of capital.” Assume most plans will change, keep serving the same users with better, newer and smarter methods, and preserve cash flow. The head of applications around Codex says every product and founder today is an option on models; Qu Kai’s point is that the goal is not to exercise the option immediately, but to keep it from going to zero and avoid blowing up before model capabilities mature.
The hype around world models, Zhipu’s market cap rising to HK$1T, and the Nvidia supply chain’s surge have led Qu Kai to wonder whether the market is nearing a frenzy or an inflection point. If Token becomes free or cheap enough, advertising and the democratization of intelligence could arrive, sending value back toward applications. He makes no call on timing, stressing only: “When most people vaguely sense that something is off about this market, then something really is off.”（当大多数人都隐隐觉得这个市场好像有些不对劲的时候，那这个市场就真的有些不对劲。）

Deep dive

1. 4 Years On, Models Have Beaten Applications Down to Their Last Gasp

Qu Kai said he returned to an AI review this year because the market had previously been boring and murky, with little to sum up; it has now entered a new phase. Looking back over the past few years:

In 2023, the market was still debating whether AI was viable and what to benchmark the opportunity against; Caret AI’s integrated model-and-application approach was briefly seen as the optimal answer. Chinese capital, however, only dipped its toes into models and applications, and neither side raised much money.
The first half of 2024 may have been the coldest stretch for China’s capital markets in the past 10 or 20 years. GPT-4o’s launch also drew a muted response; only in retrospect did Qu Kai conclude that the market had underestimated its epochal significance for reasoning and agents.
The application market saw a “return of value” in the second half of 2024, ushering in the nearly 2-year “big-tech high-P entrepreneurship era.” By early 2025, concerns that pretraining could no longer continue scaling were building, but DeepSeek and Minis somehow gave the cycle a new lease on life. ByteDance’s high-P talent could raise 3 consecutive rounds almost as soon as they emerged, on the strength of their teams and narratives.
Zhipu and MiniMax went public in succession in 2026, which Qu Kai and others once saw as a peak signal. At the same time, OpenCloud caught fire; GeekPark once noted that more than 1,000 companies and teams nationwide might be working on related projects. The market ultimately crowded into world models, with some even declaring that applications had been “abandoned altogether.” Qu Kai’s verdict: “Models have now beaten applications down to their last gasp.”

2. The Problem With Domestic Applications Is Not Just Investors

Qu Kai acknowledges that the “models eat everything” thesis and capital’s hot-theme chasing are both real forces, but the US-China results force him to put some responsibility back on founders: the US already has many application companies with $100M-plus ARR, while China can barely produce a few with $10M-plus ARR. Representatives such as Mynas and Janspark also derive most of their business from overseas markets.
The sharpest criticism is reserved for performative entrepreneurship: founders all say they will go overseas when raising money, but in execution merely circulate English-language videos inside China, while their users remain Chinese people accessing the services through VPNs. Token costs, plus willingness and ability to pay, make overseas markets a rational choice, but fewer than 1 in 10 actually stick with it.
The counterexamples are clear: D5, WorkMagic and AStudio have stayed committed to overseas markets and quietly reached $10M ARR. Once Machina founder 胡彦斌 entered overseas markets, his entire demeanor also “completely transformed.” The air overseas may not be better, but willingness and ability to pay are indeed stronger.
Even as Minus and Libleave have absorbed controversy and abuse, Qu Kai still sees them as among the few remaining survivors of the application market. He hopes they ultimately use results to prove investors wrong and even claw back the market share and mindshare that models have taken.

3. Model Winners Can Be Reshuffled Every 6 Months

For Qu Kai, the most worth studying are not the frequently discussed OpenAI, Anthropic or DeepSeek, but Zhipu and Google:

Zhipu went “from laggard to No. 1,” while Google went from being reviled to feted, its stock surging, and then abandoned by the market, alongside a wave of talent departures.
These reversals show that model competition remains in a highly turbulent early-to-middle phase, with the path to intelligence and the model landscape still resetting every 6 months. Anthropic overtaking OpenAI would not mean it could not trail again 6 months later; Google could easily return next year and become a market favorite again.
Tang Jie’s open letter treated DeepSeek R1 as marking the basic end of exploration in Chat mode, while Zhipu bet on Coding and Reasoning. Qu Kai sees the win as a combination of strategy, talent density and key foresight, with some luck mixed in; he asks whether Zhipu would have remained so focused on Coding if it had done better in multimodal or consumer applications.
OpenAI, DeepSeek and Doubao may emerge as winners in the Chat phase; Anthropic and Zhipu currently lead Coding. OpenAI and Kimi are also catching up fast. The next round is already taking shape around long-horizon tasks, autonomous agents, and competition around Cloud Tag and GPT Work overseas. The lead remains phase-specific.

4. The Application Trough Rewards Those Who Truly Understand the “Nail”

Qu Kai starts with investment method: “When things are changing fast, you are really betting on people.” The earlier the stage, the more likely the direction is to change, and the more weight people carry. Once a sector stabilizes, it makes sense to back the business and examine the data; professional managers may even perform better then. IDG is a classic case of betting on people, while Sequoia leans toward backing businesses and sectors.
Sequoia’s ability to cover every industry and stage works, Qu Kai says bluntly, “because they have money.” It can lay out sectors early and add capital with precision later, but this strategy does not suit funds with limited scale that operate in only one round.
For founders, Qu Kai instead calls today the best application window in history: competition is gone, the bubble is gone, and teams can return to user value, products and first principles. A downcycle reduces noise; real companies and founders often emerge from the trough.
Big tech’s high-P talent once relied on information arbitrage, but in the AI era the gap in understanding and deploying models is narrowing. “There is an obvious glut of hammers, but not enough nails.” Agent, proactive agent and harness engineering cannot substitute for use-case insight; the next opportunities should start with real user needs and problems.
Teams should assume they have to go with the current: ride the model wave and follow capital’s flow, accepting that only some things are worth holding onto while continuously serving the same users and solving the same class of problems in better, newer and smarter ways. This is not trend-chasing or hype, nor saying one thing to raise money and doing another; the endpoint remains the user and the problem.

5. The Token Dividend May Be Nearing a Turning Point, but Timing Remains Unknown

Qu Kai uses the fundraising experience of 安比 founder 莫子浩 to reflect on his own instincts: he had reflexively advised him to pitch the model story investors wanted to hear, but could not answer the question, “Why build a model now?” 安比 was strong on engineering and offered a strong product experience, but a capital fad alone cannot be a strategic reason to build a model.
His bottom line is “user-value fundamentalism.” Manas went through a period when it had a $50M valuation, spoke to everyone in the market and found no investor; it also went through a period when it was valued at $10M, institutions stalled and haggled it down, and it ultimately ended with a valuation below $100M. Strong teams need to preserve cash flow and keep compounding.
The head of applications around Codex says every product and founder today is an option on models. Qu Kai’s interpretation is that today’s accumulation may pay off once model capabilities mature. The key is not to monetize the option immediately, but to ensure the option does not go to zero: do not blow up, and do not sell the option prematurely.
The surge across Nvidia’s supply chain, Zhipu’s rise to a HK$1T market cap, and the return of 2021-level SaaS heat in 巨深 world models and related fields have prompted Qu Kai to ask, “It has already reached world models—it is almost the last stop. What story can capital tell after that?” When embodied intelligence and world models will actually land remains unknown.
IBM once collected fees for CPU compute under the MIPS model, while telecom operators earned from SMS and data traffic for nearly 10 years. Today, GPU and model companies charge by Token; measured from GPT-3.5, AI is approaching 4 years old. If Token becomes free or cheap enough, advertising and the democratization of intelligence could arrive, and the market may once again declare, “Now it’s applications only.”
In Technological Revolutions and Financial Capital, economist 卡洛塔 divides the installation period of a technology revolution into an eruption phase and a frenzy phase, followed by a turning point resembling the 2000 internet bubble and then a deployment phase. AI is evolving markedly faster than other technology cycles in history, but Qu Kai gave no timing call on whether it remains in the eruption phase, has already entered frenzy, or is still far from the inflection point. He left only this warning: “When most people vaguely sense that something is off about this market, then something really is off.”（当大多数人都隐隐觉得这个市场好像有些不对劲的时候，那这个市场就真的有些不对劲。）

Urgent Update- AI Sputnik Moment: Kimi K3 Released w/ Emad Mostaque | Ep. 272 The Restaurant Genius Behind Carbone, ZZ’s, Torrisi & More | Jeff Zalaznick