Summary

Qu Kai remains “extremely bullish” on AI in 2025, chiefly because capital began recovering after September 2024, foundational technology improved, and the founder pool matured. Funding has clearly recovered, multimodality continues to advance, o1 has made Agent deployment feasible, and “second-generation AI founders” are bringing more mature thinking and projects; macro and geopolitical risks may affect capital supply, but he believes “the asset itself ultimately determines the market.”
The market overestimated AI’s speed of monetization in 2023, then extrapolated disappointment into excessive pessimism in 2024. He can almost state with certainty that the primary market in the first half of 2024—especially for dollar funds—was the bottom of the past 10 or even 20 years, as well as of the next several years; funding for new companies nearly went to zero. Meanwhile, foundation models are becoming commodity infrastructure, open source is advancing faster than closed source, and applications are broadly adopting multi-model stacks. “A good product always beats everything else.”
Application revenue remains sparse, but top-end validation has arrived: around 20–30 application companies are valued above $50M. Most companies still have $0 ARR, a few have $1M, very few around $10M, and the leaders may reach $30M–$50M; most projects are still launching or looking for PMF.
The 2025 funding environment will improve, but through sharper bifurcation—not a return of mass entrepreneurship. Capital is splitting into a dollar-oriented overseas expansion track and an RMB-oriented domestic-substitution and hard-tech track; top teams’ first-round valuations have risen from around RMB50M to $20M–$30M, with isolated market rumors around $50M, but ordinary backgrounds cannot raise money simply by accepting lower valuations.
After foundation models in 2023 and embodied intelligence in 2024, the keyword for 2025 will be “application deployment.” The most consensus-backed and executable wedge today is Prosumer / small-B, large-C productivity tools: they combine consumer user and distribution attributes with some B-side willingness to pay, avoiding the delivery difficulties of domestic To B, the weak payment economics of pure C, and AI’s built-in token costs; Qu Kai expects a cohort of companies to reach $10M ARR.
The key variable for Agent is not functionality, but whether it can shift from subscription pricing to outcome-based pricing. Agent bundles software services with labor, and some believe its market value could be 10x SaaS; incumbent SaaS companies will monetize first through their existing channels, but startups that take a share of savings or outcomes would have to rebuild their sales motion and organizational structure, and would mainly mount a sustained offensive against incumbent SaaS in To B.
Qu Kai believes multimodality has even greater potential than Agent and may be the answer to AI-native To C products. NotebookLM has demonstrated a new form of multimodal interaction: an article can become an AI podcast that users can interrupt at will; the broader Any-to-Any vision is to turn text, images, video, or links into the most suitable interactive format. If video becomes a major outlet, consumption shifts from passive viewing to active participation, and production relations change, “that may be the day the real next Douyin emerges.” His advice amid the uncertainty: “Stay optimistic in the face of enormous uncertainty, keep making money pragmatically, and hold on to your dreams.”

Deep dive

1. The 2024 winter was a correction after 2023’s excessive optimism

Qu Kai sees 2023 as a collective charge by internet veterans and dollar funds: after years without a major opportunity, AI appeared capable of becoming the next electricity, internet, or cloud, and was clearly still early. Capital therefore chased technical pedigrees and foundation-model companies first; Tsinghua professors and OpenAI alumni were briefly deified, while application-layer companies received little money.
Before September 2024, the primary market was effectively dead. He can almost state with certainty that the first half of 2024—especially the primary market for dollar funds—was the worst fundraising window of the past 10 or even 20 years, and of the next several years: existing companies occasionally raised follow-on rounds, while funding for new startups nearly went to zero. An OpenAI pedigree worth tens of millions of dollars in 2023 might find no willing investor the following year.
Capital then changed lanes: foundation models were the keyword in 2023, embodied intelligence in 2024; institutions less convinced by embodied intelligence shifted part of their bets into AI hardware and consumer electronics. The market briefly decided that pre-training no longer mattered, cutting related investment; China moved further toward post-training, which also reduced the scarcity value of people focused solely on pre-training.

2. Models are becoming infrastructure while applications are only beginning to form a revenue ladder

Qu Kai believes model evolution has largely followed his early-2023 view: foundation models will ultimately become commodity infrastructure. Open source is still advancing faster than closed source, and real-world products increasingly use multiple models, routing different tasks to different models; enterprises may even train small models in-house. “Technology exists to solve problems,” and product leadership matters more than leadership on any single technical dimension.
Application-company valuations have formed a ladder ahead of revenue: the market has roughly 20–30 companies valued above $50M, most only slightly above that level and a few near $100M. But most still have $0 ARR because their products have not launched or are still searching for PMF; a few have around $1M, very few around $10M, and the leaders may reach $30M–$50M.
His review does not sidestep the miss: “All of us were too bullish on AI in 2023.” The problem is that negative feedback in 2024 pushed the market into excessive pessimism; his contrarian call is not that revenue has already arrived across the board, but that technology and supply are accumulating the conditions for the next wave of usable products.

3. The post-September technology and talent inflection restarted funding

After September 2024, high-valuation startup projects increased noticeably. Qu Kai acknowledges that some of this may reflect institutions rushing to hit year-end KPIs or adding investment records ahead of their next fundraising cycle, but he believes the more important development is that foundational capabilities were quietly getting stronger.
In his view, image generation had already surpassed humans on certain dimensions, while the voice and song generation represented by products such as Suno had “basically surpassed humans.” Video and 3D have also advanced faster than expected. China’s leading players—including Kling, MiniMax Hailuo, and ByteDance’s video models—as well as many domestic startups, are also relatively advanced internationally. He compares today’s video and 3D generation to GPT-3 and expects them to reach a “basically usable” state somewhere between GPT-3 and GPT-3.5 in 2025.
Cross-generation progress in large language models has been weaker than expected: the GPT-5 or GPT-4.5 the market was waiting for did not appear. But continued gains in efficiency and cost reduction have already allowed many applications to clear the threshold where token costs and model capabilities were each just a little short. o1’s single-use experience may be “just okay,” but its reasoning ability made Agent deployment genuinely feasible.
A more easily overlooked signal is the changing founder profile: employees of foundation-model companies, heads of AI businesses at major tech firms, and co-founders and executives from smaller AI companies are starting businesses for a second time. Qu Kai calls them “second-generation AI founders,” arguing that their grasp of AI, understanding of the field, direction, and narratives are strikingly fresh—and may differ radically from those of the previous generation.

4. A recovering funding market rewards only the best as application deployment takes over

2025 will not bring back a market of indiscriminate funding; the split will widen further. Dollar-oriented capital is looking for overseas expansion, while RMB-oriented capital is looking for domestic substitution, hard tech, and chokepoint technologies. Investors would rather assign higher valuations to strong pedigrees than spray cheap capital across projects.
Since 2023, the average first-round valuation for startups has been around RMB50M. Over the past 2 months, many projects have started at $20M–$30M, with isolated market rumors around $50M. Higher valuations do not mean a sharp increase in project volume: founders with weak backgrounds, inadequate preparation, or a willingness to accept low valuations will still find it difficult to raise money.
For people who are not yet in the AI circle but believe in their own abilities, Qu Kai does not recommend starting a company immediately. A better route may be to first join an AI startup or lead an AI business inside a major tech company, build up their understanding and track record, and then go out on their own. What capital will truly be looking for in 2025 are software or hardware assets that can actually be deployed.

5. Prosumer is currently the clearest path to application deployment

Productivity tools are focusing on Prosumer, Pro-C, or “small B, large C”: professional video and song creators, insurance brokers, securities professionals, AI Coding engineers, and similar users. They combine consumer-side user and distribution attributes with a degree of B-side payment capability, making them the most consensus-backed and executable group of the past 2 years.
Over the past 5 to 10 years, the domestic market appears to have validated how difficult To big B is: GDP per capita and labor costs are relatively low, limiting the value of replacing people with software; most leading US companies are private, while many leading Chinese companies are state-owned, and profits are often divided heavily among channels. Payment cycles, customization, and private deployments further weigh on scalability. Pure C is difficult for similar reasons: willingness and ability to pay are low, while AI incurs token costs from day 1, making the internet-era model of losing money on free users and monetizing later through advertising difficult to reuse directly.
Qu Kai therefore expects productivity tools to remain mainstream in 2025, with a batch of startups potentially reaching $10M ARR. This is the clearest revenue expectation for application deployment, rather than a blanket bet on all AI consumer products.

6. Agent will monetize first in To B and challenge SaaS through outcome-based pricing

Agent is being framed as the next iteration of SaaS: SaaS provides software, while Agent packages “software services plus labor,” with the potential to replace entire departments. Some therefore believe its market value could reach 10x SaaS. Qu Kai’s more specific conclusion is that Agent will mainly land in To B in 2025.
He uses the failed transition of traditional software companies into SaaS to explain the predicament facing incumbents: the challenge is not only cloud-migration costs, but the fact that continuous delivery and one-time sales require entirely different products, sales processes, and organizations. On one side, a company may sign a large contract worth several million or more than RMB10M after repeated meetings and dinners; on the other, a young employee may give a video demo and activate the product online for RMB500 per person per month, with functionality potentially comparable to what customers paid millions to buy.
The true disruptive condition for Agent is a shift from subscriptions to outcome-based pricing: the customer uses it first, then shares the money it saves. If that model works, customers may try multiple Agents at low risk and keep only the best-performing one; sales may no longer be essential for most companies, while SEO, marketing, and the product itself handle acquisition. Generic SaaS’s insufficient customization, tool sprawl, and disconnected data could also be addressed by multiple Agents collaborating on a shared Bot platform.
Incumbent SaaS companies will still monetize first through their existing channels: if a company with several hundred million RMB in existing revenue launches an AI product and sells tens of millions of RMB worth in 1 year, Qu Kai believes that “of course it can.” But the new business model requires sales management and organizational structure to change as well; the time incumbents spend hesitating, experimenting, hiring, and replacing people is already enough for startups to get underway.
By contrast with To C, he believes today’s average founder baseline is already very high: no one would say that today’s Zhang Yiming is behind where he was 10 years ago; he should be stronger. But To B—especially domestic To B and SaaS—still has not reached the level of To C, and remains behind some overseas markets. New companies therefore still have a great deal of room to operate.

7. From search to action, only multimodality can potentially rebuild To C

The cognitive upgrade brought by Perplexity is that search is not the endpoint. In a conversation at Stanford, its founder said that if a company could help people find the answer to every question and complete every task, it could become a great company; its launched e-commerce search is one attempt along that path. From describing a need, evaluating products, comparing brands and models across platforms, and ultimately placing an order with 1 click, the search entry point could take on task execution as well.
Following that line, Qu Kai cites a view relayed by former Microsoft CEO Satya Nadella: Bill Gates long emphasized that there is only one category in the digital realm—information management. Qu Kai guesses that Zhang Yiming may hold a similar view. In the Agent era, he reasons, the category may become “information management plus action.”
Should information-assistance products for To B and Prosumer users therefore think from the perspective of delivering an outcome? If a user is gathering information to make a video, the product should not stop at collecting source material; it should also consider whether it can generate the video automatically.
Compared with the relatively clear path for Agent, Qu Kai personally believes multimodality has greater potential and is more likely to contain the answer for AI Native products. NotebookLM turns an article into a conversation between 2 AI hosts with distinct personalities; users can interrupt at any time, ask for an explanation, or change the direction. Understanding, language, cross-modal organization, and interaction combine to create an experience that existing product forms are not equipped to support.
He extends Any-to-Any into 3 still-uncertain judgments: whether video will become multimodality’s ultimate outlet; whether content consumption will shift from passive receipt to “a combination of passive receipt and active participation”; and whether new forms of organization among AI, Bot, Agent, and people can change the relations of production. He acknowledges that this is not a complete answer, but if the answer emerges, “that may be the day the real next Douyin appears.”
Finally, he advises people to stay optimistic in the face of enormous uncertainty, keep making money pragmatically, and hold on to their dreams.

AI Engineering for Art - with comfyanonymous Jeff Clune - Agent AI Needs Darwin