Original · Unique Research · 2026-03-19 · Shanghai
English edition note: This is a complete translation of the original article, including its historical event invitation. Funding, investor participation, usage, revenue, training-cost comparisons, product dates and market-leading claims below are statements reported by the source or attributed to Xie Xuzhang, not independent verification. The source's narrative moves between dates, and its stated late-2025 timing for Seedance 2.0 remains to be checked. App retention and website bounce rate are distinct metrics; the quoted comparison does not establish a like-for-like retention ranking. Chinese names are translated or romanized and do not establish official English legal names. The original assessments and predictions are preserved as of the source date.
Unique Awards · Hangzhou AI Week
US$300 Million, Asia's Largest Single Financing Round in AI Video: In a Field Rejected by 99% of Investors, They Reached 100 Million Users
When Everyone Said "Startups Are Finished," We Chose to Keep Going a Little Longer
In March 2026, Aishi Technology completed a US$300 million Series C financing round.
The amount set a new record for the largest single financing round in Asia's AI video-generation sector.
The round was led by Dinghui Hong Kong Fund, Dinghui VGC, and Dinghui Baifu. Participating investors included industry investors China Ruyi and 37 Interactive Entertainment, alongside Yizhuang State-owned Investment, Zhongwei Capital, Guotai Junan Innovation Investment, Suchuang Investment Group, and other well-known institutions.
What many people do not know is that just three years earlier, 99% of investors were pessimistic about video generation and believed it could not be put into practice within five years.
Aishi Technology's founding team got started amid that lack of consensus. In October 2023, just 6 months after the company's founding, it launched PixVerse V1, releasing what the article describes as the industry's first AI video-generation model capable of producing 4K-quality video—before Sora.
Over the following two years, it went from V1 to V5.5, from 4K generation to rapid generation in 5 seconds, from 10 million users to 100 million, and from ARR of 0 to US$40 million.
Even more strikingly, its training costs were only 10% of those of its peers.
How did it do it?
"99% of Investors Were Pessimistic": From Non-Consensus to Consensus, and Continuing to Back Non-Consensus Ideas
When Aishi Technology was founded in 2023, video generation was still a non-consensus field in China.
Xie Xuzhang recalled that in the company's early days, 99% of investors were pessimistic about the direction and believed it could not be put into practice within five years.
"At the time, everyone was paying more attention to the progress of large language model companies. Even starting a video-model company was a non-consensus choice."
But the team firmly believed AI would fundamentally change how content was created.
It adopted a strategy of taking small, fast steps: focusing on rapid product iteration and letting actual results speak. In October 2023, PixVerse V1 launched with pioneering 4K video generation, demonstrating technical feasibility.
The real turning point came in the second half of 2024.
In October 2024, PixVerse V3.5 introduced template effects with packaged prompts. Transformation effects alone brought PixVerse more than 10 million new users worldwide within two months, creating a global ChatGPT moment for video consumption.
"At that point, we felt we had gotten both the model and the application right."
But challenges came one after another. In early 2024, Sora burst onto the scene, sending shock waves through the industry.
"Everyone was saying 'Startups are finished,' 'The tech giants will dominate everything,' and 'Startups should build applications, not large models.'"
Facing the competitive pressure brought by Sora, the company chose to increase its investment in large-model training.
"We drew up a clear technical roadmap, focused on large-model training and performance optimization, reduced the cost of trial and error through modular development, and concentrated our resources on key bottlenecks to ensure significant breakthroughs within three months."
Xie Xuzhang said that despite all the discussion outside the company, the team was fairly resolute internally: it wanted to keep control of its model capabilities.
"
There was an upside to Sora's arrival, too. We had already stuck with developing our own video-generation models from the time it was a non-consensus idea until it became consensus. Why not keep going a little longer?
Why Are Its Training Costs Only 10% of Its Peers'?
In the AI industry, training large models burns money.
But Xie Xuzhang disclosed that Aishi trains models with the same parameter counts, or even better results, using an average of fewer than a thousand accelerator cards per month. Its costs are roughly 10% of those of its peers.
How does it do that?
Xie Xuzhang's answer was straightforward:
It is a combination of strengths—model architecture, algorithms, engineering, and product capabilities. Outside observers want a simple explanation, but there is no single cause.
He added: "Some large-model companies that previously raised more money than we did lack neither people nor data. Yet most of them still cannot produce a video-generation model. That shows training video-generation models is difficult in itself."
Drawing on large-scale data-processing capabilities and deep-learning algorithm experience accumulated on platforms such as Douyin and TikTok, Aishi has distinctive advantages in video generation. These platforms helped it accumulate extensive user-behavior data and content-understanding capabilities, making it more competitive in model training and optimization.
By 2025, rapid-generation technology was continuing to improve:
📅 February 2025: PixVerse V4 could generate high-quality video within 5 seconds.
📅 November 2025: V5 Fast launched, generating 1080p video in just 30 seconds.
ToC Before ToB, Overseas Before China: The Secrets of a Global Strategy
From the start, Aishi Technology chose a strategy of ToC before ToB, prioritizing overseas markets before expanding in China.
Why?
Xie Xuzhang explained: "From the very beginning, we wanted to build a platform for users worldwide. The team itself had global experience, so early product validation and iteration focused more on the global creator community."
What did that strategy deliver?
🏆 August 2025: PixVerse, known in Chinese as Paiwo AI, surpassed 100 million users and 16 million monthly active users.
"This strategy helped us accumulate a large user base in a short time, laying the groundwork for subsequent commercialization."
In September 2025, Aishi Technology completed a Series B round of more than US$60 million, led by Alibaba. Participants included Dachen Caizhi, Shenzhen Capital Group, the Beijing AI Fund, Hunan TV & Broadcast Intermediary, Giant Network, and Antler.
The platform's annual recurring revenue (ARR) reportedly exceeded US$40 million.
It subsequently completed a Series B+ round of more than RMB100 million in October, with investors including Fosun RZ Capital, Cowin Capital, and the Shunxi Fund.
Facing Seedance 2.0: "The Market Is Big Enough; We Have Not Reached the Stage of Head-to-Head Competition"
At the end of 2025, Seedance 2.0 became a breakout sensation, putting pressure on other video-generation companies.
Was Aishi anxious?
"We are anxious. But since our founding in 2023, we have experienced the same thing so many times over these three years—Sora, Kling, Veo, and others. There have been so many 'world-changing' releases, so much 'disruption.'"
The data, however, showed no obvious impact.
"
Judging by user data, Seedance2.0 has had no obvious effect on Aishi. The video-generation market is big enough that we have not yet reached the point of needing to compete head-to-head.
Xie Xuzhang disclosed that third-party platform data showed PixVerse had the industry's highest retention for both its App and its website.
"The Sora app's 30-day retention is 8%, while our Bounce rate is lower than Sora's. A low Bounce rate means users are willing to try things on your website."
World Model R1: Exploring the Boundary Between Video and Games
In December 2025, PixVerse released the V5.5 multimodal video-generation model, supporting both storyboard-based creation and synchronized audio-and-video generation.
But what draws even more attention is the "world model R1" that Aishi Technology is exploring.
It is a video-generation model that uses an autoregressive approach to generate video in real time and let the visuals continue indefinitely.
"In simpler terms, our definition of a world model is a model that, after AI learns enough about how things work, can predictively simulate physical, causal, spatial, and temporal laws."
Why invest in real-time and unlimited generation?
We believe the boundary between video and games will inevitably become increasingly blurred. Once video becomes interactive, entirely new opportunities for content, users, and creators will emerge, so we must position ourselves ahead of time.
In R1, creation and consumption become one. The person watching a video is also a creator—for example, participating in interactions by consuming Token units. In the future, the videos people see may share the same basic framework, while their specific imagery and direction change in real time according to individual preferences.
"This will bring substantive changes to existing content-consumption models."
After R1 was released, the largest group of customers approaching the company came from the gaming industry. "Recently, Google's Genie 3 has shaken up traditional game engines. R1 could also use AI-native models to reconstruct the underlying process of game creation."
Future game development will no longer have to endure the long development cycles of the past. Whether it is gameplay, visuals, or story, AI can make it lighter and more imaginative. More importantly, it can help people who have ideas but cannot code turn those ideas into real games.
How Will the US$300 Million Be Spent? Keep Backing Non-Consensus Ideas
After the US$300 million Series C round was completed in March 2026, someone asked Xie Xuzhang how he planned to spend the money.
"We will keep investing in R&D, explore new businesses, and expand into global markets. We hope to build the best video models and make the R1 series better, too."
He disclosed that the company hoped to increase its training investment severalfold this year.
But the signal behind the financing is more noteworthy.
Xie Xuzhang said frankly that when fundraising began late last year, the initial plan was to raise US$100 million. But many investors in China and abroad wanted to participate, so the company decided to raise more and build up ample reserves.
Why were so many institutions willing to invest?
Because, in terms of both attention and revenue, video generation is growing rapidly among AI's various fields. In the United States, video-generation API usage already exceeds US$100 million in a month, amounting to RMB10 billion of API usage over a year.
More importantly, following the release of Seedance 2.0, Aishi Technology is still pursuing several non-consensus choices:
✅ Continue investing in self-developed models, rather than only building at the application layer on top of large-model companies.
✅ Continue pursuing globalization, rather than focusing solely on the Chinese market.
✅ Continue serving ordinary people, rather than only professional users.
✅ Continue investing in the video-based world model R1 to explore new forms of real-time interaction.
Xie Xuzhang said these are still non-consensus choices.
But that persistence has taken Aishi Technology from rejection by 99% of investors to the largest single financing round in Asia's AI video-generation sector today.
A Final Word
Aishi Technology's story is a story about non-consensus ideas.
From rejection by 99% of investors to Asia's largest single financing round;
From anxiety in the wake of Sora to continued commitment to self-developed models;
From doubts that "startups are finished" to 100 million users and US$40 million in ARR;
From a technical barrier built on training costs just 10% of its peers' to exploration of the real-time interactive world model R1.
Xie Xuzhang said: We have already stuck with developing our own video-generation models from the time it was a non-consensus idea until it became consensus. Why not keep going a little longer?
That is a line every entrepreneur in AI should remember.
When everyone says "startups are finished," perhaps that is precisely your opportunity to keep going a little longer.
Aishi Technology will appear at the Hangzhou Unique Awards event with its own booth. Want to experience the creative power of PixVerse and the R1 world model up close? Come and talk to the team at the event.