What is the valuation of AISHI Technology?

Around Halloween 2024:

A peculiar type of video suddenly surfaced across global social media platforms: an ordinary person stands before the camera, as black slime surges up from their feet, engulfs their entire body, and transforms them into Venom within seconds.

This special effect template, dubbed the "Venom Transformation," has officially amassed over 10 billion cumulative views on platforms like TikTok, with more than a million Chinese users participating in its creation.

The product behind this viral trend is PixVerse. Right after launching its V3 version in October that year, this single effect drove 10 million new user acquisitions in a single month.

Interestingly, the vast majority of people who tried it had no idea which company developed it. It belongs to a Beijing-based firm called Aishi Technology, founded on April 7, 2023, with Wang Changhu serving as its legal representative.

Starting as a senior research fellow at Microsoft Research Asia, he later rose to lead ByteDance's visual technology division, overseeing the end-to-end development of visual capabilities for both Douyin and TikTok from their inception.

When he left to start his own business in 2023, Zhu Xiaohu personally tried to dissuade him; he later recalled that 99% of investors at the time believed video generation technology would not see practical deployment within five years.

Ponder this: A former ByteDance employee did not chase grand, lofty narratives, but instead built a product that grew entirely through meme-style templates.

I looked into the numbers, and Aishi's user growth curve unfolded as follows:

December 2024: 12 million users. March 2025: 40 million. May 2025: 60 million. September 2025: Exceeded 100 million. Today: Over 150 million users, spanning 177 countries.

That's more than a 12x increase in 18 months. What powered this growth? A template-based mechanism.

With eight major categories and hundreds of special effect templates drawing from diverse inspirations: the bouncy balloon belly effect references a signature move from *One Piece*, while the "tears flowing like a river" and "Subject 3" effects are visualizations of viral internet memes. Users don't need to write any prompts—they just upload a single photo, tap once, and get a finished video.

This approach should feel familiar, right? That's exactly how Douyin grew, a playbook Adobe never followed.

Comparing it to peers makes the distinction clear: Runway positioned itself to serve Hollywood, Kling targets professional creators in film, advertising, and short drama production, and Jimeng builds a one-stop workflow rooted in the Douyin ecosystem. All of them sell "productivity."

What does PixVerse sell? In a single phrase: "I can post one too." Many users created their first viral short video with 10,000+ likes using this tool.

Its path to growth also went in the opposite direction of the norm.

From the very first day of January 2024, PixVerse focused exclusively on overseas markets. After launching its V4.5 version in early 2025, its app even surpassed TikTok to rank 4th on the overall iOS download chart in the United States. The domestic Chinese version, renamed "Pai Wo AI," was not launched until June 2025—roughly 500 days later.

The team's explanation at the time was: "The overseas market exploded so rapidly that we lacked the manpower, so we prioritized dominating that market first."

As a result, the company's profile is unique: a Chinese team built a product barely used within China, relying on the viral distribution methodology refined from Douyin to accumulate 150 million registered accounts across 177 countries.

Out of these 150 million users, how many have actually paid for the service?

01

Let's start with the figures the company itself has released. Aishi's last public revenue disclosure was during its Series B+ financing in October 2025, stating that its Annual Recurring Revenue (ARR) exceeded 40 million USD, primarily from membership subscriptions, with the figure marked as "as of the end of 2025."

No updated revenue numbers have been released since then.

The same holds true for Monthly Active Users (MAU). The 16 million MAU figure remained unchanged from the June 2025 launch of Pai Wo AI, through public statements in March 2026, and even in recent comments to overseas media, which only note "over 15 million MAU."

Over those nine months, total registered users surged from 60 million to 150 million, yet the MAU number stayed completely stagnant.

Putting these numbers together reveals the business's revenue density: 40 million USD spread across 16 million MAU means each active user contributes roughly 2.5 USD per year, less than 20 RMB.

Digging one level deeper: 150 million registered users but only 15 million MAU—nine out of ten people who sign up never come back.

Less than 20 RMB a year per user is the path it chose to take.

For comparison, Kling's ARR in March this year approached 500 million USD, more than ten times that of Aishi; Kling had 60 million global users at the end of last year, fewer than Aishi's total.

These two sets of figures lead to a clear conclusion: Aishi is the leading player in this video generation track with the largest user base, yet the lowest revenue density per user.

But we need to add context here. According to CICC's research report estimating market share based on 2025 projected ARR, Kling leads globally with around 20% market share, while Hailuo, PixVerse, Shengshu, and Jimeng all fall in the 4-5% range. Its low revenue density is relative to Kling, not a sign it is lagging behind the rest of the industry.

Why is its revenue density so low? Because it serves the 90% of users who are casual creators.

A person playing with the "Venom Transformation" effect and a professional using Kling to produce special effects for short dramas have vastly different willingness to pay. The former treats the tool as a toy and abandons it once bored, while the latter relies on it as an essential tool to meet their work deadlines.

Xie Xuzhang, the co-founder, once offered a candid defense of their strategy:

Most entrepreneurs prefer to directly pursue the B2B market, because clients pay for efficiency and return on investment is easy to calculate. The C2C market requires continuous investment in model development, while also addressing a full chain of challenges: user acquisition, retention, subscriptions, and overall user experience.

He believes that the biggest commercial opportunity in the video industry over the past few decades has always resided in the consumer market.

It is also exploring B2B opportunities, with nearly half of its API Token consumption coming from mobile app developers. However, the company has not staked its entire business on this segment, and subscriptions still make up the majority of its revenue.

Just how much friction exists in PixVerse's subscription business model?

Take a look at its pricing page. The basic tier gives users 120 free credits per day, while paid plans start at 10 USD per month, with the lowest tier capped at 720P resolution—users need to upgrade to Pro or higher to access 1080P. Credits are divided into three types: daily credits reset to zero at the end of the day, membership credits expire at the end of the month with no rollover, and only credits purchased directly never expire.

The real issue lies beneath this surface.

Video generation is not a deterministic process. Even with the exact same prompt, the output can be completely different between two attempts. Third-party guides have calculated that most creators need to generate 2 to 5 attempts to get one usable video, meaning the actual cost per valid video is 2 to 5 times the listed price.

And credits are deducted even for those failed attempts.

Almost all negative reviews on Trustpilot stem from this exact mechanism. Some users complain: "It can't even understand the most basic prompts, only 20% of outputs are usable, and all my free credits get wasted."

Others say they paid for an annual subscription but never received their monthly credits for two months, with customer support only sending generic, AI-generated template replies. Some users describe the Agent creation assistant, which bills by Token, as getting more expensive the more you use it, "designed to steal credits from users."

Yet the same product has a high rating on Google Play, with a huge number of casual users expressing great satisfaction with it.

This is not contradictory. Positive and negative reviews come from two distinct groups: people who treat it as a toy think the free quota is enough and the special effects are fun, while those who treat it as a professional tool get frustrated once they calculate the actual cost. The first group is the overwhelming majority, and only a tiny fraction of the latter contributes to that 40 million USD ARR.

How many paying users does it actually have? The company explicitly refuses to disclose this figure to overseas media today.

This is the full picture of its user ledger: an impressively smooth user growth curve, a stagnant MAU number that hasn't moved for nine months, an unrevised revenue number for nine months, a hidden paying user count, and a crowd of users complaining about the credit system in reviews.

It's a real business, but its revenue density is extremely thin.

02

On the evening of January 13, 2026, Aishi unveiled PixVerse R1. The official description called it "the world's first general-purpose real-time world model supporting up to 1080P resolution."

The core technology itself is legitimate.

Unlike other models that handle text, images, audio, and video through separate pipelines, R1 integrates all modalities into one unified processing logic—an Omni multimodal model. While previous video generation models required more than 50 iterative steps to produce a clip, the IRE engine compresses this process down to 1 to 4 steps.

Traditional AI video generation used to work like this:

You submit your request, wait, and then receive the final output, just like developing film. What R1 aims to achieve is entirely different: it turns video into a dynamic, living medium where users can interrupt the process at any time to adjust lighting or modify the plot.

This direction is genuinely innovative, as it is creating a whole new product category—not just a faster version of the old video generation tools.

The problem lies in its official descriptive claims.

The "1080P" resolution claim is inconsistent even within the company's own documents: the launch event advertised 1080P, but an update notice in February 2026 stated that the actual delivered resolution was "upgraded from 480P."

A developer who tested the product in March wrote that they only experienced a 720P version, and attached the official "R1 720P Real-Time Video API Partner Program" document as proof.

The "real-time" claim is similarly inconsistent, with two official standards: one statement claims the transition completes within half a second, another says the response latency is about 2 seconds; third-party technical documentation notes under 15 seconds, developers report a 5-second experience, and some prompts get no response at all.

The "general-purpose" label is the most ambiguous. As of today, R1 has no public API endpoints, only a rolling-approval partner program; in English-language materials, the R-series is defined by the company as a model "oriented toward game development and world building."

The "general-purpose" label is only used in Chinese-language communications. That said, it is indeed fun to use.

A heavy user who regularly uses Kling, Veo3, and Runway spent 50 RMB on Xianyu to buy an invitation code and became completely addicted to R1.

His comment is: All other video models are just improving existing capabilities, but R1 turns capabilities that never existed before into reality. Coming from someone who paid out of their own pocket to access it, this praise carries significant weight.

On the other side of the coin:

Media outlets tested R1 by simulating a moonwalk scene, first inputting "a UFO lands" then "several aliens step out"—the UFO appeared, but the characters that emerged were astronauts instead of aliens.

Testers concluded that the model has shortcomings in causal logic, with room for improvement in frame precision and stability. As an early internal beta product, it is still far from being a mature commercial offering.

So the real profile of R1 is: the technical direction is correct, the architecture is real, but the actual delivered performance falls far short of the promotional claims.

And it was released on January 13, 2026.

Two months later, funding arrived: a 300 million USD Series C round led by CDH Investments. Four months after that, Bloomberg reported it was preparing for a Hong Kong IPO. Six months later, Alibaba led the Series C+ round, raising a total of 2.98 billion RMB, with official announcements clearly stating that the funds would be invested in "real-time world model" development.

At the exact same time, every player in this sector started using the same buzzword.

Last December, Runway launched its first world model, and two months later closed a 315 million USD Series E round that doubled its valuation to 5.3 billion USD—with the funding earmarked for "pre-training the next-generation world model."

In July this year, Shengshu raised 500 million USD. In the first quarter alone, there were 25 financing events globally related to world model projects; one industry practitioner commented: "Within six months, every company will be calling their product a world model."

03

I sorted out Aishi's financing rhythm over the past three years.

Angel round in August 2023. Series A1 led by Dachen in March 2024, followed by Ant Group investing over 100 million USD a month later, bringing the total Series A round to over 400 million RMB. Series B round in September 2025 exceeded 60 million USD, with Alibaba leading the investment for the first time.

300 million USD Series C in March 2026, led by CDH. Most recently, Alibaba led the Series C+ round again, raising a total of 2.98 billion RMB.

Eight financing rounds across three years, totaling more than 550 million USD.

There is one detail worth noting: after leading the Series B round in September 2025, Aishi signed a comprehensive cooperation agreement on December 17 of the same year with the cloud business division under this lead investor, which provides global computing power services for Aishi.

In other words, Alibaba is both the investor and its main cost supplier.

This structure is not unusual in the AI industry; Nvidia's investment in Runway follows the same model: capital is invested, and part of it flows back in the form of computing power bills. This is standard practice for industrial capital, nothing improper about it.

But one statement now needs to be carefully re-examined.

Regarding costs, Aishi's earliest public statement was that subscription revenue already covers the vast majority of the company's expenses, with "cash flow nearly positive."

Pay attention to the word "nearly." Later, as this statement was relayed by media outlets, it gradually morphed into "achieved positive cash flow" and "subscription revenue has fully covered costs."

Let's set a reference point:

Kling's 2025 revenue was roughly 1.1 billion RMB, with a net loss of 1.9 billion RMB. Its 250 million RMB revenue in Q2, according to industry estimates, is barely enough to offset inference costs at the gross profit level. The industry leader, with revenue ten times that of Aishi, is still burning most of its cash on computing power.

I have no idea how Aishi calculated that "nearly positive" cash flow figure, and I have no evidence to prove it is untrue.

Wang Changhu himself clearly described their situation: "We don't have that luxury. One wrong step could stifle our growth. Every attempt costs huge sums of money, and we are not in the same cash-rich position as OpenAI."