---
title: "Tokens Are Going Through the Roof, but These Three Insiders Say Cost Isn't the Problem at All"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-05-27T12:37:11+00:00"
canonical: "https://ffcap.cn/en/research/src-20260527-02html"
source: "https://uniqueresearch.substack.com/p/src-20260527-02html"
language: "en"
---

# Tokens Are Going Through the Roof, but These Three Insiders Say Cost Isn't the Problem at All

_Original · Unique Research · 2026-05-27_

_Editor's note: This is a complete English rendition of the original Chinese article, including the editorial summary and the full roundtable transcript. All statements by named panelists are presented as their own views and claims, not independently verified facts. Company names, product names and dollar amounts are preserved exactly as stated in the source._

AI Industry Truth Talk

Everyone Is Asking Whether Tokens Are Expensive, but the Real Question Is What Exactly You Are Selling

At the Unique Awards Shenzhen event, a roundtable stripped the token economy down to its essentials.

"

Last week in Shenzhen, three people who are actually doing business with AI applications sat together and talked for an hour.

Li Shaohui of Kuaicece (快决测) runs market research SaaS; Ren Xinyi of Futureform Intelligence (未来式intelligence) runs an agent platform; Yu Beichuan of Zhishu Yinli (指数引力) does AI influencer marketing and going-global tools.

The moderator opened by asking: Token prices keep rising—does this affect going global?

All three said almost simultaneously: cost is not the biggest problem.

Then they spent an hour clarifying what the bigger problem is.

Who Are You Comparing Cost Against?

Li Shaohui's calculation is straightforward: he doesn't compare token spending against server bills—he compares it against labor.

If one person costs $200 a month in token fees, and the team frees up 10% of headcount, the labor savings far exceed the token cost—enough for the entire team to use tokens for a very long time.

Yu Beichuan said that when they built Laya, the first six months were heavily loss-making, precisely because tokens were expensive. But his judgment is: this cost will decline, and a large part of that decline won't come from model price cuts—it will come from engineering—cache hit rates, cross-model routing, task decomposition. This is an engineering problem, not a strategic bottleneck.

Companies that treat token cost as their biggest challenge often haven't figured out what they are selling.

A Block of Tofu Can Sell for 10 Yuan or 100 Yuan

Ren Xinyi's metaphor: tokens are raw materials, and you don't control their pricing. What you sell externally is "a dish." The same mapo tofu—at a street stall versus a high-end restaurant—can differ tenfold in price. The difference is not the cost of the tofu.

So the question becomes: are you currently selling ingredients, cookware, or the dish?

Kuaicece's answer is: selling decisions. Li Shaohui put it clearly—delivering a questionnaire or a report doesn't count as a result. The result is whether the boss, after receiving that insight, dares to make a five-million-yuan investment decision. What's valuable behind that is not tokens, but over a decade of accumulated vertical data and industry knowledge bases.

"If you really charged by token, every company would go bankrupt. Tokens have one price; the effort and expertise put into delivering that result is the core of pricing."

Yu Beichuan was even more direct: Chinese companies are reluctant to pay high prices for software tools—this is a structural issue that won't change in the short term. Their solution is to package AI as a service—AI eliminates 95% of the workload, but what the customer ultimately pays for is professional endorsement and outcome commitment. That's where the gross margin lies.

"Outcome-Based Pricing" Sounds New, but Users Are Already Accepting It

Yu Beichuan said something very clear-headed:

"Open up most software and it's still $19 a month for 3,000 credits. But those 3,000 credits are outcome-based pricing in disguise—they just don't tell users directly. Users calculate how many PPTs this money can help them make, and if it's worth it, they pay."

The real difficulty is not getting users to accept the concept of "outcome-based pricing," but whether you can define the outcome, deliver the outcome, and measure the outcome. Most AI tools haven't reached this step yet, so they can only wrap a subscription layer around it first, letting users gradually build awareness.

The Final Insight, the Most Counterintuitive One

Near the end, Yu Beichuan said: the most aggressive AI users are young people born in 2000 or 2002, because they feel they don't know anything, so they completely trust AI. On the contrary, people who have been programming for eight years think AI is nothing special—they have the most inertia and the lowest efficiency.

Li Shaohui immediately jumped in: strongly agree. The people he works with most smoothly now are 22- and 23-year-olds. People with seven to ten years of work experience always feel like they're "playing adversarial," and new things can't get through.

This is not saying young people are smarter. It's that the deeper your understanding of the old world, the higher your switching cost. Being a blank slate, at this moment in time, is an advantage.

A roundtable has no conclusion, but three questions are worth thinking seriously about:

Are you selling ingredients or a dish?

Is your pricing core token volume or outcome value?

Does your team have someone who "doesn't know anything and therefore dares to try everything"?

More Conversation Details

Zhao Liang (Abner): Today we mainly discuss the token economy. Since "token" finally has this Chinese name "ciyuan" (词元), everyone also recognizes it as a new value point in our current intelligent era. So today we've invited several founders and key company leaders who are doing very well in AI commercialization to discuss this topic. Before we start today's panel, could each guest first give a brief self-introduction, about two minutes each, introducing yourselves and your company's products and business. Shall we start with Mr. Li?

Li Shaohui: I'm Li Shaohui, founder and CEO of Kuaicece. I used to work in Marketing at Procter & Gamble. Over ten years ago I founded my first tech company, Kuaizi Tech, which is currently a leading company in the domestic AIGC field. Kuaicece was spun out of Kuaizi in 2017. We mainly do market research digitalization and AI empowerment. Currently we conduct research for leading consumer brands and internet companies in over 80 countries worldwide, serving about 400+ leading brands. We mainly provide two types of services: one is traditional research services, but AI-empowered digital online services; the other is providing our clients with online research SaaS products and AI products, including quantitative analysis, qualitative analysis, big data analysis, etc. This track is very hot globally right now, and tokens play a very important role in it. Without this kind of AI intelligence, it's hard to imagine providing clients with products that can automatically deliver professional services.

Ren Xinyi: I'm Ren Xinyi, from Futureform Intelligence. The company name is quite interesting—simply put, we help enterprises and individuals build their own Agents, so "Futureform" expresses the future way of working. Since our company was founded, we've always faced a problem: in this era, tokens are a bit like water and electricity, or you can understand them as your raw materials. But obviously, vegetables you buy at the market can't be eaten raw—you have to process them into a dish. So for a long time, what we did was provide tools, just like your kitchen has a stove, pots, and pans—we provided a complete AI agent building platform on which you could make your own dish.

After doing this for several years, we found that with so many dishes, you won't innovate differently every day. There are things we call industry know-how. Based on this industry know-how, we wondered whether we could have something ready to use out of the box, and whether while creating my own Agent, others could also use what I created. So on May 19, just two days ago, we launched our C-end product called "Profy (袋袋)"—a pocket expert team. You can find the tools you want (we call them "experts"), and you can also package your own experience as an expert and publish it on Profy. So in this process, we've indeed always had to revolve around tokens, and I very much like its Chinese name "ciyuan" (词元)—it's the raw material in our business, and we turn it into a dish that can truly be served and eaten. That's essentially what we've been doing.

Yu Beichuan: I'm Yu Beichuan. Our company has sequentially built three businesses. The first was AI influencer marketing, helping Chinese companies going global to deploy influencers. We're currently the largest in AI influencer marketing in China, and this business also experienced significant growth last year. The second business—because when doing influencer marketing we found that the most important thing isn't chatting, it's finding people—so last year we started building the product Laya. Because token costs are expensive and you need to use the world's best large models, Laya has been a Global product from Day 1, currently mainly in North America and Latin America, with many paying users. The core scenario is helping you find experts, or many overseas lead generation scenarios. The third product is something we started working on after this year's Spring Festival—a new To-C content platform. Our company's span is very large; basically every year we follow new changes and enter a new product while keeping the old products growing. That's roughly our company introduction.

Zhao Liang (Abner): As you've probably heard, both Zhishu Yinli and Mr. Li's company have going-global ambitions, or are already doing very well overseas; Futureform Intelligence may also have some going-global plans. So our first topic is a relatively open one: the current "token economy" and "token going global" are very hot, even many macroeconomists are discussing what impact these two terms might have on our entire national economy. Against this backdrop, I'd like each guest to share: in the current AI era, what opportunities and challenges will companies going global face? Since token prices have been rising since the start of this year, will this bring some impact to our business or going global? Shall we start with Mr. Li?

Li Shaohui: I think we should look at it from two sides. On one hand, it seems tokens or APIs have become more expensive recently, but from the perspective of overall industry development, tokens will definitely become cheaper and cheaper. So right now, from our own practice, we're more focused on commercially operable token applications. In product R&D, we choose to use a combination of GUI and conversational interfaces. I know many people now think that making AI Native products means making conversational products. But if we're doing To B, in my humble opinion, using conversational interfaces in many scenarios is actually not the best approach. Because when the user's intent is very clear, using a conversational interface to help interpret the intent and then waiting 30 seconds to produce something is not that smooth. This may not align with mainstream thinking, so we believe that clicking a button in a GUI can achieve it—and this approach actually saves a lot of tokens. But in the necessary links, we use the best computing power and the most advanced models, sparing no cost to meet their requirements.

If we go global, I think for us the biggest challenge is not token cost—the biggest issue is still the enterprise's own resources, such as Chinese talent versus American talent, legal compliance, etc. These challenges are bigger.

Zhao Liang (Abner): So for you, cost may be a secondary issue, and other main challenges are more important than cost.

Li Shaohui: Right, this may be related to our industry. Because we're in the research industry, the "human time" we used to sell to clients was very expensive. In this context, if we can use $200 worth of tokens to replace human work, it's much cheaper. So this cost is actually very small for us. Of course, if we were to turn the product into one where everything is delivered by AI, the cost could become very high—but we haven't made that choice now, because we calculated that it's not cost-effective, and the user experience may not be that good either. But in the future, if AI capabilities improve enough, we may migrate in that direction. Because the underlying architecture is fully capable; we just use a hybrid GUI-and-conversation packaging approach at the front end.

Zhao Liang (Abner): Cost may be just one small aspect. Let's mainly discuss more opportunities or challenges. Could Beichuan share next?

Yu Beichuan: I think it's a bit like what Mr. Li just said. When we do going global, from Day 1 what we consider is the business model—whether this thing can work. Actually, when we were building Laya—this may be a bit embarrassing to say—our first six months were consistently heavily loss-making, precisely because token costs were particularly high. But you know this cost will decline, and that decline may not come entirely from large model price cuts (because the best large models are still quite expensive); the decline may come from your engineering. For example, your cache hit rate gets higher and higher, or you can continuously use other models for cross-replacement. For instance, when we launched on Day 1, we basically used Claude for everything—very extravagant. Because at that time the best model was Claude, it was leaps and bounds ahead. But early this year, we basically shifted half of our traffic to Gemini. It's actually only slightly more expensive than Doubao, but performs better in many usage scenarios—it's the most cost-effective model (of course, this was before DeepSeek-V4 came out).

For us, what we care more about is that token is just a surface phenomenon; it's essentially intelligence, and it brings new ways to satisfy demand. So the more important thinking is: what demands in the past couldn't be done without tokens? It actually brings more new charging methods, such as what we used to call "delivering outcomes." Although many software today can't achieve one-step outcome delivery, at least it's point-to-point outcomes, and this outcome-based charging method is relatively more flexible than past SaaS software charging methods. You can charge by an image, a video, or a customer conversion. Actually, the charging rate at these nodes is significantly higher, and the average order value is much higher than past software. The industry-wide opportunity lies here—you have new thinking points for business models. Many Chinese SaaS companies didn't do well enough in the past, but in this AI wave, Chinese teams are doing relatively outstandingly compared to overseas. Today many of the best AI software are actually made by Chinese teams—this is a disruptive opportunity.

Zhao Liang (Abner): So in Beichuan's view, cost is one aspect, and the business model may be more important. But as you mentioned, we now charge by outcome, which may be more flexible than the traditional subscription model—but will it be harder to implement? Will it face more challenges?

Yu Beichuan: I think definitely, new concepts need a cycle when educating the market. For example, people have talked about outcome-based pricing for so long, but open up most software and it's still $19 a month for 3,000 credits. But actually, the conversion logic of those 3,000 credits is outcome-based (workload-based) pricing in disguise. For example, when we subscribe to this software, those 3,000 credits might be used up in a week, and what you measure is how much value this corresponds to—like it can make 5 PPTs for me, and spending these tens of dollars is worth it. So it's not directly telling users you're paying by outcome, but doing subtle conversion. As software becomes more widespread, most users have already accepted this quantitative billing model rather than unlimited usage. Only after the new model becomes popular is there an opportunity. If on Day 1 you sign up for a charging method completely contrary to past understanding, the user's payment rate may be very low. So you still need to start from a model users are more familiar with, but the core of your pricing is actually calculated by delivered outcome.

Zhao Liang (Abner): Next, Xinyi—we're very familiar with Futureform Intelligence, which has done very well on the To B side. We just heard you launched a To C product similar to a Skill Market, and we're very much looking forward to your thoughts.

Ren Xinyi: I'm quite happy—I very much agree with both viewpoints shared. The first point is "paying for outcomes." Token volume is large, but pricing power is not entirely in our hands, because it exists at the raw-material level in the industry chain. For raw materials, you can never raise the price very high to make a lot of money. What we sell externally is a dish—like a plate of mapo tofu, under different chefs and different restaurant environments, it can range from tens to hundreds of yuan. This is the same as the evaluation method Mr. Li just mentioned: if through AI you can cut two people and the cost savings break even or even profit from the AI tool investment, then this thing is worth investing in. KPIs and ROI are easy to calculate—this is tangible paying for outcomes.

Beyond calculation, what we really need to care about is how to sell mapo tofu for 100-200 yuan, which is also why we launched the product "Profy" this year. In Profy, our expectation is no longer to price in the original token way, but on top of AI, to layer everyone's know-how about these things. On the other end, we very much hope to find people with expertise to come onto Profy and put their industry knowledge up there—you are that "chef." Layering AI computing power on top of existing knowledge to help others solve problems, what you earn is also that layer of maximum value addition.

Zhao Liang (Abner): Futureform Intelligence has been proposing "digital labor paid by outcomes," and Xinyi also very much endorses this business model. Then I'll ask a follow-up: because different digital employees hold different positions and solve different problems. If token cost were infinitely low, pricing would be easy; but if token cost is high, your consumption and cost structure for different tasks are also different. So will you face challenges in pricing?

Ren Xinyi: Definitely. In the early stage, to put it bluntly, you might just lose money yourself, because you must use the most cutting-edge and expensive computing power to handle complex first-time problems. But this is a bit like "payment routing"—if you've done cross-border, you'll know that different payment channels have different prices, and you'll keep switching based on the situation. So in the process of moving toward outcome-based pricing, there will definitely be a flywheel effect. In the early stage without a good solution, you use the most cutting-edge computing power, but as scale benefits increase, you can hit some historical cache content, and overall cost will decline. A good example to understand is buying insurance: 100 people buy car insurance, and at most 10 people file claims in a year—will the insurance company lose money? No. When the base is large enough, individual high-cost situations can be diluted by the entire pool. That's our plan.

Zhao Liang (Abner): Our first round of questions is over. Next we enter the second round. I may ask each of the three guests separately, combined with their own company and product backgrounds. The first question, Mr. Li, could you share a very typical business scenario or a successful customer practice of your company? So everyone can get a closer look at the business.

Li Shaohui: A typical scenario—for example, one of China's top consumer goods brands, all their business units use the insight system we provide. They need to do product concept testing, packaging testing, price testing, and a series of tests. People in different categories have different professional capabilities, but through the AI automation empowerment of our system, different people can all achieve professional-level results: AI automatically generates the questionnaire, analyzes the data, and finally generates the report model. Another more typical scenario: if the boss says to hold an online focus group, we use an AI-empowered online focus group tool called "ezTalk." It lets bosses simultaneously hold online focus groups with 100 people across the country for communication. During the process, the various messages these 100 people flood in with are automatically categorized, analyzed, and statistically processed by AI. AI can even, based on the current viewpoints, automatically draft the next question for the boss, achieving rapid insight to aid decision-making. This is probably our more typical user scenario.

Zhao Liang (Abner): Many traditional brand companies have said they want to combine AI with their business. From your perspective, are they now willing to pay a premium for AI capability, or is it essentially for cost reduction and efficiency improvement?

Li Shaohui: I think both. All enterprises face cost reduction and efficiency improvement, and they definitely hope to reduce some costs. But because we do market insight, essentially we provide the basis for bosses to make decisions. The boss now spends tens of thousands of yuan to make decisions for subsequent investments of five million, ten million, or even a hundred million. So for him, through AI and digital technology being able to more accurately and professionally help him solve decision-making problems may be more important. In our industry, giving a questionnaire or a report doesn't count as delivering an outcome. The real outcome is: can this thing truly provide a reliable, confident decision for your business decisions? This is actually quite difficult. Tokens provide the串联 capability for professional automation, but what's more important behind it is the data and knowledge base we've accumulated in vertical fields—the cost spent there may be much higher than what we spend on tokens. If you really charged by token, every company would go bankrupt. So as Xinyi said, making the same tofu, I can sell it for 10 yuan or 100 yuan. Tokens have one price, but the effort, expertise, and investment I put into creating that outcome is the core to consider in pricing.

Zhao Liang (Abner): So is Kuaicece a company that sells tools or sells outcomes?

Li Shaohui: We provide two types of services. One is very clear pure service—you make a request, and I handle everything for you. The other is selling SaaS tool software—for example, some leading beverage brands or CHAGEE (霸王茶姬), etc., who have professional capability and are willing to operate themselves. I just came back from Silicon Valley, and everyone shares a common view: the combination of service and AI may have more potential in future business opportunities than simply selling AI tools. We use AI to make services better and gain more leadership in the industry, rather than just focusing on how much labor we can cut to save money.

Zhao Liang (Abner): Next, Xinyi, could you share your company's most typical business scenario or best customer practice?

Ren Xinyi: Our current business lines can be divided into To B and To C. The To B line is very traditional—a platform for building agents, where most Chinese enterprises do private deployment, using their own or cloud computing power to build agents. What I want to share is a more interesting scenario on the C-end product "Profy," called "intelligent customs declaration form."

Shenzhen is a typical export city. Export goods need to be declared to customs, and the forms are very cumbersome—there are many different product categories, with Chinese, English, scanned copies, handwritten copies, etc. Previously, for an excellent senior customs declarant, it might take two to three hours to process all these dozens of documents. But now on Profy, using the Agent approach, you just need to upload the materials, wait for AI to process them, then do a manual check—and it's done. When you find that AI can so significantly solve real problems, this is very typical paying for outcomes.

Zhao Liang (Abner): This question may return to the OpenClaw market that everyone was discussing some time ago. When facing the C-end market, everyone cares about data security, privacy protection, compliance, and a series of issues. How do you consider these?

Ren Xinyi: We face it very formally. Just like in reality when applying for a visa, you also give your ID and asset certificates to an agency. Since you can give materials to an agency, when giving them to an AI, you should also trust that AI has boundaries—it won't secretly take your things for post-training.

Additionally, on Profy's Skill Market, we hope people with expertise will put their experience up to empower others. Here, intellectual property protection is even more important. For example, you want to teach others how to distill information, but you don't want to "teach the apprentice and starve the master." So on the platform, we add a layer of protection to your experience: others use the results of your problem-solving, but can't see your entire underlying processing logic. You retain the most core knowledge accumulation, amplify your capabilities through the platform to gain revenue, while others also get the results they want within safe boundaries.

Zhao Liang (Abner): Then we'll launch a moderator Skill on Profy later. Last question: as a company with To B genes, why make a To C product?

Ren Xinyi: Mainly two reasons. One is that last year in To B (such as State Grid, the financial industry), we accumulated tens of thousands of agents, but always faced a problem: B-end enterprises always have a lot of personalized customization, and we need to keep modifying. On the other hand, when we stripped out the business elements customized for enterprises, we found some general capabilities that shine like polished pearls—they can be reused, which is very suitable for the C-end. Plus this year, with the explosion of products like OpenClaw, everyone suddenly discovered that AI can truly help you handle complex tasks in long chains—that's why we expanded to To C.

Zhao Liang (Abner): Understood. This is actually a dynamic solving process in the AI application track. Now let's communicate with Beichuan—could you detail the historical evolution of your company's several products, and why you went from the first product to the third?

Yu Beichuan: We founded at the end of 2023, when the capital environment was very harsh. We believed AI is essentially a productivity revolution—it can do many things that previously required humans, so we decided to make "AI employees." At that time, many going-global enterprises in Shenzhen badly needed marketing, so our first track was AI influencer marketing.

But before GPT-3.5 was released, the AI application capabilities on the market were very poor (like Dify, n8n)—AI didn't have long-chain scheduling and tool-using capabilities. In 2024 (note: should be product-line iteration node), when large models gained this capability, we realized the productization opportunity had arrived, so in May we started building Laya.

The second major change point was that after doing To B AI employees in China for a long time, we found To B is very tough. The vast majority of Chinese enterprises are reluctant to pay high prices for software tools. If you challenge the past business model and tell the client this is AI software, they'll ask if you can charge less. So the path we took is called "AI + service"—this is also similar to what Sequoia Capital said: "The next trillion-dollar opportunity is AI + service." You can use AI to eliminate 95% of the work, but what the client ultimately needs is the endorsement and professional recognition of a red-circle law firm, and they're willing to pay a premium for this professional service. So we package AI as a service to sell, and the gross margin becomes larger.

When going global, Chinese teams with Chinese faces doing To B is a disadvantage—overseas clients may prefer local white companies. So we chose the PLG (Product-Led Growth) approach, opening To B through the To C market, similar to Cursor's model—letting employees spontaneously use it and then seek enterprise reimbursement.

Why did we start a new To C business this year? Because as models further evolve and coding capability rapidly rises, the vast majority of past agents and SaaS software have been wiped out. New code generation capabilities give everyone the ability to simply make software, which衍生了 new content opportunities under the new ecosystem.

The core feeling is: at the end of 2023, our team had 10 R&D people writing code. Today we have three businesses and still these 10 people—single-line efficiency is three times higher than before. You haven't actually reduced staff; the team maintains a lean scale, but the number of businesses is increasing. Because people who can use AI to the extreme are very scarce in the market, hiring has actually become harder.

Zhao Liang (Abner): Very envious of Beichuan's team—it feels like a very young team. What's your average age?

Yu Beichuan: I was born in 1996, and I'm basically the oldest in the company. There are a lot of people born in 2000 and 2002. We found that the most aggressive AI users are these young people—they feel they don't know anything, so they trust AI very much. On the contrary, people like us who are kind of "old timers," who came out after eight years of programming, think AI is nothing special. Such people have great inertia, and their efficiency with AI is actually the lowest.

Li Shaohui: Strongly agree! I'm also an old timer. Now I personally manage the product R&D team, and after launching the AI Native transformation, I flattened the organizational structure. The people I work with best now are those 22- and 23-year-olds. In AI applications, those with seven to ten years of work experience are the hardest to push—they always feel they know better, that their own methods are better, and new things can't get through.

Yu Beichuan: Right, it's particularly interesting. The talent from the previous era who did recommendation algorithms has the greatest inertia, because they feel they know a lot and disdain to research. Young people feel they don't understand, which turns out to be the best.

Zhao Liang (Abner): Mr. Li, do you envy their team's rapid iteration and quick pivoting, even being able to cross directly from To B to making To C products?

Li Shaohui: I actually feel that our vision is to change the entire global insight industry, which requires deep immersion and doing it bit by bit. If we can still do it for 10 years without being disrupted, I'd rather stick with this thing. And To C is too, too difficult.

Zhao Liang (Abner): Due to time constraints, our panel ends here. A topic like today's won't have a definitive conclusion, but hearing the three guests share their company's thinking and strategy is more helpful for everyone. We very much look forward to the next opportunity to continue communicating with all the guests. Thank you everyone!

---

Original publication: https://uniqueresearch.substack.com/p/src-20260527-02html
On-site reading page: https://ffcap.cn/en/research/src-20260527-02html