Summary

The value anchor for AI To B is shifting from hard-to-measure “management systems” to “digital employees” whose output can be reconciled directly against the books. 翟星吉 draws a sharp distinction: SaaS restores a company from RMB800M in potential output lost to management friction back to RMB1B, while an Agent should take a company that could previously reach RMB1B to RMB2B; that is why 宇核科技 says, “We’re not selling tools—we’re selling people,” and focuses only on use cases with clearly measurable revenue gains or cost savings.
Both companies have already achieved commercial results, but their scaling paths differ. 毕昇 generated about RMB30M in revenue and RMB3M in profit last year with a 50-person team and virtually no financial funding; 宇核科技 began commercialization last April and expects roughly RMB10M in revenue this year, up about 500%, with a 20-person team whose cash flow has been healthy since last October.
宇核科技 actively rejects orders above RMB1M and sees RMB300K–500K as the sweet spot for finding PMF. Large contracts can quickly build RMB20M–30M in revenue, but they also create dependence on major accounts, collection risk and pressure to customize, while reducing the number of business scenarios the team can encounter; 翟星吉 warns, “You’ll get absorbed in your little world and start believing it is the future.”
毕昇 chose an open-source general-purpose platform, acknowledging that its net margin may be only around 10%, but views it as a self-funding “first-stage rocket,” not a transitional phase destined to be eliminated. Open source shortens the long enterprise feedback loop, brings customers to the company and lowers marketing costs; to tell a larger growth story, however, the team still needs to raise its level of standardization.
The most effective sales method is not relationship management but “insight-led selling.” 宇核 gets leads through FanRuan and ecosystem partners, then explains how Agents should be deployed and which workflows deserve to be redesigned; after the first conversation, roughly 70%–80% of customers tend to choose the company. 毕昇 similarly screens for customers who have already tried and approved the product and whose phase-one objectives are aligned: “You have to be more professional than the customer.”
The two case studies show that companies are willing to pay not for “large-model content,” but for changes in business cycles and decision quality. 宇核 reduced a ship-repair quotation for COSCO Shipping Heavy Industry that once required a senior operations representative to process for a week to roughly half an hour, charging RMB300K–500K per project; 毕昇 turned weekly PPT reporting across more than a dozen business lines at a commodities group into a daily operating view, with phase one costing several hundred thousand yuan and later phases potentially reaching RMB2M–3M and RMB5M–6M.
AI anxiety on the demand side and the retreat of capital on the supply side have jointly improved China’s To B market. CEOs and chairmen are pushing the work from the top down, reducing resistance to connecting data across departments; at the same time, capital is no longer subsidizing vendors to run loss-making projects, forcing founders back toward cash flow, products and real value. 覃睿 believes that investors’ temporary restraint “is actually good for To B.”

Deep dive

1. China’s To B market is still tough, but it no longer needs Moutai to survive

翟星吉 does not sugarcoat the industry: “It really is very difficult.” But building a consumer product, running operations or providing enterprise services is difficult too. The real question is whether you are building a project organization dependent on entertaining executives, personal connections and endless customization.
His boundary is specific: once an individual project reaches RMB3M–5M, product refinement accounts for a shrinking share of the overall lifecycle. The team has to navigate a huge decision chain and maintain complicated customer relationships, ultimately making scale impossible and forcing employees to play the role of customer flatterers for years.
翟星吉 believes the previous generation of domestic To B got its value hierarchy wrong: resource-oriented operators could win deals through customer relationships, while products and technology took a back seat. That was not a permanent fate, but a stage in China’s shift from resource-driven growth to technology and innovation.
Asked about the old-era example of hauling a case of Moutai around in a car trunk, 翟星吉 replied, “I don’t drink. I’ve never drunk.” 覃睿 said customers are wary even of ordinary business dinners. Neither sees drinking sessions as a mechanism for closing deals.

2. Both business models have already achieved commercial traction

宇核科技 provides off-the-shelf Agent digital employees for mid- to high-end manufacturing. Its current products cover presales quoting, pre-production quality-issue handling and supply-chain document processing. 翟星吉 puts it plainly: “Hire them, give them basic training, and they can go straight to work.”
The company began commercialization last April, generated “a few million yuan” in revenue last year and expects about RMB10M this year, representing roughly 500% year-on-year growth. Its 20-person team has had relatively healthy cash flow since last October. 奇绩创坛 invested in the seed round; the team is considering another round, but fundraising is not a high priority.
毕昇 is an open-source LLM application-development platform for enterprise use cases, which 覃睿 likens to “an open-source version of Coze.” The company generated about RMB30M in revenue and RMB3M in profit last year with a 50-person team.
毕昇 initially acquired equity through intellectual-property arrangements related to 第四范式, while 华泰 invested a very small amount because of procurement-service requirements. Beyond that, the company did not actively seek extensive financial funding. 覃睿 believes To B does not necessarily require huge amounts of capital, and that a small team can still produce relatively predictable results.

3. 宇核 started with the productivity lever and ultimately bet on high-end manufacturing

翟星吉 initially understood AI as a productivity lever, whose value depends on whom it serves. The team started with To C tools in 2023 and quickly accumulated users during the GPT boom, but found that ordinary users had limited productivity value. Enterprises, by contrast, are “huge organizations that exist to create value at scale,” where even small efficiency gains can produce substantial value.
He once believed To C was uncertain and To B more predictable. He now revises that view: both can be highly predictable. “Predictability comes from the DNA of the team.” 宇核’s core members came from FanRuan and have deep experience taking an organization from 1 to 100, giving the team a stronger methodological edge in enterprise services.
The team once debated between functional entry points such as SDR and marketing, and an industry-led approach focused on manufacturing. The deciding question was where expert knowledge resides: sales methodology is similar across industries and suits a functional approach, while roles in semiconductors and the automotive supply chain require heavier industry knowledge and should be approached by sector.
The team ultimately abandoned its healthcare validation. Large companies were willing to conduct free POCs in exchange for use cases and logos, which startups could not match; incumbent HIS vendors also had deep ecosystem barriers, while product and technology innovation carried too little weight in the decision chain. Manufacturing was both the team’s chosen “backbone of China’s economy” and a sector with many operational white-collar roles that had not yet been automated. Early co-development also showed that the use cases could be replicated.

4. 毕昇 had no grand starting point; open source emerged as an iterative market strategy

覃睿 admits that 毕昇 did not begin with a dream or an unshakable conviction about its direction: “Entrepreneurship is hard to plan.” The team originally worked on OCR, NLP and intelligent documents. When LLMs appeared and it saw that the ceiling of its old business was low, it decided to enter a larger market.
In the first half of 2023, the team experimented with applications aimed more at consumers, but the technology was not mature and capital was concentrated in models, making application startups difficult to build. It returned to the To B market it knew best. The real debate was not whether to build a platform, but how to take it to market and whether to open source it.
The team never expected large numbers of domestic developers to contribute code. Open source was first about creating more value from the product it had already built, and second about shortening the long enterprise-software feedback loop: “Once I open-source it, I immediately get positive feedback, and then I can actively push for improvements.”
Open source also brought visibility and low customer-acquisition costs. Enterprise customers came to the company directly, and 毕昇 had previously spent almost nothing on marketing. 覃睿 believes the company is already recognized among enterprise customers and software developers, even if it remains less prominent in mass media.

5. General platforms and vertical Agents are not mutually exclusive

Asked whether general-purpose platforms are merely a historical stage, 覃睿 rejects the premise. He calls the platform a “first-stage rocket”: enterprise software development alone is, in his estimate, a trillion-yuan market. Even a general platform that does not target specific industries or job functions can form a substantial business.
The trade-off is heavier delivery work and potentially lower overall net margins; 覃睿 says net margin might be around 10%. Raising the ceiling and improving operating efficiency will require further standardization. 毕昇 already uses a pricing bot that analyzes which parts of a customer’s requirements can or cannot be delivered and where the risks lie, then generates a reference report for presales staff.
翟星吉 viewed the platform as transitional from the start, with the goal of delivering Agent business outcomes end to end. The routes are not entirely opposed at the destination: one first uses a platform to generate cash and then increases standardization, while the other starts with manufacturing roles and then tries to abstract a method for building vertical Agents.

6. 毕昇 turned weekly reporting into a daily operating system

覃睿’s representative project came from the commodities sector, centered on market intelligence and the integration of business and financial operations. The customer previously heard one weekly briefing, with each business unit preparing PPTs and presenting in meetings. After launch, executives could see the latest information from the previous day on a large office display, and all more than a dozen business lines could access it.
One end of the system uses models to govern large volumes of unstructured information; the other connects traditional business data previously scattered across departments. The result is not generic Q&A, but support for operating decisions such as procurement prices, shipment contract value and volume, which port a logistics vessel is calling at, and which factory should receive priority supply.
Phase one cost several hundred thousand yuan, potentially toward the high end of that range. Phase two could reach RMB2M–3M, and after further proof of business value the project could expand to RMB5M–6M. The customer had previously hired several major consulting firms, which charged consulting fees before concluding that the data was incomplete and the work could not be done. 毕昇 stepped in when the project was close to collapse and completed it, though 覃睿 stresses that this is not the company’s most typical acquisition path.

7. 宇核 rebuilt a week-long ship-repair quote in half an hour

宇核 delivered a presales-quoting Agent to COSCO Shipping Heavy Industry. COSCO Shipping Group is the world’s second-largest container shipping company, while COSCO Shipping Heavy Industry handles ship repairs for both the group’s internal shipowners and external customers.
When a shipowner sends hundreds of pages of repair requirements in PDF, Word or Excel format, an operations representative must traditionally break them down by categories such as dry dock and deck, then match them against thousands to hundreds of thousands of quotation SKUs while consulting special agreements and historical quotes.
The old process took a mid- to senior-level employee a week. It was expensive and repetitive, and manual transcription created ample room for error. Junior employees could not handle it, while senior staff spent large amounts of time on low-value work. Customers, meanwhile, were unhappy with the long quoting cycle.
宇核 uses a proprietary small model to parse and extract irregular requirements, after which an Agent generates the result in roughly half an hour. 翟星吉 makes no unrealistic promise of 100% accuracy: “We only need accuracy higher than that of an employee at the same level.” The standardized project costs RMB300K–500K.

8. The next generation of sales wins through insight—and screens customers proactively

覃睿 recalls that 毕昇 once chased customers for deals in its early days, only later realizing that screening matters more than coverage. The ideal customer comes in proactively, has already tried and approved the product, and agrees with the company on the goals and expectations for phase one.
宇核 initially relied on FanRuan for enough leads, then built an ecosystem-partner network through its brand, products and core technology. 翟星吉 says the lead pool is now healthy enough that “we choose which customers to reply to; a customer’s message does not necessarily get a response.”
The actual conversion method is what he calls “insight-led selling”: explain in the first meeting how an Agent should be deployed, which scenarios the customer should prioritize redesigning, and where 宇核’s products and technical capabilities fit. After that round of output, roughly 70%–80% of customers lean toward choosing 宇核; only then does the commercial process begin.
覃睿 fully endorses the approach: “If you want a newer, healthier To B market, this is the model. You have to be more professional than the customer.” Vendors are selling more than execution capacity; they are selling business judgment formed earlier than the customer’s own.

9. An Agent is not just a new interface over every SaaS product

On the idea that every SaaS product deserves to be rebuilt with an Agent, 覃睿 gives a restrained answer. Coding, customer service, marketing and content generation will be heavily affected, but a sector-wide revolution has not yet emerged in more business-specific settings. He calls himself a “rational optimist,” continually checking whether new technology can solve scenarios that were previously handled poorly rather than assuming the answer in advance.
翟星吉’s view is more fundamental: SaaS and Agents are “two completely different things.” SaaS uses process standardization to prevent an organization’s actions from degrading as it expands from 1 person to 10,000, restoring the capability lost when management friction takes output from RMB1B to RMB800M. An Agent is more like hiring a senior employee or building a new production line: it adds organizational capabilities the company did not previously have.
宇核 therefore benchmarks against the entire labor market: every company and business process could be rebuilt with Agents. 翟星吉 agrees with Silicon Valley’s “$6B human-resources market” narrative, but adds a filter: only roles that can directly generate more revenue or save more money, with clearly calculable ROI, are worth building.

10. The “aha moment” came from repeatable wins, not one large order

宇核’s turning point came in the second half of last year. The team attended the WAIC World Artificial Intelligence Conference and, after walking through the entire exhibition, formed a broad impression: “What everyone else was doing seemed off, while what we were doing seemed right.”
More important was the micro-level evidence: after the team delivered an Agent plan to a new customer, it could win roughly 70%–80% conviction and then jointly define the real use case. The customer was ultimately paying for the core Agent business value 宇核 had wanted to sell from the beginning—not business relationships, customer access or custom development. The same core could be taken to the next customer, which was the PMF signal.
毕昇’s signal came from customers who approached it directly. They said the product was reliable, the team was dependable and they “wanted to support open source in China.” Once the value proposition aligned, internal rules and budget constraints stopped being pure obstacles; the customer-side contact instead helped the vendor push the process forward.

11. Younger customers and retreating capital are repairing the market together

Drawing on his experience from the Yitu era, 覃睿 believes the previous wave of AI To B did not lack business value; its competitors were simply too strong. Hikvision had integrated hardware and software capabilities and a nationwide sales network, making it difficult for startups to compete head-on. The change today starts with younger internal decision influencers who are “full of ideas, capable and ambitious,” and more willing to push procurement once the value aligns.
The supply side is changing too. 覃睿 says the new generation of founders has “figured things out” and is no longer willing to execute unlimited projects for customers. Some strong teams have gone overseas, reducing domestic competition; more importantly, a deteriorating capital environment has produced “more people who are actually running businesses.”
When capital stops subsidizing vendors to provide loss-making services, customers can no longer find so many teams willing to accompany them at a loss. 覃睿 says investors once demanded growth rates far above the natural pace of To B, forcing startups to distort their behavior and leaving their foundations unstable. That is one of the core reasons 毕昇 has not raised large amounts of capital.
翟星吉 attributes the ROI boundary to the founder’s values, not to a particular failure. He gave up a high income to start a company after the emergence of LLMs because he believed this was a productivity revolution: “If I’m only making money from it, it has no meaning to me.”

12. A RMB1M order is not a trophy; it may be where product focus begins to slip

翟星吉 insists that 宇核 will not accept orders above RMB1M. A super-KA deal can easily push revenue to RMB10M–30M, but it also creates dependence on major customers and causes limited product resources to follow customer budgets rather than core value: “When a pile of money is sitting in front of you, do you take it? If you do, your product loses focus.”
The more hidden risk is losing scenario diversity. A small team can take on only a limited number of projects. If several giant customers consume all its capacity, it cannot encounter enough variations in business needs. While continuously searching for and iterating on PMF, “you’ll get absorbed in your little world and start believing it is the future.”
Large customers also bring slower collections, cash-flow volatility and margins that are not necessarily higher. 宇核 prefers the RMB300K–500K range: a standardized product with room for limited configuration services, shorter sales and delivery cycles, and the ability to replicate best practices quickly.
The host asks whether Chinese customers are more willing to pay for outcomes than efficiency, and whether light delivery can really work. 翟星吉’s answer is that an Agent should deliver outcomes directly. Taking projects to support the team in the early days is fine, but the company must always remember that projects are meant to iterate the product; otherwise it is easy to lose the plot amid the excitement of closing a RMB2M deal and expanding the budget to RMB10M.

13. Differentiation comes from customer selection, product boundaries and business model

翟星吉 believes early markets, whether To B or To C, are won through innovation: whoever understands users better, solves their problems more effectively, gets paid and can replicate the result will lead. 宇核 is therefore expanding its product team rather than its delivery team. It focuses only on core Agent business modules and will not touch adjacent system integration, even when it could generate substantial revenue.
毕昇 divides its competitors into major tech companies and open-source startups such as Dify and FastGPT. 覃睿 believes “the customer you serve determines who you are”: other open-source platforms serve more overseas users or individual developers, while 毕昇 has spent years immersed in domestic enterprise customers and has built security, single sign-on, traffic control and other mundane but essential launch requirements into the product.
On delivery, 毕昇 not only provides tools but also wants to help customers create business value around use cases such as drilling reports, due-diligence reports and contract review, in the hope of winning phase-two orders. 覃睿 acknowledges that Chinese software still requires services, but sees that as a business-model difference rather than a flaw to conceal.
Compared with major tech companies, customers with budgets around RMB1M are unlikely to become an organizational priority. Big tech companies can also use businesses such as advertising to generate cash flow and may not want to keep doing difficult enterprise services over the long term. 覃睿 therefore believes they do not inherently possess a durable competitive advantage in 毕昇’s market.
Asked about the best To B founders, 翟星吉 also mentions Tony Gu and 赵磊, both of whom are building Sales Agents. What he values is not fundraising or order volume, but a founder’s depth of understanding of the vertical, the product and the future.

14. How founders allocate their time reveals two organizational paths

翟星吉 summarizes the CEO’s job as finding people, finding money and finding direction, but places fundraising at the bottom of the priority list. About half his time currently goes to interviews, performance reviews and coaching. Of the remainder, roughly one quarter goes to customer needs and scenarios, and one quarter to product, technology and commercial understanding, extracting the value chain, customer pain points and product abstractions from each project and POC.
He does not directly participate in specific projects or deal-making. 宇核 looks for geek-like curiosity, idealism, learning ability and resilience. Sales hiring is especially difficult because “if a salesperson isn’t realistic enough, they won’t do sales.” Yet 翟星吉 still refuses to attract core salespeople with a calculation that they can make several million yuan in a few years, hoping future regional commanders will carry forward the company’s values.
覃睿 divides his time roughly equally among marketing, customer communication, product development and strategic judgment, but directly intervenes in customer projects: diagnosing why results fell short, whether requirements are truly aligned, and how to adjust the plan with the team and customer. The two arrangements correspond to stronger product abstraction on one side and deeper project diagnosis on the other.

15. AI anxiety is forcing companies to catch up on old digitalization work

翟星吉 does not avoid customers’ fear of falling behind. He sees it as a normal sense of crisis when a new technology era arrives. Vendors once had to educate the market; now customers ask proactively what can be done, allowing 宇核 to plan scenarios that could increase revenue by 10%–20% or materially reduce labor hours.
覃睿 observes that many companies now face top-down pressure to deploy Agents. In the past, informationization was driven by CTOs and CIOs and often stalled once it crossed departmental boundaries. Now CEOs and chairmen are personally pushing the agenda, forcing business units to cooperate and reconnecting data and systems that had never been integrated.
He warns that LLMs are not the entire source of value. In some deeply business-oriented projects, the LLM accounts for only around 30%; more of the benefit comes from connecting system data and improving automation. This is common among the broad financial sector, central and state-owned enterprises, and their second- and third-tier subsidiaries that 毕昇 works with.
On talent, 覃睿 speculates that the smartest young people may prefer work with the greatest leverage and impact. 毕昇 combines the appeal of open source and LLMs and insists on building standardized products. The team is not trying to create original foundational models; it continuously absorbs practical advances from industry and academia that can produce real-world deployment value.

16. The vision still has to land in cash flow, organization and a sustainable life

毕昇’s vision is “Make Work Smart”: make enterprise work smarter and reduce inefficient labor. 宇核’s goal is to “truly use Agents to liberate human productivity,” rebuilding jobs one by one and having Agents deliver business outcomes end to end. 翟星吉 imagines that within 10 years, people may need to work only 3 or 4 days a week.
翟星吉 sees no endpoint for personal success; he wants to remain “on the road.” Once the domestic business is stable, he will go overseas; once general AI Agents are stable, he will build vertical AI Agents; then he will abstract construction methods and general products from those vertical Agents. For the company, stage-specific success means leading technology in vertical fields, the best commercial results and a comfortable life for the team.
覃睿 defines success as healthy cash flow, more time for the founder to spend with friends and family, lower pressure for colleagues and reasonable compensation. Given To B payment cycles, he estimates that another 1 or 2 years will be needed to build stronger positive cash flow. At that point, the “first-stage rocket” should be operating in a virtuous cycle, and the company can decide whether to pursue the next stage.

Oasis Capital's Zhang Jinjian on Opportunity in a World Splitting Fast Windsurf CEO & Co-Founder, Varun Mohan: AI's Biggest Acquisition to Date!