Megvii's 15th Anniversary: Zhao Kang's Internal Letter Discusses Strategic Convergence and Turning Profitable in 2026

In late October, market rumors circulated that Megvii was considering a Hong Kong listing and planned to raise $500 million to $1 billion. A Megvii spokesperson responded: "We have no specific IPO plans." It has been nearly two years since Megvii withdrew its STAR Market IPO application. Counting from its first filing in 2019, the company's long march to listing has lasted more than seven years; it has made two attempts at going public without success and has yet to cross the finish line.

But what is more worth watching at Megvii recently is not the IPO rumors, but the internal letter CEO Zhao Kang sent on October 8, the occasion of the company's fifteenth anniversary. In the letter, Zhao Kang summed up the core moves of the past year in four words: strategic convergence. Megvii is no longer trying to be a company covering every AI scenario, but is focusing on the single main lane of "physical space operations," using POA OS as the foundation to deploy agents in five scenarios: urban governance, transportation operations, business intelligence, workplace safety, and workplace management.

Translated, that means: over the past fifteen years, Megvii sold algorithmic capability; from now on, Megvii wants to sell replicable operational results. This is a shift in identity from "technology supplier" to "operations service provider," and it is the latest answer from a company that, after its IPO window repeatedly closed, is once again answering "what exactly does it rely on to make money."

Quality of the turnaround

The most closely watched information in the internal letter is Megvii's forecast that it will achieve full profitability in 2026.

Supporting this expectation are several growth figures: in the first half of 2026, total model releases are expected to grow 180% year on year compared with the same period in 2025; contract value in the business intelligence field is expected to grow more than 100% year on year; and both the share of agent-related revenue and gross margin have risen substantially. According to the internal letter, technical R&D staff still account for more than 60% of Megvii.

But this growth is occurring on a very small base. Megvii accumulated losses of 14.634 billion yuan from 2018 through the first half of 2021, and as of the end of June 2021 its accumulated unrecovered losses reached 16.596 billion yuan. Megvii has not publicly disclosed complete financial data since then, but its headcount has been sharply reduced and its business lines significantly contracted. The quality of the turnaround needs to be understood against this backdrop of "slimming down first, then generating blood."

The four measures Zhao Kang proposed all point to the same root problem. The direction must not waver: resources should be concentrated on the five major scenarios, without chasing short-term hot spots. The approach must converge: first break through replicable benchmarks, then talk about scale, and every project must answer "can it be replicated, can it be renewed, can the customer proactively say you should do the next scenario too." The organization must mesh: R&D, product, delivery, and sales must work in coordination, and the evaluation standard must shift from "who signed the deal" to "who made the customer successful and replicated it." Incentives should tilt toward teams that can deliver agent results.

An industry figure close to Megvii gave 36Kr a blunt interpretation: the core of these four requirements is the transition from project-based delivery to replicable product and service revenue. "Customized delivery for government and enterprise clients has always been the bottleneck in AI companies' commercialization. What Megvii wants to solve now is making every project a reusable asset."

The replication problem

"Replicable" is exactly what Megvii has lacked most over the past fifteen years. The common predicament of the "four AI dragons" lies in the diseconomies of scale of project-based delivery: every customer requires fresh development, payment cycles are long, gross margin is split with integrators, and the bigger the business gets, the more people it needs—making it essentially closer to an engineering contractor than a product company. Megvii's "strategic convergence" is essentially an attempt, after acknowledging that this model does not work, to use standardized hardware, standardized systems, and standardized applications to turn the business from "take on a project and earn a fee" into "build a product and charge many times."

The problem is that the balance between standardization and customization is a difficult issue facing every government and enterprise AI company. Among the five scenarios, urban governance, transportation operations, and workplace safety involve highly dispersed customers with vastly different needs, so the room for standardization is naturally limited. Business intelligence and workplace management are relatively closer to standardized products, but competition in these two tracks is also the most crowded. Whether Megvii can find the intersection between "replicable" and "able to satisfy customers" is the key to whether the turnaround can be sustained.

Founders diverge

Another backdrop to Megvii's fifteenth anniversary is the complete dispersal of its founding team.

Co-founder Yin Qi no longer holds any position at Megvii. In 2024 he stepped down as legal representative, and over the following two years he served as chairman of Qianli Technology and StepFun, one making cars and the other building large models. In January 2026, StepFun completed a B+ round of more than 5 billion yuan, and Yin Qi formally became chairman, responsible for advancing the commercialization of large AI models on devices. Another co-founder, Tang Wenbin, founded the embodied intelligence company Yuanli Lingji in March 2025, with a core team drawn from Megvii's original staff; in June 2026 it acquired a logistics robotics company and received investment from companies including Zhipu, SenseTime, and StepFun. Co-founder Yang Mu also, in January 2025, together with Yin Qi and Tang Wenbin, pledged all of his Megvii equity to the Chongqing Industrial Investment Mother Fund, which was interpreted by outsiders as the founding team's complete farewell to their old company.

Megvii was once the most distinctly technology-branded of the "four AI dragons"—born from Tsinghua's Yao Class, the Face++ platform, and a high market share in Android face recognition solutions. But the technology label did not translate into a commercial closed loop. According to media reports, Yin Qi said after leaving: "Technology can change the world, but it cannot change a company. The highest priority consideration for a company is always the commercial closed loop."

Placed in the context of Megvii's fifteenth anniversary, this statement carries great weight. It is not a rejection of technology, but an acknowledgment that: the true moat of an AI company lies not in the algorithm itself, but in how deeply the algorithm is embedded in the customer's workflow. Algorithms will be open-sourced, caught up with, and integrated into larger platforms. What truly retains customers is who controls the interface that customers must open every day—only they have pricing power.

Megvii's current POA OS and five major scenarios are, in essence, an attempt to become that "interface." But on this path, Yin Qi and Tang Wenbin have already chosen not to walk within Megvii's system. The person who stayed to walk this path is Zhao Kang.

The Business Test

When it was founded in 2011, Megvii represented the belief that "algorithms can change the world." When it first sprinted toward an IPO in 2019, it represented capital's frenzy for AI unicorns. When it withdrew its STAR Market application in 2024, it represented the predicament of the project-based AI business model. By its fifteenth anniversary in 2026, it represents a more plain and also harder-to-answer question: can an AI company that has raised billions of yuan, lost tens of billions, and twice failed to go public first survive on a small, replicable business?

One sentence in Zhao Kang's internal letter is the most substantive in the entire letter: "Good technology is a ticket to entry; meshing technology with scenarios so that customers are willing to renew and revenue is sustainable—that is real capability."

That statement is not wrong. But it also shows that Megvii spent fifteen years before clearly distinguishing between a "ticket to entry" and "real capability." If turning a profit can be delivered in 2026, that is an observable milestone—it proves Megvii has the ability to keep costs below revenue. But from turning a profit to sustainable profit, there is still a harder threshold in between: whether each of the five major scenarios can truly produce a replicable benchmark, and get customers to proactively renew and proactively expand scenarios after the benchmark is implemented.

Megvii is no longer waiting for an IPO. It is waiting for something more basic: to prove that it is a company that can consistently make money, not just a company that once had cool technology.