In mid-July, the company publicly acknowledged a significant issue: user demand had outpaced its computational capacity. Shortly after the launch of its K3 model, user requests surged to the limits of existing computing clusters. The firm was forced to pause new consumer subscriptions, prioritizing limited resources for existing paying users.

Ten days later, another challenge emerged. Reports indicate the company has completed an over $35 billion (approximately RMB 236 billion) Series F funding round, with a post-money valuation of $350 billion. The subscription amount exceeded the original target by more than three times, leading to an early close of the round. This level of competition for a top-tier AI project has been rare in recent years.

"The tolerance for certainty in the primary market has become extremely low. Once good allocation is released, it is absorbed very quickly, often disappearing within days," said a founding partner of a capital firm. He noted that this pace is a double-edged sword: "The window for institutions to conduct due diligence and complete internal processes is shrinking rapidly. This is a major test of investment discipline—the hotter the project, the more you must remind yourself not to be swept up by the momentum."

The market, however, appears less concerned. The imbalance between supply and demand is reflected in pricing, with far more institutions seeking entry than available allocation. Missing out could mean losing a prime opportunity, setting the stage for rapid valuation increases. There are rumors the company is preparing to launch a Pre-IPO round with a pre-money valuation of up to $500 billion. "Given the current heat, this round will likely close very quickly as well," the partner predicted.

Behind the Rapid Valuation Surge

From the end of last year to now, the company's valuation growth curve has been steep. After a Series C round in December 2025, the post-money valuation was about $4.3 billion. By early 2026, the pace accelerated significantly, with multiple rounds pushing the valuation to approximately $18 billion. In May, a Series D round pushed it past $20 billion. By the end of June, the pre-money valuation reached $31.5 billion, culminating in the July Series F at $35 billion post-money. In seven months, the valuation grew more than eightfold, from $4.3 billion to $35 billion.

"This slope cannot be explained by current profits," the partner stated. The key, he believes, is a shift in valuation benchmarks. "Large model companies are not priced linearly based on current profits, but on the probability of them entering the top tier and eventually becoming a public company. Once the market believes a company can not only continue raising funds but also has IPO potential, its pricing benchmark shifts from an ordinary startup to a potential IPO target, causing a valuation jump."

Earlier this year, strong IPO performances by peers in Hong Kong provided new valuation anchors for the primary market. Another company also saw its Pre-IPO valuation rise rapidly within a month. These cases show investors that exit channels for large model companies are genuinely forming—IPOs are no longer a distant story but a realistic expectation that can be factored into models. A new capital cycle is forming: the primary market funds model training; model capabilities and revenue growth drive the company to IPO; the secondary market establishes a valuation anchor; this anchor then raises the financing price for unlisted top-tier companies.

The fact that this round was oversubscribed by three times and closed early sends a stronger signal than the amount itself. "As the industry concentrates towards the top, the scarcity of quality leaders is rapidly increasing. This is not contradictory to the AI de-bubbling process. What is being cleared out are narrative-driven targets lacking barriers. The leaders that secure user entry points and possess foundational model capabilities are actually benefiting from this concentration," the partner explained. "Capital doesn't think this sector is risk-free, but it fears missing a key round in a leading company and never getting an ideal entry point again. If the assumption that only three or four companies will remain holds true, a $3.5 billion entry fee doesn't seem expensive in the context of an IPO."

Shifting Perspectives on the Product

The most direct catalyst for the valuation jump was the release and open-sourcing of the K3 model. This model features 2.8 trillion total parameters using a MoE sparse architecture, native support for visual multimodal understanding, and a 100 million token context window. Combined with a proprietary hybrid linear attention mechanism, overall scaling efficiency has improved approximately 2.5 times over the previous generation. On a third-party front-end coding benchmark, K3 ranked first globally, surpassing all overseas closed-source models—a first for an open-source model on this authoritative benchmark. On another intelligence index, K3 ranks third globally.

"It doesn't just push the open-source parameter scale to a new limit; more importantly, it approaches international first-tier closed-source models on high-difficulty tasks like coding, complex reasoning, agents, and long context," the partner said. Following the K3 release, the influence of Chinese models overseas saw a notable shift. The CTO of a major tech foundation began using it for daily tasks like calendar, files, and email within days of its release. Some US companies have also started adopting Chinese models to reduce API call costs. Estimates suggest that global downloads of the product surged significantly in the week after the K3 release, with notable growth in the US market.

"K3 at least proves one thing again: the company is not just about product traffic; it has the capability to continuously train frontier models," the partner stated. "Investors' perception is shifting from a long-text personal assistant to encompassing coding, agents, multimodal capabilities, and complex task execution."

As the model strengthens, revenue data is also growing rapidly. According to public information, the company's ARR first exceeded $100 million in March 2026, reached $200 million in May, and stabilized at $300 million in June. API revenue now accounts for over 70% of total revenue, as the company shifts from relying on C-end subscriptions to scaling B-end monetization. However, behind the high valuation lie long-term risks.

"The faster the valuation grows, the more future growth must be delivered. It must prove not only that the model is good, but also that users will stay, revenue will continue to grow, and ultimately form a business model matching a $50 billion valuation," the partner cautioned. "There is also external uncertainty that cannot be ignored: issues around access to advanced computing power and overseas scrutiny of training paths and compliance will directly impact its globalization and the sustainability of its exit channels. These are the real tail risks behind the high valuation."

The Evolving Business Model for Large Models

Competition in China's large model space is subtly changing. The partner's assessment is that model capability remains the entry threshold, but the next phase's true differentiator will be who can enter real production workflows and master user context. "Today, the gap in basic capabilities like general Q&A, writing, and summarization is no longer as obvious as in the early days. Leading by a few percentage points on a leaderboard may not translate into a commercial advantage. What users truly need is not a smarter chatbot, but a system that understands their work environment, business goals, historical information, and can persistently complete tasks."

Mastering context, he explains, goes beyond remembering user conversations. It means understanding their workflow, organizational relationships, data permissions, decision-making habits, and specific objectives. "Only by entering the real scenario can the model know where the problem lies and turn one-time answers into continuous task execution." He offers a more direct criterion for success: the next phase's winner won't be determined by who first gets read access to enterprise data, but who first gets write access. Read access allows a model to read emails, query databases, and summarize files. Write access means it can modify databases, send emails, create orders, adjust inventory, and embed itself into core production processes. Only with write access does a model truly enter a company's production chain.

Recent advancements in model context protocols also confirm this trend. New versions focus on solving stateless deployment, long-duration tasks, enterprise authorization, and large-scale scaling, making it easier for agents to connect to external tools and enterprise systems. Authorization specifications require standardized authorization, operation logging, and user permission when agents access emails, files, databases, or perform administrative operations.

Previously, the large model industry's infrastructure was GPUs, data, and training frameworks. In the future, permission management, task orchestration, model routing, identity systems, and audit tools could become new infrastructure. Simultaneously, model competition is expanding from text to complete multimodal tasks. Three concurrent branches are emerging: foundational models continue to scale and improve reasoning; agents enter coding, office, and enterprise workflows; and multimodal models enter advertising, e-commerce, gaming, and content production.

The company's focus is on long-cycle coding, knowledge work, and personal intelligent entry points. A competitor is betting on multimodal content and overseas consumer products. Another is closer to enterprise and government clients. Tech giants leverage their cloud computing, e-commerce, office, advertising, and content distribution scenarios. Another player continues to rely on model efficiency and developer influence to expand its ecosystem.

The Chinese large model industry once focused competition on the cost per million tokens. However, with the rise of agents, simply comparing token prices is losing meaning. "Token is fundamentally not a standardized commodity. Even calling the same model, different providers have varying latency, throughput, concurrency, stability, and caching mechanisms, leading to different prices. For enterprise users, whether a single call returns in time, is stable during peak hours, and can complete a complex task in one go is often more important than the nominal per-million-token price," the partner said.

Demand itself is also diverging: massive text classification and basic customer service require ultra-low costs; coding, scientific research, financial analysis, and enterprise decisions value reasoning quality and reliability; real-time interaction scenarios are extremely sensitive to latency. "Even the same token creates completely different values and cannot be simply compared on price. Users never buy tokens themselves; they buy tokens that solve problems."

In his view, the business model for large models will gradually evolve into: basic call prices continue to decline, but high-quality reasoning, low-latency services, dedicated deployment, industry data, agents, and task delivery will be tier-priced. When the market shifts from asking "how much per million tokens" to "how much a final task costs, how much labor it saves, and what value it creates," pure price wars will naturally lose their meaning.

This content is for reference only and does not constitute any investment advice.