Source | Bohu Finance (bohuFN)

On September 21, Zhipu's AI coding tool, ZCode, drew attention from developers due to data processing issues. Subsequently, Zhipu announced the completion of rectifications, made ZCode open-source, and invited third-party institutions to conduct security audits.

The timing of the incident is more sensitive than the incident itself.

On September 13, Zhipu unveiled a financing arrangement of approximately $5 billion, equivalent to around RMB 33.5 billion, comprising roughly $2 billion in share placements and about $3 billion in zero-coupon convertible bonds. Around 60% of the raised funds will be allocated to next-generation GLM, a 'fully self-trained' system, and related computing infrastructure.

According to the company, the next-generation GLM will be trained within the environment built by its predecessor, gradually forming recursive self-improvement, known as RSI. Previously, human researchers trained models; now, Zhipu hopes the model will participate in training its own successor.

This means that the expectations bought by capital have advanced another step: a company selling intelligence is beginning to improve the way intelligence is produced.

Behind this vision lies Zhipu's not-so-relaxed financial stance.

By the end of August, the net proceeds of HKD 4.896 billion from Zhipu's IPO had been fully utilized as planned; of the HKD 31.375 billion raised through placement in July, HKD 10.955 billion had been used, leaving HKD 20.420 billion. The company still has capital reserves, but the pace of investment is remarkable.

Zhipu's rationale is to secure computing power supply in advance. Cluster construction, network equipment, and long-term computing capacity may all require substantial prepayments. These expenditures correspond to future R&D and service capabilities, with cash flow pressures preceding some costs entering the income statement.

However, current commercial revenue cannot sustain the company's expansion rate.

In the first half of 2026, Zhipu's revenue reached RMB 954 million, up nearly 400% year-on-year, with a net loss of approximately RMB 2.07 billion. Revenue is growing rapidly, but the money earned from selling models today is still not on the same scale as the capital needed for the next phase.

As technology gradually converges, large models are essentially manufacturing under their shell (outer shell), with computing power and data as the main dimensions of competition. Zhipu needs to prove its valuation through technological pursuit.

01 The Limit of RSI

The reason RSI is expensive lies first in its potential to alter the cost curve of large models.

Currently, with each step forward in models, companies often need to add more data, researchers, and computing power. Even with continuous improvements in training methods, keeping up with the frontier remains a capital race with escalating stakes. RSI attempts to open another possibility: stronger models enhance R&D efficiency, which in turn leads to even stronger models, allowing previous investments to continue participating in the next round of production.

If this cycle is strong enough, what companies gain is not just a better-selling product but a capability to produce the next generation faster. Competitive advantages may shift from 'being stronger than you this time' to 'improving faster than you.' The imagination for valuation is thus stretched.

The core of this logic is that improvements can accumulate: the time and computing power saved in this round continue to enhance the output of the next round of R&D.

Anthropic disclosed on September 17 that as of August, around 30,000 intelligent agents were simultaneously engaged in research and engineering work on its main internal platform; according to the company's published R&D automation metrics, Claude has already led 26% of R&D tasks under human supervision. AI's participation in manufacturing the next generation of AI has entered the daily production of leading laboratories.

If AI can autonomously design experiments, validate results, and reuse effective experiences, companies can continuously reduce costs in these areas, forming cross-generational advantages.

Zhipu's newly announced Infra Agent has advanced this to production systems.

According to its technical report on September 17, GLM-5.3-driven agents participated in building and optimizing the inference infrastructure for GLM-5.3-Flash, serving a cluster of over 100,000 domestically produced AI acceleration cards. From initial adaptation to production readiness, it took less than two weeks, with end-to-end throughput increasing to about three times the initial level.

However, from an industry perspective, the standout RSI achievements so far all come from one paradigm: program evolution search with LLM proposals + external verifier scoring—in essence, these AI systems claiming recursive self-improvement still rely on human researchers to propose research directions and conduct reviews;

According to information disclosed in Zhipu's financial reports, its RSI roadmap roughly includes: automatic data generation, task environment setup, long-range reasoning, and self-validation.

It belongs to the same technical lineage as Self-Instruct, STaR, self-play, self-reward, and synthetic data reinforcement learning, but is more systematic. However, overall, the outer framework of this chain is still set by humans:

This more closely resembles an 'AI-highly-automated model factory' rather than AI autonomously designing its own successor.

Given this, can Zhipu's high valuation still support the breakthrough narrative of RSI?

02 Zhipu's Leverage

To assess whether this valuation is reasonable, we must first see if Zhipu can establish advantages in domestic competition.

Competing alongside Zhipu are DeepSeek and Kimi. All three are enhancing AI's ability to participate in complex work, but their investment focuses differ.

DeepSeek advances efficiency alongside its Agent operation framework: its V4.1-Flash, released in September, enhances text, vision, and Agent capabilities, while Harness opens up infrastructure for tool access and task execution.

Kimi's K3 emphasizes long-range task capabilities. It adopts the KDA and Attention Residuals architecture, being one of the few domestic models approaching Opus 4.8. Maintaining context over long periods, calling tools, and completing engineering tasks are precisely the underlying capabilities needed for research agents.

Zhipu's technological leverage lies in the integration of GLM-5.3 with domestic computing infrastructure: Infra Agent has already provided engineering records of inference system optimization.

Based on publicly available results, DeepSeek is advancing operation frameworks and efficiency, Kimi is challenging long-range task capabilities, and Zhipu is converting model capabilities into computing system outputs. All three have verifiable progress, with Zhipu leading in financing scale relative to its established technological advantages.

Technological positions will shift with the next iteration. Single-time benchmark records reflect current capabilities; costs, computing stacks, and iteration speeds determine how many experiments companies can conduct and how much effective improvement they can accumulate thereafter. Whoever can sustain this accumulation for several rounds has a better chance of pulling ahead.

Successful financing buys Zhipu time; technological leadership requires it to deliver higher R&D output than its competitors during this period.

Zhipu's already realized advantages are the commercial scale and financing channels that support sustained investment.

In the first half of the year, Zhipu's open platform and API business generated RMB 825 million in revenue, accounting for 86.5% of total revenue, with the gross margin for this business rising to 24.6%. This indicates that the company has clearly shifted towards providing services based on usage and has started to retain gross profits from this business. Sufficient funds improve Zhipu's chances of survival, giving the team the opportunity to continue experimenting after failures;

And this ability to sustain investment coincides with a tightening external technology supply.

On June 12, Anthropic announced that the U.S. government had requested a suspension of access to Fable 5 and Mythos 5 for foreign nationals inside and outside the country, including foreign employees;

From this point onward, the circulation of cutting-edge AI capabilities began to be directly constrained by geopolitical factors.

Typically, after a technology becomes widespread, as imitators gradually increase, the excess profits of leading enterprises are eroded by competition. However, now, restrictions on accessing cutting-edge models, anti-distillation measures, and geopolitical barriers have intercepted some low-cost pursuit paths and delayed this process. Domestic enterprises capable of continuously providing high-level models have thus gained a scarcer position.

For domestic vendors, the higher the threshold for obtaining feedback from external high-level models, the more important their own data, validation environments, and R&D systems become. Substitution demand therefore grows, and domestic competitors will continue to vie for this demand.

Zhipu thus faces a market segmented by policy boundaries.

Its realistic path to excess profits is to first establish sustainable cost, capability, or delivery advantages among domestic peers.

This is also why capital is betting on it: beyond the global technological gap, there is still an undecided competition within the Chinese market.

03 Epilogue

The key to this competition also lies in the division of labor within intelligent production itself.

Currently, the R&D of frontier models still requires substantial external capital support, but on the other hand, the price reductions of open-source models are continuously compressing commercial returns.

Under such circumstances, companies must not only purchase GPUs and hardware but also maintain research teams; whether RSI first replaces part of these costs determines how these two streams of money function.

From the perspective of the organic composition of capital, as the computing power and automation tools controlled by each researcher increase, the efficiency of the entire R&D system will also change accordingly.

Currently, the most easily automated tasks are repeatable ones such as coding, running experiments, and screening data;

However, posing valuable questions, judging routes, and explaining anomalies still heavily rely on researchers.

The more machines there are, the more the researcher's judgment influences the output of the entire system.

Therefore, at this stage, RSI first amplifies the capability differences among research teams. Choosing the right direction, an excellent researcher can mobilize a large number of intelligent agents to explore in parallel; choosing the wrong direction, even a large cluster may only accelerate ineffective experiments.

Thus, DeepSeek's efficiency engineering accumulation and Kimi's model and long-range task capabilities remain substantial assets in R&D competition, enabling investments to translate into more effective exploration.

Zhipu's opportunity lies in using its financing advantage to expand this amplification effect: stabilize the research team, secure computing power supply, and replicate local gains from Infra Agent-like initiatives to more R&D links. During the current phase where human organization still dominates research, funds can indeed buy more time for trial and error, allowing a temporarily lagging route to continue iterating.

This also explains why Zhipu needs the RSI narrative: when the profits from old-generation products are insufficient to cover next-generation R&D, the capital market becomes the bridge connecting the two rounds of investment;

And before the RSI flywheel truly drives automated research, whoever loses financing capability first may be forced to exit before technological realization.

Zhipu has secured a longer R&D cycle, with financial resilience becoming its competitive edge.

However, as computing power, equipment, and automation systems continue to expand, the pressure for capital returns will also accumulate. According to the logic of the organic composition of capital, enterprises must use higher productivity and greater sales volumes to digest escalating investments; simply expanding machine scale may instead widen the gap between revenue and costs.

Therefore, the most valuable outcome of RSI is to enable the same amount of capital to support more effective R&D and turn results into sellable capabilities.

At this stage, while Zhipu is better positioned to stay in the game, the eventual winner of Be the first to achieve (first to achieve) RSI will still emerge from scientific research organization and technical efficiency. RMB 33.5 billion has bought time for continued competition; but turning time into leadership ultimately depends on the outcomes jointly delivered by humans and machines.

The cover image and illustrations of this article are owned by their respective copyright holders. If the copyright owners believe that their works are not suitable for public browsing or should not be used without compensation, please contact us promptly, and our platform will make corrections immediately.