Zhipu Stock Selloff Deepens as the AI Price War Tests Its Valuation

Zhipu shares touched another adjustment-period low on September 25, taking the company’s reported market value back toward HK$300 billion. The Zhipu stock selloff now reflects more than a weak trading session. Investors are testing whether rapid revenue growth can justify heavy spending during an escalating AI model price war.

The Hong Kong-listed company, now branded Z.AI, briefly commanded a reported market value above HK$1.3 trillion in June. That peak followed intense enthusiasm for Chinese foundation-model developers and Zhipu’s GLM model family. Three months later, lower-cost releases from DeepSeek and other vendors have challenged the scarcity premium that supported that valuation.

Zhipu also completed a major financing package in September, adding new shares while issuing zero-coupon convertible bonds. The transaction strengthened its funding position, but it also increased dilution concerns. That conflict defines the current debate: Zhipu needs more capital to compete, yet raising that capital makes investors demand stronger commercial results.

The Zhipu Stock Selloff Has Erased Most of the June Rally

The latest decline matters because it completes a rapid shift from scarcity-driven valuation to evidence-driven scrutiny.

Zhipu’s shares weakened again on September 25 and reached their lowest level during the current correction. A Chinese financial report placed the company’s market capitalization near HK$300 billion during the session. The same market value reset followed a long retreat from June’s record.

The scale of that reversal is more important than one intraday quotation. Zhipu’s reported market value exceeded HK$1.3 trillion when its shares reached their June high. Returning toward HK$300 billion means most of that peak valuation has disappeared.

Different market-data services can show different totals because they may use different share classes or outstanding-share definitions. Recent capital issuance makes that distinction particularly important. Therefore, HK$300 billion should be treated as an approximate market level, not an accounting-period valuation certified by the company.

The direction remains unmistakable. Investors have sharply reduced the premium assigned to Zhipu’s position as a listed, specialized foundation-model developer.

That premium was substantial because public-market investors had few comparable choices. Zhipu began trading in Hong Kong in January and quickly became a prominent proxy for China’s generative AI sector. MiniMax offered another pure-play comparison, but both companies occupied a market with limited listed supply.

Scarcity can lift a stock faster than operating results alone would support. It can also disappear quickly when more shares enter circulation, competitors improve, or investors reconsider the path to profit.

All three forces have affected Zhipu.

First, previously restricted shares became eligible for trading during the summer. That increased the potential supply available to the market. Second, Zhipu completed placements that expanded its issued capital. Third, competing model providers kept lowering the cost of capable AI inference.

The June valuation also rested on expectations that Zhipu’s GLM models would capture meaningful developer and enterprise demand. Those expectations have not vanished. However, the stock now implies that investors want clearer evidence before restoring the former premium.

The correction did not begin with the September price war. Zhipu had already fallen sharply after its June peak. Summer lock-up expirations, placements, and broader weakness among AI-related shares had placed the stock under pressure.

September added a more fundamental question. If comparable models become cheaper across the market, can Zhipu convert rising usage into attractive margins?

That is harder than asking whether the company can grow revenue. A model provider can add customers while losing pricing power. It can increase API traffic while spending more on inference infrastructure. It can also report strong annual growth from a small starting base without approaching profitability.

Zhipu’s decline brings those distinctions into focus. The market is no longer rewarding growth, technical ambition, and listing scarcity as interchangeable signals.

The central issue is now whether growth creates durable economic value. That question leads directly to the latest round of model price reductions.

Cheaper Models Are Turning Cost Into a Competitive Weapon

The new price war weakens one of the advantages that helped Chinese model developers attract cost-conscious users.

API pricing determines what developers pay when applications send prompts to a model. A token is a small unit of text processed or generated by that model. Lower token costs can materially change the economics of coding agents, customer-service systems, and high-volume enterprise workflows.

DeepSeek intensified this competition in September with V4.1-Flash. The company described the model as a more efficient system with native visual understanding and higher throughput. Its Flash model update also introduced lower API pricing and retained cheaper off-peak usage.

DeepSeek said the model uses a mixture-of-experts design, which activates only part of its total network for each task. That approach aims to reduce the computing resources required for inference. The company also claimed substantial reductions in memory and storage requirements for its key-value cache.

Those technical claims require independent testing across real workloads. Yet the commercial signal is already clear. DeepSeek wants buyers to evaluate model capability and operating cost together.

A related price reduction notice showed substantial cuts for several DeepSeek services distributed through Tencent Cloud. The reductions covered input processing, output generation, and cache usage.

The specific rates matter less than the direction. Lower costs are spreading through several parts of the model-delivery chain.

That creates pressure for Zhipu in two ways.

The first pressure is direct. Developers can compare GLM services with DeepSeek, Alibaba’s Qwen models, and other alternatives before committing workloads. Switching remains difficult for some production systems, but standardized APIs have reduced that barrier.

The second pressure comes from customer expectations. Once a major provider lowers prices, enterprise buyers can demand concessions from other vendors. They may also divide workloads among several models to gain negotiating leverage.

This dynamic is particularly important for routine tasks. Buyers rarely need the most capable model for every request. They can route simple summarization, classification, extraction, and drafting work to cheaper systems.

Model routing makes that strategy practical. It sends each request to a model selected by cost, speed, or complexity. As routing improves, a provider cannot rely on one flagship model capturing every part of a customer’s workload.

Zhipu must therefore defend more than benchmark performance. It must show that GLM models deliver better results for specific workloads, stronger deployment options, or lower total operating costs.

Total operating cost includes more than token charges. It can include latency, error correction, engineering time, security requirements, and the cost of maintaining several providers. A model that appears cheaper can become expensive if it produces longer outputs or needs repeated attempts.

That qualification prevents the price war from becoming a simple race toward zero. Developers still care about reliability, context handling, tool use, and output quality.

However, lower headline prices influence initial evaluations. They also make it easier for procurement teams to question existing contracts.

Zhipu previously benefited from the idea that Chinese models could offer competitive performance at lower cost than leading Western services. That positioning becomes less distinctive when domestic rivals cut prices and international vendors introduce lower-cost variants.

The contest is shifting from national cost advantage to provider-level efficiency. Investors are responding because efficiency determines how much revenue survives as gross profit.

The Zhipu stock selloff therefore reflects a credible business concern. Cheaper inference can expand the overall market while compressing the value captured by each supplier.

Rapid Revenue Growth Has Not Settled the Margin Debate

Zhipu’s first-half results show strong demand, but they also show why investors remain focused on losses and capital intensity.

Zhipu reported revenue of RMB953.9 million for the six months ended June 30. That represented growth of 399.7 percent from the corresponding period in 2025.

Cloud-based deployment generated RMB825.2 million, compared with RMB29.1 million one year earlier. It accounted for 86.5 percent of first-half revenue. The shift suggests that APIs and hosted services have become the company’s primary commercial engine.

That change matters because cloud distribution can scale more quickly than customized, on-premise projects. A developer can start using an API without installing an entire model system inside its own infrastructure.

Cloud services also create recurring usage signals. Growing traffic can help a provider refine models, understand customer needs, and build deeper relationships with application developers.

Yet recurring usage does not automatically produce recurring profit.

Zhipu recorded a first-half loss of approximately RMB2.07 billion. Its adjusted loss was about RMB1.96 billion. Both figures were more than twice the company’s reported revenue.

The company’s published interim financial report also showed the scale of the shift toward cloud delivery. That shift created growth, but it did not eliminate the cost of model development and computing infrastructure.

Zhipu spent about RMB2.13 billion on research and development during the half. That expenditure exceeded total revenue for the period. It illustrates why access to capital remains central to the company’s strategy.

Foundation-model development requires expensive training runs, specialist employees, data work, and continuous evaluation. Serving users then adds inference costs. Those expenses continue even when competitors lower their own prices.

The result is a difficult timing problem.

If Zhipu protects margins by holding prices, developers can move workloads to cheaper alternatives. If it follows the market downward, revenue may grow while each unit of usage contributes less profit. If it reduces investment, its models risk falling behind.

This is why revenue growth alone has not stopped the Zhipu stock selloff. Investors need to understand the quality and durability of that growth.

Several indicators can clarify that quality.

One is cloud gross margin, which measures the portion of cloud revenue remaining after direct service costs. Another is revenue concentration. Heavy dependence on a small number of customers can make rapid growth less durable.

A third indicator is customer retention after promotional agreements end. Usage acquired through discounts carries different economic value from usage retained because a model performs a critical task.

Zhipu’s cloud gross profit improved substantially in the first half. That is an encouraging operational signal. It suggests scale and deployment efficiency can offset part of the cost burden.

Still, a price war can reverse those gains. Lower rates affect revenue immediately, while computing costs may decline more slowly.

Zhipu can counter that pressure by improving model efficiency. A smaller active parameter count, better caching, faster inference, or improved hardware utilization can reduce the cost of every request.

It can also prioritize workloads where customers value accuracy and reliability above the lowest possible price. Coding, complex agents, regulated deployments, and customized enterprise systems may offer stronger differentiation.

The company must prove those advantages in customer behavior. Announced benchmarks and technical claims cannot show whether enterprises will expand paid usage.

That is the new valuation standard. Investors are comparing growth with losses, and technical progress with the cost required to sustain it.

New Financing Gives Zhipu Time but Raises the Execution Bar

Zhipu has secured resources for another development cycle, but the added capital also increases dilution and repayment-related concerns.

In September, Zhipu completed a package involving newly issued H shares and zero-coupon convertible bonds. The transactions followed another large placement completed during the summer.

The September share placement issued approximately 21.97 million new H shares. That represented just over 9 percent of the issued share count before the transaction.

The convertible bonds had a principal amount of RMB20.14 billion. They are settled in United States dollars and mature in 2027. Zero-coupon means they do not make regular interest payments, although their conversion and maturity terms still affect shareholders.

Together, the transactions raised roughly HK$39 billion before considering final costs and currency effects. The company said most proceeds would support model research, training, inference capacity, and technical infrastructure.

That use aligns with Zhipu’s strategic challenge. Competing with DeepSeek, Qwen, and international model developers requires sustained investment. Model training cannot pause while management waits for current revenue to finance the next release.

The official financing record shows that the September package followed several other capital-market transactions during 2026. Zhipu has used its listing to secure a funding base unavailable to many private competitors.

That is a real advantage.

Capital can purchase computing capacity, attract researchers, and support inference demand before customer payments fully cover costs. It can also fund acquisitions or investments that expand distribution.

However, financing does not resolve the company’s economic model. It extends the period in which Zhipu can try to resolve it.

New shares dilute existing ownership because each investor represents a smaller percentage of the expanded company. Convertible bonds can create additional dilution if holders exchange them for equity.

The financing terms can also influence trading behavior. Investors may hedge convertible positions or compare conversion economics with the ordinary shares. That can add complexity to price movements beyond the company’s operational performance.

This matters when interpreting the September decline. It would be too simple to attribute the entire move to AI model pricing.

The stock was already absorbing new equity, convertible securities, and the consequences of earlier lock-up expirations. Broader weakness in Chinese AI shares also affected sentiment.

The price war intensified those pressures because it questioned what the new capital would earn. Spending more is defensible if it creates durable technical leadership or valuable customer relationships. It is less attractive if each funding round finances services that competitors quickly commoditize.

Zhipu’s management has emphasized long-term model research rather than maximizing near-term profit. That choice is understandable in a market where capability can determine future relevance.

Public investors still need measurable milestones. They cannot value long-term ambition without considering dilution, losses, and the pace of commercial adoption.

The important comparison is not simply Zhipu versus an underfunded startup. DeepSeek can compete through architectural efficiency and open releases. Alibaba can connect Qwen to an established cloud platform. Large technology companies can subsidize models through broader businesses.

Zhipu lacks those exact structures. Its specialist identity gives it focus, but it also makes model economics more visible. There is no major advertising, commerce, or gaming operation to absorb prolonged losses.

That distinction makes financing both Zhipu’s strength and its vulnerability. The company has raised enough capital to remain an important competitor. It must now convert that capital into results before the market demands another reset.

What the HK$300 Billion Figure Does Not Prove

The valuation decline signals skepticism, but it does not establish that Zhipu has lost its technical position or customer demand.

Stock movements compress several judgments into one number. They reflect expected growth, risk tolerance, trading liquidity, capital structure, and market sentiment. They do not provide a direct measurement of model quality.

Zhipu’s first-half cloud growth indicates that developers and enterprises increased their use of its services. The company also remains a significant participant in China’s foundation-model market.

A lower valuation does not erase those facts.

It also does not prove that the price war will permanently destroy margins. Model-serving costs can fall through architectural improvements, better chips, caching, batching, and more efficient scheduling.

If inference costs decline faster than customer rates, a provider can lower prices while expanding gross profit. That outcome is one reason developers should avoid treating every reduction as evidence of financial weakness.

The opposite outcome remains possible. A provider can cut rates to defend share even when its own costs have not fallen enough. That approach boosts usage but increases cash consumption.

Public disclosures do not yet provide enough detail to determine which pattern will dominate Zhipu’s next reporting period.

The HK$300 billion figure also deserves caution because market-capitalization services use different share counts. Some may include domestic shares, unlisted shares, newly issued stock, or potential dilution differently.

Recent reports have produced materially different totals for similar trading dates. The variation does not invalidate the selloff. It does mean readers should compare valuations from one consistent data provider.

Another uncertainty concerns causation.

The timing links September’s decline with lower-cost model releases. Yet Zhipu also faced dilution, large financing needs, and concerns about the wider technology market. A coding-product data controversy contributed additional pressure earlier in the week.

No single factor fully explains the move.

The stronger conclusion is that several risks became visible at once. Price competition weakened assumptions about future margins. Financing increased the share supply. Earlier gains left little room for disappointment.

That combination changed how investors interpreted good news. Strong revenue growth no longer offset concern about losses. Fresh financing no longer appeared purely supportive. New model investment also implied another period of high expenditure.

The stock’s earlier rise deserves the same skepticism. A trillion-dollar Hong Kong valuation did not prove that Zhipu had secured lasting dominance. The subsequent decline does not prove the opposite.

Both extremes reflect uncertain expectations around a young market.

For developers and enterprise buyers, this distinction is practical. A falling share price should not determine model selection. Buyers should test output quality, latency, privacy controls, deployment flexibility, and total workflow cost.

They should also avoid depending on undocumented model behavior. Rapid product changes can alter output style, context management, or tool performance. Version-specific evaluations remain essential during an active price war.

For investors, the most useful evidence will come from operating trends rather than daily quotations. Revenue retention, cloud margin, customer concentration, and cash consumption can show whether Zhipu is building a defensible business.

Until those figures arrive, the Zhipu stock selloff should be read as a warning about expectations. It is not a final verdict on the company’s models.

Three Signals Will Decide Whether the Valuation Reset Holds

The next phase depends on cloud economics, competitive releases, and evidence that Zhipu’s new capital produces measurable commercial returns.

The first signal is Zhipu’s next financial disclosure.

Investors should examine cloud revenue growth alongside gross margin and adjusted losses. Growth will carry more weight if the company retains customers without sacrificing its improving service economics.

Cash consumption also matters. Zhipu has expanded its funding base, but repeated capital raising cannot substitute indefinitely for operating leverage.

A stronger cloud margin would support the view that scale and engineering efficiency can absorb lower market prices. A weaker margin would suggest that competition is transferring more value to customers than providers.

The second signal is the next round of model releases.

DeepSeek’s V4.1-Flash announcement linked architectural efficiency directly with lower pricing. That combination establishes a demanding comparison for Zhipu and other model developers.

A credible response does not require Zhipu to offer the lowest rate. It does require a clear reason for buyers to choose GLM.

That reason might be superior performance in coding, agent workflows, Chinese-language tasks, or enterprise deployment. It might also involve more predictable latency or stronger customization.

The evidence should come from repeatable third-party evaluations and sustained customer usage. Company-selected benchmarks alone cannot settle the comparison.

The third signal is how Zhipu deploys its September financing.

Spending on a new model is not automatically productive. The company must show that infrastructure investment produces faster services, lower unit costs, or new revenue.

Business-development spending should create customers that remain after introductory incentives. Strategic investments should strengthen technical capability or distribution, rather than merely broadening the company’s narrative.

These signals will arrive at different speeds. Competitor pricing can change immediately. Customer retention and margin data take longer to appear.

That mismatch creates volatility. Investors trade on new releases before financial statements show their effect.

Developers face a related problem. Choosing a model based only on today’s rate can create migration work when providers change endpoints, retire versions, or adjust terms.

A sensible response is to maintain portable application architecture. Teams can separate prompts, evaluation data, and provider-specific code. They can then test alternatives without rebuilding an entire product.

Organizations evaluating AI-generated research can also preserve the source material behind each output. A structured AI knowledge base helps teams compare model answers against the same evidence rather than relying on memory.

The Zhipu stock selloff ultimately asks a broader question about the model market. Does value belong to the company training the model, the cloud serving it, or the application controlling the customer relationship?

Zhipu is betting that a focused model developer can retain enough value to justify sustained investment. DeepSeek’s efficiency push and the wider price war are testing that assumption.

The answer will not come from one trading session. It will come from whether Zhipu converts new capital into lower costs, retained usage, and healthier margins.

For now, the valuation reset is a demand for proof. Watch the next cloud-margin update, the next GLM release, and the first clear evidence of how September’s financing changes operating performance.