Large-scale AI models arguably represent the most 'extreme' assets in the current technology sector rally.

At its initial public offering (IPO), Zhipu was valued at a mere HK$51.8 billion based on its offering price. However, just six months later, that valuation skyrocketed to HK$1.07 trillion—marking a staggering 20-fold increase!

Yet, after reaching the trillion-dollar milestone, Zhipu's value took a nosedive. By October 7, its stock price had plummeted to HK$696, reflecting a 76.6% decline from its peak and nearing an 80% loss.

This scenario illustrates how a company can surge 20-fold in six months, only to lose nearly 80% of its value in the subsequent three.

Behind these dramatic swings in valuation lies an increasingly pivotal factor: model releases.

This year, Zhipu has undergone four major model releases. On the days of these releases alone, its stock price rose in three instances—translating to a 75% success rate—with an average gain of 12.3%.

When extending the observation window to five trading days post-release, the impact of these models on valuation becomes even more pronounced. Two out of the four releases witnessed cumulative gains exceeding 40% over the following five days.

The last industry where valuation fluctuations heavily relied on new product launches? The new energy vehicle (NEV) sector.

In recent years, the valuations of NIO, Li Auto, and XPeng have constantly been reshuffled with each blockbuster model launch. The company introducing the most competitive new vehicle model gained opportunities for fresh valuation premiums.

Today, domestic large-scale AI model firms are retracing this path.

Domestic large-scale AI model companies are transforming into 'product cycle stocks' akin to those in the NEV sector—except the driver of valuation has shifted from 'vehicle models' to 'AI models'.

More starkly, while a hit car model could sustain its momentum for years, a leading AI model might only dominate the market for months.

Today, let's delve into why large-scale AI model companies are increasingly resembling NEV firms.

/ 01 / Domestic Large-Scale AI Models Now Priced by 'Product Cycles'

This year, Zhipu's valuation has been on an extreme rollercoaster ride.

At its IPO in January, Zhipu priced its shares at HK$116.2, valuing the company at HK$51.8 billion. By June 22, its market capitalization had soared to HK$1.07 trillion—over 20 times higher.

However, the trillion-dollar valuation proved to be short-lived.

Zhipu's stock then took a sharp plunge. By October 7, its market capitalization had collapsed to HK$324.1 billion.

In essence, Zhipu took less than six months to rocket from a valuation of just over HK$50 billion to HK$1 trillion, only to shed nearly 70% of its value (over HK$740 billion) in the subsequent three months.

More intriguingly, Zhipu's commercialization efforts did not deteriorate during this period. API calls and annual recurring revenue (ARR) continued to grow rapidly, yet the capital markets had already completed a violent revaluation.

This valuation volatility strongly resembles the early days of NIO, Li Auto, and XPeng.

In January 2021, NIO's market capitalization briefly topped $100 billion, while Li Auto hovered around $33-35 billion and XPeng at $40 billion.

By February 27, 2024, Li Auto's market capitalization in Hong Kong had reached HK$349.56 billion, compared to NIO's HK$94.23 billion and XPeng's HK$68.87 billion.

In other words, one Li Auto now equaled 3.7 NIOs or 5.1 XPengs—its valuation even exceeded twice the combined value of the other two.

Auto companies' valuations are constantly reshuffled based on hit models, sales, and profit expectations.

Now, domestic large-scale AI model firms are displaying similar patterns—except the ranking driver has shifted from 'vehicle models' to 'AI models'.

This pattern becomes even clearer when comparing Zhipu's major model releases with its stock performance.

This year, Zhipu has undergone four significant model releases.

On the days of these releases alone, its stock price rose in three instances (a 75% win rate) with an average gain of 12.3%.

When extending observations to five trading days post-release, the valuation impact intensifies. Two out of the four releases saw cumulative gains exceeding 40% in the subsequent five days.

On February 12, Zhipu released GLM-5, which surged 28.68% that day and another 56.22% over the next five days. On June 17, after open-sourcing GLM-5.2, its stock rose 12.62% initially and 41.57% over five days.

Each rally shared a common trait: model releases did not mark the end of gains but triggered new valuation expansions.

More fascinatingly, as investors grew familiar with this rhythm, markets began preemptively trading model release expectations.

On August 14, Zhipu released GLM-5.3. Its stock dipped 3.57% that day and fell another 11.1% over five days.

Yet, looking backward, Zhipu had already rallied sharply. From August 3, its stock jumped from HK$940 to HK$1,317—a gain of over 40%—before the official release.

In essence, markets had already priced in the model release before it happened. When the actual announcement met expectations, profits were taken instead of new gains being realized.

Thus, domestic large-scale AI model firms now resemble NEVs—highly cyclical valuations dependent on rapid product iterations.

/ 02 / No Permanent Technological Premiums for Domestic Large-Scale AI Models

Compared to NEVs, large-scale AI model companies face two even more extreme challenges.

Firstly, their product cycles are far shorter, leading to more frequent revaluations.

A hit car model can drive sales for 2-3 years, with full model refreshes typically taking years to complete.

Large-scale AI models operate differently. From February to August, Zhipu iterated through GLM-5, GLM-5.1, GLM-5.2, and GLM-5.3—four major upgrades in just six months.

The market reassesses the technical competitiveness of these models every 1-2 months.

This shift is directly reflected in investment banks' valuation models.

In early 2024, when Morgan Stanley initiated coverage on domestic large-scale AI model firms, their valuations clustered around 54 times the projected 2027 price-to-sales (P/S) ratio.

Months later, UBS presented a completely different framework. Despite only a 50% difference in revenue projections, Zhipu's target valuation exceeded that of its peers by over five times.

Such stark valuation divergence cannot be explained by revenue alone. Breaking down the valuations of domestic large-scale AI model firms today reveals two components:

Market Capitalization = Commercialization Base + Frontier Model Options.

The first part is straightforward. ARR, token usage, client count, and gross margins determine a company's real business scale—and serve as the valuation floor.

However, stock price elasticity comes from the second part. Investors price in future pricing power, higher token usage, faster ARR growth, and even the probabilities of next-gen model success into today's valuation.

This explains the counterintuitive phenomenon: Zhipu's ARR still grows rapidly, yet its stock plunged from HK$2,980.

Secondly, technological leadership in large-scale AI models depreciates faster than in the automotive industry.

A hit car model can sustain sales for 2-3 years. An state-of-the-art (SOTA) AI model's leadership window might last only months—or even weeks.

Large-scale AI models represent relative performance assets.

Ranking among the global top three today does not guarantee that position in three months. Any major update from competitors like DeepSeek or Qwen could reshuffle the technical rankings.

While models can improve continuously, their competitive edge shrinks over time.

This implies a strong time decay in technological premiums for large-scale AI model firms.

The financial concept of 'Theta' applies here. Even if an underlying asset's price stays flat, an option's time value erodes as expiration nears. Large-scale AI models' tech premiums behave similarly.

When a model first achieves SOTA status, markets pay a premium for its leadership. But as competitors catch up, that scarce technical edge loses value.

Worse still, this process is not fully controllable by the model firms themselves.

In July, after another domestic model firm released a new version, JPMorgan slashed long-term valuation estimates for Chinese large-scale AI model companies. Zhipu's 2030 price-to-earnings (P/E) forecast dropped from 30 times to 20 times, with its target price cut from HK$2,400 to HK$1,600.

JPMorgan's rationale was blunt: frontier model leadership now rotates rapidly among Chinese firms, making sustained technical leadership far less certain than previously assumed.

Even though Zhipu's own models did not regress, a competitor's stronger release prompted a revaluation.

Thus, large-scale AI model firms' valuations face dual product cycles: their own model iterations and those of all competitors.

Revisiting the cases of NIO, Li Auto, and XPeng now offers perspective.

A few years ago, their market capitalizations differed by multiples. Today, NIO, XPeng, and Li Auto stand at HK$67.9 billion, HK$72 billion, and HK$93.4 billion respectively—gaps far narrower than before.

While sales, profits, and funding environments play roles, this reflects a broader shift: markets increasingly struggle to treat temporary product leadership as permanent valuation advantages.

Hit products rotate, technologies converge, and once-insurmountable gaps shrink through iterations.

This may prove to be the ultimate fate for domestic large-scale AI model makers.

By Aqi