Can sovereign capital overturn the Silicon Valley AI hype?

In the past, almost all of the most successful software companies in the world were created in the United States. China’s Zhipu AI stock price has surged rapidly, sparking new discussion. What do the two AI development paths represented by Zhipu and Anthropic mean?

Over the past several decades, nearly all the world’s most successful software companies have been created in the United States. Silicon Valley has not only long led technological innovation, but also remained ahead in revenue, market capitalization, and global business influence. Recently, people have found that the AI models from China’s Zhipu company are close to the U.S. frontier models, which has driven Zhipu’s stock price to surge quickly and sparked yet another discussion: can a nation-backed support system nurture a company whose technical strength and business scale are enough to challenge U.S. AI giants—and ultimately change the AI business landscape dominated by Silicon Valley? Taking Zhipu and Anthropic as examples, this article compares the differences between the two AI development paths in terms of capital sources, market quality, business models, and valuation logic, and explores whether national capital can ultimately overturn Silicon Valley’s AI myth.

1. What exactly is the capital market pricing for AI companies?

In the internet era, China produced a number of outstanding technology companies. They achieved major success in e-commerce, social media, mobile payments, and other fields, and also formed many technological innovations. However, in the most core business metrics—such as company revenue, global developer ecosystems, and capital market valuations—U.S. companies have always maintained a clear lead, as if forming a “Silicon Valley myth.” In the AI era, will this pattern be truly changed for the first time?

So, the recent surge in Zhipu’s valuation in Hong Kong’s stock market to a record high has drawn attention. Then how should a foundation model company be valued? In the past few years, AI competition has often been understood as a contest of model leaderboards, parameter scale, GPU counts, and funding amounts. But as the capabilities of frontier models in China and the U.S. continue to converge, the capital market is gradually shifting from “technology-based pricing” to “business-based pricing”: it not only cares who trained a stronger model, but also whether these models can continuously generate commercial revenue.

Zhipu’s capital story is highly symbolic. On January 8, 2026, Zhipu Huazhang listed on the Hong Kong Stock Exchange with an offering price of HK$116.2. The offering market cap was about HK$50+ billion, making it the “global first foundation model stock.” On June 17, Zhipu released its new-generation flagship model GLM-5.2. After some U.S. peers tested the model, they ranked it among the top three, which further boosted market sentiment. By June 22, the company’s total market cap first broke through HK$1 trillion, with a cumulative upside of more than 1900% from the offering price.

Such a rapid valuation jump cannot simply be explained by the model’s capabilities rising by the same magnitude within half a year. A more reasonable interpretation is that the capital market is pricing a possible future: whether Zhipu can build a commercial cycle for a Chinese foundation model company through the joint push of national support, government-and-enterprise markets, and an open-weight ecosystem.

Therefore, the real question is not why Zhipu is worth HK$1 trillion, but what value the capital market is pricing for AI companies today. The capital market is not buying today’s revenue—it is buying the ability to continuously generate revenue in the future.

2. The state provides not only capital, but also an early market

Zhipu is not a company that is supported by the nation “mobilizing all efforts,” but there is no doubt that a national support system exists behind Zhipu. It includes subsidies, capital, regulation, and industrial resources, as well as early markets formed through government-and-enterprise procurement.

From the capital side, since its founding in 2019, Zhipu has completed 8 rounds of fundraising, with total cumulative investment of about RMB 8.3 billion. The last pre-listing round’s post-money valuation was RMB 10k. State-owned assets in Beijing, Hangzhou, Chengdu, Zhuhai, Shanghai, and other places have taken equity stakes one after another. At the IPO stage, the cornerstone subscription of about HK$3 billion also includes Beijing’s core state-owned capital, indicating that state-owned capital is indeed an important participant. In addition, there are also subsidies across multiple areas such as cash, computing power, land, and taxes.

From the market side, the role of the national system is equally important. In 2024, Zhipu’s revenue was RMB 312.4 million. Among it, “government-and-enterprise customization” revenue was about RMB 130 million, accounting for about 42% of that year’s revenue, with a gross margin of about 72% (revenue of RMB 191 million in the first half of 2025, but no details of government procurement were disclosed). This means that government-and-enterprise demand has already become one of the important revenue sources in Zhipu’s early commercialization. More importantly, for foundation model companies, an early market and early capital are equally important. Government-and-enterprise customers can help the model enter real business scenarios, complete product validation, accumulate industry experience, and provide templates for entering broader commercial markets later.

But such markets also have their own characteristics. Government-and-enterprise markets can provide stable demand, but they do not equal global commercial markets. Their functional requirements, procurement cycles, budget structures, delivery methods, and revenue scalability differ from business revenue generated through developer APIs, enterprise subscriptions, and cloud platform revenue sharing. The national support system can help companies build supply capabilities faster and provide an important early market; but in the long run, valuation still depends on whether these early markets can be transformed into larger-scale, more sustained commercial revenue.

3. The real gap is not the model, but the market quality

When comparing Zhipu with Anthropic, you cannot look only at model capability. As frontier model capabilities are rapidly converging, the gap between leaders increasingly manifests as alternating leadership across different capability dimensions—not an absolute generation gap from one product to another. Model capability determines the supply ceiling, but market quality determines how much of that supply can be converted into commercial revenue.

The “market quality” referred to in this article is not only the market size, but also the ability of the market to continuously convert AI capabilities into commercial revenue. Enterprise software budgets, payment ability, subscription culture, and market coverage are just its specific manifestations. China has a large user base and rich industrial scenarios, and it also has an early market formed through government-and-enterprise procurement; Anthropic faces mainly the global enterprise AI market, where its customer structure, software budgets, and sustained subscription culture differ noticeably from China’s market.

Software for enterprises is the first “killer application” that AI large models have found. China’s enterprise software budgets are one order of magnitude lower than those in the U.S., which is a fairly reliable industry estimate. Even if this assessment is not accurate to each individual company, it still reveals a key fact: even if China and the U.S.’s leading models have similar technical capabilities, the commercial markets they can serve and customers’ willingness to pay are not the same.

That is why the prices of China’s AI models are generally lower than those in the U.S., which cannot simply be understood as China’s models being cheaper to produce. A more reasonable explanation is that they face different market qualities. Paying ability determines price; price determines revenue; and revenue then determines whether companies can continuously invest in R&D and service capabilities.

Therefore, a company’s long-term competition in AI is not about who trains a stronger model first—it is about who can convert model capability into sustained commercial revenue. The higher the market quality, the greater the commercial value that the same unit of AI capability can ultimately realize. A high-quality market, in turn, also drives further technological progress.

Table 1 Comparison of core commercial indicators of two AI business models

4. Why is Zhipu’s price-to-sales ratio much higher than Anthropic’s?

Table 1 reveals a counterintuitive fact: measured by dynamic price-to-sales ratios, Zhipu is far more expensive than Anthropic today. Based on the projected revenue for 2026, Zhipu’s current dynamic P/S ratio is about 478x, while Anthropic’s is about 20.5x. In other words, the revenue multiple the capital market assigns to Zhipu today is more than twenty times that of Anthropic.

This does not mean the capital market has already proven that Zhipu is better than Anthropic. It more likely indicates that the market is assigning different kinds of upside imagination to the two companies. Anthropic’s valuation is built more on already-formed sustained revenue arising from global enterprise APIs, Claude subscriptions, and partnerships with cloud platforms; Zhipu’s valuation is built more on expectations of future commercial revenue growth.

Valuation essentially discounts future cash flows. A 478x P/S ratio means the capital market has already priced in high expectations for many years of future commercial growth. For Zhipu, the core issue is no longer whether the model can enter the global top tier—it is whether future revenue can catch up with the expectations the capital market has already priced in.

This is also why the revenue structure matters more than the revenue numbers. For the capital market, sustained revenue typically carries higher valuation value than project-based revenue. Once-off revenue proves that products can be sold; only sustained revenue proves that the commercial cycle can exist long term. It should be noted that because Zhipu has a relatively short listing history, the actual freely tradable shares are currently still limited; as locked shares are gradually released and more similar AI companies list, the stock supply structure will change significantly, which could cause substantial pressure on short-term market prices. In the long run, investors will rationally return to financial metrics for valuation: sales, costs, profits, cash flow, and growth rates.

5. Open-weight and closed-source APIs: not a technical route, but a commercial goal

Zhipu and Anthropic also differ clearly in product strategy. In recent years, Zhipu has been more actively promoting open-weight models, hoping to attract developers, industry customers, and AI ecosystem participants with a lower barrier. Anthropic, by contrast, insists on keeping its core flagship model closed-source, mainly providing model capability through APIs, enterprise subscriptions, and cloud platform licensing.

This difference should not be simply understood as an “open versus closed” dispute. It is more like a choice of two different commercial goals. Open-weight emphasizes expanding the entry points of the commercial cycle, so more developers and enterprises enter the ecosystem first. Anthropic’s closed-source API model emphasizes increasing the cash flows generated in each round of the commercial cycle, turning every model call into billable revenue.

From Zhipu’s perspective, open-weight can help expand market coverage, lower barriers for industry applications, and create more deployment opportunities in government-and-enterprise and industrial scenarios. From Anthropic’s perspective, its closed-source API services can protect model assets, control service quality, and embed model capability into sustained revenue.

Therefore, open-weight and closed-source API services are not just a debate about technical routes; they are differences between market entry strategies and monetization strategies.

Table 2 Token price structure reveals two different market strategies

6. Token prices reveal where China’s AI price war truly happens

From Table 2’s Token price data, it can be seen that the price gap between models in China and the U.S. is not fixed—it expands as model tiers decline. The price gap is smallest for flagship models; it widens for main models; and for entry-level models, it expands to dozens of times or even hundreds of times. This structure reveals at least three important pieces of information.

First, the price gap between top-tier Tokens in China and the U.S. is not as large as rumored. The input price gap between GLM-5.2 and Claude Opus in the flagship tier is about 4.25x, while the output price gap is about 6.1x. Compared with the entry-level gap measured in dozens of times, this is already relatively close in commercial pricing.

Second, flagship Token pricing is closer to companies’ true cost and business value judgments. Training the model is only one part of the cost. The more important cost sources in long-term operations for foundation model companies come from ongoing inference, services, operations and maintenance, networking, storage, and customer support after model launch. Therefore, you cannot simply infer model costs from terminal Token prices alone. But as flagship Token prices gradually approach international levels, it at least indicates that the differences in lifecycle costs between China and the U.S.’s leading models are not as huge as the differences in prices for terminal low-end Tokens.

Third, China’s AI companies’ low-price strategy mainly focuses on low-end Tokens, more like a market marketing and customer acquisition strategy rather than a natural reflection of cost advantages. Low-end Tokens bear the task of expanding market coverage, attracting developers to try the product, and lowering entry barriers for industry customers; flagship Tokens bear more direct monetization tasks.

Therefore, China’s AI price war mainly occurs in the long-tail entry market rather than the flagship market. More precisely, Chinese AI firms are subsidizing the market entry represented by low-end Tokens, not subsidizing flagship models themselves for the long term. Using shareholder capital to subsidize is not charity—it is treating low-end Tokens as “customer acquisition cost (CAC),” betting that through ecosystem expansion, some users will eventually settle in and upgrade into flagship Tokens or high-net-worth customers for customized private deployments. Price reflects not only costs, but also the market a company wants to enter.

This pricing strategy also aligns with common practices of Chinese AI companies: expand ecosystem entry using low prices or even free products, then realize revenue transformation through higher-value models, industry solutions, and government-and-enterprise deployments. It shows that Chinese AI companies are not seeking the lowest prices across all products; rather, they allocate resources between profit margins and market share. The efficiency of switching from subsidized low-end Tokens to flagship Tokens is the key for this model to succeed. While this price war also puts pressure on U.S. AI companies, competition is currently mainly happening among Chinese companies, especially since there are real geopolitical “moats” between countries.

7. How market quality affects AI companies’ revenue ceiling

If model capabilities are converging, and flagship Token prices are also approaching each other, then the factor that truly widens the revenue gap is not just technology, but market quality.

Anthropic is facing the global high-value enterprise AI market. This market has two characteristics: first, enterprises have high software budgets, mature SaaS subscription culture, strong developer ecosystems, and API calls can be scaled quickly—every model call can potentially be converted into billable tokens, and every enterprise customer may become a source of sustained revenue. Second, in the U.S., white-collar labor costs are high, and enterprises have a strong urgency to reduce costs and improve efficiency by replacing or augmenting human work with AI, so they are willing to pay for expensive closed-source APIs.

Zhipu faces a different market structure. China’s market has abundant industrial scenarios, government-and-enterprise procurement demand, and domestic replacement momentum, but enterprise software budgets and sustained subscription habits still differ noticeably from those in the global big enterprise AI market. The government-and-enterprise market can provide important early revenues and application validation, but to support long-term revenue that matches the current valuation, it still needs to further expand the group of high-quality commercial customers. More fundamentally, the cost of knowledge workers and operational personnel in China is relatively lower. If the cost of AI Tokens or customized deployment is higher than the cost of hiring human labor, enterprises will lose their willingness to pay. This is also the deep reason why China’s market forces large models into price wars and pushes open-weight approaches.

This does not mean one market is inherently superior to the other. A national support system can accelerate technological deployment and industry applications, especially for infrastructure and critical industry scenarios; global commercial markets have advantages in scalability, and the sustainability of commercial revenues. The revenue structures, growth rhythms, and valuation logics of the two markets are not identical.

Therefore, what Zhipu truly needs to prove in the future is not whether it can train better models, but whether the early market provided by the national support system can be further converted into broader, more sustained, higher-quality commercial revenue.

8. Can national capital overturn the Silicon Valley AI myth?

So, can national capital overturn the Silicon Valley AI myth? This question cannot be answered simply with “yes” or “no.”

A national support system can do many things that market capital struggles to complete. It can provide early financing, coordinate industrial resources, drive government-and-enterprise procurement, lower barriers for domestic models to enter key industries, and improve companies’ growth efficiency through regulation and capital market channels. For foundation model companies, these supports are real and important.

But ultimately, a national support system still has to undergo market validation. It can help companies build world-class supply capabilities and provide important early markets; but what the capital market ultimately cares about is whether these advantages can be converted into sustained growth in commercial revenue.

Anthropic represents a commercial path that has already been validated by the mainstream enterprise software market worldwide; Zhipu represents a Chinese AI commercial path that is still evolving rapidly and still needs further validation. The former is not the only answer, and the latter is not destined to fail. What truly needs comparison is not just who currently has a stronger model, nor who currently has a higher valuation—rather, which system can establish a more sustainable commercial cycle.

Today, the capital market has cast a vote of confidence for Zhipu, but the real test has only just begun. What will determine whether Zhipu’s valuation is validated in the future is not the next generation of model leaderboards, but whether market quality can keep improving and ultimately establish a commercial cycle that matches today’s valuation and can keep operating relying on the market. If, in the AI era, a first batch of Chinese AI companies eventually emerges that fully outperforms U.S. peers in revenue scale, market capitalization, and global business influence, then what will truly be overturned will not be any single U.S. company, but the commercial landscape of the global information industry that has been dominated by Silicon Valley for the past decades.