This year, the biggest headache for leading large model companies is not a lack of money but a shortage of computing power. Data from the China Academy of Information and Communications Technology shows that in the first quarter of 2026, domestic AI computing power demand surged by 417% year-on-year, but effective supply growth was only 128%, widening the supply-demand gap.

Recently, Zhipu, dubbed the 'first large model stock,' has been busy explaining and raising funds.

On September 23, Zhipu responded to the earlier 'data leak' controversy, stating that product rectifications had been completed and apologizing to users. Two days earlier, on September 21, the ZCode source code was open-sourced on GitHub. This controversy highlights that competition among large models is no longer just about functionality and experience—security and trust are becoming new variables.

Earlier, on September 16, Zhipu held an analyst call, discussing computing power investments, cloud partnerships, and Co-work business progress, among other information-rich topics. At the same time, it raised its annual ARR guidance from $2.4 billion to $3 billion. On September 13, Zhipu announced the completion of a $5 billion refinancing through a 'small equity, large debt' structure.

However, market reactions to Zhipu's recent moves have been mixed, with fluctuating stock prices and divergent views among institutions. Guotou Securities lowered its target price to HK$1,351, citing the 'widening gap between Chinese and U.S. model capabilities' and 'share dilution.' Previously, JPMorgan raised its target price to HK$2,000, while CMB International raised it to HK$1,985.

Such divergent views stem from two competing logics—bulls believe in ARR growth trajectory and scaling after computing constraints are lifted; bears focus on cash burn rate and uncertainties in large model competition.

Amid the noise, some sober reflection is needed.

Is frequent large-scale financing driven by a genuine need for cash or passive involvement in a computing power arms race?

In just over eight months since its listing, Zhipu has completed three rounds of large-scale financing: a January IPO raised HK$4.896 billion net, a July 'flash placement' raised HK$31.375 billion, and this latest round brings cumulative financing to about HK$75.6 billion.

Despite raising significant funds, expenditures have been substantial. As of August 31, Zhipu's IPO proceeds had been fully utilized; of the HK$31.375 billion raised in July, HK$10.955 billion had been used, leaving HK$20.420 billion on the books.

With cash still available, why launch another large financing round? At the September 16 analyst call, management highlighted a key term—'computing capacity expansion.'

Indeed, this year, the biggest headache for leading large model companies is not a lack of money but a shortage of computing power. CAICT data shows that in Q1 2026, domestic AI computing power demand surged by 417% year-on-year, but effective supply growth was only 128%, widening the supply-demand gap.

Zhipu is no exception. After the February release of GLM-5, model call demand surged 10-fold, nearly depleting computing reserves in the release week, forcing the suspension of the main product, Coding Plan. After the GLM-5.2 release, overseas call demand spiked again, but due to insufficient computing reserves, about 90% of requests were handled by third parties. It wasn't until July's financing enabled large-scale computing expansion that the half-year suspension of Coding Plan ended, with sales growing over 15-fold.

This situation shows that the 'computing power crunch' is constraining revenue growth for large model companies. For Zhipu, financing decisions are now driven not by cash balances but by proprietary computing reserves.

Computing shortages are also driving companies to build their own infrastructure. Zhipu has built and partially operated a 1GW-scale data center using domestic chips; Deepseek plans to build a data center in Ulanqab, Inner Mongolia, adding 1GW of capacity through a 'self-owned + leased' model.

Meanwhile, the gap between rapid technological iteration and commercial return cycles has created a massive funding chasm. An industry insider told Lukou: 'This year, large model iteration has accelerated significantly. The current generation hasn't earned much before the next generation starts burning money on training, so nearly all companies are accelerating financing.'

For example, DeepSeek launched second-round financing talks in July-August and in September was reported to have hired CITIC Securities to prepare for a Sci-Tech Innovation Board IPO. Yuezhi Anmian completed a $3.5 billion financing closing in late July (post-money valuation of $35 billion) and immediately launched a pre-IPO round at a $50 billion valuation. MiniMax completed about HK$16 billion in financing on July 10 through a 'placement + zero-coupon convertible bond' structure... Nearly all large model companies are stockpiling cash for the computing power arms race.

Zhipu's current $5 billion financing round confirms this trend: about 60% of net proceeds will go toward next-generation GLM foundation model and 'fully self-trained' system R&D, as well as large-scale training, inference, and computing infrastructure deployment and upgrades.

Notably, 'fully self-trained' means the next-generation GLM trains within the environment built by the previous generation, forming a self-improving loop across three dimensions: self-produced data, self-created environments, and self-optimized infrastructure. Simply put, it lets AI train itself—a model known overseas as 'recursive self-improvement' (RSI).

'Compared to other large model companies, Zhipu currently allocates more resources to Scaling deep,' said an industry insider. Scaling deep involves post-training, reinforcement learning, and long-term trajectory and task feedback. 'This might be paving the way for fully self-training,' the insider revealed.

However, this technical narrative is still in early validation stages. 'RSI is indeed one of the latest research directions, but unconstrained recursive training can lead to system degradation—a phenomenon known in academia as model collapse, which has been empirically verified,' the insider noted.

Can new commercial growth justify $3 billion in ARR?

At the call, Zhipu management also highlighted a key metric: annual recurring revenue (ARR) guidance was raised from $2.4 billion to $3 billion.

It's important to note that ARR is not an accounting metric—it simply multiplies the most recent month's revenue by 12. In H1 2026, Zhipu reported RMB 954 million in revenue. By late August, ARR stood at $1.6 billion. A massive gap remains between these figures.

So when ARR guidance was raised to $3 billion, the market's top concern wasn't 'why the increase' but 'who will deliver it.'

Beyond computing expansion-driven growth in Coding business, the call revealed two key commercial growth drivers: revenue sharing with domestic and foreign cloud providers starting in October, and Co-work commercialization.

Let's first discuss cloud providers. As a new variable in Zhipu's revenue mix, specific partner names, per-platform revenue scales, and revenue-sharing ratios haven't been disclosed yet. However, historical parallels exist—let's examine industry precedents.

Overseas, this path has already been validated. Anthropic distributes Claude via AWS Bedrock, while OpenAI reaches enterprise clients through Azure. On the surface, model companies exchange revenue shares for sales networks, while cloud providers fill their most competitive model shelves—a win-win.

But revenue sharing means margin concessions. Cloud providers' commissions directly erode profit margins. A source revealed that about 50% of Anthropic's gross profit from AI sales through Amazon goes to Amazon. Here, 'gross profit' refers to Anthropic's revenue from Amazon AI sales after deducting Amazon cloud server operating costs.

Domestically, Yuezhi Anmian is currently negotiating cloud platform hosting for Kimi K3 with Microsoft, Amazon, and Google. On September 18, AWS officially announced Kimi K3's availability on Amazon Bedrock.

Public reports indicate Yuezhi Anmian initially sought up to 30% revenue sharing, though the industry previously considered 20% the upper limit—making 30% a market-exceeding ask.

From a profit flow perspective across the industrial chain, cloud providers' revenue certainty clearly exceeds that of large model companies. Barclays' report provides a more systematic framework: for every $100 in revenue model companies earn, about $35–40 flows to the three cloud giants (AWS, Azure, Google Cloud) as inference computing fees.

Of that $35–40, cloud providers earn about $10–20 in operating profit after infrastructure costs, corresponding to 35%–45% operating margins. This means even if large model companies secure a seemingly respectable revenue share, actual retained income after computing costs may be quite limited.

It's also worth noting that cloud providers adopt a 'supermarket logic,' offering multiple models including Claude, GPT, Gemini, DeepSeek, and Kimi simultaneously.

In this setup, cloud providers prioritize models with higher call volumes and prominent display slots. Conversely, for large model companies, customer switching costs are nearly zero—unless they have a significant leading edge, retaining users and commanding premiums will be difficult.

Now let's discuss Co-work commercialization. The call revealed that just one month after GLM-5.3's release, Co-work secured over RMB 1 billion in orders, with over 100 enterprises embedding GLM in their cybersecurity products and workflows.

This suggests Zhipu has opened a second growth curve. However, these orders will only be recognized as revenue over the next 1–2 years, and subsequent renewals and expansions depend on delivery reputation this time.

So rather than fixating on ARR figures, we should examine their underlying quality. How much of this 'future revenue' will translate into actual cash on financial statements?

Independent large model companies face pressure from both sides

A clear recent trend is cooling institutional sentiment toward independent large model companies' valuation logic.

On September 6, Jefferies released a research report stating, 'China's large language model sector is currently overcrowded. Compared to pure independent AI labs, we favor full-stack cloud platform giants with computing and data advantages and strong balance sheets.' Referencing overseas peers' valuations, the investment bank lowered Zhipu Cloud's price-to-sales multiple from 50x to 30x and cut its target price to HK$1,183.79.

The squeeze from tech giants on independent companies stems not from model capabilities themselves but from ecosystem closed loop (closed-loop) advantages in distribution and monetization. Giants can directly embed models into their products as implementation scenarios, while independent companies lack such proprietary entry points and must build channels or rely on external scenarios.

The funding gap is even more stark. ByteDance's 2026 AI capital expenditures have been raised from RMB 160 billion to over RMB 200 billion; Alibaba's 'RMB 380 billion over three years' plan is only halfway done, with management stating future five-year investments will 'far exceed RMB 380 billion.'... In comparison, despite multiple financing rounds and plans to 'invest about RMB 30 billion to build 100,000 P of computing capacity' revealed on the call, Zhipu still appears relatively under-resourced.

Meanwhile, beyond giant squeeze (squeeze), independent large model companies also face price competition from new entrants.

Another recent Jefferies research report showed that among the global top 14 language models tracked by Artificial Analysis (AA), four new models were added in September from Singapore, South Korea, the UAE, and the US.

The R&D institutions behind these models generally have limited financing, with some having raised only about $20 million cumulatively yet entering the global top tier. This indicates that distillation techniques are further lowering industry entry costs.

Simultaneously, narrowing model performance gaps and intensifying API price competition are imposing new pressures of 'high capex, low certainty returns' on the AI supply chain.

Take Zhipu's recently launched GLM-5.3-Flash version, priced even lower than the post-price-hike DeepSeek-V4-Flash.

But low prices are never a barrier, a programmer who frequently uses large models told Lukou: 'Large models have iterated for years—now basic tasks feel similar across options. Developers basically choose the cheapest option since switching costs are negligible.'

Frankly, while Zhipu doesn't lack financing ability or technical narratives, these don't constitute true moats in the hyper-competitive large model environment.

Earlier this year, Deutsche Bank predicted, '2026 will be a make-or-break year for independent AI model companies.' Except for a few like Anthropic that have stabilized with robust cash flows and enterprise products, smaller independents may struggle to bear accelerating computing costs and ultimately face acquisition by giants.

When the hype fades, the companies that remain will be those that have truly closed the commercial loop.