In September, the capital market sent dense (intensive) signals to the large model industry.

First, in the U.S. stock market, Anthropic. According to the Financial Times, Anthropic is preparing for a U.S. IPO at a valuation of $2 trillion or even higher. If successful, it would surpass SpaceX as the highest-valued IPO in history, with NVIDIA also rumored to be in talks to invest up to $10 billion.

Next, in the A-share market, DeepSeek is advancing a new round of financing, while Reuters reported that the company is in talks with CITIC Securities to prepare for a listing on the Science and Technology Innovation Board.

On September 13, it was the turn of the Hong Kong stock market. Zhipu announced the completion of approximately $5 billion in financing: about $2 billion in share placements and roughly $3 billion in zero-coupon convertible bonds, with funds earmarked for next-generation GLM models, a fully self-trained system, and computing infrastructure.

The three companies, operating within the U.S. and Chinese tech ecosystems, differ in technical approaches, shareholder structures, and commercialization rhythms—yet all chose to expand their capital reserves simultaneously.

Behind this phenomenon, competition for large models has entered a phase with higher resource concentration.

The variables determining a company's medium- to long-term position have expanded from model rankings to include computing power supply, talent density, engineering efficiency, commercial revenue, and even capital organization capabilities.

To put it more bluntly: in the first half, success depended on creating a good model; in the second half, it depends on continuously producing next-generation models through better organization.

The Second Half of Model Competition: From Single Breakthroughs to Sustained Intelligence Expansion

By 2026, the window for leadership in first-generation models has shrunk dramatically.

The Claude Code team mentioned in an interview video that Anthropic must develop products at an extremely fast iteration pace. 'You have to be very unattached to what you're building because it will soon disappear.'

Zhipu's GLM series completed six iterations from version 4.6 to 5.3 in about 11 months, with its intelligence index rising from 32 to 60.

After facing question (skepticism) over slowed iteration, DeepSeek accelerated again this year, releasing V4 in April and V4.1 in June, with just half a month between gray-scale testing and full deployment. The iteration cycle shortened to 'monthly' or even 'weekly' levels.

Launching a good model can only win one race; continuously expanding intelligence boundaries is what wins the entire season. Future iterations will accelerate in new ways.

Anthropic co-founder and CEO Dario Amodei believes AI is approaching the critical point of recursive self-improvement (RSI) within 6 to 12 months. In Anthropic's latest research, Claude even had the less capable Claude Sonnet 5 train an early version of Claude Opus 4.8 that had not yet completed safety training.

The RSI trend is not unique to Anthropic; it is emerging among leading models.

In its September 12 financing announcement, Zhipu allocated about 60% of funds—approximately HK$23.5 billion—to a 'fully self-trained' system.

The technical roadmap for the next-generation GLM involves self-produced data, no longer relying entirely on human annotation; self-created environments, where agents collect and transform real-world tasks; and self-optimized infrastructure, where coding-capable models On the contrary (in turn) optimize the training and inference systems they rely on.

Once recursive self-improvement is achieved, the logic of iteration changes.

In the past, model iteration was bottlenecked by humans. Data relied on human annotation, experiments on researcher design, and a single version could take months to refine.

Under the RSI framework, models produce their own data, build environments, and optimize infrastructure, removing human bottlenecks. Stronger models enable faster self-improvement, which in turn produces even stronger models. The cost of continuously expanding intelligence boundaries rises with each generation.

Iteration capability thus becomes a compound proposition of 'technology × investment.' Being good at the game is not enough; you must also keep paying the tuition.

Hence, the common subtext of this financing wave is to ensure companies can sustain technological innovation.

None of the three AI companies are short on cash in the short term.

Anthropic completed about $30 billion in Series G financing in February and $65 billion in Series H in May. Its valuation surpassed OpenAI, making its equity the most sought-after asset in the primary market.

In DeepSeek's first funding round, founder Liang Wenfeng himself was the largest single investor.

Zhipu's July share placement raised approximately HK$31.375 billion net, and now another round worth about HK$39.3 billion has landed.

AI companies flush with cash are still raising funds aggressively, with capital expenditures aimed at securing a spot in the next race.

Three Divergent Capital Routes Converge: Organizing Resources for Next-Generation Models

Continuously expanding intelligence boundaries requires more than just money.

What truly determines positions in the second half is the system that organizes computing power, talent, clients, and capital. The influx of funds into these leading AI companies reflects confidence in their ability to organize resources and maintain leadership.

This September, leading companies presented three distinct financing routes.

Anthropic pursued a 'global industrial capital + U.S. public market' route.

Its shareholder list reads like the AI industry itself: Google, Amazon, Microsoft, and NVIDIA have all invested, deeply binding capital with computing power supply and client channels.

DeepSeek took a 'founder-controlled + local industrial capital' route.

In its first funding round's unique structure, external funds did not enter the company directly but were injected into a limited partnership managed by Liang Wenfeng. Investors had no voting rights, only priority financial information and first refusal in subsequent rounds. Liang ultimately controlled over 84% of shares and nearly 100% of voting rights.

Zhipu leveraged the Hong Kong stock market to connect public markets with long-term capital.

It launched a global offering of H-shares in January, completed a share placement in July, and in September, used a 'small equity, large debt' approach, raising approximately HK$75.5 billion net in eight months since listing.

This September's financing combined equity and debt tools: stocks accounted for about 40%, with about 60% from zero-coupon convertible bonds issued at a premium to face value, with conversion prices implying double-digit premiums to market prices. The company minimized immediate dilution while locking in long-term funds, with investors paying a premium for potential conversion value.

The announcement also included a caveat: conversion prices would not adjust due to expected share issuances for a Science and Technology Innovation Board listing. An A+H strategy is already embedded in the terms.

All three companies secured funds on relatively favorable terms, but their underlying goal was highly consistent: to organize sufficient R&D, computing power, and talent resources for next-generation models in advance.

Anthropic signed massive computing contracts with SpaceX, AWS, and Google, poaching Andrew Tulloch from Meta and Schulman from OpenAI.

DeepSeek invested about RMB 11 billion in AI infrastructure in the first seven months, nearly tenfold year-on-year. After financing confirmed its valuation, employee options received clear pricing, reversing the passive situation of core researchers being poached.

Zhipu allocated about 60% of funds to next-generation GLM and computing infrastructure, advancing computing procurement and leasing while investing in chip adaptation, operator development, and cluster interconnection. Aligned with the new goal of full self-training, model iteration, computing construction, and capital investment are further integrated, creating greater space for next-generation technological exploration.

This is where leading companies begin to distance themselves from followers.

Falling one model version behind still offers chances to catch up through open-source models, distillation, and engineering optimization. However, gaps in resource organization systems simultaneously affect model iteration, service stability, talent flow, and client choices. For top talent, choosing a company means selecting the AI future they believe is most achievable.

Gaps widen with each model generation, ultimately translating into differences in the pace of intelligence expansion.

Market Consensus Forms: AI Companies' Scarcity Determined by Three Factors

From a capital perspective, financing details reveal a common signal: to enter the game, capital is willing to cede rights.

Anthropic's equity is highly sought after in the primary market; DeepSeek's investors accepted no voting rights and a five-year lockup; Zhipu's convertible bond investors subscribed at above-par prices, receiving no coupons during holding and accepting conversion prices with double-digit premiums to market prices.

The allocation of chips is even more noteworthy: Zhipu's top 20 investors received over 85% of equity and 88% of convertible bonds, with most allocations going to institutions with deep research, capacity, and willingness to hold long-term.

These capitals undoubtedly believe they have acquired scarce assets. But what is truly scarce?

What is truly scarce are foundational model assets capable of sustaining technological progress while maintaining independent commercialization paths.

Specifically, three factors matter: technology flywheel, capital efficiency, and commercialization prospects.

Anthropic's technology must translate into revenue, and revenue into profit, to justify discussions of a $2 trillion valuation. According to Bloomberg, its Q2 revenue exceeded $11.5 billion, surging over 14 times year-on-year and up 143% quarter-on-quarter, with adjusted operating profit turning positive for the first time.

DeepSeek turned extreme cost control into its narrative, achieving 'comparable performance at low cost.' According to The Information, its revenue reached about $70 million in the first seven months, nearly ten times 2025's full-year total. Its open-source strategy secured an ecological niche, with efficiency advantages supporting profitability, prompting capital to cede voting rights for entry.

Zhipu also outlined its path:

Technologically, the GLM series advanced to code generation, complex reasoning, and long-cycle agents, with full self-training pointing to the next generation. Its All-in-Infra strategy for domestic computing adaptation created another layer of differentiation—'intelligence ceiling × cost frontier' directly determines developer choices and usage scale.

Commercially, it achieved growth in both volume and price. H1 MaaS and API revenue reached RMB 825 million, up about 2736% year-on-year, accounting for 86.5% of total revenue; token usage surged over 40 times from the start of the year, with API average prices rising about 101%.

For these AI newcomers, the coming quarters are critical post-financing: whether new models maintain leadership, whether domestic computing power continues to reduce inference costs, whether enterprise revenue becomes high-quality recurring revenue, and whether new capital drives revenue to outpace cost curves.

If capital inflows, technology flywheels, and commercialization progress form a virtuous cycle, AI core assets will underpin long-term market confidence in the next era.

Conclusion

What long-term and industrial capitals seek is not short-term returns but entry tickets. Financing reflects capital's comprehensive judgment on technology, teams, commercial prospects, and industrial positioning. The ability to continuously secure large-scale, low-cost funds and convert them into model progress and client revenue is itself a difficult-to-replicate capability.

Zhipu's $5 billion terms make this explicit: zero-coupon bonds issued at a premium, with chips concentrated among institutions with deep research, capacity, and long-term holding Will (willingness). This is a form of shareholder structure screening.

Leading AI companies have secured more than just money—they've secured the right to continuously allocate resources. Anthropic may use NVIDIA's money to buy NVIDIA's computing power; Zhipu will keep pushing intelligence boundaries while laying infrastructure foundations.

Looking ahead, securing large funds grants a tech company entry to the second half, but that's just the beginning. Only by transforming money into next-generation models, more enterprise clients, and higher-quality revenue can a company remain at the center of the game.