The era of complimentary AI services appears to be concluding. Doubao, China's AI application with the highest monthly active users, has quietly introduced a subscription plan. The service now offers three pricing tiers: a Standard plan at 68 yuan per month, an Enhanced plan at 200 yuan, and a Professional plan at 500 yuan. This move signals a shift away from the industry's previous reliance on subsidies and free access.

For the past two years, Doubao aggressively promoted its large model to hundreds of millions of users through heavy subsidies. However, soaring computing power costs have become a significant burden. Data indicates that Doubao's daily token usage doubled in just three months, exceeding 120 trillion tokens. This surge is supported by ByteDance's massive capital expenditure, which reached hundreds of billions of yuan, alongside a year-on-year net profit decline of over 70%.

No major AI model developer can avoid this financial pressure indefinitely. Recently, both domestic and international firms have been releasing frequent updates. Beneath the surface of these technical advancements lies a deep anxiety about sustainability under the weight of computing bills.

In late April, OpenAI launched GPT-5.5, which is twice as expensive as its predecessor. Domestically, Zhipu increased its prices twice within a single month, while Kimi raised its input token price by 58%. Price hikes and the introduction of fees are becoming unavoidable realities for the industry.

The industry is now focusing on profitability. The past month could be termed a "super update month" for large models. Unlike previous cycles focused solely on benchmark performance, the latest releases are more pragmatic, targeting specific applications and commercial returns.

OpenAI's GPT-5.5 is designed for complex tasks and AI agents, with input pricing set at $5 per million tokens and output at $30 per million tokens. A Pro version is priced even higher. Shortly after, OpenAI released two more variants: GPT-5.5-Cyber, optimized for cybersecurity, and GPT-5.5 Instant, a free model for consumers, alongside a beta ChatGPT ad manager.

OpenAI's strategy involves serving high-value paid users and enterprises with its flagship Pro service, while using the free Instant version to maintain a broad user base and explore monetization through advertising. Each new model carries explicit commercial objectives.

Anthropic, which primarily serves the enterprise market, released Claude Opus 4.7 in April, showcasing significant improvements in software engineering just over two months after its previous update. Facing intensified competition from OpenAI and others in the business sector, Anthropic is compelled to accelerate its pace.

In China, Zhipu announced its third price increase this year in early April, raising prices by 10% with the launch of its new flagship model GLM-5.1. This followed a previous 30% hike just one month earlier. In late April, Kimi launched K2.6, increasing its API input price from $0.60 to $0.95 per million tokens, a 58% jump.

Despite these price increases, market demand remains strong. According to disclosures from the National Data Bureau in March, China's daily token usage has surpassed 140 trillion, a more than 1000-fold increase since early 2024. During Zhipu's 2025 earnings briefing, CEO Zhang Peng noted that despite an 83% rise in API pricing for Q1 2026, call volume still grew by 400%, indicating a supply shortage.

In the consumer market, the chorus for "free" services and "subsidies" is fading. Doubao, as the top AI app in China by monthly active users, has begun testing a paid subscription model with its three-tier plan.

ByteDance's approach is not merely about generating revenue. Details show that basic chat functions remain free, with fees applied to computationally intensive tasks like PPT generation, data analysis, and video production.

A Morgan Stanley report suggests that Doubao's move, as the most aggressive subsidizer in China's consumer AI sector, sends a strong industry signal: the user cultivation phase is largely complete, and the focus is shifting from subsidies to sustainable business models.

However, a disconnect exists between Chinese users' willingness to pay and their income levels. The 500-yuan monthly fee represents approximately 16% of the average monthly disposable income per capita in China.

This creates a delicate balancing act for domestic commercialization efforts. Setting prices too low fails to cover costs, while setting them too high risks driving users to competitors still offering loss-leading subsidies. Finding the right equilibrium will be a key challenge for companies.

Why is the large model industry accelerating commercialization in 2026? The pressure comes from multiple fronts simultaneously.

The primary driver is computing cost. An AI industry insider noted that while internet services traditionally had near-zero marginal costs, large models have constant or even increasing marginal costs, especially as context lengths grow.

In 2026, four tech giants—Google, Microsoft, Amazon, and Meta—are projected to spend a combined $725 billion on AI, a roughly 77% year-on-year increase. OpenAI disclosed that its computing expenditure for 2026 will reach $50 billion.

Domestically, ByteDance's 2025 net profit fell by over 70% year-on-year, with its net profit margin also declining sharply. Zheshang Securities estimates ByteDance's 2025 capital expenditure at approximately 160 billion yuan, with 90 billion yuan allocated to AI computing procurement.

A significant time lag exists between incurring these computing costs and generating revenue. Amazon stated in a shareholder letter that 2026 capital expenditures will only gradually monetize from 2027 to 2028. The industry widely faces short-term profit margin pressure from these substantial investments.

A situation of "revenue growth without profit growth" and "excessive investment" is unsustainable in capital markets. Following their earnings reports, the stock performances of the four tech giants diverged, with only Google, which has the clearest monetization path, convincing investors.

Structural changes in token consumption further amplify cost pressures. Zhipu's Zhang Peng previously explained that complex tasks now require long reasoning chains, repeated debugging, and error correction, consuming tens or even hundreds of times more tokens than simple Q&A sessions.

For example, Doubao's daily token usage exceeded 120 trillion in March 2026, doubling in three months and growing 1000-fold since its launch in May 2024. This is seen as the direct reason for its recent subscription model test.

Concurrently, the venture capital market is increasingly reluctant to fund these capital-intensive ventures. Although total funding in China's AI sector grew over the past year, the average deal size has shrunk. Computing investments requiring tens or even hundreds of billions of yuan exceed what most VCs can bear.

This has triggered a "race to go public." Zhipu and MiniMax are already listed in Hong Kong, while Moonshot AI (Kimi) and Stepfun are reportedly preparing for Hong Kong IPOs.

For listed companies or those planning listings, sustainable commercial metrics are necessary to support market valuations. Before the "burn cash for growth" model completely fails, large model developers must articulate a new narrative focused on profitability.

Anthropic's case is instructive. Its implied valuation in private markets once surpassed OpenAI's, not due to superior technology, but because of its higher penetration in the enterprise sector—paying customers accounted for 30.6% in a sample of US companies. Capital is willing to pay a premium for companies that can generate revenue.

In China, leading AI companies aspire to emulate Anthropic's success. Morgan Stanley predicts that top-tier model companies could achieve an Annual Recurring Revenue (ARR) of $1-1.5 billion by the end of 2026, potentially rising to $2.5-5 billion by the end of 2027, representing a 3-5x annual growth rate.

The commercialization acceleration that began in the spring of 2026 marks merely the start of a new survival game for the industry.