First-half 2026 financial results have revealed an unusual trend: both Z.AI and MiniMax more than doubled their revenue, yet simultaneously reduced selling expenses. Z.AI reported revenue of 954 million yuan, a 399.7% year-on-year increase, while its sales and marketing expenses dropped 14.8% to 178 million yuan. MiniMax delivered a similar performance, with revenue of 786 million yuan representing a 283.1% increase, and sales and distribution costs falling 17.9% to 181 million yuan. This curve, showing revenue soaring while sales costs decline, is something traditional industries or SaaS sectors could not chart, and it highlights a pivotal shift in the growth engine of large model companies.

Part 1: Shifting from Human-Driven Promotion to Model-Driven Sales

The reduction in selling expenses is not a result of cost-cutting discipline, but rather a fundamental change in growth mechanics. Z.AI's confidence to trim sales costs stems from an evolution in how it acquires customers through its platform. In the first half of last year, Z.AI's revenue was primarily driven by local deployments of enterprise-grade general models, which involved bundling models into client data centers with transaction values ranging from millions to tens of millions of yuan. Revenue recognition depended on delivery and acceptance, while repeat business relied on relationship management. This business model inherently required a large sales force to manage projects and cultivate relationships. However, in the first half of this year, open platform and API revenue surged from 29.1 million yuan in the same period last year to 825 million yuan, now accounting for 86.5% of total revenue. Clients no longer need sales representatives to pitch them; they can directly self-serve on the platform, paying based on token usage. As the transaction method has changed, the role of sales has naturally diminished. Z.AI's financial report states this plainly, noting the transition from one-time revenue recognition to recurring revenue, and highlighting that for the first time, the company has gained predictable recurring income. This trend is further confirmed by Z.AI's subsequent actions, such as opening an official flagship store on Tmall on September 2nd to sell its GLM Coding Plan subscription packages, with personal plans starting at 118 yuan per month, Pro at 538 yuan, and Max at 1078 yuan. Placing large model subscription packages on an e-commerce platform was unimaginable in the traditional software era, but it is happening today. The customer acquisition cost for models is moving closer to standard e-commerce metrics.

MiniMax follows a similar logic, though its path has been different. Initially known for consumer-facing products like Hailuo AI and Xingye, MiniMax was often labeled a "C-end company." Yet, in the first half of this year, its open platform and enterprise services revenue grew 703.1% to 497 million yuan, accounting for 63.4% of total revenue, surpassing its AI-native products (287 million yuan) for the first time as the largest revenue source. MiniMax's user growth is primarily driven by product strength and word-of-mouth, rather than paid acquisition or field promotions; it is the model's own capability that attracts developers and enterprises to join organically. Its enterprise clients and developers have exceeded 2 million, a tenfold increase from the end of 2025. Token consumption in July reached 20 times the January level, and ARR surpassed $800 million in August. During the earnings call, MiniMax disclosed that To B business constitutes about 80% of its ARR, with To C accounting for 20%. A year ago, the revenue structure was primarily consumer-oriented; today, B-end business has become the dominant force. This pace of structural adjustment cannot be achieved through sales efforts alone. When a model's capability is strong enough and inference costs are low enough, developers and enterprise clients will proactively migrate, not because a salesperson made a call, but because failing to use the model would leave them lagging in competitiveness. This form of organic growth driven by product capability is fundamentally reshaping the sales expense structure of AI companies.

However, both companies optimizing their selling expenses simultaneously also reflects a deliberate strategic choice to redirect resources from high-cost, low-marginal-utility ground promotions toward model R&D and infrastructure upgrades. This is not just an improvement in technical capability, but also a strategic pivot and cost structure optimization, with both elements influencing each other. The decline in selling expenses is not merely a natural outcome of having a superior product that doesn't need promotion; it is also a proactive management decision to allocate limited resources toward more efficient areas.

Section 2: Deep Adoption vs. Service Gaps

While both companies have cut marketing budgets, the underlying business models driving these decisions differ. Z.AI derives 86.5% of its revenue from its open platform and API. The advantage of the API business lies in self-service recharging, pay-as-you-go pricing, and recurring revenue, but it also carries a clear disadvantage: customer switching costs have layers, and significant effort is needed to drive users toward deeper adoption. Developers and small to medium-sized enterprises using models at a superficial level can indeed switch by altering just a few lines of code. These customers are the most price-sensitive and eager to experiment with new models, showing almost no loyalty. If a competitor offers a comparable product at a lower price, this revenue segment could quickly evaporate. However, for enterprise clients that have deeply integrated AI into their core production processes, the situation is entirely different. According to a Zapier survey, in 2026, only 42% of enterprises attempting to switch AI vendors reported a smooth process; the remaining 58% encountered failures or costs that far exceeded expectations. This is because AI systems involve vendor-specific APIs, proprietary training data, custom deployment tools, and deep workflow integrations, none of which can be seamlessly migrated between vendors. In other words, while the cost of switching API calls is low, once an enterprise client embeds a model deep into its core business, a de facto technology lock-in is created. This lock-in effect partially offsets the risk of zero switching costs, making customer stickiness stronger than expected and adding logic to the decision to reduce sales staff.

Z.AI's API customer base is experiencing a divergence: shallow-tier customers face extremely low switching costs and may churn with any shift in price or performance. Deep-tier customers, meanwhile, benefit from strong technology lock-in due to private training data, custom deployments, and workflow integrations. Token call volume on Z.AI's MaaS open platform has grown over 40 times since the beginning of the year, paid daily active users have increased 603%, and the average daily call volume from its top ten customers has surged 98 times. These figures confirm the trend of deep integration, with leading clients embedding Z.AI's models into their core operations. But the real test lies in whether Z.AI, after optimizing high-cost promotions, can drive deep customer conversion at a lower cost. Deep-tier users require API stability guarantees, fine-tuning support, industry solution references, and a robust community ecosystem. These capabilities were previously partially coordinated by the sales team; in the future, they will need to be delivered through more efficient self-service tools, comprehensive technical documentation, and an active developer community. Z.AI's potential risk is whether the pace of this functional transition can keep up with customers moving from shallow trials to deep integration. If the transition is slow, the stickiness advantage of deep customers may not materialize before price-sensitive shallow customers have already churned.

MiniMax's revenue structure has shifted from a consumer focus to a B2B emphasis, and its expenses have declined due to organic user growth, making this a noteworthy structural change. Where has the money saved from reduced sales costs gone? If the cuts were aimed at inefficient field marketing and brand advertising, while redirecting resources to customer success, solutions architects, and after-sales technical support teams, then this would be a precise reallocation, matching lower acquisition costs with higher customer service density. However, if the expense reduction represents an overall contraction, relying solely on organic user growth while neglecting after-sales capabilities, problems will accumulate: B2B enterprise clients, especially large overseas accounts, require not just a good model, but also SLA commitments, security audits, customized fine-tuning, and dedicated responsiveness. These services are not inherently bundled with the API; they require human delivery. Among MiniMax's 2 million global enterprise clients and developers, the number of head customers requiring dedicated service may be limited, but as the share of B-end revenue continues to rise, the absolute number and service depth required from top clients are growing. Sales teams can be reduced, but delivery and service capabilities cannot be; otherwise, higher revenue will lead to a faster accumulation of renewal risks. The real risk is not the reduction in sales staff, but whether the cuts target unnecessary sales overhead or essential customer success functions. This depends on management details beyond the financial statements, and public information currently cannot provide a definitive answer.

Part 3: The Model as Salesperson, Yet Sales Beyond the Model

The competition among large model companies has evolved from selling software to selling intelligence. Traditional software companies typically have sales expense ratios between 20% and 40%, requiring sales forces to convince clients of the need for their software. But the logic for large models is different: clients do not adopt them because they are convinced, but because failing to use them would put them at a competitive disadvantage. When a model's intelligence level is sufficiently high, the decision cost of adopting it approaches zero, and the value of a salesperson naturally approaches zero as well. This explains why both companies can simultaneously cut sales expenses while revenue continues to soar. It also explains why their R&D expenses remain over ten times higher than their sales budgets. Z.AI's R&D spending of 2.131 billion yuan is 12 times its sales costs of 178 million yuan, while MiniMax's R&D expenses of 2.0 billion yuan are 11 times its sales costs of 181 million yuan. The core asset of a large model company is not the sales team, but the model itself. The company that can push its model to a better experience level in the shortest time will acquire more customers at a lower cost. The decline in sales expenses is both a natural result of improved model capability and a strategic choice by management to prioritize resources for R&D.

Large models no longer require the traditional sales force. Or rather, they no longer need traditional sales in the conventional sense. As clients shift from asking "should we use AI" to "how can we use AI to maximize value," the role of sales transforms from convincing clients to sign contracts to helping clients succeed. This is more akin to a hybrid role combining engineering, consulting, and sales. Consequently, the function of sales gradually moves from pre-sales to during-sales and after-sales support. OKRs shift from user volume to user retention. The model itself serves as the best salesperson, with people more like workers serving the customers that the model naturally attracts.