Written by | Wu Xianzhi

Edited by | Wang Pan

Driven by a significant surge in internal demand for AI, a certain company reached out to every domestic supplier it could find in the first half of this year. After evaluating performance and pricing, they ultimately signed a contract with one supplier. Surprisingly, the supplier returned in the afternoon, seeking to revise the agreement by removing the pay-per-use restriction and switching to a pure Token-based billing model. Essentially, they conveyed one core message:

"We can deliver, but it'll cost you more."

According to the supplier's revised terms, the procurement cost would skyrocket fivefold. The AI lead at the company commented, "Last-minute changes in commercial cooperation are a cardinal sin. Our internal processes strictly forbid such situations, so we had no choice but to blacklist this supplier and seek another."

The supplier's abrupt change highlights the ongoing exploration and refinement of AI payment models in China's commercial landscape. Suppliers must navigate various costs, including computing power allocation, the ecological gap between different products, and their own commercial inertia. While the marginal costs of traditional IT and SaaS are nearly negligible, the costs associated with large AI models and applications, whether for productivity coding or desktop agents, escalate non-linearly.

Anthropic has largely succeeded in establishing a ToB business model through coding, achieving notable success in stages.

The high-value task monetization model, built on enterprise subscriptions and API calls, is being emulated by domestic AI giants. Since May, AI cloud providers such as Baidu, Alibaba, Tencent, Huawei, and ByteDance have all emphasized their focus on delivering results and targeting high-value scenarios at their respective conferences. However, this path is not as glamorous as the visual presentations in the main conference hall suggest.

Model Giants Grapple with Mounting Costs

From 2023 to the present, AI has evolved from chatbots to Copilot assistance, and then to Agents and Cloud Agents.

The tasks that models can handle have grown increasingly complex, involving multi-round reasoning, tool calls, and long-chain execution, with Token consumption per call growing exponentially. As AI can accomplish more, demand has surged, meaning that users' single-use computing power consumption is on the rise, and the scale of use is continuously expanding.

A ten thousandfold increase in Token consumption does not translate to a corresponding ten thousandfold surge in demand from enterprises. Model capabilities have advanced rapidly, and model manufacturers, burdened with substantial physical computing power costs, are eager to monetize these capabilities to recoup their investments.

TechInsights data reveals that global data center GPU shipments only grew about 2.5 times during the same period. Meanwhile, the expansion of HBM high-bandwidth memory and CoWoS advanced packaging was only completed last year. The expansion cycles for heavy assets like wafer fabs, packaging lines, and memory production lines all increase linearly.

An investor believes that capital expenditures by tech giants will continue to rise in the race for computing power supply. Endless investment is unsustainable, and just as Anthropic took the lead in unlocking commercial value at the B-end, AI companies saw a potential path through the fog.

Anthropic discovered a viable path that allows clients to directly see results—coding. After signing a large-scale computing power cluster lease agreement with SpaceX in May, it experienced explosive growth, with its annualized revenue surpassing $45 billion.

The release of computing power supply has unleashed Anthropic's commercial potential and demonstrated the commercialization space of AI at the B-end. However, potential does not equate to achieving a balance between revenue and expenses in the short term; rather, it's about whether enterprises' AI expenditures can yield quantifiable benefits. AI, as a variable, causes costs to fluctuate with usage, which in turn is difficult to fix due to varying scenarios.

Faced with new business opportunities and ambiguous business models, many suppliers have established FDE teams under the guise of helping enterprises implement AI, engaging in in-depth co-creation with benchmark clients. For instance, Microsoft invested $2.5 billion to establish Microsoft Frontier Company's frontier deployment engineering, integrating 6,000 engineers, technical consultants, and sales teams to provide services directly within client enterprises.

In the era of traditional SaaS, suppliers would provide a price list and corresponding annual fees, and clients would deploy the system themselves after signing the contract. However, in the AI era, expenditures vary greatly across different industries, scales, and implementation scenarios. Clients struggle to articulate their needs clearly, and suppliers, burdened with costs, are unsure of the value of specific scenarios, leading them to dispatch personnel to track the entire process.

Tan Dai from Volcano Engine once proposed a framework for explanation: "The price per Token is rising, but the value created is rising even faster." While this logic holds, it leaves unresolved the question of who verifies that the value is rising faster, especially when sellers are under cost pressure and buyers lack a benchmark for evaluating effectiveness. This issue becomes even more thorny.

Perhaps considering the above situation, Tan Dai set an entry threshold: "If the revenue target is less than 1 billion, it's best not to do Agents; just do Skills instead." On the surface, suppliers are screening clients, but this reflects the difficulty for both model manufacturers and enterprise clients to calculate their accounts clearly, leading them to focus on implementation effectiveness first.

The confusion of "not knowing how much it's worth" is even more direct and practical in the actual business operations of large companies.

A Meituan insider revealed that the company spends 2 to 3 billion yuan annually on AI data procurement, while the entire R&D department's annual budget is only 1 billion yuan. This massive data investment has not yielded the desired results. For example, AI's accuracy in Meituan's core road network recognition scenario is only maintained between 60% and 70%, with a long way to go before implementation.

Despite the difficulties in implementation, AI represents the future, and the company's determination to invest remains unchanged. The insider told Photon Planet that Meituan's Wang Xing stated at a small-scale executive meeting this year, "We don't know when AI will take off; if we don't pursue it, we might not survive three years."

Suppliers cannot calculate costs clearly, and clients cannot calculate benefits clearly. When both buyers and sellers cannot calculate their accounts, the market spontaneously shifts toward the present reality, pursuing relatively controllable costs.

Tan Dai previously complained that the rumored revenue figures for Seedance were all incorrect and excessively high, causing him significant pressure. The source of this pressure can be glimpsed from the video generation product Jimeng, based on Seedance.

An informed source told Photon Planet that Jimeng consumes at least half of ByteDance's internal computing power resources. After actions such as canceling discounts and launching VIP models, only 10% of the costs have been recovered. The intense consumption of computing power without cost recovery affects business sustainability. Perhaps seeing that acclaim does not translate to profitability, some tech giants' leaders have abandoned the idea of competing in model development.

A Baidu insider told Photon Planet, "Robin said Baidu would not focus on model capabilities but only on distribution and product packaging." Previously, Baidu's benchmark product "Baidu Mirror" at the Baidu Creat conference was a typical case. The team utilized external model capabilities like Kling, focusing only on product packaging. Nevertheless, the commercialization of Mirror is still in its early stages.

The Calculations of Enterprise Clients

Model manufacturers respond by continuously increasing investments, introducing new products, setting new payment points, and accelerating commercialization. However, when it comes to specific implementation scenarios, the accounts do not add up.

It is not easy for enterprises to calculate their own cost accounts. When the responsibilities and relationships in implementation business are unclear, frontline business can only turn to controllable aspects, seeking certainty in input-output within this infinite AI loop.

Previously, J&T Express launched a work order quality inspection system, significantly reducing the secondary complaint rate for work orders by 23% in the Chinese market. However, during the promotion in other countries, local teams were relatively cautious, "worried that the cost of calling large models might not be recoverable," preferring to gradually promote it after the revenue model became clear. There is a deviation between the global efficiency and local sunk cost accounts, so the execution layer sometimes prefers to follow its own logic.

A leading AI short drama company, fearing the loss of business dominance, also focuses on control.

This company uses AI to produce short dramas, with an efficiency of up to one drama per day. AI covers many aspects, including script understanding, storyboarding, card selection, review, and editing. Even though there are off-the-shelf products like ByteDance's Xiaoyunque Agent on the market, the company still insists on in-house development and attempts to productize it for external services.

"Remote departments of large companies often underperform due to insufficient investment," the company's lead believes. While the production process can fully utilize AI, the toolchain must remain in-house. "Using someone else's Agent means the efficiency belongs to them and could be taken back at any time."

In fact, large companies are even more persistent than small companies in controlling costs.

Considering the continuous and massive investment black hole of AI, Meituan only uses small models of 4B and 35B internally, resorting to the "most expensive and best" large model only for testing and verification. Their logic is simple: "If the top models can't handle it, others have no chance either."

Photon Planet learned that, with unclear returns, Meituan believes in controlling investments first and has summarized a set of implementation experience based on this: use the most expensive model to explore the capability ceiling, then use small models for daily tasks, thereby keeping Token expenditures within a controllable range.

This aligns with the approach of the most cost-conscious logistics companies.

In data annotation, J&T Express has developed its own fine-tuning model with billions of parameters for local closed-loop verification. Only in scenarios involving global scheduling and direct user interaction do they call external large models to ensure effectiveness. Expensive models are prioritized for essential uses, while other aspects are handled by the technical department.

Interestingly, industries with lower expenditures, such as online education, are also clear about allocating resources effectively.

Yangcong Xueyuan told Photon Planet that diagnosis, feedback, and personalized recommendations are currently widespread AI applications. "Mistakes have controllable impacts and can be corrected promptly." Introducing verification mechanisms in high-risk scenarios avoids situations where models operate independently without oversight.

The education industry has an extremely low tolerance for errors, as a single incorrect answer could lead to parent refunds. Of course, the unspoken reason remains cost. The cost of improving from 80 points to 95 points grows exponentially. Instead of pursuing an expensive 100 points, it's better to place AI at a financially balanced 70 points.

The emergence of these control strategies not only indicates that enterprises are still exploring certainty in return on investment but also reveals that model manufacturers lack effective means to bind clients when providing AI services.

We inquired with multiple companies and received similar answers. Some companies directly stated, "Suppliers have indeed expressed the idea of locking us in, but we don't want to be locked in."

Model iteration speeds are fast, with ever-changing SOTA models, and clients are unwilling to sign long-term contracts. Moreover, the business model of AI in productivity lacks the three conditions for traditional SaaS to bind clients: precipitated data, solidified workflows into systems, and high switching costs due to integrating upstream and downstream.

Yangcong Xueyuan did not develop its own general-purpose large model but integrated Volcano Engine's Doubao and DeepSeek. If the base model's price increases, they switch to another; if capabilities improve, they switch accordingly. Any dependency built by suppliers is reset to zero with each version iteration.

As a productivity tool, AI Coding and Coding Plans all adopt quota-based packages, lacking effective means to bind clients. "Coding Plans are essentially distribution channels for model capabilities; we procure Coding Plans from multiple manufacturers ourselves."

After price increases this year, more usage is provided at the same price, which should suffice. However, as business volumes increase, costs continue to rise. To control costs, many clients simultaneously use ByteDance, Alibaba, and Tencent, comparing prices and dynamically adjusting accounts internally. When asked how they would respond to the expectation of continuous Token price increases, manufacturers stated that they might consider setting up their own servers.

Suppliers are acutely aware of this issue and are striving to bridge the chasm between technological capabilities and business model integration by deploying human-centric services through the establishment of FDE (Full-Service Digital Enablement) teams. Concurrently, certain manufacturers have embarked on restructuring their product portfolios. For instance, Alibaba recently consolidated three enterprise-level Agent products to streamline resource allocation.

Two Accounts, One Door

The discrepancy in account management between suppliers and clients poses a significant obstacle to AI-driven productivity transformation.

Model manufacturers must capitalize on product deployment to offset the costs associated with physical computing power. Tokens represent the intellectual consumption of chips, models, and engineering prowess. The linear increase in physical supply fails to keep pace with the exponential surge in demand. Focusing solely on video models, industry leader Jimeng consumes vast computing resources but only recoups approximately 10% of its costs. If this is the situation for the market leader, the prospects for other players are even more dire.

Clients' accounts are geared towards achieving specific outputs through sustained investment. Meituan's substantial data acquisitions aim to replace an internal scenario whose value remains unquantified. J&T Express's work order quality inspection system has seen successful implementation in the Chinese market but still grapples with ROI (Return on Investment) calculation challenges when expanding internationally. Both companies employ different metrics to assess AI output and must strive for greater universality.

Anthropic offers a valuable reference point. The company views Coding as a foundational element and expands into the broader white-collar and traditional software markets (encompassing traditional, SaaS, and cloud infrastructure), providing clear replacement costs and establishing a unified metric for both buyers and sellers.

There are emerging signs in China as well. AI-powered customer service has reached a relatively mature stage but only replaces customer service representatives earning modest hourly wages. When it comes to internal processes lacking a standard price, suppliers and clients find themselves entangled in endless negotiations.

Once enterprises can identify tangible value, their investments will become more certain. J&T Express has reported that AI has significantly reduced manual intervention rates in problem resolution, customer service, and other areas, resulting in labor cost savings. AI-optimized intelligent routing has surpassed expectations, enhancing operational efficiency. The accumulation of these tangible benefits has demonstrated to enterprises the concrete value of AI.

Tan Dai's assertion that companies with revenues under 1 billion should refrain from developing Agents, Baidu's emphasis on distribution and product packaging over model development, and Meituan's use of the most expensive models solely for testing are merely interim strategies to avoid escalating costs.

While technology can already provide answers regarding feasibility, the business world has yet to determine its worth.

If enterprises can clearly define the cost of a specific link, or if suppliers can offer a list of alternative values and incorporate them into contracts, the day will come when they can confidently sign agreements in the morning without regretting them in the afternoon.

And at this juncture, when AI truly integrates into productivity, it will set the business flywheel in motion.