The commercialization of AI remains a muddled mess for major tech giants.
Driven by a sharp surge in internal AI demand, a certain company approached every accessible domestic supplier in the market during the first half of this year.
After comprehensively weighing performance against cost, they finally signed a contract with one vendor. Unexpectedly, mere hours after the agreement was finalized in the morning, the supplier reached out again in the afternoon, requesting to remove the per-payment usage cap and switch entirely to a pure token consumption billing model. Their entire message boiled down to one core point:
"We can deliver, but you'll have to pay more."
Under the supplier's proposed adjustment, procurement costs would jump 5 times. The company's AI lead stated, "Reneging on a commercial agreement at the last minute is a cardinal sin. Our internal policies strictly prohibit this kind of behavior, so we had to blacklist this supplier and switch to another partner."
This last-minute reversal by the vendor reflects that domestic AI players are still exploring and refining their monetization models for commercial use cases. Suppliers must account for a wide range of costs, including the allocation of computing power expenses, ecosystem gaps across different products, and their inherent business inertia. Traditional IT and SaaS have near-zero marginal costs, but for large language models and AI—whether productivity-focused coding tools or desktop agents—the underlying costs grow non-linearly.
Anthropic has essentially validated a viable B2B business model centered on coding, achieving phased success.
The high-value task monetization model built on enterprise subscriptions and API calls is now being replicated by major domestic AI players. Since May, AI cloud providers including Baidu, Alibaba, Tencent, Huawei, and ByteDance have all outlined their strategies focused on outcome delivery and high-value scenarios at their respective industry conferences. However, this path is far less glamorous than the polished stage presentations suggest.
Model Giants Dancing Under Cost Pressures
Since 2023, AI has evolved from basic chatbots to copilot assistance, and now to agents and cloud-native agents.
Models are handling increasingly complex tasks: multi-turn reasoning, tool invocation, and long-sequence execution have caused token consumption per call to grow exponentially. As AI capabilities expand, demand has exploded—simply put, individual users consume far more computing power per session, while overall adoption scales rapidly.
A ten-thousand-fold surge in token usage does not mean enterprises have corresponding ten-thousand-fold growth in actual demand. With model capabilities outpacing current market conditions, model vendors burdened by rigid physical computing costs are urgently seeking transparent pricing for their offerings to recoup cash flow.
According to TechInsights data, global data center GPU shipments only increased by roughly 2.5 times over the same period. Meanwhile, high-bandwidth HBM memory and advanced CoWoS packaging only completed capacity expansions gradually last year. Expansion cycles for capital-intensive facilities like fabs, packaging lines, and memory production lines follow a strictly linear growth trajectory.
One investor believes that major tech companies will continue ramping up capital expenditures to compete for limited computing power supply. Unsustainable infinite spending cannot last long, and Anthropic's early success in unlocking commercial value for the B2B sector has illuminated a potential path forward for AI vendors navigating uncharted territory.
Anthropic identified a clear, results-driven path to value—coding. After signing a massive computing cluster lease agreement with SpaceX in May, the company saw explosive growth, with its annualized revenue surpassing $45 billion.
Expanded computing power supply unleashed Anthropic's commercial potential, proving the vast B2B monetization opportunity for AI. This potential does not guarantee near-term profitability, but rather depends on whether enterprises can find quantifiable returns on their AI investments. AI introduces a variable that makes costs fluctuate with usage, while usage itself remains unpredictable due to inconsistent scenario requirements.
Facing this new business landscape and ambiguous monetization models, many suppliers have established dedicated FDE (Field Deployment Engineering) teams, claiming to help enterprises implement AI solutions through deep co-creation with benchmark clients. For example, Microsoft invested $2.5 billion to launch its Microsoft Frontier initiative, integrating 6,000 engineers, technical consultants, and sales staff to deploy teams directly on-site at customer premises.
In the traditional SaaS era, suppliers would provide price lists with corresponding annual fees, allowing clients to deploy solutions independently after contract signing. But in the AI era, costs vary dramatically across different industries, enterprise sizes, and implementation scenarios. Clients cannot clearly articulate their exact needs, while vendors burdened by heavy costs struggle to quantify the value of specific use cases—hence the need for full-time on-site support.
Tan Dai from Volcano Engine once proposed a framework: "While per-token prices are rising, the value generated by each token grows even faster." This logic holds theoretical merit, but it avoids the critical question of who verifies that value is actually increasing faster—especially when vendors face intense cost pressure and clients lack clear metrics for measuring outcomes, this problem becomes far more intractable.
Likely in response to these challenges, Tan Dai set a practical threshold: "If your annual revenue hasn't reached 1 billion RMB, you'd better hold off on building full agents—focus on developing a single core skill instead." Ostensibly this guidance helps clients prioritize their efforts, but it also reveals that both model vendors and enterprise customers struggle to accurately calculate ROI, forcing them to focus first on tangible implementation results.
This confusion of "not knowing how much something is worth" manifests even more directly in the day-to-day operations of large enterprises.
An internal source at Meituan revealed that the company spends 2 to 3 billion RMB annually on AI data procurement, while the entire R&D department's total annual budget is only 1 billion RMB. This massive data investment has not delivered the expected results: for example, in its core road network recognition scenario, AI accuracy remains stuck between 60% and 70%, leaving significant gaps before full production deployment.
Despite implementation challenges, AI represents the future, and the company remains fully committed to its investment strategy. The source told Photon Planet that Meituan's founder Wang Xing stated during a recent small executive meeting: "We don't know when AI will achieve mass adoption, but if we don't invest now, we may not survive the next three years."
Suppliers cannot accurately calculate their costs, and customers cannot clearly define their returns. When both sides cannot reconcile their books, the market naturally gravitates toward practical, near-term realities, prioritizing relatively predictable cost control.
Tan Dai previously complained that external reports about Seedance's revenue figures were all incorrect and overinflated, creating unnecessary pressure for him. This pressure is evident from the performance of Dreamina, a video generation product built on Seedance.
An insider told Photon Planet that Dreamina consumed at least half of ByteDance's total internal computing resources. After a series of measures including repeated price hikes and reduced free tier benefits, the product only recouped roughly 10% of its total costs. Excessive computing power burn and insufficient returns threaten business sustainability—and this underwhelming performance has led some major tech leaders to abandon their ambitions of competing in raw model development.
A Baidu employee told Photon Planet: "Robin Li has stated that Baidu will no longer focus on foundational model capabilities, instead prioritizing distribution and product packaging." A prominent example is "Baidu One Shot," a flagship product showcased at Baidu's Creat Conference. The team leverages external model capabilities like Keling, only handling the product layer themselves. Even so, the commercialization of One Shot remains in its very early stages.
Enterprise Customers' Calculated Strategies
Model vendors' approach is to continuously increase investment, launch innovative new products, introduce new monetization touchpoints, and accelerate commercialization. However, when these efforts translate to real-world implementation scenarios, the math no longer adds up.
Calculating clear ROI for AI initiatives is no easy task for enterprises. When implementation responsibilities and accountability lines are blurred, frontline teams naturally shift focus to areas they can control, searching for predictable returns within this endless AI implementation cycle.
J&T Express previously launched an intelligent work order quality inspection system that reduced secondary complaint rates by 23% in the Chinese market. However, during international rollout, local teams remained cautious, "worried that large model invocation costs would never be recouped," preferring to wait until the return model was fully validated before expanding globally. Misalignments between global efficiency goals and local sunk cost considerations often lead implementation teams to follow their own pragmatic logic.
A leading AI short drama company, concerned about losing control of its core business, has similarly prioritized full ownership of its tech stack.
This company uses AI to produce short dramas, achieving maximum throughput of one full episode per day. AI automates many steps including script comprehension, storyboarding, asset generation, review, and editing. Even though off-the-shelf solutions like ByteDance's Skylark Agent are available on the market, the company insists on building its entire system in-house, and is even attempting to productize its tools for external clients.
"Large corporations' peripheral teams cannot execute well enough—they lack sufficient dedication," the company's founder argues. "AI can handle the entire production workflow, but the toolchain must remain fully under our control. If we rely on someone else's agent, their efficiency becomes our bottleneck, and access could be revoked at any time."
In reality, large enterprises are far more obsessive about cost control than smaller companies.
Given AI's massive and ongoing resource consumption, Meituan only uses small 4B and 35B parameter models internally, reserving the "most expensive, highest-performing" large models exclusively for testing and validation purposes. Their logic is straightforward: "If even the top-tier models cannot solve a problem, no other model stands a chance."
Photon Planet learned that while returns remain unclear, Meituan's internal strategy prioritizes controlling upfront investment, leading to a proven implementation playbook: first use the most powerful model to identify the upper limits of capability, then deploy smaller, cost-effective models for daily operations, bringing total token spending under manageable thresholds.
This philosophy aligns perfectly with the cost-conscious strategies of logistics leaders.
For data annotation tasks, J&T Express fine-tunes its own in-house multi-billion parameter models to maintain fully closed-loop local validation. External large models are only invoked for global scheduling and customer-facing scenarios to guarantee optimal outcomes. Premium models are reserved strictly for use cases that absolutely require their capabilities, while engineering teams optimize all other workflows internally.
Interestingly, industries with naturally low token consumption like online education also demonstrate a clear understanding of allocating resources strategically.
Onion Academy told Photon Planet that AI is most widely deployed for diagnostic assessments, real-time feedback, and personalized learning recommendations. "Errors in these areas have manageable impacts and can be corrected quickly." They implement validation mechanisms for high-risk scenarios to prevent models from operating unsupervised.
The education industry has an extremely low tolerance for errors—even one incorrect answer could drive parents to request refunds. The unspoken underlying reason remains cost: advancing from 80% to 95% accuracy causes costs to rise exponentially. Instead of chasing the prohibitively expensive 100% accuracy threshold, they prioritize deploying AI at the 70% level where costs and returns are balanced.
The emergence of these cost-control strategies not only shows that enterprises are still searching for predictable ROI, but also reveals that model vendors lack effective mechanisms to lock in customer loyalty when delivering AI services.
We surveyed multiple companies and received nearly identical responses, with one enterprise stating explicitly: "Suppliers have clearly expressed their desire to lock us into long-term arrangements, but we have no interest in being locked in."
Models iterate at an extremely rapid pace, with state-of-the-art performance becoming outdated quickly, making customers reluctant to sign multi-year contracts. Furthermore, productivity-focused AI business models lack the three core lock-in factors that traditional SaaS leverages: persistent data accumulation, hardened workflow systems, and high switching costs created by deep integration with upstream and downstream ecosystems.
Onion Academy has not built its own general-purpose large model, instead integrating services from Volcano Engine's Doubao and DeepSeek. When one base model raises prices, they simply switch providers; when another model improves its capabilities, they migrate their workloads. Any minor vendor dependency that develops can be completely eliminated with a single version upgrade.
For productivity-focused AI coding tools, coding plans are typically structured as fixed-allowance packages, lacking effective customer retention mechanisms. "Coding plans are essentially distribution channels for model capabilities—we ourselves purchase coding plans from multiple different vendors."
After this year's price increases, the same subscription tier now offers higher usage quotas, which should theoretically be sufficient. But as business volumes grow, total costs continue to climb. To rein in expenses, many customers simultaneously use services from ByteDance, Alibaba, and Tencent, continuously comparing prices across providers and dynamically allocating usage across multiple internal accounts. When asked how they would respond to expectations of ongoing token price hikes, some enterprises noted they are not ruling out building their own on-premise server infrastructure.
Suppliers are well aware of this dynamic, attempting to bridge the gaps in technical and business model lock-in by expanding their FDE teams with dedicated on-site personnel. Meanwhile, some vendors are restructuring their product portfolios—for example, Alibaba recently consolidated three separate enterprise agent products to concentrate resources more effectively.
Two Sets of Books, One Unopened Door
The misalignment between suppliers' cost calculations and customers' return expectations is slowing AI's transformative impact on productivity.
Model vendors must leverage real-world product implementations to amortize their physical computing infrastructure costs. Behind every token lies the intellectual investment of chips, model development, and engineering talent. The linearly expanding physical supply of computing hardware cannot keep up with exponentially growing demand. Even in the specialized domain of video generation, Dreamina—an undisputed market leader—has burned through massive computing resources while only recouping roughly 10% of its costs. If top players face these challenges, smaller competitors fare no better.
Customers, meanwhile, want to identify predictable, measurable outputs through sustained investment. Meituan's massive spending on data procurement aims to replace internal workflows whose value cannot yet be clearly quantified. J&T Express's work order inspection system delivered strong results in the Chinese market, but still struggles with ROI calculations during international expansion. Both companies use different metrics to measure AI's impact, and both are searching for greater universal applicability.
Anthropic provides a valuable reference point. The company built its foundation on coding, then expanded into broader white-collar productivity and traditional software markets (combining legacy systems, SaaS, and cloud infrastructure), creating clear replacement cost benchmarks that unify measurement standards for both buyers and sellers.
Early signs of this trend are emerging domestically: AI customer service solutions are relatively mature, but they only replace human agents earning modest hourly wages. When it comes to internal workflows with no established market pricing, suppliers and customers get trapped in endless cycles of negotiation.
Once enterprises can clearly identify tangible, predictable value, their investment levels become equally predictable. J&T Express reported that AI has significantly reduced human intervention rates for exception handling and customer service workflows, cutting labor costs. AI-powered intelligent routing optimization has exceeded performance expectations, driving operational efficiency gains. These accumulating, verifiable returns demonstrate concrete, real-world value for enterprises.
Tan Dai's advice to avoid agent development under 1 billion RMB in revenue, Baidu's strategic shift away from foundational model development toward product packaging, and Meituan's practice of limiting top-tier model usage to testing are all pragmatic workarounds to prevent runaway costs.
Technology has already proven that AI can deliver results, but the business world has not yet answered whether those results are worth the investment.
The day enterprises can precisely calculate costs for every workflow step, or suppliers can deliver clear value replacement commitments written directly into contracts, will be the day when no supplier reneges on agreements signed in the morning. At that moment, AI will fully integrate into core productivity workflows, and its commercial flywheel will truly begin spinning.