Alibaba Navigates by the Dark Side of the Moon, Large Models Weigh the Scales

Author｜Xinjian

Editor｜Xiaobai

Illustration｜AI Generated

Produced by｜Qiangdiao Next On August 6, DeepSeek added a note in its official API documentation: It plans to raise API service pricing across the board in the near future, with a "significant expected increase," and will announce specific plans later.

A day later, Reuters, citing two sources familiar with the matter, reported that Alibaba plans to introduce new commercial licensing conditions for its next-generation flagship model, Qwen3.8-Max: Model weights will remain open, but large commercial users may need to share a portion of their revenue with Alibaba. The revenue-sharing ratio is still under discussion, and Alibaba has not yet officially released the licensing terms.

The "price hike" logics of the two companies differ. DeepSeek is raising the unit price for model calls, while Alibaba is attempting to alter revenue distribution after opening up model weights. However, both moves reveal that domestic large models are crossing the same threshold in commercialization: Low prices and open access got models into more products; now, fees and price adjustments target those already doing big business with models.

Over the past two years, model companies have competed to sell Tokens at lower prices. Next, as cloud providers, inference platforms, and application companies earn money using these models, the question becomes: How much can model companies take back?

───

01

　Low Prices Aren’t Over, but Who Gets Charged Has Changed　■

DeepSeek’s currently announced prices for V4-Flash are 1 yuan per million Tokens for cache-miss inputs and 2 yuan for outputs. The corresponding prices for V4-Pro are 3 yuan and 6 yuan. The new prices have not yet been announced officially, except for a clear warning that the overall increase will be substantial.

A "significant increase" could easily be interpreted as the end of the price war, but it depends on the starting point.

According to research firm Artificial Analysis, the average cost for V4-Flash to complete a standard test is approximately $0.03, compared to $0.86 for Kimi K3, $1.86 for OpenAI GPT-5.6 Sol, and $3.15 for Anthropic Claude Fable 5. Even with a substantial price hike, DeepSeek may not lose its low-cost advantage.

Thus, DeepSeek seems to be repairing its unit revenue from extremely low prices rather than abandoning cost-effectiveness. The company has not explained whether the price hike is driven by demand, computing power, new model costs, or a proactive move to improve commercialization.

Alibaba’s approach goes further. Previously, Alibaba could charge for Qwen calls deployed on Alibaba Cloud, but once clients downloaded the open weights to their own data centers, Alibaba typically received no ongoing revenue. The new license reported by Reuters aims to close this commercialization gap: Weights can still be downloaded, deployed, and modified, but if clients package the model into a service and generate significant revenue, renegotiation is required.

This means the charging logic extends from "how many Tokens were used" to "how much money was made with the model." The former sells computing power and calls, while the latter competes for the model’s value share in the industrial chain.

───

02

　Alibaba Navigates by the Dark Side of the Moon　■

Kimi K3 from the Dark Side of the Moon has already provided a more complete example.

Kimi K3’s public license stipulates that if the licensee operates a MaaS (Model as a Service)—offering model inference or fine-tuning capabilities to third parties as a service—and its total revenue, including affiliates, exceeds $20 million over 12 consecutive months, a separate agreement with the Dark Side of the Moon is required before commercial use.

The license does not specify a fixed revenue-sharing ratio. Reuters, citing anonymous sources, reported that the Dark Side of the Moon demands up to 30% revenue sharing in actual negotiations. The license also sets boundaries: Simply embedding the model into end-user products with specific functions is not automatically considered MaaS. Pure internal use, calls through official products or certified inference partners, are exempt. Commercial products with over 100 million monthly active users or monthly revenue exceeding $20 million must prominently display "Kimi K3" on their interfaces.

The focus of this design is not to charge all developers but to preserve the long-tail ecosystem while intercepting the commercial channels most likely to bypass official APIs: third-party inference platforms and model service providers.

If Alibaba adopts a similar approach, the change will be more profound than a simple API price hike. The current Qwen3 main force models use the Apache 2.0 license, allowing royalty-free use, modification, and distribution. If Qwen3.8-Max switches to a custom license with commercial thresholds, Alibaba can still benefit from the dissemination and developer scale enabled by open weights but will no longer promise permanent free use for large-scale commercial applications.

Open weights thus shift from a product philosophy to a tiered pricing tool: Individuals and small-to-medium teams contribute to the ecosystem, while large clients contribute revenue; self-deployment expands coverage, while commercial licenses recover value once scale is achieved.

Strictly speaking, "open weights" does not equal "open source and permanently free." Weights being downloadable only determines whether the model can be deployed locally. Whether it can be used commercially without conditions, whether MaaS can be provided, and at what scale payments are required are still determined by the license.

───

03

　Alibaba Seeks to Regain Channel Bargaining Power　■

Alibaba is now in a position to attempt this step because Qwen is no longer just a model project reliant on subsidies to acquire clients.

Qwen3.8-Max, unveiled on August 3, boasts 2.4 trillion parameters and uses a Mixture of Experts architecture, activating approximately 95 billion parameters per request. It subsequently became the highest-ranked Chinese text model on Arena.AI and ranked second globally in visual benchmarks.

In tests more focused on enterprise workflows, the results were equally strong. Artificial Analysis’ Agentic Index, which evaluates tool invocation, task planning, and multi-step execution capabilities, placed Qwen3.8-Max in the global top tier with a score of 58 as of August 9.

Leaderboards do not directly determine enterprise procurement, but they strengthen Alibaba’s bargaining position with large clients on commercial terms.

A more direct signal comes from the cloud business. In the quarter ending March 2026, Alibaba Cloud Intelligence Group’s revenue grew 38% year-over-year to 41.63 billion yuan. AI-related products accounted for 30% of external client revenue. Alibaba also stated that its AI investment over the next three years would exceed the previously announced 380 billion yuan plan, with management prioritizing market share expansion over short-term profit margins.

Thus, Alibaba clearly cannot settle for just exchanging open weights for influence. The next competition is not just about download counts but about who controls access points, enterprise clients, and settlement relationships.

Third-party inference platforms are the most sensitive layer. They download weights, optimize inference on their own clusters, and sell them to clients via APIs. The model capabilities come from Alibaba, but the computing power and client relationships remain with the platforms. Under traditional permissive licenses, the larger the platforms grow, the more potential cloud revenue Alibaba loses.

Revenue sharing attempts to recapture this overflow value for the model company: earning from Qwen calls on competing clouds and potentially directing some clients to official clouds and certified partners. The license is not just a legal document but also a channel policy.

However, Alibaba’s bargaining power is not unlimited.

If commercial terms become too onerous, large clients can continue using older Apache-licensed models or switch to alternatives like DeepSeek, which remains fully free under Apache 2.0. Zhipu’s GLM-5.2 uses the MIT license with no revenue thresholds, and Meta’s Llama imposes only a 700 million monthly active user threshold without revenue sharing.

Alibaba must carefully consider whether its fee structure will drive clients away. Qwen’s leading capability, migration costs, and tool ecosystem will ultimately determine how much Alibaba can collect.

───

04

　Revenue Sharing Is Harder to Enforce Than Price Hikes　■

The logic of API price hikes is straightforward: Model companies announce unit prices, and clients settle based on Tokens used. Revenue sharing, however, requires first determining how much revenue the model has generated.

For pure MaaS platforms, this question is relatively clear. But in code tools, customer service systems, and enterprise software, the model is only part of the product. Should revenue sharing be based on the entire product’s revenue or just the model-related portion? How should fine-tuning, distillation, and multi-model routing be attributed? Each interpretation could become a contract negotiation point.

The sharing ratio also affects whether interests align. Too low a ratio fails to cover model investments; too high a ratio leads application companies to believe the model provider is taking value that belongs to the product, channels, and customer service. If Reuters’ claim of a maximum 30% sharing ratio for Kimi K3 becomes a reference point, disputes will arise over how much value the model truly contributes to a commercial service.

As of August 8, the public only knows that Alibaba is discussing commercial thresholds similar to Kimi K3. The specific applicable objects, revenue thresholds, sharing bases, ratios, audit methods, and exemption scopes remain unclear. Any change to these terms will alter the actual impact.

Thus, it is premature to say that domestic large models have ended price wars. At best, price wars are stratifying: Basic calls remain low-threshold, while top-tier capabilities see price increases; weights stay open, but large-scale reselling triggers renegotiation.

DeepSeek’s next official price list will reveal the cost floor for low-priced models. Qwen3.8-Max’s final license will answer a more crucial question: After growing an open ecosystem, can model companies recapture some of the value taken by channels without driving away developers?

These two documents matter more than slogans about "open source versus closed source" in marking the true inflection point for domestic large model commercialization. Cover image: Quentin Massys, "The Moneylender and His Wife" Note: Data in this article comes from public sources and does not constitute investment advice.

- END -