Alibaba's Qwen3.7-Max Runs 35 Hours Autonomously as AI Labs Race to Proprietary Models

Alibaba's Qwen Team today released Qwen3.7-Max, a proprietary large language model capable of approximately 35 hours of continuous autonomous execution, marking the Chinese tech giant's definitive retreat from the open-source AI strategy that defined its earlier work. The model, which supports external harnesses including Anthropic's Claude Code, represents a direct competitive response to the escalating pricing war between OpenAI and Anthropic. OpenAI's GPT-5.4 now costs $17.50 per million tokens, while Anthropic's Claude Opus 4.7 commands $30.00 per million tokens, a premium that Alibaba is undercutting by offering Qwen3.7-Max exclusively through its own API endpoints. The model's 35-hour autonomous runtime positions it as a tool for long-duration agentic workflows, a category that both OpenAI and Anthropic have targeted with their own proprietary releases. However, Qwen3.7-Max is only accessible from Chinese-based endpoints, a limitation that will severely constrain its appeal among US and European enterprises facing data sovereignty and compliance requirements. This launch arrives months after the departure of key Qwen Team leaders earlier this year, raising questions about the group's internal stability. The strategic shift from open source to proprietary licensing signals that Alibaba is betting on API revenue over ecosystem influence, a bet that will test whether China's largest cloud provider can compete in a market increasingly defined by hyperscaler-scale compute deals and aggressive pricing.

Where the $30-Per-Million-Token Pricing Floor Collapses

The pricing dynamics of the frontier AI market have crystallized into a three-tier structure that Qwen3.7-Max now enters. At the top sits Anthropic's Claude Opus 4.7 at $30.00 per million tokens, a price point that reflects the model's benchmark-leading reasoning capabilities and enterprise-grade safety features. OpenAI's GPT-5.4 occupies the middle tier at $17.50 per million tokens, a deliberate undercut that has already forced Anthropic to defend its premium through superior agentic performance and longer context windows. Qwen3.7-Max enters this landscape without a publicly disclosed per-token price, but Alibaba's history of aggressive cloud pricing, coupled with the model's exclusive Chinese-endpoint distribution, suggests a significantly lower price point designed to capture domestic market share while avoiding direct price competition on Western soil. The 35-hour autonomous runtime creates a fundamentally different cost calculus than per-token pricing alone. An enterprise deploying Claude Opus 4.7 for a continuous 35-hour agentic workflow would incur costs that compound with every token generated during that period, while Qwen3.7-Max's pricing structure likely bundles the runtime into a fixed API fee. This pricing architecture mirrors the shift from compute-by-the-hour to compute-by-the-job that defined the transition from on-premise data centers to cloud computing. The model's support for external harnesses like Claude Code further complicates the pricing picture, as enterprises can now mix and match Alibaba's inference engine with Anthropic's orchestration layer, a scenario that neither company's pricing teams likely modeled.

The $1.25 Billion Monthly Compute Deal That Reshaped the Market

Anthropic's $1.25 billion monthly compute deal with SpaceX's AI infrastructure unit, structured with a 90-day cancellation clause, has become the benchmark against which all frontier AI compute agreements are now measured. This deal, which has driven the debate over whether SpaceX's AI infrastructure business is worth $700 billion or $1.75 trillion, reveals the extraordinary capital intensity required to train and serve frontier models. Alibaba's decision to move Qwen3.7-Max to proprietary licensing is a direct response to this capital calculus. Open-source models generate no direct revenue to offset the compute costs of training and inference, while proprietary models, even at aggressive price points, create a recurring revenue stream that can be used to secure additional compute capacity. The 35-hour autonomous runtime feature is particularly capital-intensive, as it requires sustained GPU allocation for each inference session, tying up compute resources that could otherwise serve multiple shorter requests. Alibaba is effectively betting that enterprises will pay a premium for guaranteed, long-duration compute access rather than competing for spot inference capacity. This bet mirrors the strategy that drove AWS's EC2 Reserved Instances business, but applied to the inference layer rather than the compute layer. The 90-day cancellation clause in Anthropic's SpaceX deal also signals that even the largest AI companies are unwilling to lock themselves into long-term compute commitments, preferring the flexibility to shift workloads as pricing and performance evolve.

How the Competitive Reshuffle Hits OpenAI, Anthropic, and Google

The proprietary shift at Alibaba reshapes the competitive landscape in three distinct ways. First, it removes one of the few remaining high-quality open-source alternatives to the frontier models, forcing enterprises that relied on Qwen's open-source releases to either pay for Qwen3.7-Max API access or migrate to competing proprietary models. This dynamic benefits OpenAI and Anthropic in the short term, as enterprises evaluating alternatives will find fewer free options. Second, the Chinese-endpoint restriction creates a bifurcated market where US and European enterprises cannot legally or practically use Qwen3.7-Max for production workloads, effectively ceding the Western enterprise market to OpenAI, Anthropic, and Google. This geographic limitation is not merely a compliance issue: it prevents Alibaba from competing for the highest-value enterprise contracts in financial services, healthcare, and defense, where data sovereignty is non-negotiable. Third, the model's support for Claude Code as an external harness creates an unusual competitive dynamic where Alibaba is simultaneously competing with Anthropic on model performance while integrating with Anthropic's tooling. This integration strategy suggests that Alibaba views the orchestration layer as a commodity and the inference layer as the source of competitive advantage, a bet that runs counter to the industry trend of vertical integration, where companies like OpenAI bundle model, tooling, and deployment into a single platform. Google's Gemini ecosystem, which integrates model, search, and cloud services, represents the opposite strategic pole.

The Hyperscaler and Supply Chain Implications of 35-Hour Inference

The 35-hour autonomous runtime requirement creates second-order effects across the AI infrastructure supply chain that extend far beyond Alibaba's own data centers. Sustained inference sessions of this duration require GPU clusters with guaranteed availability, meaning that Alibaba must either reserve capacity on its own cloud infrastructure or negotiate dedicated compute deals with hyperscalers like NVIDIA's cloud partners. This demand profile favors hyperscalers that can offer reserved GPU instances with SLAs on uptime and latency, a market where AWS, Microsoft Azure, and Google Cloud currently dominate. The model's reliance on NVIDIA hardware, likely the GB300 Grace Blackwell Ultra Desktop Superchip or equivalent datacenter GPUs, ties Alibaba's Qwen3.7-Max capacity to NVIDIA's supply chain and allocation decisions. Any disruption to NVIDIA's production or allocation priorities directly impacts Alibaba's ability to serve the 35-hour runtime feature. The model's Chinese-endpoint restriction also means that Alibaba must build out domestic inference capacity, potentially straining China's access to advanced NVIDIA GPUs given ongoing export controls. This supply chain vulnerability creates an opening for Chinese GPU alternatives, though none currently match NVIDIA's performance for long-duration inference workloads. The packaging and HBM memory requirements for 35-hour inference sessions are also more demanding than standard inference, as sustained compute generates heat that must be managed through advanced cooling solutions and high-bandwidth memory that can maintain throughput over extended periods. Corsair's entry into the AI workstation market with its CORSAIR PRO portfolio, including systems powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, shows that the hardware ecosystem is adapting to these sustained inference requirements.

What Alibaba's Proprietary Pivot Signals About the Market's Trajectory

Alibaba's retreat from open source is the strongest signal yet that the open-source AI model strategy has failed as a business model for frontier AI development. The departure of key Qwen Team leaders earlier this year, combined with the shift to proprietary licensing, indicates that Alibaba's leadership concluded that giving away frontier models for free did not generate sufficient cloud revenue or strategic advantage to justify the compute costs. This conclusion validates the strategy that OpenAI and Anthropic have pursued from the beginning: proprietary models monetized through API access. The pricing war between OpenAI at $17.50 per million tokens and Anthropic at $30.00 per million tokens is now entering a phase where the marginal cost of inference, driven by hardware improvements and model optimization, is falling faster than prices, creating margin expansion for the winners. Alibaba's entry with an undetermined price point and geographic restrictions shows that the company is targeting the Chinese enterprise market where it already has cloud relationships, rather than attempting to compete globally. This regional strategy mirrors the broader fragmentation of the AI market along geopolitical lines, with US and Chinese models increasingly serving distinct customer bases with limited overlap. The 35-hour autonomous runtime feature also signals that the next frontier of AI competition is not just model quality but sustained agentic execution: the ability to deploy AI systems that operate independently over hours or days rather than responding to individual prompts. This shift from conversational AI to autonomous AI will drive demand for new infrastructure, pricing models, and enterprise deployment patterns that the current generation of API pricing does not fully address.

The market should expect more proprietary model launches from Chinese AI labs in the coming months, each with geographic restrictions that reinforce the bifurcation of the global AI market. The open-source era of AI is giving way to a proprietary phase where the winners will be determined not by model quality alone but by the ability to secure compute capacity, manage pricing pressure, and navigate geopolitical constraints. Alibaba's bet on 35-hour autonomous execution is a bet that enterprises will pay for duration, not just intelligence, a thesis that will be tested against OpenAI's and Anthropic's competing visions of how autonomous AI should be priced and deployed. The 90-day cancellation clause in Anthropic's SpaceX deal serves as a reminder that even the largest AI companies are hedging their bets, keeping the flexibility to shift compute providers as the market evolves. Alibaba's Qwen3.7-Max launch, with its proprietary licensing and Chinese-endpoint restriction, is the clearest signal yet that the AI industry is entering a phase of regional consolidation, proprietary monetization, and sustained capital deployment that will separate the hyperscaler-backed players from the rest.