Inside Alibaba’s Turn from E-Commerce Empire to Token Industrial System
The most important change is reorganizing e-com, cloud computing, foundation models, inference platforms, video generation, instant retail, and local fulfillment into the same AI commercial loop
The next stage of China’s AI competition will not be decided by who leads by a few points on model leaderboards, but by who can connect Token consumption, inference cost, application scenarios, commercial pricing, and real transaction systems into a continuously operating industrial machine.
This essay is part of the series What China’s Industry Media Is Really Talking About.
Executive Summary
Alibaba’s latest AI shift is an organizational restructuring around technical power, not just another model release. The creation of the Alibaba Group Technology Committee, the upgrade of Tongyi Lab into a large-model business unit, and the reassignment of key technical executives show that AI has become a group-level command structure.
Token consumption is becoming Alibaba’s new operating unit. Chinese industry media reported that China’s daily Token calls rose from 100 billion in early 2024 to 140 trillion by March 2026, while Alibaba Bailian’s MaaS platform saw Token consumption grow sixfold in three months.
Alibaba is trying to convert its cloud business into an AI industrial platform. Alibaba Cloud, Bailian MaaS, Tongyi, Qwen, T-Head chips, and inference platforms are being reorganized into a system where model calls, inference services, enterprise APIs, and multimodal generation become commercial revenue streams.
HappyHorse and Wan2.7 show Alibaba moving from model capability to productized AI pricing. HappyHorse’s video-generation pricing, including 720P generation as low as RMB 0.44 per second after discounts, signals that Alibaba is turning AI video from a technical demo into a usable commercial product.
E-commerce and instant retail remain Alibaba’s largest AI scenario base. Taobao Flash Sale, Ele.me, Taobao, and Tmall give Alibaba high-frequency transaction, fulfillment, merchant, and consumer scenarios that can generate sustained inference demand if AI is successfully embedded into real workflows.
Alibaba’s real test is whether it can turn a heavy organizational and commercial system into an AI flywheel. The company’s advantage lies in its combination of cloud, models, commerce, fulfillment, enterprise customers, and consumer scenarios, but the challenge is coordinating these assets into one system without being dragged down by complexity, subsidies, and margin pressure.
1. Alibaba Has Entered an AI Wartime Organization Mode
Over the past month, what Chinese industry media has really focused on is not another slogan about Alibaba “embracing AI,” but the recentralization of technical power inside Alibaba.
Around April 8, several Chinese technology and financial media outlets noticed a key organizational adjustment at Alibaba. Eddie Wu issued an internal letter announcing the establishment of the Alibaba Group Technology Committee, with himself as the head. The committee members include Jingren Zhou, Zeming Wu, and Feifei Li. At the same time, Tongyi Lab was upgraded into the Tongyi Large Model Business Unit, with Jingren Zhou taking full responsibility. Feifei Li became CTO of Alibaba Cloud, responsible for Alibaba Cloud’s technology and AI cloud infrastructure. Zeming Wu shifted his focus to his role as Group CTO, coordinating the group’s business technology platforms and AI inference platform. Taobao Flash Sale’s CEO role was handed over to Yanqun Lei.
This was not a normal personnel adjustment. When Jiemian republished a Hefan Finance report, it framed the issue around the phrase “Alibaba does not need a CTO who sells goods.” That expression captured the real question behind Alibaba’s adjustment: once AI becomes a group-level main battlefield, the technology leader can no longer be trapped inside a single business line, and the group can no longer allow models, cloud, inference platforms, e-commerce, instant retail, and application layers to move forward separately.
36Kr’s reporting on Alibaba’s AI organizational changes pushed the issue further toward “AI centralization.” It emphasized that Eddie Wu’s core solution is to make Token the key measurement unit of Alibaba’s AI business, running through model development, platform services, and application deployment. According to the data cited in the report, China’s daily Token calls rose from 100 billion at the beginning of 2024 to 140 trillion by March 2026, a more than thousandfold increase in two years. Alibaba Bailian’s MaaS platform saw Token consumption increase sixfold over the past three months. By the end of February 2026, Alibaba Cloud’s external commercialized revenue had exceeded RMB 100 billion. T-Head’s self-developed GPU chips had reached 470,000 units in scaled delivery.
Put together, these numbers show that Alibaba’s change is no longer simply an “internet company doing AI.” It looks more like an attempt to build a new AI production system: models provide intelligence, cloud provides compute and inference services, the MaaS platform handles external commercialization, chips supply part of the underlying infrastructure, e-commerce and local life services provide high-frequency application scenarios, and Token becomes the commercial measurement unit connecting all these links.
This is why Alibaba’s organizational adjustment belongs in a Chinese industry-media observation rather than a normal company news item. Chinese industry media is not capturing a single isolated change. It is capturing a large internet platform entering an AI wartime organization mode: technical power is being pulled upward, the model unit is being upgraded and made more independent, cloud and inference platforms are being strengthened, business responsibilities are being redistributed, and the group CEO is personally leading the technology committee.
2. From a Traffic Platform to a Token Measurement System
Over the past two decades, Alibaba’s core assets were e-commerce traffic, merchant ecosystems, payment and fulfillment networks. Taobao, Tmall, Alipay, Cainiao, Ele.me, local life services, and Alibaba Cloud together formed a vast digital commerce empire. Its underlying logic was to monetize through traffic, advertising, transactions, payments, logistics, and merchant services.
But in the AI era, Alibaba is trying to change its measurement unit.
In the traditional e-commerce era, the most important platform metrics were GMV, active buyers, advertising revenue, merchant numbers, conversion rates, and fulfillment efficiency. In the AI era, increasingly important metrics are becoming Token calls, inference cost, model-service revenue, enterprise API usage, agent-call frequency, multimodal generation cost, and scenario-loop efficiency.
When 36Kr discussed the growth of Token consumption, it noted that the Token consumption of Alibaba Bailian MaaS platform’s public model service market had increased sixfold in one quarter, and that commercialized MaaS revenue could become one of Alibaba Cloud’s most important revenue products. The key point is not the “sixfold” figure itself. It is that Alibaba Cloud’s growth logic is changing: cloud revenue used to come mainly from computing, storage, databases, and enterprise IT infrastructure. Now cloud is increasingly carrying model calls, inference services, agent workflows, and multimodal generation.
More importantly, Token is not an abstract technical metric. It is a commercial unit that can be directly priced, directly tied to cost, and directly linked to gross margin. Every user request, every enterprise call, every image or video generation, and every continuous agent execution consumes Tokens. The larger the Token consumption, the greater the revenue opportunity for cloud providers. The lower the inference cost, the stronger the platform’s price competitiveness. The more application scenarios there are, the more likely models and cloud infrastructure can form a positive feedback loop.
This is what makes Alibaba different from many pure model companies. Alibaba is not only building models; it has cloud. It does not only have cloud; it has e-commerce. It does not only have e-commerce; it has local life services and instant fulfillment. It does not only have consumer-side scenarios; it also has enterprise customers and developer platforms. What Alibaba really wants to do is recombine these dispersed assets into a closed loop running from AI infrastructure to models, from models to MaaS, from MaaS to applications, and from applications to transactions and fulfillment.
This structural change is heavy and difficult. It requires Alibaba’s internal departments to stop competing for resources around old business boundaries, and instead redistribute power around AI infrastructure and Token consumption. The upgrade of Tongyi Lab into the Tongyi Large Model Business Unit, the change in Alibaba Cloud’s CTO role, the Group CTO’s focus on inference platforms and business technology platforms, and Eddie Wu’s direct leadership of the technology committee are all dealing with the same problem: Alibaba wants to turn its old traffic-platform organization into a Token production and consumption system for the AI era.
3. Video Generation Is Moving Model Capability into Productized Pricing
If one only looks at organizational adjustment, Alibaba’s AI transformation can still be understood as an internal management reform. What makes the change concrete is the productization signal from HappyHorse and Wan2.7.
On April 27, Sina Tech reported that Alibaba’s video generation model HappyHorse 1.0 had begun gray testing. Built on a native multimodal architecture and an audio-video joint generation approach, it targets content production scenarios such as advertising, e-commerce, short dramas, and social-media creative work, offering an integrated capability from intelligent generation to editing. The most valuable part of the report was not the phrase “gray testing,” but the pricing: on the HappyHorse website, the listed prices for 720P and 1080P video generation were RMB 0.9 per second and RMB 1.6 per second respectively. After professional monthly membership and limited-time discounts, the prices were RMB 0.44 per second and RMB 0.78 per second.
This pricing signal is very important. It shows that Alibaba’s video generation model is no longer just a technical demo. It has entered a commercial stage where it can be priced, tested, and delivered to creators and enterprise customers. For advertising, e-commerce, short drama, and social-media creative industries, the core question of AI video generation has never been whether a model can produce an attractive sample clip. The real questions are cost, speed, stability, consistency, editability, and the ability to support production at scale.
The practical test by National Business Daily also grounded HappyHorse in real usage. According to the report, a journalist used the Qwen app and HappyHorse to generate an 8-to-10-second video, which took about three minutes. A technology professional interviewed by the publication said that when the first-frame image is clear and the prompt is precise, the output is already usable, and in many cases, transitions involving manhua-style characters can generate decent results.
At the same time, Leiphone and other technology media noticed that Alibaba’s video generation model Wan2.7 ranked first on DesignArena’s Video to Video leaderboard with an Elo score of 1334, 68 points ahead of second-place Grok Imagine. This data can certainly demonstrate model capability, but it should not be the endpoint of the analysis. The real point is that Wan2.7 provides technical capability, HappyHorse provides the product entry point, Alibaba Cloud Bailian provides the enterprise platform, and the Qwen app provides the mass-user entry point. Alibaba is pushing video generation from model leaderboards into creator tools, enterprise APIs, content production platforms, and commercial pricing systems.
In global AI narratives, China’s AI competition is often compressed into a few simple questions: whether China has the strongest models, whether it has access to Nvidia chips, and whether it has open-source weights. But the granularity provided by Chinese industry media shows that model capability is only the first step. The real competition is entering a second stage: who can package model capability into products, embed those products into real industries, turn industry demand into Token consumption, reduce costs through scaled inference, and convert cost advantages into lower prices and higher usage frequency.
4. E-Commerce and Instant Retail Remain Alibaba’s Largest Scenario Base
The easiest way to misunderstand Alibaba’s AI transformation is to view it as a standalone story about Alibaba Cloud and Tongyi. In reality, Alibaba’s greatest advantage and greatest pressure both come from the fact that it still owns one of China’s most complex consumer-internet and commercial-fulfillment scenarios.
Taobao Flash Sale is the window into this change.
Taobao Flash Sale was originally led by Zeming Wu. In this organizational adjustment, Wu stepped down as CEO of Taobao Flash Sale and shifted his focus to his role as Group CTO and AI inference platform construction. Yanqun Lei became the new CEO of Taobao Flash Sale. Sina Finance, citing related reports, noted that Taobao Flash Sale faces a dual test of AI and profitability. This is an accurate framing. Instant retail is not a light-asset traffic business. It is a high-intensity operating system that depends heavily on fulfillment density, merchant supply, delivery efficiency, order structure, and subsidy intensity.
A previous Economic Observer report on Taobao Flash Sale provided a set of meaningful data points. Through coordination with Ele.me, Taobao Flash Sale pushed daily peak orders to 60 million within two months and reached 120 million by August 2025. Monthly transacting buyers increased by 200% in four months to 300 million. Non-tea-drink and non-food-delivery orders had risen to more than 75%. On-time order delivery remained stable at 96%. Management expected the instant retail business to achieve overall profitability in fiscal year 2029.
These numbers show that Taobao Flash Sale is not a marginal business. It is an important experimental field for Alibaba to embed AI into real transactions and local fulfillment. AI can enter search, recommendations, customer service, subsidy allocation, merchant operations, inventory matching, route planning, fulfillment dispatching, user repurchase, and advertising. The higher-frequency instant retail orders become, the more opportunities models and inference systems have to access real commercial scenarios. The more complex the fulfillment system becomes, the greater the value of AI optimization. The higher the share of non-food orders rises, the closer the platform moves toward local integrated retail infrastructure, rather than merely a supplement to food delivery.
This is why Alibaba cannot be understood simply as a cloud company. Many AI companies have strong models but lack sufficiently large transaction scenarios. Many cloud companies have computing infrastructure but do not have e-commerce, local life services, payments, and fulfillment loops. Many application companies have users but do not have their own cloud, models, and inference platforms. Alibaba’s structural advantage is that it has the opportunity to place these links inside one commercial system.
Of course, this is also its structural burden. Taobao Flash Sale requires subsidies. E-commerce profits are under pressure. Local life services are fiercely competitive. Cloud and AI require long-term capital expenditure. Video generation and MaaS platforms still face price wars and inference-cost pressure. What Alibaba is doing is not a light transformation. It is trying to turn a mature e-commerce empire into a heavy-asset, technology-heavy, organization-intensive AI industrial system.
5. China’s AI Competition Has Moved from Model War to Industrialization War
Alibaba’s case reveals a deeper change taking place in China’s AI competition.
The first stage of AI competition looked like a model race. Companies compared parameter scale, leaderboard rankings, open-source capability, multimodal performance, context length, and reasoning performance. The second stage of AI competition began to look like an infrastructure race. Companies compared compute, chips, data centers, electricity, cloud platforms, inference cost, and model-service capability. The third stage of AI competition is increasingly becoming a contest of industrial organization: who can combine models, cloud, chips, scenarios, customers, products, pricing, and real workflows.
Alibaba is standing at the entrance to this third stage.
36Kr’s report on Alibaba, ByteDance, and Tencent beginning a new round of AI infrastructure construction pushed the competition from model capability toward agent engineering, Harness engineering, Computer Use, Skill stores, and Token consumption. This observation is very important, because it shows that Chinese industry media is no longer only watching model launches. It is beginning to track whether AI can enter stable engineering systems, whether it can be continuously called by enterprises and consumers, and whether it can generate large-scale commercial value.
From this perspective, Alibaba’s advantage does not lie in any single model’s isolated lead. It lies in its attempt to control multiple key links at the same time: Alibaba Cloud provides infrastructure, Tongyi provides model capability, Bailian provides the MaaS platform, T-Head provides part of the chip capability, Qwen provides the user entry point, HappyHorse provides the video generation product, Taobao and Tmall provide e-commerce scenarios, and Taobao Flash Sale and Ele.me provide local fulfillment networks.
This is not a lightweight AI application story. It is the reindustrialization process of a large platform company.
The mainstream narrative of global AI competition is still highly concentrated on OpenAI, Anthropic, Google, Meta, Microsoft, and Nvidia. That narrative is centered on models, chips, cloud capital expenditure, and developer ecosystems. But Alibaba’s case reminds us that China’s AI competition has another dimension: Chinese platform companies are not only training models. They also possess massive consumer scenarios, merchant networks, fulfillment systems, payment systems, enterprise customers, and local service entry points.
This means China’s AI commercialization path may not fully replicate the American path. American AI looks more like diffusion from models and cloud into applications. Chinese AI may move more quickly in the opposite direction: real commercial scenarios pull models, cloud, and inference infrastructure forward. The core problem in the former is how to find a sufficiently large application loop. The core problem in the latter is how to re-AI-ize already existing massive transaction and fulfillment systems.
6. Alibaba’s Real Test: Can It Turn a Heavy-Asset Loop into a Flywheel?
The biggest question Alibaba now faces is not whether it has an AI strategy. It is whether it can turn that strategy into a flywheel.
A sixfold increase in Token consumption is important, but whether Token consumption can be converted into high-quality revenue is more important. Alibaba Cloud’s external commercialized revenue exceeding RMB 100 billion is important, but whether cloud and AI commercialization revenue can support more aggressive future growth is more important. HappyHorse 720P video generation falling to RMB 0.44 per second after discounts is important, but whether low prices can bring large-scale usage, stable retention, and enterprise-level payment is more important. Taobao Flash Sale’s daily peak orders and buyer growth are important, but whether instant retail can move toward profitability under intense competition, heavy subsidies, and high fulfillment costs is more important.
This is the contradiction of Alibaba’s AI transformation: it owns scenarios that many other AI companies do not have, but it also bears organizational complexity that many other AI companies do not face.
Alibaba must solve four problems at the same time. First, model capability must continue to keep up, or the cloud and MaaS platform will lack attraction. Second, inference costs must continue to fall, or greater Token consumption will create greater cost pressure. Third, e-commerce, local life services, and content production scenarios must genuinely generate AI-driven increments, rather than treating AI as a marketing label. Fourth, the group organization must be able to coordinate cloud, models, e-commerce, local life services, chips, developer platforms, and enterprise customers, rather than falling back into departmental walls.
This is where the past month of Chinese industry-media reporting on Alibaba becomes most valuable. These reports did not simply portray Alibaba as an “internet company whose stock is recovering,” nor as a “technology company releasing a new model.” Through the technology committee, the ATH business group, the upgrade of Tongyi, HappyHorse’s gray testing, Wan2.7’s leaderboard position, Taobao Flash Sale’s leadership change, Token consumption growth, and Bailian platform commercialization, they presented a more realistic Alibaba: a company trying to transform itself from an e-commerce platform into an organizer of AI infrastructure and commercial scenarios.
Alibaba’s future will not be determined only by model leaderboards, nor only by cloud revenue growth. It will depend on whether Alibaba can turn one of China’s largest consumer-internet scenarios into one of China’s largest AI usage scenarios; whether it can upgrade cloud computing from an infrastructure service into a Token industrialization platform; and whether it can turn e-commerce, video generation, instant retail, and enterprise services into sustained sources of inference demand.
If this works, Alibaba will no longer be merely “an e-commerce company doing AI.” It will become one of the most important examples of China’s AI industrialization path.
If it fails, Alibaba will not fail because it lacks models. It will fail because it cannot twist models, cloud, scenarios, organization, and commercialization back into one system.
This is what this set of Chinese industry-media reports really reveals: China’s AI competition is leaving the launch-event stage and entering the deep waters of organizational structure, cost curves, Token consumption, scenario density, and commercial closed loops. Alibaba is the largest, most complex, and most important sample in that deep water.
Source note: This essay is based on recent Chinese media and industry-media reporting on Alibaba’s AI organization restructuring, Alibaba Group’s technology committee, the upgrade of Tongyi Lab into a large-model business unit, Alibaba Cloud and Bailian MaaS token consumption, HappyHorse 1.0 gray testing, Wan2.7’s DesignArena ranking, and Taobao Flash Sale’s leadership and operating pressures. All translations of short quoted phrases are my own.