How will Alibaba respond to Google's complete closed-loop shopping experience in the Agent era?

Written by | Landong Business Zhao Weiwei

When you open the official website of Alibaba Cloud's new AI product, the first thing you see is a row of skill installation instructions, which are readable by an Agent. After invoking it, your AI agent can instantly load the native invocation capabilities of the Tongyi Qianwen large model.

This marks the first time in 17 years that Alibaba Cloud has established an independent product site outside its main official website, but it has not followed the conventional practice of creating a showcase website for users. Its first screen is designed specifically for Agents, with the logic being: when your users are AI, you don't need a Banner; you need an executable instruction.

The underlying logic is not difficult to understand. In the Agent era, the service objects are not just humans but also autonomously decision-making Agents. Alibaba Cloud has clearly transformed its cloud products through Skill modularization, MCP standardization, and CLI instructionalization, turning each cloud product into a standardized capability module that can be directly activated by agents like invoking program functions.

At the same time, Agent-based transformations are also taking place across the ocean. Google also announced full-stack technology and product updates, ranging from chips and models to applications, at its I/O conference. The two major AI full-stack companies in China and the United States are once again laying out in the same track (translated as "track" or "arena") in the same direction.

Especially at the application layer, Google's latest Antigravity 2.0 platform is the core environment for developing and managing autonomous AI Agent clusters. It can autonomously write a complete operating system within 12 hours, focusing on core agent dialogue, artifacts generated by agents, and multi-agent orchestration. "We are making Antigravity the only platform you need for Agent-first development."

Agent-first, with agents taking over everything—similar development trends are emerging simultaneously at Alibaba Cloud and Google.

This summer, when Google began rolling out smart shopping carts to facilitate shopping while users browse the web or chat with Gemini, automatically finding discounts and price reductions, and relying on underlying architectures such as the Universal Commerce Protocol (UCP) to partner with giants like Amazon, Meta, and Microsoft, this cross-platform shopping experience is destined to make shopping carts in the Agent era even smarter.

Facing Google's complete closed-loop shopping experience in the Agent era, how should Alibaba respond and establish industry foundational rules in the field of agent-based shopping?

Alibaba Cloud Becomes More Open

As previously mentioned in "Defeating the Enemy: Google Cloud Teaches Alibaba Cloud a Lesson," it is important for cloud vendors to win competitions, but even more important to ensure that competitions occur within their own ecosystems. Model freedom is one of Google Cloud's key advantages, and it is an important lesson Google has taught Alibaba Cloud.

Now, Alibaba Cloud is also learning from Google's strengths and striving to become "the most open cloud in the AI era." As an enterprise-level large model application development platform, Alibaba Cloud's BaiLian has also begun to open up access to third-party models.

In addition to Alibaba's self-developed Qianwen model matrix, the BaiLian platform will also integrate third-party models such as Zhipu GLM-5.1, MiniMax M2.7, Yuezhia'an Kimi K2.6, Kling, and Vidu Q3.

On the official Qianwen Cloud website, over 150 model series and more than 480 various models are already available, covering mainstream models from both domestic and international sources, supporting synchronous comparison of multiple models. Developers can quickly complete experience, evaluation, and selection according to their own needs.

At the same time, Qianwen Cloud encapsulates the core capabilities of model services into Skills and CLI tools, meaning that Agent tools like OpenClaw can learn the full capabilities of the entire platform with just one instruction and autonomously plan, allowing image tasks to invoke visual models, image generation tasks to invoke image generation models, and video tasks to invoke video models, all without human intervention or the need to write integration code.

For cloud vendors' customers, how to transparently consume Token resources is a very real issue.

Qianwen Cloud's solution is an intelligent and transparent management mechanism. An AI agent can retrieve model usage data in real-time, analyze data trends, identify abnormal usage, and assist in cost optimization. At the same time, it can pull logs, Key activities, and other data via CLI to achieve anomaly identification and task traceability.

This is also a common trend for Alibaba Cloud and Google Cloud. They are no longer just selling models but transforming themselves into AI factories that provide computing power and mobilize infrastructure.

Google's advantage lies in global developer density, while Alibaba's advantage lies in local ecological depth.

At this I/O conference, Google announced that the number of tokens processed per minute via APIs has reached 19 billion, with 8.5 million developers using Google's AI models to build applications each month. Internally, Google processes over 3 trillion tokens daily through AI development tools, and this number doubles every few weeks.

These are not just model computing power capability data but also core data on the volume of infrastructure bear (translated as "capacity" or "load").

So, under this logic, it is not difficult to understand why the pricing of Gemini 3.5 Flash has tripled. Google's own calculations show that although the unit price is higher, this model is more efficient and can help enterprises save more than $1 billion in AI costs annually.

It is not selling cheapness but making every cent's worth of processing capacity count. This is completely different from the traditional model price reduction logic. Previously, price reductions were used to capture users, acting as an entry ticket logic. Now, price increases are accompanied by efficiency improvements. Who can produce higher-quality Tokens at lower chip costs is the infrastructure logic.

The infrastructure logic means that when an Agent needs to invoke language capabilities, the agent will prioritize which capabilities to use and which invocation paths to take. This is the true goal of all technical announcements at the summits of both Alibaba and Google.

Google Remains Alibaba's Teacher

Google announced many things at I/O 2026, ranging from the Gemini Omni world model at the model layer to the first audio smart glasses with built-in Gemini based on the Android XR platform at the hardware layer. It can be said that Agents have been thoroughly integrated into all of Google's businesses, building its own ecosystem in scenarios such as search, office work, and shopping, making it difficult for all competitors to surpass.

More importantly, there is the investment behind it. Google's annual capital expenditure this year is between $180 billion and $190 billion, with a key portion spent on custom chips.

Google has previously released the TPU 8t optimized for pre-training and the TPU 8i optimized for inference, indicating that chips have reached a fork in the road with further differentiation in direction. Training requires extreme computing power density and large-scale parallelism, while inference requires extreme low latency and memory bandwidth. There is a fundamental design tension between these two goals, and pursuing both simultaneously on the same chip results in neither being extreme enough.

Alibaba's newly released Zhenwu M890, on the other hand, has 144GB of HBM memory, an inter-chip interconnection bandwidth of 800GB/s, and overall performance three times that of the previous generation Zhenwu 810E.

128 chips form the Panjiu AL128 super node, with P2P latency below 150 nanoseconds. Gao Hui, Vice President of T-Head, positions this chip as follows: When an Agent is executing a task, it may initiate dozens of model invocations consecutively within milliseconds, requiring tight coordination between CPUs, GPUs, networks, and storage, rather than simply stacking computing power.

The Zhenwu M890 is designed for both training and inference, forming a stark contrast with Google's approach of separating training and inference. The two choices reflect different judgments on the current main bottlenecks.

On the path of integrated chip functionality R&D, Alibaba's T-Head stands alongside NVIDIA and Baidu's Kunlun Core, while Google's TPU and Huawei's Ascend belong to the "complete differentiation" faction. Such technological route divergences are an inevitable trend after the scalable development of the computing power industry, as providing enterprise customers with the simplest and most cost-effective all-in-one solution or a solution with clear divisions of labor are both products of responding to different market demands.

Alibaba and Google have different directions on the chip path, but coincidentally, the planning and landing times of Google's eighth-generation TPU and Alibaba's Zhenwu V900 chips are similar, both aiming for the end of 2027.

This may be a consistent bet by both companies. The next main battlefield in the AI performance competition is not about whose model has more parameters but who better meets market demands and can produce high-quality tokens with the least energy consumption.

From this perspective at the chip R&D level, Google remains Alibaba's teacher. Liu Weiguang, Senior Vice President of Alibaba Cloud Intelligence Group and President of the Public Cloud Business Unit, believes that the combination of Google's TPU and Gemini achieves the highest performance, with the underlying logic being that one's own chips and one's own models can definitely achieve the best cost-effectiveness.

Who Will Rewrite Smart Shopping?

What is most worth Alibaba's proactive layout (translated as "layout" or "planning") and anticipation is that Google has recently introduced a universal shopping cart feature, targeting e-commerce consumption in the Agent era.

This is a new AI scenario created by Google, named "Universal Cart." Users can add products anytime while searching, on YouTube, or in Gmail. The shopping cart automatically checks for discounts, monitors price reductions, and sends restocking reminders in the background, then uses Google Wallet for payment, automatically calculating which payment card offers the best deal. Even if users do not pay with Google, they can revert to the retailer's website for checkout.

This is actually an attempt to transform Google into a one-stop shopping website, with Google acting as a "matchmaker" in users' shopping and consumption, and currently not charging commissions.

More importantly, Google's underlying Universal Commerce Protocol (UCP) and the AP2 protocol for securing payments are establishing a new set of e-commerce rules, which is something worth anticipating for the entire e-commerce industry.

UCP can be understood as a set of open standard protocols for AI-powered shopping. From searching for products, adding them to the shopping cart, purchasing, paying, to obtaining after-sales service, the initiators of this set of rules include large retailers such as Google, Walmart, Shopify, and Target. In April, companies like Amazon, Microsoft, and Meta also joined this open standard.

In other words, in the future, it will not be humans placing orders on e-commerce websites but specific Agents placing orders on behalf of humans, comparing prices and placing orders across numerous shopping websites, rather than being limited to a specific shopping website.

This stands in stark contrast to the current e-commerce Agents in the Chinese market. Doubao can place orders on Douyin E-commerce, and the Qianwen App can connect to Taobao for ordering, but neither can achieve cross-platform consumption. Therefore, the capabilities of each Agent will be limited to their respective scopes.

Google aims to promote this smart shopping experience to a larger market. The "Universal Cart" Agent consumption experience will be launched on Google Search and Gemini this summer. The UCP checkout experience will also be available in Canada, Australia, and the UK in the coming months and gradually expand to vertical industries such as hotel reservations and local food delivery.

In addition, Google's AP2 protocol is also a underlying rule for securing smart shopping, aiming to allow Agents to safely represent users for payment within set limits.

The underlying mechanism of AP2 is to establish a transparent, verifiable connection between users, merchants, and payment processors, using encryption technology throughout to protect user data. The protocol also includes tamper-proof digital records to ensure that Agents always act on behalf of users and provide permanent audit trails for both buyers and sellers to refer to in case of returns or disputes.

In other words, Agent shopping must meet various limit (translated as "restricted" or "limited") conditions, including specifying the desired brand and product, as well as the consumption limit. When these conditions are met, the Agent will automatically complete the purchase.

A2A covers communication between Agents, UCP covers commercial behavior of Agents, and AP2 covers payment authorization of Agents. With these three layers superposition (translated as "stacked" or "layered"), Google is not writing just a product in the Agent era but a set of foundational procedures for cross-platform shopping and consumption. This is also an industry trend that domestic e-commerce giants urgently need to anticipate:

The battlefield is no longer about which platform users buy from but which Agent they use to place orders.

For Chinese users, the consumption habits and trust foundation for shopping on e-commerce platforms will not change in the short term. However, in the long run, if the set of smart shopping underlying protocols led by Google matures globally, the logic of "shopping entrances" will be rewritten.

E-commerce platforms such as Alibaba, JD.com, and Pinduoduo are destined to face choices: whether to independently build a new set of rule systems or choose to be compatible with this set of global universal protocols?