Summary

The competitive focus in Agent Infra is shifting from “getting models to run” to “making Agents actually complete tasks.” 旭卿 characterizes traditional AI Infra in terms of token throughput, time to first token, distributed training and inference efficiency, and cost; Agent Infra demotes model service to one component and adds memory, tool use, task planning, sandboxing, multi-agent systems, security, and identity. “Every component is mission-critical,” and any weak link ultimately shows up in task success rates.
AgentBay chose to build a complete platform rather than compete for the top spot in one vertical component. The team wants every module to be “solidly above average,” then combine code, browser, and mobile environments, persistence, MCP, and eventually memory, context, and higher-level agent services into a system; the target customer is also expanding from top developers who can assemble the pieces themselves to SMBs that need low-code and API-based capabilities. AWS AgentCore’s subsequent product architecture and interfaces showed “an extremely high degree of similarity” with that thesis.
The strongest demand signal is concurrency: one domestic Agent company asked for roughly 200K concurrent subtasks during the day; another domestic foundation-model company needed 200K to 400K concurrent virtual machines. The former does not mean 200K users each sending one task; a small number of primary tasks may branch into 100–200 subtasks each, like Manus Wide Research. The host therefore believes Agents could turn token and compute consumption from a concern about hitting a ceiling into 10x or even 100x growth. The team describes the demand curve as “highly exponential” and has already begun considering global resource scheduling.
AgentBay currently cares more about adoption than revenue, but its long-term pricing could shift from selling sandboxes to selling the full service. 安陈 said there is still no clear monetization model and that the immediate priority is building the product; 旭卿 said pricing could eventually expand from compute and sandbox resources to the costs of multimodal inference, browser perception, state consistency across hundreds of sandboxes, and more—at which point “compute may be only a very minor cost” within the package. More notable for investors: Alibaba Cloud has added people, money, and resources without imposing an incremental revenue KPI. Management asks instead, “How many customers are using this?”
The cloud-versus-local sandbox debate is fundamentally about trust, isolation, and concurrency—not merely where data is stored. Responding to concerns that putting passwords and identity information in the cloud “could cause a major incident,” 旭卿 pointed to the public cloud’s continuous iteration, security teams numbering in the thousands, and DDoS defenses, comparing it with the choice of keeping money at home or in a bank. The host retained the convenience of a local browser with existing cookies and no need to log in again. The cloud’s more direct advantage is that failures can be discarded: a model deleting a directory on a local machine causes real damage, while the same action in an isolated sandbox can be rebuilt without touching personal files.
Visible Agent commercialization is emerging first in Coding Agents and general-purpose Agents; the larger transformation may be hidden inside enterprises. 安陈 sees e-commerce price comparison, product listings, hero-image design, operations, finance, OA, and internal customer service as traditional RPA workflows becoming Agent-driven. External-user opportunities are concentrated in HR, financial investment advice, and healthcare. Finance is crowded because it is “close to money,” data is predominantly numerical, and feedback loops are easy to build, but the program offered no specific revenue or penetration forecasts.
AgentBay is not a replacement for Wuying Cloud Computer; it extends it into an elastic execution environment that AI can drive. Wuying continues to grow at a triple-digit rate, and the two are defined as sibling businesses. The host called 2025 “the year of the super-individual” and “the year of personal cloud computing”; the next-generation logical cloud computer would combine persistence, instant replication into 100 copies, cross-device roaming, and natural-language control, ultimately pointing toward “silicon-carbon symbiosis” between people and continuously online digital counterparts.
旭卿’s organizational bet is to infer the infrastructure of 2–3 years from unusual signals today; the biggest risk is not getting a planned item wrong but missing a module nobody had imagined. In 2016, starting from the 3%–5% performance loss still associated with virtualization, he pushed for hardware-software co-design and the Shenlong server; in 2015, demand from K2-powered Pailitao image search prompted him to position early for AI Infra. The overnight pivot into Agent Infra after GTC followed the same playbook. His central warning is: “Not knowing what you don’t know—that’s the scariest thing.”

Deep dive

1. Agents Have Already Pulled Technical Managers Back Into the Code

安陈 uses Onlook in daily work to rapidly generate frontend interaction designs in natural language. Another frequent tool is the Alibaba Taobao team’s internal product 星流, which he uses to read foreign-language papers quickly. The two use cases map to executable product design and high-density information processing.
旭卿 uses Alibaba’s internal One Day more often and is also a Cursor user. He last wrote code continuously with his own hands around 2017–2018, but Coding Agents have given technical managers who moved into strategy and architecture a chance to rediscover “the joy of writing code.”
The shift is not confined to frontline engineers. 旭卿 said 张建锋, president of DAMO Academy, also uses Cursor to write code and once showed him a Windows Minesweeper game. Early Agent adoption is following a simple path: turning natural language into running output.

2. Agent Infra Turns Model Service from the Finish Line into the Foundation

旭卿’s definition of traditional AI Infra is clear: optimize token throughput, time to first token, large-scale distributed training and inference efficiency, and cost—“how to squeeze every last bit of efficiency out of the GPU.” The final deliverable is the model service required for training, deployment, and inference.
In the Agent era, model service is merely a base component. For an application to actually complete work, it also needs tool use, short- and long-term memory, task planning, execution environments, multi-agent collaboration, and security mechanisms. The infrastructure scorecard shifts from single-inference efficiency to end-to-end task success.
旭卿 uses a traditional computer analogy: the foundation model is the CPU, while short- and long-term memory are like memory and storage. 安陈 adds that code execution and web operations require compute, while storage maps to long-term memory; the network connects more capabilities, and MCP takes models from “focused on thinking” to actually taking action.

3. Six Component Categories Have Become the De Facto Agent Stack

The guests identify the core modules as memory, tool use, task planning, sandboxing, multi-agent collaboration, and security and privacy, with context engineering, memory management, and task orchestration running through the stack. No one formally prescribed the answer, but developers have “quietly and independently” converged on similar architectures.
In memory, 安陈 cites Mem0 and MemGPT as distinctive and relatively deep implementations. In execution environments and tools, he points to E2B, Browserbase, and product forms such as computer use and browser use.
It is still difficult to say that any one company leads clearly in task planning, because performance remains heavily dependent on the model, prompts, and degree of automation. Multi-agent orchestration already has frameworks such as LangGraph and OpenAI Agents SDK; Alibaba also uses a fully self-developed collaboration framework internally, but has not released it externally.
At the collaboration-protocol layer, Google A2A is attempting to define an interaction model between Agents, in what 安陈 describes as “a bit more aggressive.” The security layer covers code fencing, sensitive-information protection, multi-tenant isolation, and zero local intrusion; authentication and identity form another independent chain.

4. Every Module Is Mission-Critical, but Procurement Happens at the Pain Point

Asked which link matters most, 安陈 refused to rank them: “Actually, they’re all important.” He considers every component mission-critical. A coding sandbox may be invisible to end users, but memory problems are nearly impossible to hide because context windows are limited while users demand higher task success rates.
Before Agent Infra emerged as a concept, developers generally tried to build the entire stack vertically. It was like asking a new employee to prepare the desk, network cable, and computer after joining the company. “You’re trying to run while also building the road,” which slowed application growth and created room for third-party components and platform services.
One domestic foundation-model company initially built its own virtual coding environment, browser use, and mobile use. After seeing AgentBay, it switched quickly because what it actually needed was 200K to 400K concurrent virtual machines—capability outside a model company’s core strengths.
Some automakers are less concerned with extreme concurrency than with having an Agent remember what a driver asked in the past. The team can help build long-term memory on top of vector databases. 旭卿 emphasizes that this is not bespoke development for each customer; the same general-purpose capabilities simply surface different priorities across customers.

5. Identity and Security Are Long-Term Modules Most Easily Masked by Efficiency Goals

安陈 adds identity management as a particular focus. When a computer-use Agent accesses multiple websites, it must authenticate automatically and handle different site-level permissions without allowing credentials to escape into the workflow. The team has already begun laying out this capability.
Early teams are naturally inclined to prioritize execution efficiency and defer identity and security. Cloud providers, by contrast, have built a different kind of muscle memory over time. 旭卿’s view is that a failure of data privacy or permission boundaries could be catastrophic, so these cannot be patches added after launch.

6. AgentBay Emerged from Manus’s Breakout and an Overnight Pivot after GTC

After the 2025 Spring Festival, Manus shifted the industry conversation from foundation models and chat capabilities toward general-purpose Agents. The team then attended Nvidia’s GTC and saw the rapid rise of U.S. Agent companies; companies backed by Nvidia were “almost all talking about their own Agents.”
安陈 concluded that the team should not build another Manus competitor, but should “empower companies like Manus.” They held an overnight meeting to change direction, reorganizing Alibaba Cloud’s existing compute, storage, and network capabilities into AgentBay.
The team took roughly 4–5 months from the decision to the launch during the World Artificial Intelligence Conference. The early product offered execution capabilities such as computer use, mobile use, and browser use. Only after deepening the product did the team discover that customers also needed long-term memory, identity management, and security—now the focus of the second phase.

7. The Value of a Full Platform Comes from Higher-Level Abstraction, Not a Single-Component Championship

The host challenged the strategy: Would pursuing 10 capabilities dilute attention and leave the sandbox vulnerable to specialized startups? 旭卿’s answer is that To B is inherently a long game. The goal is not to lead every category in the short term, but to make each module “reasonably balanced” and “solidly above average,” then win at the system level.
旭卿 draws a further distinction in positioning. Vertical startups mainly serve top AI developers, who can assemble complicated building blocks themselves. Cloud providers serve a universal market: they need to let small and midsize startups build Agent applications through low-code tools and simple APIs.
AgentBay therefore supports MCP first and exposes its capabilities through APIs. The underlying layer covers code, mobile, and browser environments; the upper layer provides Agent Service, with unified persistence below it and memory and context to come later. The end product is closer to an Agent-building platform than an atomic sandbox.
The abstraction must continue upward into the application layer. A DOM tree cannot fully identify elements such as iframes and video, while pure text input leaves the model “flying blind.” The service provider must assemble multimodal models and provide visual perception and action output, so customers do not have to manage the Agent layer inside the sandbox themselves.

8. Cloud Providers Are Converging on the Same Agent Infra Product Map

AgentBay first introduced its concept and initial capabilities at a conference around April 7, then became commercially available during the World Artificial Intelligence Conference. After AWS AgentCore launched, the team analyzed it and found that its capabilities, interface definitions, and overall layout had “an extremely high degree of similarity.”
旭卿 did not comment on which cloud provider is investing the most. He only judged that Agents are an important AI application direction and that every provider will take them seriously. Inside Alibaba Cloud, 吴泳铭 highlighted Agents and Agent Infra at an all-hands kickoff early this year, and received another briefing before the commercial launch.
The team continues to receive people, money, and resources. 安陈 adds that management has never asked, “How much revenue?” Instead, it asks, “How many customers are using this?” There is currently no target tied to selling more Alibaba Cloud resources.
AgentBay’s role in Alibaba Cloud’s AI strategy is defined as “connecting what comes before and after”: Agent Infra opens model capabilities externally, while the upper layer supports Agent development frameworks and applications. 旭卿 even believes that, for now, the top decision-maker may not view it primarily as a business, but as a capability a cloud provider must build.

9. The Local-versus-Cloud Sandbox Debate Is First a Debate over Trust

The host cited the route chosen by 周鸿祎 when launching the “Nano AI Agent Swarm”: put the sandbox locally because sending user passwords and authentication information to the cloud “could cause a major incident.” The concern captures the most sensitive adoption barrier for cloud-based Agents.
旭卿 acknowledges that he once also believed self-built data centers were safer, but changed his view after joining Alibaba Cloud. He noted that attackers once needed very little money—just 5G of traffic purchased with the Bitcoin of the time—to potentially bring down a game company’s self-built center, while public-cloud DDoS defenses continue to evolve.
His core rebuttal is economies of scale. Alibaba Cloud has a security team of thousands protecting its systems in real time; most enterprises could neither hire engineers at the same level nor afford comparable investment. “Is it safer to keep your money at home or in a bank?” is his analogy for cloud hosting.
The host’s counterpoint is worth preserving: a local browser already has the cookies, so users do not need to hand over their account passwords again. 旭卿 responds that using a local system likewise means trusting operating-system and browser vendors such as Microsoft or Google, while local malware and viruses can make the environment messier. Whether to trust cloud providers is a choice of route, not a belief that “local means trusted by default.”

10. Cloud Isolation Makes Failed Environments Disposable and Large-Scale Execution Possible

On the capability boundary, 旭卿 cites Manus founder 肖弘’s description of simultaneously scheduling 100 Agents for Wide Research. Such a workload is difficult to run on a personal computer. As tasks become more parallel, “running locally is actually not very realistic”; compute and scheduling will ultimately rely on the cloud.
Cloud sandboxes also isolate destructive Agent actions. 旭卿 initially thought the idea of “the model executing a command and deleting a directory” was just a joke, until he learned that “this really happens.” On a personal computer, real files disappear; in a sandbox, the environment can simply be restored or rebuilt.
AgentBay therefore designs code execution, multi-tenant isolation, zero local intrusion, and end-to-end data protection together. It provides a way to confine permissions, data, and failures within a relatively contained, recoverable environment—not a guarantee that software has no vulnerabilities.

11. Concurrency Expansion Is Multiplying Compute Demand by Subtask Count, Not User Count

The host observes that the market had been worried about token consumption nearing a ceiling and GPU demand potentially cooling. As Agent capabilities improve, however, the compute required per task could rise 10x or even 100x. The key variable is no longer just active users, but how many execution branches each primary task generates.
安陈 cites a customer requirement of roughly 200K concurrent sessions during the day from one domestic Agent company. Further clarification showed that the number referred to concurrent subtasks, not 200K users each sending one task; another customer had previously requested 200K–400K concurrent virtual machines.
Manus Wide Research demonstrates the multiplier effect: one primary task can “branch out” into 100–200 subtasks. Even with only 10K tasks at the front end, splitting each into 20 execution paths would quickly push underlying concurrency to 200K.
安陈 says there is still no clear pricing model, and that the immediate priority is building the product; money is not the main consideration. Long-term pricing could still follow the cloud-computing model of charging for compute and resources—for example, billing for the corresponding resources when a customer starts 100 sandboxes—then adding multimodal inference, state consistency, and environment control as managed services. The team describes demand as “an extremely exponential curve” and was discussing global resource scheduling on the day of the recording.

12. Coding and General-Purpose Agents Monetize First; Enterprise Transformation Is Less Visible

安陈 observes that general-purpose Agents and Coding Agents are growing first. Coding Agents have broad coverage and relatively mature business models; general-purpose Agents became a major area of concentrated industry investment after Manus launched, and consumers can already feel the difference directly.
A broader but less visible demand comes from turning traditional applications into Agents. In e-commerce, for example, operators once had to compare prices across platforms, list products, and design hero images. These repetitive RPA workflows are gradually being handed to Agents.
Many large customers connect to AgentBay not to launch a new consumer product immediately, but to automate internal OA, operations, finance, and customer service. The “quiet” productivity gains may initially appear as lower costs rather than as standalone product revenue.
Customer-facing use cases cluster around HR, finance, and healthcare. Financial investment advice is especially crowded because it is “close to money”: people want broad access to news and other data for AI analysis, while the data is mostly numerical and readily forms a feedback loop. Healthcare is seen as an area where foundation models could create more value, but the guests offered no specific deployment scale.

13. AgentBay Has Not Replaced Wuying; the Organization Is Absorbing a Genuine Two-Track Investment

旭卿 uses “10 years” to define To B. Before the 2016 Spring Festival, Jack Ma cited Taobao’s path from its 2003–2004 beginnings to its 2014 IPO, Alipay’s development from roughly 2006 to 2016, and Alibaba Cloud’s progression from 2009 to 2019 to emphasize long-term investment. “Whatever I do, I basically think in 10-year units.”
He corrects the host’s use of “transformation”: Wuying has not been abandoned and continues to grow at a triple-digit rate. After becoming a first-level business unit, Wuying and foundation models were designated the two most important strategic products of Alibaba Cloud and the Cloud Intelligence Group. AgentBay is an extension of its capabilities—a “sibling relationship.”
The real pressure is to maintain rapid growth in the mature business without missing the new track. After GTC clarified the direction, the organization still had to reallocate people: some volunteered to join AI, while others believed their existing work remained equally important and did not want to be transferred directly.
Asked whether people staying with Wuying would feel disappointed, 旭卿 acknowledged that “of course some do,” but stressed that both are strategic businesses. Wuying is already deployed and growing rapidly; AgentBay is the longer-term bet. The management challenge is not choosing one side, but making both curves work at the same time.

14. Personal Cloud Computing Is Rewriting “One Computer” as a Replicable, Persistent Environment

The host’s “hot take” is that Wuying previously pushed edge compute into the cloud mainly for specialized needs: designers using Stable Diffusion and FLUX, or simulation users temporarily calling on highly elastic GPUs. Ordinary consumers may not have felt the need for a cloud computer.
Agents change that premise. A personal task may need 100 environments at once, beyond the capacity of a local PC. The host therefore calls 2025 “the year of the super-individual” and “the year of personal cloud computing,” arguing that cloud computers may finally enter the everyday workflow of the mass market.
The team’s logical cloud computer has four features: state and data can persist; a large task can be mirrored into 100 concurrent instances; the environment can roam across phones, tablets, computers, and proprietary hardware; and it can ultimately be driven directly by natural language.
AgentBay is about “letting AI use a computer.” A seven-year-old child or an eighty-year-old adult could operate it through natural language, while a person’s digital counterpart continues working during breaks and conversations. 旭卿 describes the relationship between people and cloud-based Agents as “digital employees” and “silicon-carbon symbiosis.”

15. Shenlong and Early AI Infra Show the Same Pattern of Betting Early

Early in his career, 旭卿 worked on systems virtualization at Intel, by his account spending 3 years as an intern and 6 years as a full-time employee. At the time, fewer than 100 people worldwide may have been researching virtualization; his team had roughly 27–28 people, and nearly 10 were later directly recruited by U.S. companies. Technologies such as Xen and KVM were subsequently adopted by cloud computing.
After joining Alibaba in 2014, he proposed building a computing platform serving tens of millions of customers and scrapping the legacy architecture that hindered scale. The team completed preparations before the mobile-internet application boom in 2015, validating his view that “once the wave arrives, it’s too late to start defending against it.”
By 2016, virtualization’s performance loss had fallen to 3%–5%, but 张建锋 still demanded zero loss. The container boom at that year’s trade shows, along with performance problems in overlayfs and software-defined virtual networks, led 旭卿 to bet on hardware-software co-design. The Shenlong server launched in 2017; the broader industry later followed its standards for DPUs, bare metal, and virtualization.
The AI Infra bet came even earlier. In 2015, he saw an external market in the Group’s need for Nvidia K2 GPUs to run machine learning and support Pailitao image recognition; the product was ready by the end of 2016. Computer vision and speech then entered a boom, and 旭卿 says more than 80% of Chinese technology companies’ AI compute demand ran on Alibaba Cloud. In 2017, he began building GPU supercomputing clusters.

16. The Real Career Leverage Is Technical Depth, Awareness of the Unknown, and Persistent Persuasion

旭卿 summarizes his forecasting method as looking for “traces” in today’s difficulties, especially refusing to treat a seemingly impossible request from a boss as evidence that the boss does not understand technology. Technical managers need to prepare in advance for demand 2–3 years out. In Agent Infra, his concern is not getting planned modules wrong, but “what if there is a piece I haven’t thought of?”
He directly rejects the host’s romanticized question about whether he found something effortless for him but difficult for everyone else: “Quite the opposite.” When he first joined the virtualization team, colleagues spoke Chinese; he could understand the literal words but had no idea what they were discussing. His goal was to truly understand them within 6 months.
His advice to young engineers is not to “know a little about everything.” Foundation models are already sufficiently broad; people need to build depth in one field, then connect adjacent technologies with the business. Without a base of expertise, simply expanding one’s range of knowledge may constrain long-term development.
While pushing the Shenlong chip, he explained its importance to then-leader 李津 almost every morning, taking roughly half a month to secure the investment. The experience shaped what he expects from his team: “When you run into these difficulties, be willing to face them and persuade me.” AgentBay continues to recruit both graduates and experienced AI developers.

Lovable CEO, Anton Osika: The State of Foundation Models, Grok vs OpenAI, and Replit vs Bolt Vol.71 A Quick Chat About GPT-5 and U.S. Power—with 刘一鸣/101 Weekly