D-Robotics and Alibaba Cloud Interview: Eve of the Robot Breakout

Summary

• 秦玉森’s timeline for the embodied-intelligence breakout is “thirty-six months”: hardware itself has a 12–18 month manufacturing cycle, and “two 18-month cycles will let products iterate through two generations, so a large-scale breakout will definitely appear three years from now.” The underlying logic is that oversaturated investment compresses the trial-and-error cycle: one company used to have to make five attempts alone, whereas “today there are ten companies, and one or two of them will have guessed correctly after just two rounds”; he also believes “the companies with meaningful scale that will ultimately remain at the table have already appeared,” while niche vertical markets are “developing deeper underwater, where we still can’t see them.”
• The hardest signal of an inflection point is where talent is going: last year, more new graduates chose embodied intelligence over autonomous driving, marking a historical cycle repeating itself. Around 2018, autonomous-driving companies poached robotics talent at 1–2x pay, pushing starting salaries for graduates in SLAM, motion control, planning, mapping and perception up by more than 50%; the reversal today reflects the fact that “an early, high-conviction outline of the commercial loop for embodied intelligence has emerged.” 秦’s functional defense of bubbles is worth remembering: “A bubble is the gap between expectations and reality… capital bubbles allow an industry to rapidly make mistakes—and rapidly get things right,” while in 2025 “the month-by-month technological progress, and the technological progress of the entire industry, have both been faster than every preceding year.”
• D-Robotics is positioning itself as the robotics industry’s “mother ecosystem,” seeking to fill the integration gap in Nvidia’s ecosystem: “The world has suffered under Nvidia for too long.” 秦 believes Nvidia operates upstream from a “God’s-eye view,” ships 60-point products once they seem usable, and repeatedly changes its simulation-software versions without forward compatibility, forcing the industry to integrate the same pieces over and over; D-Robotics could turn a 60-point product into an 80- or 90-point one. In revenue terms, he says autonomous driving and robotics once sat inside the smallest “Others” category and may have totaled just over 1%; autonomous driving may now account for 3%, while embodied robots “may originally also have been 3%.” He adds that “too many new, fast-moving customers are in China, while Nvidia is too far away from China.”
• 张献涛 sees terminal intelligent computing as Alibaba Cloud’s second major compute market beyond data centers, with Tokens serving as the key unit of pricing and consumption: “Token delivery is already no different from household electricity and water usage.” Alibaba established Token Hub last month, bringing together the group’s AI-related organizations; it “may be the first organization among major tech companies to place Tokens in such an important position.” From chips and AI Infrastructure to model-inference efficiency, “everything is on the same line as Tokens.”
• The cloud-plus-edge combination has caused the developer barrier to collapse, with the most quantifiable evidence coming from a single simulation competition: roughly two WeChat groups, 190 teams, and about 60 using Wuying Cloud, with “a few hundred yuan per person enough to run an entire competition.” Environment setup fell from 4–5 hours to 4–5 minutes out of the box. 秦’s summary: “Five hours can turn a lot of people away, but five minutes can motivate many more.” Large models have also sharply reduced the learning burden for robotics engineers moving across mechanics, mechatronics, embedded systems, software, algorithms and cloud computing—“fifty papers can be read in a single afternoon”—which 卫诗婕 calls “equality of knowledge and equality of experience.”
• A useful metaphor for locating the cycle is that the mobile-internet breakout was signaled by “join an iOS team and get an iPhone,” while today’s equivalent is “join here and get unlimited Tokens, or a top-tier Coding Plan package.” 张献涛 confirms the shift: AI talent asked how many GPU cards a company had over the past two years; this year they ask whether Tokens are unlimited. He thinks 2026 “has that feeling,” with this year at least the “detonation point” for Agents. 秦 adds the generational marker: once Tokens and Coding Plan become default equipment rather than perks, the breakout phase will be over.
• The medium-term risk differs across the three startup camps: autonomous-driving veterans bring scenarios into the market and try to “finish the whole job in one shot,” generating the most volume and growth in the short term, but they may follow a fast-moving consumer-goods logic—“they arrive fast and big, but they also leave faster.” The industry is duplicating too many components, with everyone developing its own motor while testing agencies “make a fortune.” 秦’s verdict: “When someone thinks they can do everything and wants to do everything, they probably don’t know their real mission and vision.” Unitree’s answer is to find the “only one” in motion control rather than stop at “I can.”
• Looking three years out, 张献涛 expects industrial and commercial applications to scale first, with household use potentially later; the Agent “soul” may be integrated into embodied robots and break out over the next 1–2 years, while the cost of the robot body should be at least 50% lower than today. 秦 hopes robots will achieve online-reinforcement-learning-style self-evolution—“really becoming better after having a dream”—while weak-intelligence “lending a hand” in tasks such as mixing baby formula, handing over clothes or closing the washing-machine door, combined with highly granular needs, could create a breakout window for small developers.

Deep dive

1. 秦玉森’s Foundation: Five Master’s Degrees and More Than Two Decades in Robotics

• His career path: five master’s degrees from three European schools in two and a half years, a brief stint at the drone company Zero Zero Robotics after returning to China, then six and a half years at Segway-Ninebot, where he rose to dean of the AI Robotics Research Institute. He left Segway-Ninebot in 2022 and later joined D-Robotics, where he oversees infrastructure R&D across the entire robotics field.
• His starting point was high-school competitions around 2000: “Learn to program first, then touch robotics.” His first robot was two wheels and a board, equipped with only basic ultrasonic sensors, a basic IMU and the laser-ranging equipment of the time—“in 2000, cameras were still a rarity.”
• The original impulse has never changed: Astro Boy and Doraemon, one representing heroism and the other companionship. “Men are boys until they die.”

2. Segway Robotics: The World’s First Balancing-Robot Companion, Killed by Timing

• Around 2015, he worked with his senior colleague Dr. Pu on a companion robot that put “a head on a self-balancing scooter,” using Intel’s first-generation mobile processor and first-generation RealSense. They tested and wrote a substantial amount of RealSense’s calibration and correction code. Microsoft, Intel and Apple were customers, and top U.S. universities bought the robot for teaching and experiments; it was ultimately discontinued and reached EOL after Segway-Ninebot shifted toward commercial delivery and food-delivery robots.
• 秦’s postmortem does not dwell on regret: “You have to respect the objective laws of technological development.” Compute and sensors were not mature, and the consumer moment had not arrived. The real legacy was that “the people we trained later became the backbone and technical leaders of Segway-Ninebot.”

3. From To Lab to To C: The Fixed Script of Informatization

• 秦 uses PC history to frame the anxiety facing embodied-intelligence companies today: the internet initially served laboratories, and “the first users of computers were developers, not users.” Only when Macintosh brought the GUI to individuals did the industry move from To Developer to To User. Today’s leading embodied companies are moving from To Lab toward To B and To C along a similar historical curve.

4. “Technology Itself Is Not What Limits Technology”: An Engineer’s View of the Golden Age

• The core proposition of the episode is that what constrains technical capability includes team-management ability, the company’s financial, commercial and operational condition, the market, the upstream and downstream supply chain, and insight into the business itself. Management, business and product are leverage points for technology, but “many people can’t find the fulcrum and discover that they are using brute force.”
• The environment matters. Slow industries encourage “technical self-indulgence”; in fast industries, technology quickly stops being the constraint, and “constraints appear everywhere.” Only that kind of environment produces genuine engineer growth.
• He borrows from Ray Dalio’s Principles: “Personal growth is not a series of lines extended forward, but a cycle of growth, decline, reflection and growth again.”

5. From Engineer to Architect: The Mythical Man-Month and Design That Fits the Moment

• The original formulation of The Mythical Man-Month remains intact: “One woman can spend ten months carrying one baby, but ten women cannot spend one month carrying one baby.” More resources do not guarantee technical output; but preparing too early can also lead to overdesign and wasted upfront resources.
• An architect’s solution is to break the infinite game into a staircase of Stage One and Stage Two: “Deliver the right thing at the right time, and let it continue to exist until you deliver the next right thing.” The mental model shifts from linear—technology and better technology—to three-dimensional: technology, people, time, collaboration, and even architectural decomposition in the face of time complexity.

6. Seven Years at Segway-Ninebot: The SLAM-Navigation Era; Perception Now Leads

• 2015–2022 was the incubation period for navigation robots. Logistics and warehousing, robot vacuums and DJI-style drone navigation all took off during that period; some of the most profitable commercial robotics scenarios and products today “basically all emerged then.”
• The paradigm has shifted from motion capability as the primary layer with perception as support to “perception and strong planning in the lead.” The underlying reason is straightforward: “Sensors have matured, and compute is much more powerful. If you took a phone back ten years, it couldn’t run any of today’s applications.”

7. Leaving Segway-Ninebot in 2022: When Waiting Became a Faith, and GPT Rewrote the Story

• The IPO exposed him to EBIT and other financial figures. “Excessive pre-research and waiting for robotics technology” had turned into “a faith rather than a business,” while “we have to be accountable to shareholders” became a standard refrain.
• More importantly, he recalls that around June or July 2022—GPT-2.5 should have just arrived—his undergraduate training in NLP made him realize that “the entire story had changed.” The shift from fill-in-the-blank tasks to Next Token was “a change in paradigm.” “You will never wait for the wave of an era to arrive; you have to pursue it.”

8. Clearing His Mind for Eight Months: Treating Graduate-Entrance Exams as Meditation and Reading Strategic Rhythm

• When ChatGPT emerged, he was already working on Agents: memory-first systems, what are called RAG systems today, but with an additional layer of inference and control beyond knowledge-base retrieval. “Letting it hallucinate the right thing within a limited range of hallucinations—that is what inspiration is.”
• After years as a manager, his “head was filled with all kinds of multithreaded noise,” preventing high-quality thinking. His solution was to apply to the inaugural Industrial Innovation MBA at Tsinghua SEM, spending 4–5 hours each day revisiting high-school and university textbooks. After eight months, “the noise was gradually removed, and only then could I enter high-quality thinking.”
• His strategic question changed from optimizing under finite conditions to, under infinite conditions, “choosing what not to do and choosing what to do.” He also moved from passively receiving timing to actively choosing it. The catalyst was his mentor 朱海源’s course Strategic Rhythm.

9. The Inflection Point Has Arrived: Structural Holes, the Spring Festival Gala and “The Computers Used for Training Were Definitely Not Built by Us”

• He uses the “structural hole” framework: when demand remains massively unmet while supply capacity begins to overflow, a new industry inevitably emerges. From last year to this year’s Spring Festival Gala, “not every motor was definitely developed in-house by each company… I’ll put it bluntly: the computers everyone used for training were definitely not built by themselves.” Mature upstream infrastructure has allowed this generation of engineers to show its ingenuity.
• Attitudes reversed within a year: “I remember saying last year that the embodied industry required sitting on a cold bench, but this year the bench is too hot—maybe they are even sofas, and much more comfortable to sit on.” Another rule: inflection points always emerge before their social effects. “The golden age of PCs was not the wave after 1990 at all; it had already arrived with Macintosh in the 1980s.”

10. Talent Flows Signal the Inflection Point: Embodied Intelligence and Autonomous Driving Repeat History

• 秦 dates the inflection point to last year: “Young practitioners and new graduates were entering embodied intelligence in greater numbers, rather than autonomous driving.”
• 卫诗婕 identified the cycle and 秦 confirmed it. It may look as though embodied intelligence is now taking autonomous-driving talent, but around 2018 autonomous-driving companies were poaching robotics engineers at 1–2x pay, pushing graduate salaries in SLAM, motion control, planning, mapping and perception up by more than 50%. The return today reflects the fact that “an early outline of the commercial loop for embodied intelligence has emerged,” and people are pursuing their own work “within an updated, larger story” built around a bigger commercial vision.

11. Why Mega-Financings Are Happening Now: A Functional Defense of Bubbles

• There are two reasons: “once a trend takes shape, everyone simply chases it,” and this year’s Spring Festival Gala rapidly pushed the technology into mainstream awareness and created higher expectations. Last year’s yangge dance, viewed after this year’s gala, “looked rather primitive.”
• His defense of bubbles deserves to be preserved in full: “A bubble is the gap between expectations and reality… capital bubbles allow an industry to rapidly make mistakes and rapidly get things right.” Each company in the market takes a different route; “when it makes a mistake, more resources converge on the right areas; when it gets it right, others put in more money.” Consensus forms quickly, and the tuition is no longer paid by a single company alone.
• The timeline runs from the 2022 transition period to rapidly forming consensus and execution in 2023–2024, then to 2025, when “the month-by-month technological progress, and the technological progress of the entire industry, have both been faster than every preceding year.” “Ten times the resources may even produce 100 times the progress.”

12. Large Models Lower the Barrier: Years of Robotics-Engineering Accumulation Compressed

• The traditional path runs through mechanics, mechatronics, embedded systems, software, algorithms and cloud computing. “You only sample each field superficially; a project takes a year, and completing the whole sequence takes 8 years.” Today, a large model lets people target the knowledge they need: “Throw it fifty papers and you can read them all in one afternoon.” 卫诗婕 summarizes this as “equality of knowledge and equality of experience.”
• Why can young people become Game Changers? “What makes you will inevitably destroy you.” While veteran engineers are still thinking through the first, second, third, fourth and fifth validation stages and inspecting four key points—A, B, D and E—the large model has already told you that only four of the five stages matter: “Just look at four.” Young people have “less noise in their heads, fewer experience-based constraints… and fewer fixed frameworks, so they naturally move faster.”
• The cognitive framework is crystallized intelligence versus fluid intelligence, with 35 as the dividing line. “For making newer, more advanced and better things, 25 to 35 is the golden age.” That echoes his own good fortune in encountering a rapidly growing company during his “golden engineer” years.

13. Rejecting Another Startup to Build Infrastructure: “Only Those Who Have Been Rained on Know How to Hold an Umbrella”

• Asked whether any top-tier embodied company had recruited him, he replied: “I’ve already seen everything I wanted to see… there’s no need to run a tape recorder over my own understanding and insights again.” The pain in his earlier work was the lack of usable tools: “Every time you wanted to write code, you wished you could start by building the computer”; debugging code took 3 minutes, while fixing the computer and robot took 3 days.
• The key question is: “What is OTA for the robotics field?” The shorter the feedback cycle between writing the first line of code and seeing the first change in a robot, “the faster people improve.” In social-recruiting interviews, he often asks how to repair a failed graphics-card driver without reinstalling the operating system. He is testing whether candidates have “really done it themselves,” because “robots need hands-on people.” The theorist who has read many papers but freezes when touching the machine may not solve real problems.
• He defends academics who start companies: “There’s no need to criticize it. Everyone has a different way and paradigm of working… scholars are also exploring the boundaries of what humanity does not yet know.”

14. Connecting with 王丛: One Like on WeChat and Three or Four Months to a Decision

• He first met 王丛 through Segway-Ninebot’s overseas product selection of Horizon Robotics’ X3, having used the Matrix Two platform even earlier. 王丛 was then general manager of Horizon’s AIoT business, and 秦 still keeps his business card. In October or November 2023, 王丛 liked a post in his WeChat Moments, and the two quickly aligned, meeting over the next 3–4 months. “He said he wanted to build the infrastructure of this era and accelerate its arrival. I said, ‘This can work. I’m in.’” The decision was made one month before his MBA entrance exam.

15. Three Startup Camps: Academics Open the Road, Autonomous-Driving Veterans Bring the Scenarios

• Why the first wave had to come from academia: embodied intelligence was something that had never appeared in human commercial society. Only scholars who explore boundaries without treating current conditions as constraints could see it. Even industry insiders were “half-convinced and late to understand it.”
• The camps entered from fundamentally different angles. Engineering veterans follow a “patching and mending for another 3 years” model, building something imperfect but testable in the market and iterating alongside it. Autonomous-driving veterans try to “finish the whole job in one shot,” bringing a scenario rather than technology into the market, anchoring on a scenario and a product and perfecting it in one go—the perspective of end-state product competition.
• 秦’s medium-term risk assessment is that the autonomous-driving camp will lead in both near-term volume and growth, “but this kind of speed may show fatigue in the medium term.” The scenario is already there, while newer technology “may not be better, but it will definitely rise along some dimension,” potentially creating a dimensionality-reduction attack. “If you are only making one product in one scenario, you may end up following the logic of a fast-moving consumer good: it arrives fast and big, but leaves faster.” Moving from the second stage to the third is difficult: What is the next product? Where is the next scenario? “Focus itself suddenly stops working at that moment.”

16. The Beauty of Robots Lies in Their Clumsiness

• Asked what is beautiful about robots, 秦 gave an unexpected answer: “They are beautiful because they are clumsy, because they are immature… the smarter they are, the weaker your sense of accomplishment in transforming them.” He uses crayfish as an example: the pleasure comes from the early feedback of something moving from very dumb to slightly smarter, one step at a time.
• Robots have remained both clumsy and slow for a long time. “As long as you find one technical point, you discover a qualitative and quantitative transformation,” creating a powerful sense of achievement. 卫诗婕 joked that “interest comes from human vanity,” and 秦 corrected her: “Or from the feedback of achievement.”

17. D-Robotics as the Mother Ecosystem: Air, Soil and Water

• The name is not a joke about “a melon on the horizon.” It evokes something “growing quietly underground, with tenacious life, and able to provide life-saving food and nourish more life when everyone is in a barren period.” He also stresses that the real foundation is chips and the higher-education system that supplies talent. D-Robotics’ role is to “make capabilities visible and lower the barriers.”
• The business model is To B plus To C, but inside the product it is “only To D, meaning To Developer.” The largest developers are CTOs and CEOs; in the middle are the technical leaders for algorithms, software and complete machines; at the smaller end are startup founders, university makers and engineers at large enterprises. D-Robotics takes the common denominator across these needs “to prevent everyone from reinventing the wheel.”
• His development-board analogy: a computer without an operating system is unusable. D-Robotics builds system-layer tools above the board so users no longer have to “install the machine, install the operating system, write software, use software, then iterate the software, slowly moving forward.” After an hour of training, “a junior, sophomore or even freshman can learn in one hour what an algorithm engineer would have spent 3–5 years doing.”

18. Why Infra Is Scarce: Only Fast Industries Pay for Speed

• The logic is straightforward. When an industry moves slowly, algorithm teams build their own scaffolding—“output in the first half of the year, capacity in the second half”—and companies choose the option with the lowest TCO over the full life cycle. “Only when speed is high and has direct competitive value will anyone serve speed or pay for it.” The automotive industry’s real-world feedback loop and the window for training large-model clusters were both short; tacit knowledge was concentrated in a few people, leaving a major supply gap. D-Robotics’ answer is to “package the dirty, hard and exhausting work, do it once, and serve the entire industry.”
• He is candid about the boundaries of embodied intelligence: frontier technology “has not converged, and we are not entirely sure” what the final form will be. Embodied intelligence is one state of breaking through a capability ceiling, with excess capability flowing back into unmet demand—just as large models spilled over into lower-quality simultaneous interpreting. “People who only write functional code have already been replaced at high speed by large models, but architects have not been replaced nearly as much.”

19. From Traditional Customers to New Geeks: “Existing Knowledge Inventory Has Become Obsolete”

• The strategic shift is clear. Last year the company targeted the intelligent upgrading of traditional robots—the team had just been formed, and “the best challenge is one only slightly beyond your own capabilities.” This year it is targeting emerging markets and innovative geeks with RoboGo, an all-in-one development platform designed to “quickly give one person the integrated capabilities of an entire team,” covering algorithms, data, simulation, training, models and development-board management.
• The clearest example of obsolete thinking is that people obsessed with end-to-end systems cannot see that “remote control can also make a robot walk very well.” Many commercial demonstrations over the past year were remotely controlled behind the scenes. “Those who believe a technology must be perfect before entering the market will definitely miss and never see this market, and they will miss the technology dividend.” That also explains why autonomous-driving veterans entered in the third wave: their mental model is already end-to-end and L4, while a large portion of robotics today is still debating between L2 and L3.
• To customers unwilling to change, 秦 offers a calm but cutting line: “We respect every customer’s fate.” Organizational inertia and “learned path dependence” can both accelerate and constrain a company.

20. Reinventing the Wheel and the “Only One” Thesis

• The industry reinvents the wheel “quite a lot”: everyone develops its own motor. “A motor is just three things: winding the wire, making the coil, writing the electronic controls, then putting a shell around it for certification.” Motor contract manufacturers and CE and FCC testing agencies consequently “make a fortune.” The root cause is not a desire to serve capital, but engineers’ pursuit of perfection. 秦’s verdict is harsher: “When someone thinks they can do everything and wants to do everything, they probably don’t know their real mission and vision.”
• Core competitiveness means finding the “only one,” not proving “I can.” Unitree is “number one in motion control,” while many others are Me Too. As for the argument that self-developed technology creates a moat: “It develops in-house because it has nothing to use, but today everyone has something to use.” Otherwise, every company building an embodied brain would have to start with chips. 卫诗婕 adds that Unitree not only found its point of differentiation but proved it. 秦’s reply is telling: “Or perhaps the era made it the answer that was proved.”
• The ideal partner relationship is a two-way fit. D-Robotics and the domestic autonomous-control engine company Moxianfei empower each other: “If the upstream does not form a combined force, the downstream has to do a lot of integrative work. If the two of us integrate once, every customer pays the cost of integration one fewer time.”

21. Benchmarking Against Nvidia: “The World Has Suffered Under Nvidia for Too Long”

• The industry complaint is direct: “It does everything, but never quite fully… it stands at an extremely high upstream vantage point, gets to 60 points, decides it is usable, and throws it out.” The example is simulation software whose version numbers have changed repeatedly and which is “one of the few pieces of software that does not maintain forward compatibility,” forcing the entire industry to rebuild after every upgrade. D-Robotics’ role is potentially to take a 60-point product to 80 or 90 points, reduce secondary integration and make version updates smoother.
• The structural opportunity is visible in the numbers. 秦 says autonomous driving and robotics once sat in the smallest Others category and may have totaled just over 1%; autonomous driving now contributes somewhat more and may account for 3%, while embodied robotics “may originally also have been 3%.” In addition, “too many new, fast-moving customers are in China, while Nvidia is too far away from China,” creating gaps in internal coordination and customer service. The difference in posture is summarized in one line: “Nvidia looks at the world from a God’s-eye view. We are still willing to be a member of this industry and push it forward—so why not D-Robotics?”

22. The Breakout Point: Thirty-Six Months, Because Ten Companies Are Testing in Parallel

• The timeline is driven by hardware’s 12–18 month manufacturing cycle plus the time needed for mass-market acceptance: “Two 18-month cycles let a product iterate through two generations, so a large-scale breakout will definitely appear three years from now.” This echoes 王兴兴’s shift from a 5–10 year view to a shorter timeline. Under oversaturated investment, paths converge quickly: “Previously, one company had tried only two things after 36 months, even though it needed to try five. Today there are ten companies, and one or two of them will have guessed correctly after two rounds.”
• The assessment of who remains at the table echoes investor 陈玉’s autonomous-driving-window thesis: “The companies with meaningful scale that will ultimately remain at the table have already appeared.” But small vertical markets are still incubating: technology has not yet spilled over to them. This resembles the late surge of note-taking, personal-knowledge-base and enterprise-knowledge-base companies during the large-model era. Because foundation models are advancing so quickly, “no one dares to build” in these niches for fear of being swallowed after years of work. “The niche vertical markets should be developing deeper underwater. We still can’t see them.”

23. China’s Edge: The World’s Largest Engineer Base and the Confidence to Imagine

• The core of China’s supply-chain advantage is the organization of talent: “China has the world’s largest technical engineering population. Their greatest need is to be organized, attracted by a vision and a long-term horizon, and brought together.” From DeepSeek and 王兴兴 to the leaders of several large-model companies, the common thread is “believing they can build the best thing, rather than believing they should only build a Me Too product.” “When the U.S. starts copying from China, everyone should understand that we are among the best engineers in the world.”

24. 张献涛 Enters: Twelve Years from Elastic Computing to Terminal Intelligent Computing

• His career and bet: he joined Alibaba Cloud in 2014 and will complete 12 years next month. For his first 10 years he worked on elastic computing, a foundational cloud-computing product; since 2024 he has led terminal intelligent computing. He divides compute into data-center and terminal markets. In 2019, he saw the opportunity: “If you could bring the tens of thousands of cores of compute in the cloud to a single terminal,” you could break the fixed local limits of “four cores and 8G, or eight cores and 16G.” That led to the launch of Wuying Cloud Computer in 2020.
• At the beginning of this year, the Wuying business unit was renamed the Terminal Intelligent Computing business unit because terminals such as cars and embodied robots “are relatively weaker in compute.” Integrating large-model capabilities into terminal compute “should bring about a much bigger change.”

25. Alibaba Cloud’s Customer Generations: From Small Webmasters to the Hundred-Model War

• The phases are clear: before 2014, Alibaba Cloud served small and midsize webmasters and hosted corporate sites; after 2013–2014, the mobile internet drove e-commerce, games and internet finance, which could require tens of thousands or hundreds of thousands of cores in a short period; after 2016–2017, the previous AI wave moved training and inference to the cloud; by 2023, “almost every company in the hundred-model war was on Alibaba Cloud.” “From the internet to the mobile internet and then to the AI internet, the required compute model changes with each phase.”
• Why terminal cloud computing only became viable in 2019: data centers could be configured by a few operations engineers logging in through a Terminal, while a cloud computer had to “maintain the experience of using a local computer.” This meant matching latency and responsiveness, not just compute. Network upgrades, together with a cloud-to-terminal transmission protocol Alibaba began tackling in 2016 and had brought to initial maturity in early 2019, made that year the “starting point, or inaugural year,” of terminal cloud computing.

26. Cloud Plus Edge Solves the Developer Pain: Five Hours Discourage, Five Minutes Motivate

• 秦 lists the pain points students face when building robots: dual-booting laptops and switching back and forth, universities unable to afford powerful GPUs even though students only need them for 8–10 hours of lab work, and the fear of the command line—“the first time you face a black window with only a cursor blinking.” Wuying Cloud desktops with preconfigured images make the environment ready to use: “Configuration time fell from the most painful 4–5 hours of environment setup to 4–5 minutes after opening the box. Five hours can turn a lot of people away, but five minutes can motivate many more.”
• The data is concrete. A single simulation competition involved roughly two WeChat groups and 190 teams, with about 60 using Wuying Cloud; “a few hundred yuan per person was enough to run the entire competition.” Training labs cost less than RMB200. 张献涛’s product logic is simple: “So many customers were asking us for these things; I had no reason not to build a product that was easy for them to use.” The company began roughly two years ago to build an AI workstation that could be integrated into webpages and apps and used out of the box.
• 秦 describes the collaboration spillover from cloud desktops in practical terms: keep the R&D environment in the cloud so colleagues in other locations can log in, see the current state and continue the work; give interns permission to explore in a non-production environment; and let employees leave the computer at home and use a phone in an emergency. The three-layer structure—cluster-scale GPUs, cloud desktops and smaller-scale cloud computers—“completes the triangle of support” from large enterprises and midsize development teams to individual developers.

27. The Token Economy: AI’s Water and Electricity, and Alibaba’s Token Hub

• Alibaba established Token Hub last month, bringing together AI-related organizations across the group, including the Qwen foundation-model team at Tongyi Lab, the Wukong business unit and the Qwen consumer business. This “may be the first organization among major tech companies to place Tokens in such an important position.” 张 put it plainly: “Token delivery is already no different from household electricity and water usage. An Agent holds an API Key; powered by Tokens, it can make a person or an intelligent agent do what you want it to do.”
• The Token economy has several layers: “AI Infrastructure efficiency determines the efficiency and cost at which Tokens are produced,” while model-inference efficiency matters just as much. “From chips to infrastructure to models, everything is on the same line as Tokens.” The high-level Agent architecture is Token plus cloud computer; beneath that are four major blocks: model engineering, memory engineering, sandbox engineering and prompt engineering.

28. Crayfish and the Victory of the Individual Developer

• 张’s reading of OpenClaw is that it was built by an individual developer, with “no big-tech position,” and is open-source and neutral. It “may be a foundational project on the road to future AGI. It may not be the only one, but it should be foundational.” Alibaba recently launched ClawCloud, allowing people to “raise a crayfish” through a mobile app. “Only with more developers participating can this industry become more prosperous.”
• The comparison is mobile internet. In 2013, perhaps 80% of users still participated in Singles’ Day through computers; that figure declined in 2014, prompting Alibaba to go all in on wireless. “Some companies may have gradually declined simply because they failed to integrate into the mobile internet.” The developer population surged in 2014–2015, with mobile games providing the defining demand: compute elasticity, storage capacity and network packets processed per second all rose exponentially.

29. From “Join and Get an iPhone” to “Unlimited Tokens”: Is 2026 Like 2014?

• 秦’s most quotable cycle indicator is that the mobile-internet breakout was signaled by “join an iOS team and get an iPhone”—both work equipment and a disguised perk. Today’s equivalent is “join here and get unlimited Tokens, or a top-tier Coding Plan package.” Once Tokens and Coding Plan become default configurations rather than benefits, “the breakout phase will be over.”
• 张 confirms the shift in talent questions. AI recruits asked how many GPU cards and how much high-quality data a company had during the prior two years; this year they ask whether Tokens are unlimited. He thinks 2026 “has that feeling”: last year was the Year One of Agents, and this year “should count as the year of the Agent breakout,” or at least its detonation point. 秦 gives the more precise periodization: it was really 2014, 2015 and 2016—the spread of iPhone 6 and the transition from 3G to 4G. After 2016, “nobody mentioned joining and getting an iPhone” anymore; it was already the post-mobile-internet era, the era of short video.

30. Cloud-Edge Co-Evolution: The Robot’s “Dream System”

• 张’s division of labor is straightforward: large models in the cloud, with higher latency but greater intelligence and multimodality, handle task decomposition, planning, decision-making and environmental perception; many small models on the edge act like “muscle memory,” executing balance, walking and real-time joint force control. 秦 compares it to learning to ride a bicycle: at first the brain says, “If you fall left, turn left,” but once the edge model is good enough, “just look forward; you will definitely ride the bicycle well.” The two mainstream technical paths are distilling cloud models to the edge and adding supervised learning based on cloud feedback.
• The deployment examples are already dense. The open-source project URDF Studio converts robot drawings into standardized description files for simulation reconstruction; its GitHub group “quickly swelled to 300–400 people,” and algorithm engineers “finally no longer have to ask mechanical engineers to do this for them.” Combined with AgentBay to schedule edge chips, robots can learn skills during the day, while cloud agents supervise during rest periods, send control signals and accumulate knowledge in reward functions. “This is the robot’s ‘brain-dreaming system,’” and cloud-edge co-evolution may become a core feature delivered to many developers later this year.
• On perception specifically, a robot can use Qwen’s VLM to call cloud compute and Tokens once to identify an unfamiliar object, store the image locally and remember “this is a microphone.” “Tomorrow it won’t need to call the cloud’s Tokens again.” As 卫诗婕 summarizes, the Agent’s job is “to translate one form of intelligence into another so the edge can absorb it efficiently.”

31. AI Ready: Infrastructure’s User Is the Large Model, Not the Human

• Asked what infrastructure the Agent era needs, 秦 gives one sentence: “The user of infrastructure is the large model, not the human. That’s it.” The practical task is to provide “AI Ready training” for Agents covering embedded systems, software, algorithms, project management and hardware product management—injecting RAG and reinforcement learning “so it forgets capabilities it should not remember.” Human in the Loop keeps raising its level of expertise. The fundamental question remains: “The documents were originally written for humans. Have you ever considered AI’s feelings?”
• 张 offers two directions. First, build the Agent itself by making AI Infra more efficient and models more intelligent. Second, “make the digital and physical worlds friendly to Agents” by opening software and data capabilities to Agent self-evolution through MCP and CLI. His observation is that “self-evolution has already become a standard feature of Agents”; install the right Skill, and “the crayfish you raise gets smarter and smarter.”

32. Altruistic Ecosystems, the Perpetual 59% and Three Years Out

• The answer to what makes a good ecosystem is one word: “altruism.” D-Robotics’ position is naturally consistent with that principle: “We are upstream by nature and do not truly compete with anyone. The earlier, faster and more forcefully the robotics era arrives, the better D-Robotics develops. Altruism is self-interest.” 张 sees a long industrial chain in which dexterous hands, joints, robot bodies and models each have room to specialize; the industry has not yet reached a zero-sum phase. But he also believes that “when it truly reaches the end, it will look like what happened in the automotive industry over the past 2–3 years. That is inevitable in industry development.”
• Infrastructure’s Sisyphean answer is: “Every day we climb from 59% to 61%… we are always between 59% and 60%. Good enough is enough; then hurry forward.” Infrastructure forever runs behind the business. Once everyone has explored the infinite possibilities, the industry can “converge and solve for the largest common denominator.”
• The three-year forecast closes the episode. 秦 hopes “robots can truly self-evolve—really becoming better after having a dream,” with online reinforcement learning allowing a robot to perform better the second time it encounters something it saw that day. 张 expects industrial and commercial scenarios to scale within 3 years, with household use potentially later; the Agent “soul” may be integrated into embodied robots and break out over the next 1–2 years, while robot-body costs should be at least 50% lower than today after 3 years. 秦 adds a product thesis around weak intelligence “lending a hand”: mixing baby formula, handing clothes up from a basket, or closing the washing-machine door after someone leaves. “Once the cost comes down, weak intelligence will also have room to perform.” 卫诗婕’s closing line is the episode’s theme: weak intelligence plus a huge range of niche needs belongs to small developers—“we are about to enter an era of developer breakout.”