Author | Liu Wei, Editor | Lin Juemin
Within a single week, Edison has become the darling of the AI circle.
On September 16, at the Intelligent Economy Forum held at the People's Daily, Shen Dou of Baidu used an analogy from the electricity era to position the current stage of the AI industry.
Six days later, at the Alibaba Cloud Apsara Conference, Wu Yongming also talked about electricity, stating that tokens are the new electricity and the cloud is the power grid. The two most important annual speeches by China's AI clouds seemed to have been coordinated, telling the same story from afar.
Behind this coincidence lies a consensus.
Over the past year or so, the large model industry's attention was still focused on leaderboards, who released a new model, and whose benchmark scores increased by a few points. But this September, the themes of the two conferences coincidentally shifted from "how smart the models are" to "how intelligence can be transformed into productivity".
Behind this is a shared judgment: models are no longer the primary bottleneck. In the next phase of AI industrialization, the competition is about who can deliver electricity to thousands of industries first. Regarding the judgment of industry trends, China's two major AI clouds are already standing on the same starting line.
However, on the same starting line, the directions the two are running towards are not the same.
01. Full-Stack Is Not a New Term; Some Have Already Laid Out the Strategy
The analogy of the electricity era has a subtext: if you want to be a power generator, you must first have a complete power grid. Applied to the AI industry, this is called full-stack layout.
This is not a new buzzword that only became popular in 2026. As early as 2015, "Xiantong," Baidu's highest award-winning team, was working on chips. By 2018, Baidu had already formed an AI full-stack technology layout "from chips to deep learning frameworks, platforms, and ecosystems," when the large model wave was not even in sight.
In 2020, Baidu AI Cloud took the lead in proposing "Cloud-AI Integration," merging cloud and AI into an infrastructure system, forming an intelligent closed loop of "chip—framework—large model—industry applications," and laying a solid foundation for the AI full-stack technology layout.
In 2023, Robin Li defined the IT technology stack in the AI era as four layers: "chip—framework—model—application." In October of the same year, the Baidu AI Cloud Super Factory, covering full-stack services for computing power, models, and applications, made its debut.
In 2024, this capability was upgraded to an "AI+" full-stack enterprise-level service system, beginning to penetrate the industry. The computing power platform and model platform form the core, supporting industry capability building and scenario application innovation from the inside out.
In 2026, the industry focus shifted from models to AI agents. Robin Li proposed the "chip-cloud-model-agent" full-stack AI, and Baidu AI Cloud, with the "new full-stack AI cloud," is ready to embrace the explosion of large-scale AI agents.
Alibaba is also laying out the full stack, and doing so rapidly. Alibaba's strategy is to place the three cornerstones of chips, cloud, and models in equally important positions, pushing them forward together, and waiting for a new species to grow.
02. The Same Flywheel, Different Starting Points
Regarding the future, Shen Dou gave his judgment at the forum: the four layers of chips, cloud, models, and AI agents amplify each other, which is the growth flywheel of AI as the underlying infrastructure. Alibaba's three cornerstones of chips, cloud, and models interlock with each other, and when they spin, it is the same machine.
For the same flywheel, different starting points mean the flywheel spins in two different ways.
The electricity era also experienced the same divergence. Some built power stations first and waited for users to come; others first found factories that absolutely needed electricity and then built power stations according to the factories' needs. Over a hundred years have passed, and both strategies survived in the end, but the ways they survived are completely different.
Baidu places the starting point of the flywheel on real business value. Industrial AI agents first enter the front lines of enterprises to do real work, solve practical problems, and create efficiency and growth. Doing real work generates real business data, data drives model iteration, and model iteration drives larger-scale investments in cloud and chips. As the scale expands, costs further decrease, allowing more enterprises and more scenarios to afford and use them well, thereby releasing new demands and driving the continuous acceleration of the flywheel.
This "chip—cloud—model—agent" chain is closely linked, but the ignition point of the whole machine is only one: the real landing of industrial AI agents and the realization of value. Whether the first circle of the flywheel can spin depends not on the parameters on the PPT, but on whether enterprises are willing to integrate it into core processes, and whether it can effectively solve problems and improve business results.
Shen Dou's statement is that industrial AI agents are driving AI to become a real economic force. The weight of this statement lies in the fact that, for the first time, AI is not talked about as a technology, but as an economic variable.
Alibaba's flywheel relies on supply. As computing power goes up, the model (Qwen RSI) becomes stronger accordingly; as capabilities become stronger, it waits for representative products to emerge; as products emerge, it waits for the explosion of demand.
Placing the starting point on the supply side is based on a judgment: "representative products" will definitely emerge in the future, thereby triggering large-scale demand. Building 20 GW of data centers before 2032 is the bet placed on this judgment. This strategy is in the same lineage as Alibaba's values: because we believe, we see; build the power station first, and wait for the users to come.
The two flywheels each have their own soft spots. The business flywheel fears that the landing will fall short of expectations; if AI agents cannot do real work in enterprises, every subsequent link cannot be driven; the supply flywheel fears that representative products will not come for a long time; the power station is built, but the users have not yet appeared.
The two flywheels are also influencing each other. The faster Baidu's AI agents run, the greater their appetite for computing power, which objectively also raises demand on the supply side.
Once the representative products Alibaba has been waiting for emerge, the first batch of enterprises to scale up will still choose the cloud that can realize value first. Flywheel against flywheel, the competition is not about who spins first, but who spins to scale first.
One pulls supply with demand, and the other stimulates demand with supply; one bets that value will be realized, and the other bets that the future will arrive. The two routes each have their own accounts.
03. Charging for "Thinking" or Charging for "Results"
The divergence of routes will ultimately fall on the pricing method. What cloud vendors price is what customers pay for; these are two kinds of economics.
Alibaba is about Token economics, supply first, charging by the volume of thinking; the more tokens consumed, the longer the bill. This is very much like the early electricity meters in the electricity era, charging by kilowatt-hours first, and then scaling up the electricity consumption.
What Wu Yongming presented at the Apsara Conference is the imagination space: in the future, the total volume of machine thinking will reach more than 1,000 times that of humans. If the thinking volume increases by 1,000 times, token consumption will also increase by 1,000 times. This is the demand curve that the supply side most hopes to see. What customers buy is the possibility of capabilities.
Baidu is about productivity economics, demand-driven, tying revenue to business value. Shen Dou provided a set of global metrics at the Intelligent Economy Forum: over the past three years, model capabilities have been increasing, but GDP has not changed significantly; the proportion of AI capital expenditure to GDP is only 0.4%, while the electricity system is close to 5%.
In terms of these figures, this round of investment is still far from the electricity era. Shen Dou expects that by 2030, there will be 3 billion industrial AI agents running continuously, bringing more than 100 GW-level inference demand. Every bit of infrastructure investment is backed by real scenarios.
Charging for "thinking" or charging for "results," China's AI clouds are walking out two different routes. The former bets that the exponential growth of thinking volume will generate demand on its own, while the latter believes that value must be realized in real production and operation.
And when the two routes converge in real industrial practice, the common answer of China's AI clouds will emerge.
04. The Customer's Ledger Is More Honest Than the Press Conference
The customer's ledger best illustrates the difference between the two routes.
Customers who choose Baidu are buying certainty of value. IAT, a leading domestic complete vehicle R&D enterprise, worked with Baidu to optimize automotive aerodynamic drag design, entrusting the task to FM Agent, the self-evolving super AI agent on Baidu AI Cloud.
It autonomously evolved for over 120 rounds, compressing a single CFD simulation from several hours to the second level.
China Southern Power Grid uses Baidu's ontology-building AI agent Shengsuan to solve problems in real business scenarios of the power grid; after deploying the AI agent development platform in April this year, Taikang launched more than 3,500 AI agents, with a daily average token consumption of 20 billion, deployed in core scenarios such as medical health, risk control and compliance, and sales operations.
The commonality of these projects is that the acceptance criteria are not model benchmark scores, but whether problems have been solved, costs have been reduced, and new growth has been achieved.
Customers who choose Alibaba are buying possibilities. A total machine thinking volume of 1,000 times; if the usage cost drops to the level of electricity prices, the economic benefits brought will be incalculable. It's just that we don't know how long it will take for this day to come.
In the supply-first strategy, customers and cloud vendors are more like partners, waiting together for a new species to grow.
The simultaneous existence of the two pricing models precisely shows that this market has not yet reached the point of fixing the pattern. And the only two that can sit at both ends of the poker table at the moment are these two.
05. Conclusion
Back to the electricity era. The greatness of that era lies not in Edison building power stations, nor in factories replacing with electric motors, but in the power stations and the electricity users achieving each other, pushing the whole world into electrification.
The power station waited for the electric motor, and the electric motor also made the power station cost-effective; this is a two-way rush that has lasted for decades.
The electricity era of AI will also reach this step. The value of the first landed AI agents is being continuously realized in the real business cycle, and the first laid out computing power is waiting for its representative products. The starting points of the two routes are different, but both are driving the foundation of China's AI industrialization deeper.
The real winning hand is not today, but when the next industrial wave hits, whose flywheel spins to scale first.
What converges from different routes is that China's two major AI clouds are jointly defining the Chinese path for AI industrialization.