The US is building “Stargate” and China is refining the “Token Factory”
When the revenue growth rate of leading AI cloud vendors in North America broke through three digits in a single quarter and on-hand orders exceeded 100 billion US dollars, market habits interpret this as “demand for computing power is far from peaking.” However, demand is only an indication; what really determines the number of players at the table is supply-side systemic barriers — countries that can simultaneously control 10,000 card clusters, GW class electricity, autonomous chip ecology, and model iteration capabilities. China and the US are the only countries left in the world. The Oxford University and Stanford University reports confirm this: the two countries have concentrated the vast majority of large-scale computing power factories in the world, and the performance gap of the leading edge model is only 2.7%, and they have already entered the “parallel running” stage; the rest of the countries either only have supervision or capital, lack a complete chain, and are already at different tables.
What defines the end of this game is the “Token Factory” theory proposed by Hwang In-hoon at the GTC conference: an intelligent computing center is no longer a storage warehouse, but a factory that produces terms; commercial value = throughput per watt × usable power. Electricity comes in, and the word comes out — whoever first runs this transformation chain into an industrial assembly line that can be measured, priced, and traded, will get the pricing power for the next generation of “general commodities.”
1. How America Can Play: Outsource National Strategy to Market Alliances
The symbol of America's answer is “Stargate” (Stargate). This superproject, jointly sponsored by OpenAI, SoftBank, Oracle, and the Sovereign Fund, is worth 500 billion US dollars in four years, and the planned capacity is close to 7 GW. The Abilene Flagship Park in Texas is 1.2 GW in a single phase. It was once known as “a nuclear reactor-grade computing power factory.” Its significance is not limited to scale — for the first time, it changed “national computing power infrastructure” from government bidding to market behavior coordinated by giants, replacing the pace of administrative examination and approval with the speed of capital.
The style of play can be classified as three points: the giants join forces and exalted capital; bind a single technology ecosystem to achieve vertical integration and pursue the ultimate computing power density; under the closed loop of heavy capital, engineering logic and capital logic are always wrestling. Leading cloud vendors have signed long-term agreements one after another, essentially an extension of the same set of logic: use deterministic orders to lock in electricity and cabinets for the next three to five years, and turn CAPEX into discounted cash flow.
The one that best explains American logic is still Nvidia's actions. Over the past 16 months, it has completed more than 145 mergers and acquisitions and spent a total of about 90 billion US dollars. From the inference chip team to the model open source platform to the code generation technology license, it has taken almost every card “from energy to basic models to applications” into its own hands. In the US, full stack is a kind of capital will. The “token factory” concept is the ultimate measure of this vertical integration: no longer has more GPUs than who, but how many terms can be “refined” per watt of electricity. Computing power changed from a cost center to a profit center, and a data center from an “electronic warehouse” to a “production workshop” — this is not only a new story for Wall Street, but also the underlying operating system for the American route.
2. How to play in China: Officials set the tone, private enterprises compete for the lead
The government first treated the most difficult external conditions into deterministic supply: “East Digital and Western Arithmetic” defined the framework, “Computing and Electricity Collaboration” was incorporated into the top-level design, and “the end of computing power is electricity” became a clear consensus. The country paved “surplus water, sufficient electricity, and usable land” as a public foundation, and then let market players form their own rules.
At the same time, “vocabulary” is changing from technical jargon to industrial language. In March 2026, the National Data Bureau officially defined the Chinese translation of Token as a “word element”, clarifying its “measurable, priceable, and tradable” product attributes; on the other hand, China's average daily word call volume has exceeded 500 trillion dollars, an increase of nearly 5,000 times over two years ago. When the “etymology economy” is incorporated into local industrial plans, China's “token factory” is no longer just a corporate act, but an industrial chain formed from the top down, from infrastructure to settlement standards.
Runze Technology (300442.SZ) is a model for “base first”. In the first half of 2026, its AIDC business revenue increased 126% year over year, surpassing traditional IDC for the first time. It replicates exactly the logic of its North American counterpart: first solve the physical bottlenecks of computing power deployment, power supply and cooling, and high-density deployment, and then load high value-added services upward. More importantly, it has already obtained multiple 10,000 card cluster delivery contracts and is beginning to touch the edge of computing power scheduling — moving from “turnkey” to “communication element”, only one layer of software away.
Co-Creation Data (300857.SZ) and Litong Electronics (603629.SH) represent the collective actions of “completers”: one extends from server manufacturing to the intelligent computing center and both ends of the optical module to try to turn the “stuck neck” into “self-made” at the hardware level; one shifts from leasing computing power to self-built production capacity to increase the underlying computing power independently and controllable. They confirm the same consensus — although the full stack is difficult, no one wants to be absent; “making up” itself is tightening every screw in the word industry chain.
The Zhitong Finance App learned that Guangdong-Hong Kong Bay Intelligent Computing (01396) has made the “full stack” more complete. The Hong Kong stock company's revenue in the first half of the year increased more than tenfold year-on-year, with on-hand orders exceeding 37 billion yuan (RMB), and FP16's dense computing power scale reached 50,000 PFLOPS. Its “full stack” is not a simple superposition of links; it is based on the self-developed cloud platform “Quantum Power”, which not only has the ability to independently schedule 10,000 card clusters, but also links energy bases, computing power clusters, term delivery, and application incubation into a closed loop: internally, it can dynamically allocate power and computing power according to the training inference load, output token services externally using a standard interface, and reverse optimize upstream resource allocation. This shift from “selling computing power” to “syntactic elements” just falls on the logic of a “token factory.” While most of its peers are still figuring out “whether or not,” it has quietly outlined the delivery outline of the Token factory in China.
3. A special path, possible return
The best perspective was never to look for a Chinese copy of an overseas giant. What is really worth watching is that the two routes are leading to the same end: sedimentation of astronomical CAPEX into sustainable cash flow and token production.
The end of electricity is computing power, and the end of computing power is pricing power. Whoever first turns “electricity” into a “word” that can be measured, traded, and can go to sea will hold on to the next round of tickets. In this batch of responses in China, the first person to show this card has already taken a higher position.