While leading North American AI cloud vendors push quarterly revenue growth past triple digits and backlogs cross the $100 billion mark, markets habitually read this as proof that compute demand is far from peaking. But demand is only the surface; what truly determines the seating order at the table is the systemic moat on the supply side—only China and the US remain globally capable of simultaneously commanding multi-10,000-GPU clusters, gigawatt-level power, proprietary chip ecosystems, and model iteration capacity. Reports from Oxford and Stanford confirm this: the two nations concentrate the overwhelming majority of the world's scaled compute plants, and the performance gap in frontier models has narrowed to just 2.7%, entering a head-to-head phase. Other countries either have regulation alone or capital alone, lacking the complete chain—present at the venue, yet not seated at the table.

What defines the endgame of this contest is the "Token Factory" thesis Jensen Huang introduced at GTC: intelligent computing centers are no longer storage warehouses but factories that produce tokens; business value equals token throughput per watt multiplied by available power. Electricity goes in, tokens come out—whoever first turns this conversion chain into a measurable, priced, and tradable industrial assembly line captures the pricing power of the next-generation "universal commodity."

How the US plays it: outsourcing national strategy to a market alliance

The symbol of the American answer is "Stargate." This megaproject, jointly lifted by OpenAI, SoftBank, Oracle, and sovereign wealth funds, carries a $500 billion four-year price tag with planned capacity approaching 7GW. Its flagship campus in Abilene, Texas, houses a single-phase 1.2GW build—once dubbed a "nuclear-reactor-grade compute plant." Its significance goes beyond scale: it marks the first time "national-level compute infrastructure" has shifted from government tenders to a market-driven alliance of giants, replacing administrative approval rhythms with the speed of capital. Its playbook can be distilled into three points: giant alliances with heavy capital firepower; vertical integration anchored to a single technology ecosystem to chase extreme compute density; and under a capital-heavy closed loop, engineering logic and capital logic perpetually wrestle for control. Top cloud vendors are signing multi-year long-term agreements, all extending the same logic: locking in power and racks for the next three to five years with deterministic orders, converting CAPEX into discountable cash flow.

The clearest expression of American logic still comes from Nvidia. Over the past 16 months, it has completed more than 145 M&A deals and investments totaling roughly $90 billion—from inference chip teams to open-source model platforms to code-generation technology licenses—nearly gripping every choke point from energy to foundation models to applications. In the US, full-stack is an act of capital will. The "Token Factory" concept is precisely the ultimate yardstick for this vertical integration: no longer comparing who owns more GPUs, but how many tokens each watt of electricity can "refine." Compute shifts from a cost center to a profit center; data centers evolve from "electronic warehouses" to "production workshops"—this is both a new story for Wall Street and the underlying operating system of the American route.

How China plays it: official direction set, private enterprises racing ahead

The state first transforms the hardest external conditions into deterministic supply: "East Data, West Computing" draws the skeleton, "computing-power coordination" enters top-level design, and "the end of compute is electricity" becomes a declared consensus. The nation lays down "surplus water, sufficient power, and usable land" as a public foundation, then lets market players carve their own paths. Meanwhile, "tokens" are shifting from a technical term into industrial language. In March 2026, the National Data Administration officially designated "词元" as the Chinese translation for Token, clarifying its commodity attributes of "measurable, priced, and tradable." On the other hand, China's daily token calls have surpassed 500 trillion, a nearly 5,000-fold increase over two years. As "Token Economy" enters local industrial planning, China's "Token Factory" is no longer just corporate behavior but an industrially forming chain from top to bottom—infrastructure to settlement standards.

Range Intelligent Computing Technology Group Company Limited (300442.SZ) exemplifies the "foundation-first" playbook. In H1 2026, its AIDC business revenue grew 126% year-over-year, with its share exceeding traditional IDC for the first time. It replicates the front-end logic of North American peers: first solving the physical bottlenecks of compute placement, power supply, cooling, and high-density deployment, then layering high-value-added services on top. More critically, it has secured multiple whole-cabinet delivery contracts for 10,000-GPU clusters and is beginning to touch the edge of compute scheduling—from "turnkey delivery" to "token delivery," only one software layer away.

Sharetronic Data Technology Co.,Ltd. (300857.SZ) and Jiangsu Lettall Electronic Co.,Ltd. (603629.SH) represent the collective action of "gap-fillers": one extends from server manufacturing toward both intelligent computing centers and optical modules, attempting to turn "bottlenecks" into "self-made" at the hardware level; the other pivots from compute leasing to self-built capacity, using private placements to reinforce bottom-level compute autonomy. They both validate the same consensus—full-stack is difficult, yet no one is willing to sit out; and "gap-filling" itself is tightening every screw in the token industry chain.

Per reports monitored by Zhitong Finance, GBA AI COMP (01396) has brought "full-stack" to a more complete form. This Hong Kong-listed company saw H1 revenue surge over tenfold year-over-year, with in-hand intent orders surpassing RMB 37 billion and FP16 dense compute capacity reaching 50,000 PFLOPS. Its "full-stack" is not a simple stacking of segments but centers on its self-developed cloud platform "QuantumPulse," which both independently schedules 10,000-GPU clusters and threads the energy foundation, compute clusters, token delivery, and application incubation into a closed loop: internally, it dynamically allocates power and compute based on training and inference loads; externally, it outputs token services through standard interfaces and reverse-optimizes upstream resource allocation. This shift from "selling compute" to "delivering tokens" lands precisely on the logic of the "Token Factory." While most peers are still solving "whether they have it," it has quietly outlined the delivery blueprint of China's Token Factory.

Divergent paths, possible convergence

The best perspective is never seeking a Chinese replica of some overseas giant. What truly matters is that both routes point to the same destination: sinking astronomical CAPEX into sustainable cash flow and token output. The end of electricity is compute; the end of compute is pricing power. Whoever first refines "electricity" into "tokens" that are measurable, tradable, and exportable holds the ticket to the next round. Among China's answer sheets, the one who has already shown this card has positioned itself further ahead.