The true ignition of the AI revolution in the physical world is a wave of computing infrastructure construction measured in gigawatts (GW).
By compiling the publicly disclosed global projects from the past few months, the contours of this supercycle become unmistakable—it is no longer about piecemeal data center expansions but involves continuous, massive, and intercontinental capacity lock-ins.
On June 2, 2026, SoftBank proposed building a 5GW AI data center in France and a 10GW facility in Ohio, USA.
On June 5, 2026, Damac Digital announced plans to expand its data center footprint by 6GW across four continents.
Also on June 5, AirTrunk signed a $21 billion agreement with the Indian state of Maharashtra for a 3GW data center project.
On June 3, GDS-SW (09698) signed a strategic cooperation agreement with the Ulanqab Municipal Government in Inner Mongolia, planning to build multiple GW-scale data center parks with over 80% green power coverage, which forms part of its three-year capital expenditure plan exceeding RMB 30 billion.
On the same day, Oracle and OpenAI launched an approximately 1GW project in Michigan, USA.
Token Consumption Emerges as a New Barometer for AI Industry Vitality
Previously, industry discussions on large models focused on parameters, training costs, and capability frontiers.
Now, a metric more closely aligned with commercialization has surfaced—Token consumption.
This is not merely a technical parameter but a result of real AI application usage, consumption, and billing.
When consumption rises, what is being mobilized at high frequency is not just user queries but an entire production system comprising compute power, servers, power supply, cooling, networking, and data center assets.
Public data indicates that by March 2026, China's average daily Token consumption had surpassed 140 trillion, a growth of over 40% from the end of 2025.
The steep trajectory—from 100 billion in early 2024 to 100 trillion by late 2025, and then to the 140-trillion level by March 2026—clearly signals that AI applications are moving from demos and trials to scaled operation.
Signals from the enterprise side are equally clear: the Doubao large model saw daily Token usage exceed 120 trillion in March, while model companies like MiniMax and Zhipu reported several-fold to over tenfold growth in consumption for AI programming and coding plan scenarios.
This growth is increasingly driven by production scenarios like AI programming, intelligent agents, multimodal generation, and enterprise process automation, rather than simple chat-based Q&A.
This distinguishes the current AI infrastructure cycle from the previous IT build-out: demand is no longer about "how many enterprises are moving to the cloud" but about "every inference, every API call, and every task executed by an agent continuously consumes compute power."
Thus, Token consumption has become an external indicator of AI application activity and a leading indicator of pressure on the underlying infrastructure.
Token Consumption Reshapes the Production Function of AI Infrastructure
Strictly speaking, Token consumption cannot be simply converted into a fixed amount of power consumption, nor does it represent the actual effectiveness of AI deployment.
However, factors like model architecture, parameter scale, context length, inference precision, concurrency scheduling efficiency, KV Cache usage, quantization compression levels, and hardware utilization all tangibly affect the actual compute and energy footprint per Token.
Yet, the industry trend is clear enough.
As consumption leaps from billions to trillions of Tokens, demand propagates down a relatively stable chain: increased application requests lead to higher inference loads, which boost GPU/NPU server utilization, expand cluster scale, and subsequently necessitate higher-density racks, more complex east-west networking, larger power capacity, more demanding cooling systems, and longer-term capacity reservations.
Therefore, the Token surge is fundamentally altering the production function of AI infrastructure, not just a single technical component.
In the past, training clusters often expanded in a phased, project-based, and peak-driven manner.
Now, inference workloads are becoming continuous, real-time, high-concurrency industrial loads.
Once applications such as AI coding, agents, enterprise knowledge bases, intelligent customer service, office automation, marketing, and video generation are embedded into business processes, their compute consumption becomes more routine and rigid.
This is the fundamental reason the data center industry is back in the capital markets spotlight—growth in application-side Token consumption will ultimately translate into medium- to long-term reservations for high-performance data center resources by cloud providers, internet platforms, and model companies.
Demand Materialization: Competition Shifts from 'Space' to 'Power and Delivery'
The most direct signals from the demand side come from the capital expenditures of leading tech firms and the construction of massive power infrastructure.
Overseas, Google and Intersect Power have broken ground on a data center and over 1GW of clean energy infrastructure in Texas; Meta signed a power supply agreement for up to 1GW with long-duration storage firm NoonEnergy; Oracle procured up to 2.8GW of fuel cells from Bloom Energy and is deploying a microgrid with up to 2.45GW of Bloom Energy fuel cells in New Mexico for backup power.
In China, backed by robust grid infrastructure, significant capital has also accelerated investment in AI infrastructure in recent years.
Alibaba has announced plans to invest over RMB 380 billion in cloud and AI hardware infrastructure over the next three years.
Tencent's Q1 2026 capital expenditure was approximately RMB 31.9 billion, primarily directed towards IT infrastructure, data centers, and AI-related areas.
Multiple media reports also point to ByteDance's continued ramp-up in AI infrastructure, with its 2026 capital expenditure plan reportedly exceeding RMB 200 billion, with about 65% allocated domestically.
The substantial investments by Chinese tech giants are also fueling the rapid rise of domestic computing power, propelling China's AI industry into a "supercycle" characterized by high rigidity, long duration, and exponential growth.
These investments will not remain as mere Capex line items on financial statements.
For the data center industry, they will ultimately translate into three more concrete demands: larger-scale power capacity, delivery of higher-density AI-ready facilities, and longer-term resource reservations.
Consequently, the core of competition in the IDC industry is shifting from a "space race" to a "power race" and a "delivery race."
In traditional cycles, clients made decisions based on rack count, geographic location, bandwidth, and cost.
In the AI cycle, the primary concerns have become: Can they secure sufficient, continuous power capacity in one go? Can the infrastructure support AI racks with power densities of tens of kilowatts or higher? Can delivery be completed within the agreed timeframe? Can long-term SLAs, PUE targets, liquid cooling retrofits, and green power coverage be guaranteed?
While floor space remains important, it is no longer the scarcest production factor—power quotas, supply stability, cooling architecture, network quality, construction speed, and long-term operational capabilities are collectively determining the asset value of AI data centers.
Case Study: Triple Validation of Orders, Capital, and Regional Strategy for GDS-SW
A more insightful way to observe this cycle is to view corporate financial reports as a lens on industry shifts—examining order growth, early capacity lock-ins by clients, construction acceleration, utilization ramp-up, sufficiency of capital for future Capex, and the functionality of the asset recycling model.
The Q1 2026 financial report of GDS-SW provides a representative case study.
First, Orders & Operations: The risk profile is shifting from "build first, lease later" to "production-to-order." The company added approximately 200MW of new contracts in Q1, a record high for any single quarter.
As of March 31, 2026, total contracted and pre-contracted area reached 725,485 sqm (up 11.7% YoY), operational area was 674,269 sqm (up 10.4% YoY), and revenue-generating area in service was 520,929 sqm (up 12.7% YoY), with the utilization rate rising to 77.3%.
Area under construction was 118,411 sqm (up 60.0% QoQ), with an operational area contracted rate of 92.8% and a pre-contracted rate for area under construction of 84.4%.
Existing assets continue to ramp up, new projects are accelerating, and most capacity under construction is already pre-sold—this indicates the asset risk structure is shifting from the inventory risk of "build first, lease later" to the delivery risk of "production-to-order."
The real test is no longer securing orders but converting them into billable capacity on schedule.
Second, Finance & Capital: Beyond reported profits, capital deployment capability is key. Q1 net revenue was RMB 3.367 billion (up 23.6% YoY), or RMB 2.938 billion (up 7.9% YoY) excluding certain one-time items.
Adjusted EBITDA was RMB 1.949 billion (up 47.2% YoY), or RMB 1.430 billion (up 8.0% YoY) on the same adjusted basis.
Net profit was RMB 2.652 billion (up 247.1% YoY).
Evaluating AI data centers requires looking beyond reported profits to also consider operating EBITDA, utilization ramp-up, order backlog, capital expenditure capability, and asset recycling capacity.
In Q1, GDS-SW divested part of its stake in DayOne, raising $385 million, and as of April 29, 2026, still held approximately 19.9% (with a remaining market value exceeding $2.2 billion based on Series C financing price).
It also completed a private placement of $300 million in Series B convertible preferred shares.
Cash and cash equivalents (including term deposits) stood at RMB 19.23 billion at quarter-end, while the company maintained its full-year 2026 Capex guidance of approximately RMB 9 billion.
Third, Region & Energy: From "All in on China" to Compute-Power Synergy. On June 3, GDS-SW signed the strategic agreement with Ulanqab, planning to develop multiple high-density and GW-scale data center parks, achieving over 80% green power coverage through direct connections and trading, forming a scaled zero-carbon data center cluster.
This is only part of its three-year capital expenditure plan exceeding RMB 30 billion.
When viewed alongside its domestic investment plan of RMB 30-50 billion over the next three years (a record high in the company's 25-year history), the capital allocation direction is clear: the company is refocusing its new capital, order intake, and regional layout on China's AI infrastructure.
The "All in on China" strategy is a business reality constituted by capital expenditure, customer contracts, projects under construction, and energy resources.
This reflects an evolution of the "East Data, West Computing" narrative—mature markets (Beijing-Tianjin-Hebei, Yangtze River Delta, Guangdong-Hong Kong-Macao) continue to host online inference, financial transactions, real-time interaction, and low-latency services, demanding high network quality, delivery certainty, and operational stability.
Emerging hubs (Ulanqab, Hohhot, Zhongwei) leverage land, power, climate, and green energy resources to accommodate larger-scale training, offline inference, batch computing, and multimodal generation tasks.
The key phrase is shifting from "resource scheduling" to "compute-power synergy": future data center competition depends not just on proximity to clients, but on proximity to stable, low-carbon, scalable power.
Valuation Reassessment: From Static EBITDA to a Multi-Stage Dynamic Model
Amid the AI infrastructure cycle, the valuation logic for data centers is also evolving.
Traditional valuation anchors revolve around EBITDA, utilization rates, stable cash flows, asset depreciation, and regional supply-demand dynamics.
While this framework remains important, it is no longer sufficient to fully capture the growth attributes of AI data centers—their value is not only reflected in currently operational assets but also in the deliverable power capacity, contracted orders, client reservations, construction progress, and capital recycling efficiency over the coming years.
A multi-stage dynamic model is better suited to the current cycle: short-term focus on order intake capability and delivery cadence; medium-term focus on utilization ramp-up and EBITDA realization; long-term focus on free cash flow, return on capital, and asset recycling capability through REITs, ABS, funds, sale-leasebacks, or asset sales.
Under this model, the market will re-examine: How many contracted MW does the company hold? Is the pre-contracted rate sufficiently high? Can projects under construction be delivered on schedule? Are the investment cost per MW and payback period controllable? What is the client move-in speed? Will financing costs erode project returns?
This is why GDS-SW's Q1 report warrants analysis within an industry framework: the ~200MW of quarterly new contracts, over 340MW of year-to-date signings, an 84.4% pre-contracted rate for projects under construction, a 77.3% data center utilization rate, RMB 19.23 billion in cash (including term deposits), combined with a three-year investment plan of RMB 30-50 billion, collectively sketch the雏形 of an "order-construction-capital-delivery"闭环.
No single indicator alone proves a central position in the cycle, but their simultaneous presence indicates that industry momentum is permeating corporate operations.
Navigating Risks Within the Supercycle
However, a "supercycle" does not mean all IDC assets will automatically benefit; multiple risks still stand between industry prosperity and corporate returns.
First, high Token growth does not guarantee perpetually同步 growth in per-unit computing demand; model compression, inference optimization, operator efficiency gains, domestic chip iteration, and scheduling algorithm improvements could all reduce the compute cost per Token.
Second, client contracts do not equate to immediate revenue recognition; there is a lag between order signing and revenue realization due to construction, delivery, equipment installation, rack deployment, and utilization ramp-up.
Third, AI infrastructure has higher Capex intensity; if demand growth slows, financing costs rise, or deliveries are delayed, the asset payback period will lengthen.
Fourth, high-density data centers impose higher demands on power, liquid cooling, networking, and operational stability; securing power quotas is just the first step, continuous supply, PUE control, green power integration, and meeting SLAs are the long-term competitive differentiators.
Fifth, industry prosperity could also lead to localized supply expansion and price competition; competitive dynamics among small-to-medium players, local platforms, telecom operators, and cloud providers building their own data centers remain to be seen.
Therefore, the红利 will not be evenly distributed.
True beneficiaries are more likely to be firms possessing four key capabilities simultaneously: the ability to secure long-term client orders, to lock in scarce power and location resources, to execute the engineering delivery and long-term operation of high-density data centers, and to support continuous expansion through diversified financing and asset recycling.
The strengths of GDS-SW should be understood within this framework—it does not gain inherent certainty from a single financial report, but rather its order book, capital position, regional布局, and delivery capabilities all showed signs of aligning with the AI infrastructure cycle in Q1 2026.
It serves as a significant观察样本 for this cycle, not the sole answer.
What Drives IDC Value Post-Token Surge?
The Token surge has altered how we observe the AI industry and how we assess the value of the data center sector.
In the past, IDC was largely seen as infrastructure assets承载 internet traffic and cloud computing demand.
Now, it is becoming the physical foundation for AI inference and training workloads.
Clients are no longer purchasing just rack space, but power capacity, cooling capability, network quality, delivery certainty, and long-term stable operation.
Corporate valuation should not be anchored solely on current EBITDA, but must consider future deliverable capacity, order quality, utilization ramp-up, capital recovery, and asset recycling capability.
The essence of this IDC industry transformation is not merely an increase in power density, but AI's role in repositioning data centers back to the core of the digital economy's production function.
Those who can translate the Token surge into stable, green, billable, financeable, and sustainably deliverable computing infrastructure are more likely to become long-term beneficiaries in China's AI infrastructure supercycle.
Comments