The expansion of AI agents is transforming data centers from places that house servers into spatial assets capable of continuously producing AI inference outputs. The core value of enterprise data centers and traditional IDCs has traditionally been rack deployment, colocation, and general-purpose computing. AIDCs, however, are designed for AI training and AI inference, requiring higher rack power density, more advanced liquid cooling systems, stricter power quality management, and more robust UPS and backup power systems.
As AI inference demand becomes the fastest-growing source of compute consumption in the AI era, data centers are gradually evolving from facilities that simply provide space and electricity into production systems that continuously generate intelligence. Going one step further, AI Factories are no longer merely a data center concept. They package standardized nodes, power distribution systems, cooling infrastructure, software stacks, and orchestration systems into replicable and scalable compute units. Token Factories, meanwhile, measure operational efficiency through metrics such as token throughput, cost per token, and performance per watt.
Against this backdrop, GoodVision AI aims to become more than just a compute service provider. It seeks to position itself as both a “compute developer” and a “compute operator” for the AI era. On one hand, it standardizes and scales compute assets through modular AI Factory deployments. On the other hand, it enables these assets to continuously generate value through intelligent orchestration systems, multi-cloud routing, and edge node networks.
Data Centers Have Undergone Their Own Industrial Revolution
From the perspective of the AI infrastructure value chain, an AI inference request does not begin with the model itself, but with power supply, power distribution, cooling, networking, and compute nodes.
The International Energy Agency estimates that global data center electricity consumption will reach approximately 945 TWh by 2030, representing a significant increase from current levels, with AI being one of the most important drivers. Meanwhile, the U.S. Department of Energy, citing research from Lawrence Berkeley National Laboratory, notes that data center loads in the United States have tripled over the past decade and could potentially double or even triple again by 2028. For AI agents, this means that scarcity no longer lies solely in model capabilities, but in the ability to secure stable power, sufficient cooling capacity, and acceptable network latency in compliant locations.
As a result, the evolution of data centers is also changing. Early Enterprise Data Centers (EDCs) were primarily self-built facilities focused on local control, information security, and regulatory compliance. Later, IDCs and colocation providers commercialized “space + power + networking + operations” through rack hosting. After 2006, cloud infrastructure represented by AWS introduced pooled and on-demand computing resources. Subsequently, hyperscale infrastructure leveraged scale to support global cloud services. IBM’s definition of AI data centers is clear: AIDCs are no longer general-purpose IT facilities. They are computing, networking, storage, energy, and cooling systems specifically redesigned for training, deploying, and delivering AI applications.
The Differences of Definition Between AIDC, AI Factory, and Token Factory
In recent years, the terms AIDC, AI Factory, and Token Factory have appeared frequently. Many people mistakenly assume that they represent three different types of infrastructure. In reality, they are better understood as different ways of describing the same AI infrastructure from different dimensions.
First, AIDC (Artificial Intelligence Data Center) represents the hardware infrastructure layer. It refers to a new generation of data centers purpose-built for AI training and AI inference, including GPU clusters, high-speed networking, advanced cooling systems, and high-density power distribution infrastructure. Compared with traditional IDCs, AIDCs are more like a new type of factory built specifically for AI.
Later, NVIDIA introduced the concept of the AI Factory. Jensen Huang has repeatedly emphasized at GTC that future data centers should no longer be viewed as warehouses for storing servers, but as factories capable of continuously producing intelligence. As a result, AI Factories place greater emphasis on system capabilities and operating models. They encompass not only power systems, GPUs, networking, and cooling infrastructure, but also software stacks, orchestration systems, and standardized delivery capabilities. If AIDCs represent the facilities and machines, then AI Factories represent the entire intelligent production line.
With the arrival of the inference era, the metrics used to evaluate infrastructure have changed once again. In the past, data centers focused on rack counts, available power capacity, and PUE. Today, attention has shifted toward metrics such as token throughput, TTFT, tokens per second, cost per token, and performance per watt. Jensen Huang has even argued that data centers are evolving from places that store files into factories that produce tokens. As a result, Token Factory has become a new operational language for the inference era.
Simply put, AIDCs represent facilities and equipment, AI Factories represent production systems, and Token Factories represent final output capacity. Fundamentally, they describe the same AI infrastructure at different stages and from different perspectives.
The Differences Between AIDCs and Traditional IDCs
Returning to AIDCs themselves, one of the biggest differences between AIDCs and traditional IDCs lies in power density and thermal management. According to Eaton, the average rack density of many traditional data centers has long remained around 10–20kW, while AI workloads have already pushed a large number of racks beyond 100kW. Its 2025 data center report also shows that 21–50kW racks have become increasingly common, while the proportion of racks exceeding 100kW continues to rise. Google has publicly discussed power architectures designed for next-generation AI workloads, with the goal of expanding IT rack capacity from 100kW to 1MW. As a result, thermal design is shifting from air cooling to liquid cooling. Both IBM and Eaton have pointed out that high-density AI workloads require more advanced cooling technologies, including direct liquid cooling and, in certain scenarios, immersion cooling systems that submerge equipment in dielectric fluids. PUE remains the standard metric for infrastructure energy efficiency, but in the AIDC era, PUE alone is no longer sufficient to determine competitiveness. Available power capacity, thermal limits, bandwidth, and system reliability must also be taken into consideration.
The second difference lies in the power chain. The role of UPS systems goes far beyond providing backup power for a few minutes after an outage. They also protect equipment against grid fluctuations, power surges, and momentary interruptions, while ensuring continuous power before backup systems take over. Eaton’s official documentation states that UPS systems provide battery backup and surge suppression during outages, while Vertiv notes that large data centers typically adopt online double-conversion UPS architectures because they continuously provide the highest level of power quality and integrate more effectively with backup generator systems. For AIDCs, UPS systems, energy storage, backup power supplies, and high-density power distribution are not ancillary components, but part of service-level agreements (SLAs), availability, and long-term operational capabilities. As a result, AIDC capital expenditures (CAPEX) are increasingly concentrated in high-density power distribution, UPS systems, liquid cooling, and rack-level networking, while operating expenditures (OPEX) are more heavily influenced by electricity costs, cooling, and automated operations efficiency.
If traditional IDCs primarily sell “rack space + bandwidth + colocation,” AIDCs sell AI-ready capacity. IBM Cloud emphasizes local deployment, low latency, and certified security across its global data center footprint, while AWS’s distinction between data residency and data sovereignty highlights how the physical location of data and the laws governing it directly influence cloud architecture design. Combined with the IEA’s and DOE’s projections on electricity demand growth and infrastructure constraints, it becomes easier to understand why compute is evolving from an abstract service into a form of “spatial asset” defined by location, power, cooling, networking, and compliance boundaries. Its value lies not only in computing speed, but in whether a specific location can continuously, reliably, and compliantly support AI inference workloads.
This trend has also given rise to a new generation of operators focused on AI infrastructure. For example, TDTC (Tokenomics Digital Tech Corp.) specializes in data center and AI compute infrastructure development, with operations centered around power, cooling, networking, and high-density compute resources. To some extent, this reflects how data centers are evolving from traditional colocation facilities into productive assets for the AI era. Referring to AIDCs as “real estate for the AI era” is essentially a way of describing this asset characteristic.
AI Factories therefore represent the next stage of asset organization. Rather than simply placing GPUs inside data centers, they package prefabricated power distribution systems, liquid cooling, racks, networking, software stacks, and operations systems into standardized units. Schneider Electric describes modular data centers as “fast, flexible and predictable,” emphasizing their suitability for deployment using standardized racks and pod architectures. Public materials from Eaton, IBM, and Cloudflare also show that demand for edge nodes and localized deployments is increasing due to latency, performance, security, and sovereignty requirements. For AI agents, this is particularly important: not every inference request should be sent back to massive remote clusters. Many workloads must be processed closer to users, devices, or where data is generated. From this perspective, AI Factories can be understood as “standardized compute real estate” — with fixed construction logic, replicable delivery models, and scalable asset forms.
Looking further up the stack, AIDCs, AI Factories, and Token Factories represent three different ways of measuring value. AIDCs are closer to real estate and industrial infrastructure, with core metrics including available megawatts, rack density, PUE, and deployment cycles. AI Factories begin to treat “intelligent output capacity” as a product, and NVIDIA’s public definition of an AI Factory identifies token throughput as a key output metric. At the Token Factory stage, operational priorities shift further toward metrics such as TTFT, inter-token latency, tokens per second, cost per token, and energy-efficiency indicators such as token/kWh derived directly from performance per watt. It should be emphasized that token/kWh is not a separately standardized industry term, but rather an operational metric derived by converting tokens per watt or throughput per megawatt into an energy-based measurement framework, making it a direct extension of publicly available efficiency metrics.
AI Compute Grid: Connecting Distributed Intelligence Capacity
In the past, websites needed servers. Today, AI agents require operating environments with power, cooling, networking, orchestration capabilities, regulatory compliance, and the ability to deliver services consistently based on tokens or outcomes. Whoever can turn these capabilities into replicable nodes, and then transform those nodes into operational networks, will come closer to defining the next generation of AI infrastructure.
AIDCs provide the foundation, AI Factories represent the standardized asset model, and Token Factories define the operational logic. Within this framework, GoodVision AI serves as the operational layer that connects assets, orchestration, and delivery. On one hand, it builds underlying AI inference compute assets through rapidly deployable modular AI Factories. On the other hand, through intelligent routing, multi-cloud orchestration, and edge node networks, it dynamically allocates different requests to the most appropriate locations. Ultimately, individual AI Factory nodes are connected together to form a global AI Compute Grid.
References
- NVIDIA GTC 2024 Keynote
- NVIDIA GTC 2025 Keynote
- NVIDIA — What Is an AI Factory?
- International Energy Agency, Energy and AI (2025)
- Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report
- IBM — What Is an AI Data Center?
- Eaton — 2025 Data Center Trends Report
- Google — Powering the Future of AI Infrastructure
- Vertiv — The Future of High-Density AI Infrastructure
- Schneider Electric — Modular Data Centers
- AWS — Data Residency vs. Data Sovereignty
- Cloudflare — Edge AI
- Microsoft Research — Designing Datacenter Power Delivery Hierarchies for the AI Era